content
stringlengths 1
15.9M
|
|---|
\subsection{Fisher Information}
A crucial aspect of our approach relies on understanding how the parameters of the model influence its output distribution. To do so, we leverage tools from information geometry. We can view the family of distributions defined by the model $\mathcal{P}(\bm{\theta})$ as defining a statistical manifold with coordinates $\bm{\theta}$. The Riemannian metric for this manifold is the Fisher information matrix
\begin{align}
F}%{\bm{\mathcal{I}}_{\bm{\theta}}(\bm{\theta}) &= \mathbb{E}_{\bm{y}} \left[ ( \nabla_{\bm{\theta}} \log p_{\bm{\theta}}(\bm{y}) )(\nabla_{\bm{\theta}} \log p_{\bm{\theta}}(\bm{y}))^\T \right] \label{eq:fisher_def} \\
&= \mathbb{E}_{\bm{y}} \left[ - \frac{\partial^2}{\partial \bm{\theta}^2} \log p_{\bm{\theta}}(\bm{y}) \right], \quad\text{(see note)\footnotemark}
\end{align} \footnotetext{This equality holds under mild regularity conditions on $\mathcal{P}(\bm{\theta})$}
where $p_{\bm{\theta}}(\cdot) = \mathcal{P}(\bm{\theta})$, the pdf of the probability distribution on $\mathcal{Y}$ defined by parameters $\bm{\theta}$.
The Fisher information matrix (henceforth referred to as the Fisher) represents the second-order approximation of the local KL divergence, describing how the output distribution of a model changes with small perturbations to the distribution parameters $\bm{\theta}$:
\begin{align}
D_\mathsf{KL}(\mathcal{P}(\bm{\theta}) || \mathcal{P}(\bm{\theta} + d\bm{\theta})) &\approx d\bm{\theta}^\T F}%{\bm{\mathcal{I}}_{\bm{\theta}}(\bm{\theta}) d\bm{\theta} + O(d\bm{\theta}^3). \label{eq:fisher-kl}
\end{align}
The subscript on $F}%{\bm{\mathcal{I}}_{\bm{\theta}}$ serves to make explicit the Fisher's dependence on the model's parameterization.
For many common parametric distributions, the Fisher can be computed analytically.
For example, for the family of Gaussian distributions with fixed covariance $\Sigma$, parameterized by the mean vector $\bm{\theta} = \bm{\mu}$, the Fisher information is simply the constant $F}%{\bm{\mathcal{I}}_{\bm{\theta}}(\bm{\theta}) = \Sigma^{-1}$. For a categorical distribution parameterized such that $\bm{\theta}_i$ represents the probability assigned to class $i$, $F}%{\bm{\mathcal{I}}(\bm{\theta}) = \left(\mathsf{diag}(\bm{\theta})\right)^{-1}$.
In cases where this analytic computation is not possible or difficult, one can compute a Monte-Carlo approximation of the Fisher by sampling $\bm{y} \sim p_{\bm{\theta}}(\cdot)$ to estimate the expectation in \eqref{eq:fisher_def}.
\subsection{Fisher for Deep Neural Networks}
For DNNs, we can also consider the Fisher in terms of the network weights $\bm{w}$ using a change of variables. Since $\bm{\theta} = f(x, \bm{w})$, we have
\begin{align}
F}%{\bm{\mathcal{I}}_{\bm{w}}(\bm{x}, \bm{w}) &= \bm{J}_{f,\bm{w}}^\T F}%{\bm{\mathcal{I}}_{\bm{\theta}}(f(\bm{x}, \bm{w})) \bm{J}_{f,\bm{w}}, \label{eq:reparameterization}
\end{align}
where $\bm{J}_{f,\bm{w}}$ is the Jacobian matrix with $[\bm{J}_{f,\bm{w}}]_{ij} = \frac{\partial f_i}{\partial w_j}$, evaluated at $(\bm{x}, \bm{w})$. Note that $F}%{\bm{\mathcal{I}}_{\bm{w}}$ is a function of both $\bm{w}$ and the input $\bm{x}$. We will henceforth use the shorthand $F}%{\bm{\mathcal{I}}^{{(t)}}_{\bm{w}^*} := F}%{\bm{\mathcal{I}}_{\bm{w}}(\bm{x}^{(t)}, \bm{w}^*)$ to denote this weight-space Fisher evaluated for a particular input $\bm{x}^{(t)}$ and the trained weights $\bm{w}^*$.
From \eqref{eq:fisher-kl}, we see that the Fisher defines a second-order approximation of how perturbations in the weight space influence the DNN's probabilistic predictions:
\begin{align*}
\delta_\mathsf{KL}(\bm{x}^{(t)}) &:= D_\mathsf{KL}\left(p_{\bm{w}^*}(\cdot \mid \bm{x}^{(t)}) || p_{\bm{w}^* + d\bm{w}}(\cdot \mid \bm{x}^{(t)}) \right) \\
&\approx d\bm{w}^\T F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*} d\bm{w}.
\end{align*}
We can also use the Fisher to consider the impact of weight perturbations on the predictions over the entire dataset as
\begin{align*}
\delta_\mathsf{KL}(\mathcal{D}) = \frac{1}{M} \sum_{i=1}^M \delta_\mathsf{KL}(\bm{x}^{(i)})
&= \frac{1}{M} \sum_{i=1}^M d\bm{w}^\T F}%{\bm{\mathcal{I}}^{(i)}_{\bm{w}^*} d\bm{w} \\
&= d\bm{w}^\T \underbrace{\left( \frac{1}{M} \sum_{i=1}^M F}%{\bm{\mathcal{I}}^{(i)}_{\bm{w}^*} \right)}_{:= F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}} d\bm{w}.
\end{align*}
\subsection{Connection to the Hessian}
As is evident from \eqref{eq:fisher_def}, there are strong connections between the dataset Fisher $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$ and the Hessian with respect to $\bm{w}$ of the log likelihood of the training data. If we define $L(\bm{w}) = \sum_{i=1}^M \log p_{\bm{w}^*}(\bm{y}^{(i)} \mid \bm{x}^{(i)})$,
the Hessian of $L$ evaluated at $\bm{w}^*$ can be well approximated by the dataset Fisher\footnote{By application of the chain rule, $H_L = \hat{F}%{\bm{\mathcal{I}}}^\mathcal{D}_{\bm{w}^*}+ C(\bm{w}^*)$, where $\hat{F}%{\bm{\mathcal{I}}}^\mathcal{D}_{\bm{w}^*}$ is the empirical Fisher where the expectation in \eqref{eq:fisher_def} is replaced by an empirical expectation over the dataset, and $C(\bm{w}^*)$ is a term involving first derivatives of the log likelihood and second derivatives of the network $f$. For trained networks, we expect the Fisher and the empirical Fisher to be closely aligned, and $C(\bm{w}^*) \approx 0$ since the first derivatives of the log likelihood are near zero at the end of training. Furthermore, for piece-wise linear networks (e.g., with ReLU activations), the second deriviatves of $f$ are 0, and so, $C(\bm{w}^*) = 0$.} \citep{Ritter2018ASL,martens2015optimizing}:
\[
H_L = \frac{\partial^2 L}{\partial \bm{w}^2} \biggr|_{\bm{w}=\bm{w}^*} \approx M F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}.
\]
Unlike the Hessian, the Fisher is always guaranteed to be positive semidefinite. For this reason, this approximation is common in the field of second-order optimization, where preconditioning gradient steps with the inverse Fisher tends to be more efficient and numerically stable than using the Hessian.
\subsection{The Laplace Approximation of Epistemic Uncertainty}
The Hessian of the log-likelihood has strong connections to Bayesian ideas of \textit{epistemic uncertainty}, the uncertainty due to lack of data. From a Bayesian perspective, one can choose a prior over the weights of a DNN, $p(\bm{w})$, and then reason about the posterior distribution on these weights given the dataset, $p(\bm{w} \mid \mathcal{D})$. Often, due to the overparameterized nature of DNNs, many values of $\bm{w}$ are likely under this dataset, corresponding to different ways the DNN can fit the training data \citep{azizan2019stochastic}. By characterizing the posterior, and then marginalizing over it to produce a posterior predictive distribution $p(\bm{y} \mid \bm{x}) = \int p(\bm{w} \mid \mathcal{D}) p_{\bm{w}}(\bm{y} \mid \bm{x}) d\bm{w}$, one can hope to detect atypical data by incorporating uncertainty due to lack of data into the network's probabilistic predictions.
While computing this posterior is intractable for DNNs, due to their nonlinearity and high-dimensional weight space, many approximations exist, leveraging Monte-Carlo sampling, or by assuming a distributional form and carrying out variational inference. One approximation is the Laplace approximation \citep{mackay1992practical}, which involves a second-order approximation of the log posterior $\log p(\bm{w} \mid \mathcal{D})$ about a point estimate $\bm{w}^*$. This quadratic form yields a Gaussian posterior over the weights. If the prior on the weights is $p(\bm{w}) = \mathcal{N}(\cdot; 0, \epsilon^2 I_N)$, the Laplace posterior is given by
$\Sigma^* = \frac12 \left(H_L + \frac{1}{2\epsilon^2} I_N \right)^{-1}$.
The Laplace approximation is attractive as it uses local second-order information (the Hessian $H_L$) to produce an estimate of the Bayesian posterior for any pretrained model. However, even computing this approximation to the exact Bayesian posterior can be challenging for large models, where estimating and inverting an $N \times N$ matrix to compute $\Sigma^*$ is demanding.
\subsection{Efficiently computing the Uncertainty Metric}
While \eqref{eq:oodmetric_simple} is a clean expression for an uncertainty metric, its na\"ive computation is intractable for typical DNNs, as both $\Sigma^*$ and $F}%{\bm{\mathcal{I}}^{(t)}_{\mathcal{D}}$ are $N \times N$ matrices. To make this computation amenable for real-time operation, we exploit the fact that both matrices are low-rank. From their definitions, it follows that $\mathsf{rank}(F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*}) \le d$, and thus $\mathsf{rank}(F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}) \le Md$. The number of weights in a neural network $N$ is always greater than the output dimension $d$, and for large models and typical dataset sizes, often also greater than $Md$.
Thus, we choose instead to express both Fisher matrices in factored forms
\begin{align}
F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*} = L^{{(t)}}_{\bm{w}^*} L^{{(t)}\T}_{\bm{w}^*}, \qquad
F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*} = U \mathsf{diag}(\vec{\lambda}) U^T,
\end{align}
where $L^{(t)}_{\bm{w}^*} \in \mathbb{R}^{N \times d}, U \in \mathbb{R}^{N \times Md}$, and $\vec{\lambda} \in \mathbb{R}^{Md}$.
Note that if we write $F}%{\bm{\mathcal{I}}_{\bm{\theta}}(f(\bm{x}^{{(t)}}, \bm{w}^*)) = L^{{(t)}}_{\bm{\theta}^*}L^{{(t)}\T}_{\bm{\theta}^*}$, then from \eqref{eq:reparameterization}, we can see that $L_{\bm{w}^*}^{(t)} = \bm{J}_{f,\bm{w}}^\T L^{(t)}_{\bm{\theta}^*}$. For many common choices of parametric distributions, $L_{\bm{\theta}^*}^{(t)}$ can be computed analytically. Furthermore, leveraging the linearity of the derivative, we can compute each row of $L^{(t)}_{\bm{w}^*}$ efficiently via backpropagation (see Appendix~\ref{app:backprop-tricks} for details and examples for common distributions).
Given these factored forms, an application of the Woodbury matrix identity allows us to simplify the computation to
\begin{align}
\begin{split}
&\mathrm{Unc}(\bm{x}^{(t)}) = \\
&~~\epsilon^2 \left\| L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F} -
\epsilon^2\left\| \mathsf{diag}\left( \sqrt{ \frac{\vec{\lambda}}{\vec{\lambda} + \nicefrac{1}{\left(2 M \epsilon^2\right)}}} \right) U^\T L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F} .
\end{split}
\end{align}
A derivation is provided in Appendix \ref{app:metric-derivation}.
In this new form, the computation is split into computing the factor $L_{\bm{w}^*}^{(t)}$, carrying out the matrix product with $U^\T$, and then computing Frobenius norms.
The main bottleneck in this procedure, both in terms of memory and computation, is the matrix multiplication with the $N \times Md$ matrix $U$. However, several empirical analyses of neural network curvature have found that the Hessian and dataset Fisher $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$ exhibit rapid spectral decay \citep{sagun2017empirical, madras_detecting_2019}).
Indeed, we see that if $\lambda_i << 1/(2M\epsilon^2)$, the corresponding element in the diagonal matrix tends to 0. Thus, we can approximate this computation by only considering the top $k$ eigenvalues and eigenvectors of the dataset Fisher, drastically reducing memory and computation requirements at test time. Notably, making this low-rank approximation gives us a strict \textit{over-estimate} of the exact quantity, which is well-suited for safety-critical settings, where being under-confident is more desirable than being over-confident. An error bound is provided in Appendix~\ref{app:metric-derivation}.
\subsection{Tractably approximating the Dataset Fisher via Matrix Sketching}
In order to compute $\mathrm{Unc}(\bm{x}^{(t)})$ online, we require the top $k$ eigenvalues and eigenvectors of $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$. While this computation can happen offline, it is still intractable to carry out exactly for common DNN models and large datasets, since simply representing $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$ exactly requires storing an $M d \times N$ factor, which can easily grow beyond the capacity of common GPU memories for large perception networks and datasets with tens of thousands of parameters.
To alleviate this issue, prior work has considered imposing sparsity patterns on the Fisher, e.g., diagonal (which ignores correlations between weights), or layer-wise block-diagonal \citep{Ritter2018ASL} (which ignores correlations between layers).
Instead, we note that only the top eigenvectors of the datastet Fisher are important to the computation of our uncertainty metric, and thus we turn to tools from matrix sketching to tractably estimate a low-rank approximation of the Fisher \textit{without} imposing any sparsity structure on the matrix.
The key idea in matrix sketching, to avoid working with a large matrix directly, is to apply a randomized linear map $\mathsf{S}$ to the matrix of interest \citep{tropp_practical_2017}. By appropriately randomizing this map (the \textit{sketching operator}), we obtain high-probability guarantees that the image of the original matrix produced by the map (the \textit{sketch}) encodes sufficient information about the original matrix.
Given a bound on the desired approximation error, the size required for the sketch depends on the desired rank of the approximation $k$, and not the original size of the large matrix. Thus, this can enable our technique to be applied to arbitrarily large datasets.
The linearity of the sketching operator allows us to form this sketch iteratively, using a single pass over the dataset, without storing the full dataset Fisher in memory.
Specifically, from the definition of $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$, we can compute its sketch as the sum of smaller sketches:
\begin{align}
\mathsf{S}\left(F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}\right) &= \frac1M \sum_{i=1}^M \mathsf{S}\left( F}%{\bm{\mathcal{I}}^{(i)}_{\bm{w}^*} \right)
= \frac1M \sum_{i=1}^M \mathsf{S}\left( L^{(i)}_{\bm{w}^*} L^{(i)\T}_{\bm{w}^*} \right). \label{eq:total-sketch}
\end{align}
Following \cite{tropp_practical_2017}, we choose $\mathsf{S}$ to independently left- and right-multiply the Fisher by random sketching matrices. Specifically, the sketch of each component is computed as $Y^{(i)},W^{(i)} \gets \mathsf{S}( L^{(i)}_{\bm{w}^*} L^{(i)\T}_{\bm{w}^*} )$, where
\begin{align}
Y^{(i)} &= \left( \left(\Omega L^{(i)}_{\bm{w}^*} \right) L^{(i)\T}_{\bm{w}^*}\right)^\T, &
W &= \left( \Psi L^{(i)}_{\bm{w}^*} \right) L^{(i)\T}_{\bm{w}^*},
\end{align}
where $\Omega \in \mathbb{R}^{r \times N}$ and $\Psi \in \mathbb{R}^{s \times N}$ are random sketching matrices, with $T = r+s$ defining the total size of the sketch. Note that following the operation order indicated by the parentheses avoids instantiating any $N \times N$ matrix. The memory complexity of this operation is $O(T(N + d))$. \citet{tropp_practical_2017} suggest splitting the budget to $r = (T-1)/3, s = T - r$, and suggest choosing $T=6k+4$ as a minimal value of $T$ for a given $k$ to minimize a high-probability bound on the approximation error of the sketch. We refer the reader to \citep{tropp_practical_2017} for a discussion of these theoretical results.
These sketching matrices can be as simple as matrices with i.i.d. standard Gaussian entries. However, to further reduce the memory and computation overhead of sketching, in this work, we use Subsampled Randomized Fourier Transform (SRFT) sketching matrices \citep{woolfe2008fast}
\begin{align}
\Omega &= P_1 F_{N} \mathsf{diag}(\vec{d}_1), & \Psi &= P_2 F_{N} \mathsf{diag}(\vec{d}_2), \label{eq:sketch-map}
\end{align}
where $\vec{d}_1, \vec{d}_2 \in \mathbb{R}^N$ are vectors with entries sampled from independent Rademacher random variables\footnote{A Rademacher random variable has a value of $+1$ or $-1$ with equal probability.}, $F_{N}$ is the linear operator which applies the discrete cosine transform on each $N$-length column, and $P_1, P_2$ are matrices which each select a random subset of $r$ and $s$ rows respectively. The SRFT sketching matrices offer similar approximation performance when compared to Gaussian matrices, but adding only a $(T+2N)$ parameter overhead, as opposed to $TN$ of the Gaussian case \citep{tropp_practical_2017}.
\begin{algorithm}[t]
\caption{SCOD{} Offline\label{alg:offline}}
\begin{algorithmic}[1]
\Require{Dataset $\mathcal{D} = \{ \bm{x}^{(i)}, \bm{y}^{(i)} \}_{i=1}^M$, DNN architecture $f, \mathcal{P}$, trained weights $\bm{w}^*$}
\Function{SketchCurvature}{$f, \mathcal{P}, \bm{w}^*, \mathcal{D}$}
\State Sample $\Omega, \Psi$ as in \eqref{eq:sketch-map}. \Comment{construct sketching map}
\State $Y, W \gets 0, 0$ \Comment{initialize sketch}
\For{$(\bm{x}^{(i)}, \bm{y}^{(i)})$ \textbf{in} $\mathcal{D}$}
\State $\bm{\theta}^{(i)} \gets f(\bm{x}^{(i)}, \bm{w}^*)$ \Comment{forward pass}
\State Compute $L^{(i)}_{\bm{w}^*}$ from $\bm{\theta}^{(i)}$ \Comment{$d$ backward passes}
\State $Y \gets Y + \frac{1}{M} \left( \left(\Omega L^{(i)}_{\bm{w}^*} \right) L^{(i)\T}_{\bm{w}^*}\right)^\T$ \Comment{update sketch}
\State $W \gets W + \frac{1}{M} \left( \Psi L^{(i)}_{\bm{w}^*} \right) L^{(i)\T}_{\bm{w}^*} $
\EndFor
\State $U_\mathrm{top}, \vec{\lambda}_\mathrm{top} \gets$ \Call{FixedRankSym}{$\Omega, \Psi, Y, W$}
\State \Return $U_\mathrm{top}, \vec{\lambda}_\mathrm{top} $
\EndFunction
\end{algorithmic}
\end{algorithm}
Armed with these tools, we can now follow the procedure detailed in Algorithm \ref{alg:offline} to produce a low-rank approximation of the dataset Fisher.
Our approach uses one pass over the dataset $\mathcal{D}$, incrementally building the sum in \eqref{eq:total-sketch} by applying the SRFT sketching matrices to the $L^{(i)}_{\bm{w}^*}$, the factor for the Fisher for a single input, which can be computed with $d$ backward passes for each input. Having constructed the sketch, we can use this much lower-dimensional, $T \times N$ representation of the dataset Fisher to extract a low-rank, diagonalized representation $F}%{\bm{\mathcal{I}}^{\mathcal{D}}_{\bm{w}^*} = U_\mathrm{top} \mathsf{diag}(\vec{\lambda}_\mathrm{top}) U_\mathrm{top}^\T$ by following the \texttt{FixedRankSymmetric} algorithm detailed in \citet{tropp_practical_2017}.
\subsection{Choosing the Sketch Budget and Rank of Approximation}
A key aspect of SCOD{} is using matrix sketching to approximate the dataset Fisher as a low-rank matrix. This presents the practitioner with two key hyperparameters: the memory budget $T$ to allocate for the sketching, and the rank $k$ of the approximation used in online computation. While memory budget $T$ is generally set by hardware constraints, it impacts the quality of low-rank approximation attainable through sketching. Indeed, the sketching procedure produces a rank $2(T-1)/3$ approximation, and theoretical results suggest keeping only the top $(T-4)/6$ eigenvalues and eigenvectors from this approximation \citep{tropp_practical_2017}.
We explore this trade-off empirically on the Rotated MNIST domain. We choose a range of values for the sketching budget $T$, and sketch the dataset Fisher. For each sketch size, we choose $k$ from a range of values from $1$ all the way to the theoretical maximum $2(T-1)/3$, and test the AUROC performance of $\mathrm{Unc}$. Figure \ref{fig:auroc-vs-rank} shows the results of these experiments.
These results show two key trends. First, we see that increasing the rank $k$ improves performance, but with diminishing returns. Second, for a fixed rank $k$, we see that the performance is insensitive to the sketch budget, especially if $k < (T-4)/6$. For larger $k$, we start seeing benefits from increasing the sketch budget, consistent with the theory; performance starts to plateau as $k$ increases beyond $(T-4)/6$.
Beyond the rank, $\mathrm{Unc}$ is also impacted by the scale of the prior $\epsilon^2$. In our experiments, we found performance to be insensitive to this hyperparameter and thus use $\epsilon=1$ in all the experiments. See Appendix \ref{app:Meps-sweep} for more details.
\subsection{Further improving efficiency for large DNNs}
These results suggest that the best classification performance is achieved by setting $T$ as high as memory allows, and subsequently choosing $k \ge (T-4)/6$. While sketching enables us to avoid the quadratic dependency on $N$, the memory footprint of the offline stage of SCOD{} is still linear, i.e., $O(N T)$. As a result, GPU memory constraints can restrict $T$, and thus $k$, substantially for large models. To alleviate this issue, we study how performance is impacted if we simply restrict our analysis to the last few layers of the network, thus lowering the effective value of $N$. For convolutional networks, the first layers tend to learn generic filter patterns which apply across many domains, while later layers learn task-specific representations \citep{zeiler2014visualizing}. Thus, it is possible that the curvature on the later-layer parameters is more informative from an OoD detection standpoint.
To test the impact of this, we compare to \textbf{SCOD{} (LL)}, an ablation which applies this analysis only to the last layers of the network. For these experiments, we chose to limit analysis to the last $L$ layers, where we chose $L$ for each network architecture such that a minimum of 2 layers were considered, and at least 1 Conv layer was considered. For the larger models in TaxiNet and CIFAR10, we restricted our analysis to the last 15\% of the layers. In these latter two domains, we also increased the sketch budget $T$ and associated rank $k$ up to the capacity of our GPU. Full details of the setup are included in Appendix \ref{app:domains}. Results are included in Figure \ref{fig:runtime-vs-auroc}. For an in-depth look into the impact of focusing on the last layer, see Appendix \ref{app:last-layers}.
\subsection{Performance relative to Baselines}
We compare SCOD{} against several baselines.
First, we compare with the two closely-related approaches, namely, \textbf{Local Ensembles} \citep{madras_detecting_2019} and \textbf{KFAC Laplace} approximation \citep{Ritter2018ASL}, which use curvature estimation to augment a trained model with uncertainty estimates.
Next, while \textit{not} a method applicable to a pre-trained model, we compare against \textbf{Deep Ensembles} \citep{lakshminarayanan2017simple} as a benchmark, as it has shown strong OoD detection performance across regression and classification examples. For this baseline, we retrain $K=5$ models of the same architecture from different initializations.
Both KFAC Laplace and Deep Ensembles output mixture distributions, which we turn into a scalar uncertainty measure by computing the total variance (summed over output dimension) for regression problems, or the predictive entropy for classification, as in \citep{lakshminarayanan2017simple}.
Finally, we compare to a \textbf{Naive} baseline which produces an uncertainty measure directly from the output of the pre-trained model. For regression models which output only a mean estimate, extracting an uncertainty estimate is not possible, and thus this baseline outputs a constant signal $\mathrm{Unc}(x) = 1$. For classification models, we use the entropy of the output distribution as a measure of uncertainty, as in \citep{lakshminarayanan2017simple}.
We compare the performance of these baselines across three regression domains (Wine, Rotated MNIST, and TaxiNet), as well as three classification domains (Binary MNIST, MNIST, and CIFAR10).
For each domain, we create a dataset known to be semantically different from the training data. Where possible, we consider OoD data that is realistic --- for example, data from white wines as OoD for a model trained on data form red wines; or in TaxiNet, images from a wing-mounted camera in different times of day and different weather conditions for a model trained only on data from the morning with clear weather conditions. We also consider using the model's own accuracy to label points as in- or out-of-distribution in Appendix \ref{app:error-based}. For all domains except CIFAR10, we train all networks, and then apply the post-training algorithms. For the CIFAR10 domain, we test on a pre-trained DenseNet121 model \citep{huang2019convolutional, huy_phan_2021_4431043} to highlight the fact that SCOD{} can be applied to augment any pre-trained model, independent of training methodology, with uncertainty estimates.
The results are summarized in Figure \ref{fig:runtime-vs-auroc}. Overall, we see three key trends. First, SCOD{} consistently provides the most informative uncertainty measures out of the methods applicable on a pre-trained model. In fact, its AUROC often matches or exceeds that of Deep Ensembles. While KFAC Laplace also matches Ensemble performance in many settings, our approach tends to dominate it on a runtime/AUROC Pareto frontier. Computing uncertainty metrics using KFAC Laplace requires the computationally-intensive repeated evaluations of the DNN with different sampled weights. Furthermore, this sampling process is very sensitive to the regularization hyperparameters. In contrast, our uncertainty metric does not require any sampling, and thus adds very little overhead. Local Ensembles similarly add little overhead, but have a worse AUROC on many domains, especially for large models and datasets, e.g., in the TaxiNet domain, where $M = 50,000$. It is likely that sketching the Fisher yields a better low-rank curvature estimate than using stochastic mini-batching to estimate the top eigenspace of the Hessian.
On regression problems, the output of the network provides no estimate of uncertainty, and, unsurprisingly, the Naive baseline performs poorly. In contrast, on classification problems, the Naive strategy of interpreting the predictive uncertainty as epistemic uncertainty works quite well \citep{hendrycks_baseline_2018}. Nevertheless, we see that, SCOD{} is generally able to exceed the Naive performance on these classification domains, and match the performance of Deep Ensembles. Importantly, unlike Deep Ensembles or even KFAC Laplace, we do not directly use the base DNN's output uncertainty as part of our score, and only consider how weight perturbations may change the output. Given that the output uncertainty is a strong baseline for softmax classification problems, connecting these ideas is an interesting avenue for future work.
In general, we see that restricting analysis to the last layers often leads to better AUROC performance as well as faster runtime. This is particularly evident on CIFAR10, where without restricting analysis to the last layers, SCOD{} performed worse than the Naive baseline (although still achieving AUROC > 0.9). We suspect this is due to the first layers representing generic features that are suited to most natural images, and thus, the curvature of the later layers is more useful for OoD detection. This is supported by a case study on the CIFAR10 model, with results in Appendix \ref{app:last-layers}.
\section{Computing per-input Fisher via backprop}
\label{app:backprop-tricks}
We work with the factorization $F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*} = L^{(t)}_{\bm{w}^*}L^{{(t)}\T}_{\bm{w}^*}$
Recall that $L^{{(t)}}_{\bm{w}^*} = L_{\bm{w}^*}^{(t)} = \bm{J}_{f,\bm{w}}^\T L^{(t)}_{\bm{\theta}^*}$.
Leveraging the linearity of the derivative, we can avoid carrying out this matrix multiplication on $\bm{J}^{{(t)}}_{f,\bm{w}^*} \in \mathbb{R}^{d \times N}$, and instead perform the matrix multiplication prior to computing the Jacobian via backpropagation. Defining the function $\vec{g}(A, \bm{x}, \bm{w}) = A f(\bm{x},\bm{w})$, which applies a linear transformation $A$ to the output of the DNN, we have
\begin{align}
L^{{(t)}\T}_{\bm{w}^*} &= \left[ \begin{array}{c}
\frac{\partial}{\partial\bm{w}} g_1 (L^{{(t)}\T}_{\bm{\theta}^*}, \bm{x}^{(t)}, \bm{w}^*) \\
\vdots \\
\frac{\partial}{\partial\bm{w}} g_d (L^{{(t)}\T}_{\bm{\theta}^*}, \bm{x}^{(t)}, \bm{w}^*)
\end{array} \right],
\end{align}
where each row can be computed by backpropagation from the corresponding output dimension of $\tilde{\bm{\theta}}$, but ignoring the gradients that flow through the dependence of $L^{{(t)}\T}_{\bm{\theta}^*}$ on $\bm{w}$.
Alternatively, for models with large output dimensions, exactly computing the Fisher as outlined above can be expensive. For such settings, it is possible to exploit the definition of the Fisher as an expectation over $\bm{y} \sim P(\bm{\theta})$, and turn to numerical integration techniques such as Monte-Carlo estimation. In our experiments, we use the exact Fisher in both the offline and online phases.
Below, we provide analytic forms of $L^{(t)}_{\bm{\theta}^*}$ for common parameteric distributions:
\begin{itemize}
\item \textbf{Fixed Diagonal Variance Gaussian.} If $\mathcal{P}(\bm{\theta})= \mathcal{N}(\bm{\theta}, \mathsf{diag}(\vec{\sigma}))$, then
\[
L^{(t)}_{\bm{\theta}^*} = \mathsf{diag}(\vec{\sigma})^{-\nicefrac12}.
\]
\item \textbf{Bernoulli with Logit Parameter.} If the output distribution is chosen to be a Bernoulli parameterized by the logit $\bm{\theta} \in \mathbb{R}$, such that the probability of a positive outcome is $p = 1/(1+\exp(-\bm{\theta}^*))$, then, $F}%{\bm{\mathcal{I}}^{(t)}_{\bm{\theta}^*} = p(1-p)$. So,
\[
L^{(t)}_{\bm{\theta}^*} = \sqrt{p(1-p)}.
\]
\item \textbf{Categorical with Logits.} If the output distribution is a categorical one parameterized by the logits $\bm{\theta} \in \mathbb{R}^d$, such that $p(y = k) = \frac{\exp(\bm{\theta}_k)}{\sum_{j=1}^d \exp(\bm{\theta}_j)}$, then,
\[
L^{(t)}_{\bm{\theta}^*} = \mathsf{diag}(\vec{p})^{\nicefrac12} (I_d - \vec{1}_{d} \vec{p}^\T),
\]
where $\vec{p}$ is the vector of class probabilities according to $\bm{\theta}^*$.
\end{itemize}
\section{Derivation of the Simplified Uncertainty Metric}
\label{app:metric-derivation}
If we substitute the eigenvalue decomposition of the dataset Fisher $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*} = U \mathsf{diag}(\vec{\lambda}) U^T$ into the expression for the posterior covariance \eqref{eq:posterior-cov}, and apply the Woodbury identity, we obtain
\begin{align}
\Sigma^* &= \epsilon^2 \left( I - U \mathsf{diag}\left( \frac{\vec{\lambda}}{\vec{\lambda} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}} \right) U^\T \right),
\end{align}
where the operations in the diagonal are applied elementwise.
Now, if we plug this expression into our expression for $\mathrm{Unc}$, we obtain
\begin{align}
&\mathrm{Unc}(\bm{x}^{(t)})\notag\\
&= \trace \left( \epsilon^2 \left( I - U \mathsf{diag}\left( \frac{\vec{\lambda}}{\vec{\lambda} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}} \right) U^\T \right) F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*} \right) \\
&= \epsilon^2 \trace\left( F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*} \right) - \epsilon^2 \trace\left( U \mathsf{diag}\left( \frac{\vec{\lambda}}{\vec{\lambda} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}} \right) U^\T F}%{\bm{\mathcal{I}}^{(t)}_{\bm{w}^*} \right) \\
&= \epsilon^2 \left\| L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F} - \epsilon^2 \trace\left( L^{{(t)}\T}_{\bm{w}^*} U \mathsf{diag}\left( \frac{\vec{\lambda}}{\vec{\lambda} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}} \right) U^\T L^{(t)}_{\bm{w}^*} \right) \\
&= \epsilon^2 \left\| L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F} - \epsilon^2 \left\| \mathsf{diag}\left(\sqrt{ \frac{\vec{\lambda}}{\vec{\lambda} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}}} \right) U^\T L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F},
\end{align}
where we make use of the cyclic property of the trace and the fact that $\| A \|_F = \sqrt{\trace\left(AA^\T\right)}$.
\subsection{Low-rank Approximation and Error Bounds}
We notice that the elements of the diagonal $\frac{\lambda_j}{\lambda_j + \nicefrac{1}{2 M \epsilon^2}}$ tend to 1 for $\lambda_i >> \nicefrac{1}{2 M \epsilon^2}$, and 0 for $\lambda_i << \nicefrac{1}{2 M \epsilon^2}$. Therefore, only the top eigenvectors of the dataset Fisher are relevant to this posterior. We see that assuming a fixed spectral decay rate of $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$, more eigenvectors are relevant if we choose $\epsilon^2$ to be large (wider prior weights), or $M$ to be large (more data points collected). The number of eigenvectors we keep influences memory and compute requirements. So, alternatively, we can choose a fixed rank of approximation $k$, and then choose $\epsilon^2$ as appropriate.
Thus, we see that the dataset Fisher characterizes how the weights of the DNN are determined by the dataset. In fact, we see that this posterior distribution on the weights has a wide variance in all directions expect in the directions of the top eigenvectors of $F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}$, for which $\lambda_j / \left(\lambda_j + (2 M \epsilon^2)^{-1}\right)$ is non-negligible.
We can characterize the error made by keeping only the top $k$ eigenvalues and eigenvectors. Let $\vec{\lambda}^\T = [\vec{\lambda}^\T_\mathrm{top}, \vec{\lambda}^\T_\mathrm{bot}], U = [U_\mathrm{top} U_\mathrm{bottom}]$, where $\mathrm{top}$ selects the top $k$ eigenvalues. If we define $\tilde{\mathrm{Unc}}(\bm{x})$ as using the low-rank approximation, we have
\begin{align*}
&\tilde{\mathrm{Unc}}(\bm{x}^{(t)})\\
&= \epsilon^2 \left\| L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F} - \epsilon^2 \left\| \mathsf{diag}\left(\sqrt{ \frac{\vec{\lambda}_\mathrm{top}}{\vec{\lambda}_\mathrm{top} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}}} \right) U_\mathrm{top}^\T L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F}.
\end{align*}
The error in the approximation can then be characterized as
\begin{align}
&\tilde{\mathrm{Unc}}(\bm{x}^{(t)}) - \mathrm{Unc}(\bm{x}^{(t)})\notag\\
&= \epsilon^2 \left\| \mathsf{diag}\left(\sqrt{ \frac{\vec{\lambda}_\mathrm{bottom}}{\vec{\lambda}_\mathrm{bottom} + \nicefrac{1}{\left( 2 M \epsilon^2 \right)}}} \right) U^\T_\mathrm{bottom} L^{(t)}_{\bm{w}^*} \right\|^2_\mathrm{F} \\
&= \epsilon^2 \sum_{j=k+1}^{\mathsf{rank}\left(F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}\right)} \frac{\lambda_j}{\lambda_j + \nicefrac{1}{\left( 2 M \epsilon^2 \right) }} \left\| L^{{(t)}\T}_{\bm{w}^*} \vec{u}_j \right\|_2^2 \\
&\le \epsilon^2 \left\| L^{{(t)}\T}_{\bm{w}^*} \right\|^2_2 \sum_{j=k+1}^{\mathsf{rank}\left(F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}\right)} \frac{\lambda_j}{\lambda_j + \nicefrac{1}{\left( 2 M \epsilon^2 \right) }}\\
&\le \epsilon^2 \left\| L^{{(t)}}_{\bm{w}^*} \right\|^2_F \sum_{j=k+1}^{\mathsf{rank}\left(F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}\right)} \frac{\lambda_j}{\lambda_j + \nicefrac{1}{\left( 2 M \epsilon^2 \right) }}\\
&\le \epsilon^2 \left\| L^{{(t)}}_{\bm{w}^*} \right\|^2_F \left({\mathsf{rank}\left(F}%{\bm{\mathcal{I}}^\mathcal{D}_{\bm{w}^*}\right)}-k\right) \frac{\lambda_k}{\lambda_k + \nicefrac{1}{\left( 2 M \epsilon^2 \right) }}.
\end{align}
\section{Further Experimental Discussion}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/taxinet.png}
\caption{Example input images from the TaxiNet domain, in clear morning (left), and cloudy evening (right) conditions.}
\label{fig:taxinet}
\end{figure}
\input{results_table}
\subsection{Experimental Domains}
\label{app:domains}
We evaluate SCOD{} on several problem settings ranging from classification to regression. Here, we provide implementation details for all the domains. For all results, we report 95\% confidence bounds on AUROC and AUPR computed by bootstrapping. We measure performance on a system with an AMD Ryzen 9 3950X 16-core CPU and an NVIDIA GeForce RTX 2070 GPU.
Within regression, we consider the following datasets.
\begin{itemize}
\item \textbf{Wine} is a dataset from the UCI Machine Learning Repository, in which inputs are various chemical properties of a wine, and labels are a scalar quality score of the wine. We train on the dataset of red wine quality and use the white wine dataset as OoD. The network architecture is a 3-layer fully connected network with ReLU activations, and each hidden layer having 100 units, yielding $N = 11401$. Here
$
T = 604,
k = 100.
$ for SCOD{}. For Local Ensembles, we use $k=20$ using all $M=1000$ datapoints to compute exact Hessian-vector products.
\item \textbf{Rotated MNIST}, where the input is an MNIST digit 2, rotated by a certain angle, and the regression target is the rotated angle. We consider two types of OoD data: the digit 2 rotated by angles outside the range seen at train time, as well as inputs that are other digits from MNIST. The model starts with three conv layers with $3\times3$ filters with a stride of 1. The number of channels in the conv layers is 16,32, and 32, with MaxPooling with a kernel size of 2 between each conv layer. The result is flattened and then processed by a linear layer with hidden dimension of 10 before the linear output layer, yielding a total of $N = 16949$. We use ReLU activations. Here we use
\(
T = 304,
k = 50
\) for SCOD{}. For Local Ensembles, we use $k=50$ eigenvectors and all $M=5000$ datapoints to compute exact Hessian-vector products.
\item \textbf{TaxiNet} which is a network architecture designed to process $3\times260\times300$ RGB input images from a wing-mounted camera and produce estimates of the aircraft's distance in meters from the centerline of the runway as well as its heading in radians relative to the runway, both of which can be used for downstream control during taxiing. It was developed by Boeing as part of the DARPA Assured Autonomy program\footnote{\url{https://www.darpa.mil/program/assured-autonomy}}. The model is based on a ResNet18 backbone, pre-trained on ImageNet, with the last layer replaced with a linear layer to the 2 output dimensions. This yields a total of $N = 11177538$ weights. We fine-tune the network on data collected in the X-Plane 11\footnote{\url{https://www.x-plane.com/}} flight simulator with clear weather and at 9am. Here, we tested against realistic OoD data, by changing weather conditions to cloudy and changing the time-of-day to the afternoon and evening, which change the degree to which shadows impact the scene. Figure \ref{fig:taxinet} visualizes inputs under different conditions. Here $M=30,000$. SCOD{} processes all 30,000 datapoints in its sketch in under 30 mins.
Here,
\(
T = 46,
k = 7
\), for SCOD{}, and for SCOD{} LL,
\(
T = 124,
k = 20
\). For Local Ensembles, we use $k=14$ eigenvectors of the Hessian.
\end{itemize}
For classification, we consider:
\begin{itemize}
\item \textbf{BinaryMNIST}. We consider a binary classification problem created by keeping only the digits 0 and 1 from MNIST. As OoD data, we consider other MNIST digits, as well as FashionMNIST \cite{xiao2017/online}, a dataset of images of clothing that is compatible with an MNIST architecture. The network architecture we use has a convolutional backbone identical to that in the Rotated MNIST, with the flattened output of the conv layers processed directly by a linear output layer to a scalar output, for a total of $N = 14337$ parameters. This output $\bm{\theta}$ is interpreted to be pre-sigmoid activation for logistic regression; i.e. the output parametric distribution is a Bernoulli with the probability of success given by $\mathrm{sigmoid}(\bm{\theta})$. Here, we use
\(
T = 304,
k = 50
\) for SCOD{}. For Local Ensembles, we use $k=50$ eigenvectors of the Hessian, computed with exact Hessian-vector products using all $M=5000$ datapoints.
\item \textbf{MNIST}. We also consider a categorical classification example formulated on MNIST, this time interpreting the output of the network as logits mapped via a softmax to class probabilities. We train on 5-way classification on the digits 0-4. We use digits 5-9 as OoD data, along with FashionMNIST. The network architecture is identical to that of BinaryMNIST, except here the output $\bm{\theta} \in \mathbb{R}^5$ represents the logits such that the probability of each class is given by $\mathrm{softmax}(\bm{\theta})$. This yields $N = 15493$. Here, we use
\(
T = 604,
k = 100
\) for SCOD{}. For Local Ensembles, we use $k=100$ eigenvectors.
\item \textbf{CIFAR10}. To test on larger, more realistic inputs, we consider the CIFAR-10 dataset \citep{krizhevsky2009learning}. Here, unlike the previous experiments, we use a pre-trained DenseNet121 model from \cite{huy_phan_2021_4431043}, and do not train the model from scratch, to highlight that SCOD{} can be applied to any pre-trained model. While this model is trained on all 10 classes of CIFAR-10, we use only the first 5, and thus keep only the first 5 outputs as $\bm{\theta} \in \mathbb{R}^5$, and use them as the pre-softmax logits. This yields a total of $N = 6956426$. We process $M = 5000$ images sampled from the first 5 classes of the train split of CIFAR10, and use the val split as in-distribution examples to test on. For the experiments in the body of the paper, we use as the out-of-distribution dataset TinyImageNet\footnote{\url{http://cs231n.stanford.edu/tiny-imagenet-200.zip}}, a scaled down version of the ImageNet \citep{imagenet_cvpr09} dataset, keeping only 200 classes, and resizing inputs to $32\times32$ RGB images. In the tests in Appendix \ref{app:last-layers}, we also consider OoD data from the 5 held-out classes from CIFAR-10, as well as from the Street View House Numbers dataset \cite{Netzer2011} which has a compatible size.
Here, we use
\(
T = 76,
k = 12
\) for SCOD{}, and
\(
T = 184,
k = 30
\) for SCOD{} (LL). For Local Ensembles, we use $k=20$ eigenvectors.
\end{itemize}
\subsection{Baselines}
We outline details of the implementation of each baseline here:
\begin{itemize}
\item \textbf{Local Ensembles.} We reimplement this baseline in PyTorch, using the \texttt{pytorch-hessian-eigenthings} library \citep{hessian-eigenthings} to compute the top eigenspace of the Hessian. We implement Local Ensembles using a stochastic minibatch estimator for the Hessian-vector product for all domains except for Wine and the MNIST-based domains, where the models are small enough that using the full dataset to compute the Hessian vector product remains computationally feasible. At test time, we use the prediction gradient to compute the extrapolation score, while for multivariate regression and classification problems, we use the loss gradient, sampling possible outputs, computing the gradient on each, projecting it, and then aggregating by taking the minimum over the resulting scores.
\item \textbf{KFAC Laplace.} We use the KFAC Laplace implementation found at \url{https://github.com/DLR-RM/curvature}. We use $30$ samples from the posterior prediction to estimate the Fisher, on a batch size of $32$. We found that in many cases, choosing default values for the norm and scale hyperparameters would lead to singular matrices making. For each experiment, we performed a coarse sweep over these hyperparameters, and chose the best performing set on a validation set to use for the results. Indeed, especially on the classification examples, we found that we required large values of the norm parameter to obtain accurate predictions, which regularizes the posterior towards a delta distribution centered around $\bm{w}^*$.
\item \textbf{Deep Ensemble.} We train $K=5$ models of identical architecture from random initializations using SGD on the same dataset. In all domains where Deep Ensembles are used, the first member of the ensemble is the same model as in the post-training and Naive approaches.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/ablations/AUROC_vs_Meps.pdf}
\caption{AUROC on RotatedMNIST as a function of prior scale $\epsilon^2$.}
\label{fig:Meps-sweep}
\end{figure}
\subsection{Sensitivity to Scale of Prior}
\label{app:Meps-sweep}
We test the impact of the prior scale $\epsilon^2$ on the performance of SCOD{} by performing a sweep on the Rotated MNIST domain, with $T=604, k=100$. The results, shown in Figure \ref{fig:Meps-sweep}, suggest that this has little impact on the performance, unless it is set to be very small corresponding to very tight prior on the weights. This is unsurprising, as, apart from linearly changing the scale of our metric, the term only enters in the diagonal matrix, and has a significant impact if $1/(2M\epsilon^2)$ is of comparable magnitude to the eigenvalues of the Fisher. This is rarely the case for the first $k$ eigenvalues of the Fisher, especially when $M$ is large (here, $M = 5000$).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/lastlayer_tests_cifar.pdf}
\caption{Impact of restricting analysis on the last layers of a network. We see that depending on the OoD dataset, there is a different optimal choice for which layers to consider.}
\label{fig:last-layers}
\end{figure}
\subsection{Limiting SCOD{} to the Last Layers}
\label{app:last-layers}
To explore the impact of only considering the last layers of a network, we consider restricting SCOD{} to different subsets of the layers of the DenseNet121 model used in the CIFAR experiments. We hold the sketching parameters constant, and only vary how much of the network we consider. We denote this by the fraction after LL. For example, LL 0.5 means we restrict SCOD{} to the last 50\% of the network. With this notation, LL 1.0 represents considering the full network, which is simply SCOD{}. At the other extreme, we consider restricting our analysis to just the last linear layer of the model, which we denote ``SCOD{} (only linear).'' We consider performance with several different OoD datasets. Figure \ref{fig:last-layers} shows the results. The first OoD dataset is simply the other classes of CIFAR10 which we did not use to create the sketch. We see that here, using just the last linear layer performs best. Indeed, as the network was trained on all 10 classes, it is not surprising that only the last layer analysis provides any meaningful separation between the classes. Including the earlier layers ``dilutes'' this signal, and, for a fixed sketch budget, considering more layers leads to worse performance. On SVHN and TinyImageNet, we see that the optimal choice of the layers to consider is somewhere in between the two extremes. Here, it is possible that as we increase the fraction of the network that we analyze, we first gain the benefits of the information stored in the last layers, and then start to suffer the consequences of approximation error in the sketching, which depends on the size of the original matrix. We also visualize on this plot the performance of the version of SCOD{} (LL) used to produce the results in the body of the paper. As is evident, we did not use the results of this sweep to optimally choose the number of layers to consider for the experiments in the body.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/error_based.pdf}
\caption{Error-based OoD detection on TaxiNet, where inputs where the DNN made an error greater than 0.5 in Mahalanobis distance are deemed to be out-of-distribution, i.e., out of the DNN's domain of competency.}
\label{fig:error-based}
\end{figure}
\subsection{Error-based OoD detection}
\label{app:error-based}
In Figure \ref{fig:runtime-vs-auroc}, we measure how informative an uncertainty measure is with respect to how well it can classify examples that come from an altogether different dataset as anomalous. However, ideally, we would like this measure of uncertainty to also correspond to the network's own accuracy. To test this, we use the TaxiNet domain as a test case, where we construct a dataset which includes images from all day, morning and afternoon, on a clear day. The training dataset consists only of images from the morning, so this all-day dataset shares some support with the training dataset. Moreover, the differences in these inputs are quite subtle, as the images remain brightly lit, though some shadows appeare in the afternoon. We choose an error threshold, and consider inputs for which the network has an error (measured in Mahalanobis distance) less than this threshold to be ``in-distribution'' and those where the network has a high error to be ``out-of-distribution.''
Figure \ref{fig:error-based} shows the results of this experiment. We see that even in this setup, SCOD{} produces very high AUROC scores, similar to Deep Ensembles, and outperforms all post-training baselines. In this experiment, we find that applying SCOD{} to a single pre-trained DNN almost perfectly characterizes the DNNs domain of competency.
\subsubsection*{References}}
\usepackage{mathtools}
\usepackage{booktabs}
\usepackage{tikz}
\newcommand{\swap}[3][-]{#3#1#2}
\include{macros}
\title{Sketching Curvature for Efficient Out-of-Distribution \\ Detection for Deep Neural Networks}
\author[1]{\href{mailto:Apoorva Sharma <<EMAIL>>?Subject=SCOD}{Apoorva Sharma}{}}
\author[1,2]{Navid Azizan}
\author[1]{Marco Pavone}
\affil[1]{%
Stanford Universit
}
\affil[2]{%
Massachusetts Institute of Technolog
}
\begin{document}
\maketitle
\begin{abstract}
\input{0_abstract}
\end{abstract}
\section{Introduction}
\label{intro}
\input{1_intro}
\section{Problem Statement}
\label{prob_statment}
\input{2_problem_statement}
\section{Background: Curvature and Laplace Approximation}
\label{prelim}
\input{3_preliminaries}
\section{Proposed Method: Sketching Curvature for OoD Detection{}}
\label{approach}
\input{4_approach}
\section{Related Work}
\label{relatedwork}
\input{5_related_work}
\section{Experimental Results}
\label{experiments}
\input{6_experiments}
\section{Conclusion}
\label{conclusion}
\input{7_conclusions}
\begin{acknowledgements}
A.S. and M.P. were supported in part by DARPA under the Assured Autonomy program and by NASA under the University Leadership Initiative program. N.A. was supported by MIT. The authors wish to thank Robin Brown and Edward Schmerling for helpful input and discussions during the development of these ideas.
\end{acknowledgements}
|
\section{Introduction}
\label{sec::intro}
\IEEEPARstart{C}{urrent}
mobile robot navigation methods can navigate a robot from one point to another safely and reliably in structured and homogeneous environments~\cite{quinlan1993elastic, fox1997dynamic}, such as indoor hallways or outdoor paved surfaces. These consistent environments allow the robots to use simple kinodynamic motion planners independent of the environment, thanks to the limited environment disturbances and stochasticity.
Dating back at least to DARPA's Grand Challenge~\cite{seetharaman2006unmanned} and LAGR (Learning Applied to Ground Vehicles)~\cite{jackel2006darpa} program, researchers have also looked into applying autonomous navigation in unstructured outdoor environments. Challenges arise from multiple fronts in those natural spaces, but most off-road navigation work focused on perception, e.g., detecting natural obstacles~\cite{jackel2006darpa}, classifying underlying terrain types~\cite{bai2019three, shi2020laplacian, mei2019comparative}, or building semantic maps~\cite{maturana2018real, wolf2020advanced, sharma2019semantic}. For motion control, most off-road robots simply travel at low speeds to minimize uncertainty and to maximize safety~\cite{rabiee2019friction, tian2014control, rogers2012aiding, seyr2006proprioceptive}.
While recent advances in deep learning provide roboticists with a different avenue to investigate those perception problems in off-road navigation, researchers also started to combine perception, planning, and motion control using end-to-end learning in unstructured environments~\cite{muller2006off, giusti2015machine}. These systems do not require a heavily-engineered navigation pipeline, and can react to natural environments in a data-driven approach. Although these methods can enable successful navigation, they are data-intensive and and generally do not lead to better navigation than their classical counterparts.
Focusing on the motion control side of off-road navigation on unstructured terrain, the contribution of this paper is to use learning to capture the effect of complex and unknown environmental factors on the robot's low-level kinodynamic model.
We use onboard inertial observation to encode environmental factors and learn an inverse kinodynamic model to produce fast and accurate low-level control inputs. Using our method, even with extensive disturbances caused by high-speed terrain-vehicle interaction, the robot is still able to perform fast and accurate off-road navigation (Fig. \ref{fig::backyard}), compared to a kinodynamic model based on ideal assumptions and a learned model that does not consider environmental factors.
\begin{figure}
\centering
\includegraphics[width=1\columnwidth]{contents/figures/backyard.jpg}
\caption{The UT Automata scale 1/10th autonomous vehicle drives an eight-turn (T1--T8 on the red path) outdoor race track with unstructured terrain. Close-ups of some terrain (black inserts) are shown, including cement, mud, grass, which are covered by leaves, stalks, and/or twigs at different densities. For high-speed terrain-aware navigation, the robot has to interact with these unstructured terrain (red insert). }
\label{fig::backyard}
\end{figure}
\section{Learning Inverse Kinodynamic Models}
\label{sec::approach}
Most navigation systems either assume kindodynamic models to be independent of the environment, or that there exists a discrete set of environment classes~\cite{siva2019robot}, e.g., one model for paved terrain and another for grass. In this work, we relax these assumptions by learning a single continuous inverse kinodynamic model that accounts for environmental factors across different terrain types without having to perform discrete terrain classification, or analytic modelling. The learned model takes as input inertial observations that make the impact of environmental factor on kinodynamic motion observable (e.g., how bumpiness from gravel will result in understeer at high speeds). This inverse kinodynamic model is learned in a data-driven manner.
\subsection{Problem Formulation}
Given vehicle state $x$,
control input $u$, and world state $w$, the state dynamics and observation $y$ are given by
\begin{equation} \label{eqn::dynamics}
\dot{x} = f(x, u, w),
\qquad y = g(x, w),
\end{equation}
where $f(\cdot, \cdot, \cdot)$ is the system's forward kinodynamic function, while $g(\cdot, \cdot)$ is the observation function. Note that in most cases, $w$ is not directly observable and cannot be easily modeled.
A navigation planner generates a global plan $\Pi: [0,1] \rightarrow X$ mapping from a unitless progress variable $s\in [0,1]$ to planned vehicle state $x \in X$, incorporating both global (\textit{e.g.,}{} traversable map) and local (\textit{e.g.,}{} sensed obstacles) information to take the robot from the start state $\Pi(0)$ to the goal state $\Pi(1)$. A projection operator $\rho: X\rightarrow [0,1]$ maps the robot state $x$ (e.g., from localization) to infer the progress variable $s$ (i.e., the robot’s progress along the global plan so far), such that the closest state in the plan to a robot state $x$ is $\Pi(\rho(x))$. For simplicity of notation, we represent the projected state at any time as $x_\Pi=\Pi(\rho(x))$. We also omit the explicit time-dependence of variables $x(t)$, $u(t)$, and $y(t)$, denoting them simply as $x$, $u$, and $y$. The objective of our controller $u$ is thus to minimize the total navigation time $T$ while following the plan precisely, as represented by the joint cost function
\begin{align}
J = T + \gamma \int_0^T ||x(t) - x_\Pi(t)||^2 dt.
\label{eqn::objective}
\end{align}
Here, $\gamma$ is a hyperparameter that trades total navigation time for execution accuracy.
We formulate the solution to this optimal control problem as a receding-horizon controller $u^*$ over a unitless progress horizon (along the global plan) $\Delta$ and corresponding time-horizon $\Delta t$ such that the control input drives the robot state from $x$ to the receding horizon plan state $\Pi(\rho(x)+\Delta)$ over time-period $\Delta t$:
\begin{align}
&u^* = \argmin_u \left( \Delta t + \gamma \left\lVert \Delta x_\Pi - \int_0^{\Delta t} f(x, u, w)dt \right\rVert^2 \right), \nonumber \\
&\Delta x_\Pi = \Pi(\rho(x)+\Delta) - x, \label{eqn::controller}
\end{align}
where $\Delta x_\Pi$ is the state change between the receding horizon plan and current vehicle state.\footnote{In general, the minus signs in Eqn. \ref{eqn::controller} is the generalized difference operator $\ominus$ over Special Euclidean Group $\mathrm{SE}(n)$ and the corresponding Lie Algebra.} The optimal control $u^*$ can be solved using the receding horizon inverse kinodynamic model $f^{-1}$ as
\begin{align}
u^* = f^{-1}(\Delta x, x, w), \label{eqn::ikd}
\end{align}
that takes as input the desired relative state change $\Delta x$, the current robot state $x$, and world state $w$. Unfortunately, it is hard to express $f^{-1}$ accurately via analytical models, and even if it could be expressed accurately, computing $u^*$ is error-prone since the world state $w$ is not directly observable.
\subsection{Learning Inverse Kinodynamics}
In this work, in order to enable fast and accurate navigation under the influence of different terrain interactions from the world state $w$, we adopt a data-driven approach to capture the effect of $w$. Specifically, we introduce the function
\begin{equation}
f^{+}_\theta(\Delta x, x, y) \approx f^{-1}(\Delta x, x, w),
\end{equation}
parameterized by $\theta$, as an approximation for the original receding horizon inverse kinodynamic function (superscript $^+$ denotes pseudo inverse).
The key insight in this approximation is that the impact of $w$ on $u^*$ becomes predictable given observations $y$ related to high-speed terrain-aware navigation---in our case we use onboard inertial sensing to capture speed-dependent terrain interaction.
To learn $f^{+}_\theta$, a training dataset $\mathcal{T}$ with $N$ samples
\[
\mathcal{T} = \{\langle \Delta x^i, x^i, u^i, y^i\rangle_{i=1}^{N}\}
\]
is desired, using the optimal but unknown receding horizon inverse kinodynamic function $u^i=f^{-1}(\Delta x^i, x^i, w^i)$ and observation function $y^i = g(x^i, w^i)$ from Eqns.~\ref{eqn::dynamics} and \ref{eqn::ikd}. Unfortunately, we neither know $f^{-1}$, nor do we know the world states $w^i$. However, we do have access to $f$ as a black-box function via real-world execution: we can simply pick arbitrary sample controls $u^i$ at corresponding starting states $x^i$, and observe the resulting state change $\Delta x^i$ after the chosen receding horizon $\Delta t$, including the impact of the unknown $w^i$:
\[
\Delta x^i = \int_0^{\Delta t} f(x^i,u^i,w^i)dt.
\]
Thus, the original chosen control $u^i$ is
the control\footnote{Here, we assume the optimal control that follows the global path (second term in Eqn. \ref{eqn::objective}) also implicitly minimizes navigation time (first term).}
for the resulting state change $\Delta x^i$ from the original state $x^i$, for the corresponding but unknown (and hence unrecorded) world state $w^i$. Along with the corresponding sensor observation $y^i$, we generate each sample $i$ for dataset $\mathcal{T}$.
To ensure that the learned parameters $\theta$ of $f^{+}_\theta$ approximate $f^{-1}$ accurately at states that the robot will encounter during execution, $\mathcal{T}$ must include representative samples for $x$, $u$, and $y$.
With the collected dataset $\mathcal{T}$, we formulate deriving $f^{+}_\theta(\cdot, \cdot, \cdot)$ as a learning problem by minimizing a supervised loss:
\begin{equation}
\begin{split}
\theta^* &= \argmin_{\theta} \sum_{(\Delta x^i, x^i, y^i) \in \mathcal{T}} \lVert f^{-1}(\cdot, \cdot, \cdot) - f^{+}_\theta(\Delta x^i, x^i, y^i)\rVert_H \\
&= \argmin_{\theta} \sum_{(u^i, \Delta x^i, x^i, y^i) \in \mathcal{T}} \lVert u^i - f^{+}_\theta(\Delta x^i, x^i, y^i)\rVert_H,
\end{split}
\label{eqn:bc}
\end{equation}
where $||v||_H = v^THv$ is the norm induced by positive definite matrix $H$, used to weigh the learning loss between the different dimensions of the control input $u^i$. We represent $f^{+}_\theta(\cdot, \cdot, \cdot)$ as a neural network and can therefore use gradient decent to find an approximately optimal $\theta^*$. In this work, we collect raw 6-DoF readings from an onboard IMU, and construct $y$ by feeding inertial data through an autoencoder to encode relevant terrain-vehicle interaction at different driving speeds. More details are provided in Sec. \ref{sec::experiments}.
\subsection{Online Execution}
The learned inverse kinodynamic model $f^{+}_{\theta^*}(\cdot, \cdot, \cdot)$ provides a means to approximately account for $w$ using the onboard observation $y$ for fast, terrain-aware, and precise navigation. At each time step $t$ during online execution,
we compute the desired change of state $\Delta x_\Pi = \Pi(\rho(x)+\Delta) - x$ with $x$ from localization, the projection operator $\rho(\cdot)$, projection horizon $\Delta$, and global plan $\Pi(\cdot)$.
Along with onboard observation $y$ and current vehicle state $x$, we use the learned inverse kinodynamic model $f^{+}_{\theta^*}$ to produce system control input:
\begin{equation}
u(t) = f^{+}_{\theta^*}(\Delta x_\Pi, x, y),
\end{equation}
and repeat this process for every time step.
\section{Conclusions}
\label{sec::conclusions}
In this paper, we present a data-driven approach to learn an inverse kinodynamic model for accurate high-speed navigation on unstructured terrain. To capture the elusive and stochastic world state caused by vehicle-terrain interaction at different high speeds, we use an inertia-based observation embedding as an input to the learned inverse kinodynamic function. This approach is tested on a physical robot on seen and unseen terrain with different global plans at different high speeds. The experimental results show that the learned model can significantly outperform an ideal baseline model without consideration of world state. Our ablation study also shows our observation embedding is useful to enable fast and accurate off-road navigation on unstructured terrain. For future work, better ground truth linear velocity estimation needs to be investigated: in addition to wheel odometry alone, other sources of perception, e.g., vision, point cloud, and/or inertia, can be leveraged. Better linear velocity estimation can account for significant wheel slippage on more challenging terrain, e.g., on ice, and enable even faster navigation. Adding vision-based observation also has the potential to enable the robot to prepare for future interactions, e.g., to reduce the failures at Turn 8. Another interesting direction to investigate in the future is generalization from easier to harder environments.
\section{Related Work}
\label{sec::related}
In this section, we review related off-road navigation work in terms of perception and motion control.
\subsection{Off-road Perception}
The first challenge arises from off-road navigation is perception. In unstructured off-road environments, perception is no longer simply in the geometric sense (e.g., free vs. obstacle), but also requires semantic information (e.g., grass vs. mud).
A plethora of research in terrain classification has leveraged vibration-based signals to classify terrain types~\cite{bai2019three, shi2020laplacian, mei2019comparative}. Vision-based sensors, e.g., LiDAR and camera, combined with current deep learning methods, have also been used to build semantic maps~\cite{maturana2018real, wolf2020advanced, sharma2019semantic}.
These perception methods assign costs to discrete terrain classes for planning, but do not consider robot's kinodynamic model when moving on these terrain. Our work does not distinguish among discrete terrain classes and uses observations collected during interactions with different terrain to enable fast and accurate kinodynamic planning.
\subsection{Off-road Motion Control}
Although research thrust for off-road navigation has been primarily focused on perception, roboticists have also investigated off-road navigation from the motion control side. Many wheel slip models~\cite{rabiee2019friction, tian2014control, rogers2012aiding, seyr2006proprioceptive} have been developed and used to design controllers to minimize slip. Most of these models treat slip only as a function of the vehicle kinematics. But to achieve high-speed off-road navigation, slip is inevitable and also highly dependent on the underlying terrain.
Researchers have also used machine learning for motion control in off-road navigation. A recent survey~\cite{xiao2020motion} pointed out that learning is most efficient when targeting at navigation components, e.g., learning local planners~\cite{wangagile, xiao2021agile, xiao2020toward, liu2020lifelong, chiang2019rl, kontoudis2019kinodynamic}, costmaps~\cite{wigness2018robot, richter2017safe}, or planner parameters~\cite{xiao2021appl, xu2020applr, wangapple, wang2020appli, xiao2020appld}.
Research on using learning for motion control in off-road scenarios is scarce. Pan, et al.~\cite{pan2020imitation} enabled high-speed navigation with end-to-end imitation learning from RGB input in a closed circular dirt track. The expert demonstrator is a model predictive controller with access to high-precision sensors including GPS-RTK. The end-to-end learning approach most likely does not generalize well to other terrain and tracks. Aiming at a variety of terrain, Siva, et al.~\cite{siva2019robot} used imitation learning from human demonstration to navigate five discrete terrain types (concrete, grass, mud, pebble, rock).
In contrast, our method only targets at learning a kinodynamic model and can navigate any global path. We also do not intentionally separate terrain into discrete types, and treat different terrain characteristics in a continuous manner. Reinforcement learning has also shown potential in learning motion control for navigation, but at the cost of extensive training overhead: Brunnbauer, et al. \cite{brunnbauer2021model} took 8 million time steps (37 hours of interactions) to learn a simple \emph{kinematic} model. In comparison, our approach only requires 30 minutes to learn a \emph{kinodynamic} model, which would require far more training by such reinforcement learning approaches.
\section{Experiments}
\label{sec::experiments}
In this section, we present experimental results using a learned inverse kinodynamic model $f^{+}_{\theta^*}(\Delta x, x, y)$, which considers unobservable world state $w$ by taking $y$ as input, and can precisely track different global plans by different $\Pi$.
We denote the baseline forward and inverse kinodynamic functions, which do not consider world state $w$, as $f_B(x, u)$ and $f^+_B(\Delta x, x)$, respectively. As an ablation study to test the effectiveness of capturing the world state $w$ with $y$, we also learn an ablated inverse kinodynamic model $f^+_{A\phi^*}(\Delta x, x)$, which is parameterized by $\phi^*$ and does not take observation $y$ as input to represent $w$.
Note that the baseline represents classical model-based local planners \cite{fox1997dynamic}, while the ablation is equivalent to existing learning-based local planners, e.g., Imitation Learning \cite{pfeiffer2017perception}.
We show that the learned $f^+_{A\phi^*}$ can outperform the baseline $f^+_B$. Adding the learned observation $y$ as another input, $f^{+}_{\theta^*}(\Delta x, x, y)$ can outperform both $f^+_{A\phi^*}(\Delta x, x)$ and $f^+_B(\Delta x, x)$.
The learned inverse kinodynamic model for online execution is also agnostic to different online plans and unseen terrain. We show that the model learned through a global planner $\Pi_1$ (a randomly exploring policy driven by a human operator) can also generalize well to a complicated global planner $\Pi_2$ on an outdoor race track, and a simple global planner $\Pi_3$ on an indoor track.
We also test the generalizability with respect to unseen world states through the encoded $y$ and $f^{+}_{\theta^*}(\Delta x, x, y)$ on unseen terrain.
\subsection{Implementation}
\subsubsection{Robot Platform} Our learned inverse kinodynamic model is implemented on a UT Automata robot, a scale 1/10th autonomous vehicle platform (Fig. \ref{fig::ut_automata}). The Ackermann-steering four-wheel drive vehicle is equipped with a 2D Hokuyo UST-10LX LiDAR for localization, a Vectornav VN-100 IMU for inertial sensing (200Hz), a Flipsky VESC 4.12 50A motor controller, and a Traxxas Titan 550 Motor. Although the platform has individual suspensions, the relatively short-travel suspensions are not specifically designed for off-road navigation. We specifically pick this platform because its small size and weight and the lack of designed off-road capability can maximize the difference in accuracy between navigating with and without the learned terrain-aware inverse kinodynamic model. The robot has a NVIDIA Jetson onboard, but only the CPU is used during deployment.
\begin{figure}
\centering
\subfloat[]{\includegraphics[width=0.5\columnwidth]{contents/figures/ut_automata.jpg}
\label{fig::ut_automata}}
\subfloat[]{\includegraphics[width=0.5\columnwidth]{contents/figures/graph_nav_cut.jpg}%
\label{fig::graph_nav_cut}}
\caption{(a): A UT Automata robot, scale 1/10th autonomous vehicle platform used in the experiments. (b): The sampling-based baseline approximates $\Delta x_\Pi$ by rolling out the optimal $u_B$ among 100 samples with $f_B$.}
\label{fig::ut_automata_and_graph_nav}
\vspace{-10pt}
\end{figure}
\subsubsection{Environment} The environment comprises of cement, grass, and mud; and some patches are covered by different artifacts, such as leaves, stalks, and/or twigs, with different densities (Fig. \ref{fig::backyard}). For a small vehicle like the UT Automata robot, these artifacts can cause significantly different world states (Fig. \ref{fig::backyard} red insert). Note that the terrain also changes due to environmental factors such as sunlight, wind, and moisture, and is also affected by the robot's wheel and chassis. To minimize the effect of artifacts being pushed off the course during extended experiments, we frequently shuffle and redistribute the artifacts. We do not specify discrete terrain classes and treat the terrain characteristics in a continuous manner.
\subsubsection{Model Implementation} During autonomous navigation, the robot uses Episodic Non-Markov Localization (ENML)~\cite{biswas2017episodic} with a pre-built map of the environment to derive vehicle state $x$. A global planner $\Pi$ includes a pre-generated global path for the robot to follow and uses line-of-sight control (similar to~\cite{xiao2017uav, xiao2015locomotive}) to generate desired receding horizon plan state $\Pi(\rho(x)+\Delta)$ on the global path 1m away from the robot.
In a model predictive control manner, the robot uses the baseline forward kinodynamic function $f_B$ and samples candidate velocity and curvature control inputs $u\in U$ evenly distributed within a physically-feasible window to jointly find the desired state change $\Delta x_\Pi$ and control input $u_B$ (shown in Fig. \ref{fig::graph_nav_cut}).
More specifically, to compute control input, the baseline inverse kinodynamic model $f^+_B$ produces the curvature input, which results in the desired $\Delta x_\Pi$ and drives the robot as close to $\Pi(\rho(x)+\Delta)$ as possible:
\begin{equation}
\begin{split}
u_B & = f^+_B(\cdot, \cdot)\\
& = \argmin_{u}\lVert\Pi(\rho(x)+\Delta) - \int_{t=0}^{\Delta t} f_B(x, u)dt \rVert,
\end{split}
\end{equation}
for the second term in Eqn. \ref{eqn::objective}, and selects the fastest possible velocity for the first term $T$, considering the robot's acceleration limit and a safety distance to decelerate in case of obstacles.
For our learned ablated and final model, $f^+_{A\phi^*}$ and $f^{+}_{\theta^*}$, we utilize the $\Delta x_\Pi$ from the baseline kinodynamic model (corresponds to $u_B$), but instead of using the baseline's control input, we query our learned models to produce
$u = f^+_{A\phi^*}(\Delta x_\Pi, x)$
or
$u = f^{+}_{\theta^*}(\Delta x_\Pi, x, y)$. In practice, we use the baseline control input $u_B = \{v, c\}$ (linear velocity and steering curvature) to represent the desired state change rate $\Delta x_\Pi$.
\subsubsection{Data Collection} To collect training data, the robot is teleoperated with a joystick in an open environment with linear velocity $v \in [0, 3]\mathrm{m/s}$ and steering curvature $c \in [-1.35, 1.35]\mathrm{m^{-1}}$ for 30 minutes (24418 data points). The teleoperator randomly varies both linear velocity and steering curvature ($\Pi_1$).
In our specific implementation, the robot reasons in the robot frame and therefore the state $x^i$ in the training trajectory $\mathcal{T} = \{\langle \Delta x^i, x^i, u^i, y^i\rangle_{i=1}^{N}\}$ becomes the origin in the robot frame.
The ground truth $\Delta x^i$ is represented as real $\{v^i_r, c^i_r\}$, where $v^i_r$ is from vehicle odometry and $c^i_r = \omega^i_r/v^i_r$ ($\omega^i_r$ is the sensed angular velocity around the vertical $z$ axis from the IMU). Currently, we take $v^i_r$ from wheel odometry only, which can be further improved by adding visual, point cloud, and/or inertial information in future work. The commanded control input $u^i=\{v^i_c, c^i_c\}$ is recorded from joystick input. For $y$, we collect the 6-DoF raw IMU signal, including 3-DoF accelerometer and 3-DoF gyroscope, as a sliding history window.
\subsubsection{Network Architecture} As a function approximator for $f^{+}_{\theta^*}$, we use a two-layer neural network with 32 neurons each layer (shown in purple in Fig. \ref{fig::nn}) and we empirically show that such a small network is very efficient to train and suffices to improve off-road navigation.
The neural network takes realistic/desired $\{v^i_r, c^i_r\}$ (as a proxy for $\Delta x^i$, Fig. \ref{fig::nn}, orange) and observation $y$ as input, and outputs to-be-commanded control input $u^i=\{v^i_c, c^i_c\}$.
For observation $y$, we concatenate the last 100 IMU readings (0.5s) into a 600-dimensional vector, and feed it into two 256-neuron layers as an autoencoder (Fig. \ref{fig::nn}, blue).
The final embedding for $y$ is a two dimensional vector, then concatenated with $\{v^i_r, c^i_r\}$, and finally trained in an end-to-end fashion.
The entire network architecture is shown in Fig. \ref{fig::nn}. For the ablated model $f^+_{A\phi^*}$, the two-dimensional $y$ embedding is removed (only the orange and purple components remain). Training both models takes less than five minutes on a NVIDIA GeForce GTX 1650 laptop GPU. During runtime, the trained model is used onboard the robot's Jetson CPU with \texttt{libtorch}.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{contents/figures/nn.jpg}
\caption{Neural Network Architecture. Input: blue IMU encoder and orange desired state change (desired velocity and curvature in practice); Output: purple learned function approximator $f^{+}_{\theta^*}$ as the inverse kinodynamic model. }
\label{fig::nn}
\vspace{-10pt}
\end{figure}
\begin{figure*}
\centering
\subfloat[1.6m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/1.6.jpg}%
\label{fig::1.6}}
\subfloat[1.7m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/1.7.jpg}%
\label{fig::1.7}}
\subfloat[1.8m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/1.8.jpg}%
\label{fig::1.8}}
\subfloat[1.9m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/1.9.jpg}%
\label{fig::1.9}}
\subfloat[2.0m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/2.0.jpg}%
\label{fig::2.0}}\\
\subfloat[2.1m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/2.1.jpg}%
\label{fig::2.1}}
\subfloat[2.2m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/2.2.jpg}%
\label{fig::2.2}}
\subfloat[2.3m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/2.3.jpg}%
\label{fig::2.3}}
\subfloat[2.4m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/2.4.jpg}%
\label{fig::2.4}}
\subfloat[2.5m/s]{\includegraphics[width=0.4\columnwidth]{contents/figures/2.5.jpg}%
\label{fig::2.5}}
\caption{Results of Outdoor Experiments on Seen Terrain: Localized robot positions of all 300 laps are plotted around the pre-defined global path (black line segments) on the map (grey lines). The size of the circles at each turn denotes the number of failures at that turn. Red: baseline $f^+_B(\Delta x, x)$, blue: ablated model $f^+_{A\phi^*}(\Delta x, x)$, green: learned model $f^{+}_{\theta^*}(\Delta x, x, y)$.}
\label{fig::all_data}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{contents/figures/failure_per_velocity.jpg}
\caption{Failure Rate Per Target Speed}
\label{fig::failure_per_velocity}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{contents/figures/failure_per_turn.jpg}
\caption{Failure Rate per Turn}
\label{fig::failure_per_turn}
\vspace{-10pt}
\end{figure}
\begin{figure*}
\centering
\subfloat[Unseen Test]{\includegraphics[width=0.25581395348\columnwidth]{contents/figures/living_room.jpg}%
\label{fig::living_room}}
\subfloat[2.4m/s]{\includegraphics[width=0.3488372093\columnwidth]{contents/figures/generalizability_2.4.jpg}%
\label{fig::2.4_unseen}}
\subfloat[2.5m/s]{\includegraphics[width=0.3488372093\columnwidth]{contents/figures/generalizability_2.5.jpg}%
\label{fig::2.5_unseen}}
\subfloat[2.6m/s]{\includegraphics[width=0.3488372093\columnwidth]{contents/figures/generalizability_2.6.jpg}%
\label{fig::2.6_unseen}}
\subfloat[2.7m/s]{\includegraphics[width=0.3488372093\columnwidth]{contents/figures/generalizability_2.7.jpg}%
\label{fig::2.7_unseen}}
\subfloat[2.8m/s]{\includegraphics[width=0.3488372093\columnwidth]{contents/figures/generalizability_2.8.jpg}%
\label{fig::2.8_unseen}}
\caption{Experiment Results in Unseen Environment}
\label{fig::all_data_unseen}
\vspace{-10pt}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{contents/figures/failure_per_velocity_unseen.jpg}
\caption{Failure Rate per Target Speed on Unseen Terrain}
\label{fig::failure_per_velocity_unseen}
\vspace{-10pt}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{contents/figures/failure_per_turn_unseen.jpg}
\caption{Failure Rate per Turn on Unseen Terrain}
\label{fig::failure_per_turn_unseen}
\vspace{-10pt}
\end{figure}
\subsection{Navigation on Seen Terrain}
We first test the inverse kinodynamic model's performance on the same terrain where the training data is collected, but with a different global planner $\Pi_2$. After collecting the training data with $\Pi_1$, an outdoor race track is constructed using plastic panels and wooden posts (Fig. \ref{fig::backyard}). Starting from the origin (robot location in Fig. \ref{fig::backyard}), eight turns are created (T1--T8, $\Pi_2$). While Turn 1, 2, and 3 are relatively gentle \text{left-,} right-, and left-hand turns, Turn 4 and Turn 8 are roughly 90$^{\circ}$ left-hand turns. Turn 5, 6, and 7 are sharp 180$^{\circ}$ left-, right-, and left-hand turns.
Ten different target speeds (the maximum speed the robot targets at reaching while maintaining safety tolerance to decelerate and avoid potential collisions) are tested, ranging from 1.6m/s to 2.5m/s with 0.1m/s intervals. For the three models, the baseline $f^+_B(\Delta x, x)$, the ablated $f^+_{A\phi^*}(\Delta x, x)$, and learned $f^{+}_{\theta^*}(\Delta x, x, y)$, we repeat ten trials/laps each for statistical significance. A total 300 laps are executed. The localized robot position from ENML are shown in Fig. \ref{fig::all_data} in the subplots corresponding to the target speeds.
At lower target speeds, the green trajectories by the learned model $f^{+}_{\theta^*}(\Delta x, x, y)$ are much closer to the pre-defined global path, compared to the baseline $f^+_B(\Delta x, x)$, because the latter model fails to consider the world state caused by the unstructured terrain.
With increasing speed, the robot trajectory becomes more scattered around the global path due to increased stochasticity from vehicle-terrain interaction. But overall speaking, the green trajectories are always closer to the global path than other alternatives. The blue trajectory is generated by the ablated model $f^+_{A\phi^*}(\Delta x, x)$. Like $f^{+}_{\theta^*}$, it learns from actual terrain interactions, but it does not consider the current observation $y$. So $f^+_{A\phi^*}$ is roughly an averaged model over the continuous spectrum of terrain. Therefore, $f^+_{A\phi^*}$ outperforms the baseline $f^+_B$, but underperforms $f^{+}_{\theta^*}$ because it fails to consider the current world state.
At each turn, the size of the red, blue, and green circles represents the number of failed turns (collision or getting stuck) in the ten attempted turns. Turn 6 and 7 cause a lot of trouble for the baseline even at lower speeds. With increasing speed, more turns cause failure for other models as well, but in general, the baseline fails more frequently at most turns than the ablated and learned models. Fig. \ref{fig::failure_per_velocity} and Fig. \ref{fig::failure_per_turn} show the percentage of failed turns per target speed and per turn, respectively. In Fig. \ref{fig::failure_per_velocity}, failure rate increases with faster speed, while within each speed, the learned model achieves the lowest failure rate, while the baseline fails most frequently. In Fig. \ref{fig::failure_per_turn}, Turn 6 is the most difficult for all three alternatives, and at most turns, the learned model outperforms the ablation and the baseline. Since Turn 8 is immediately after the terrain change from grass to cement, the robot sometimes oversteers (to compensate for slip on grass) and does not react quickly enough to understeer (for higher friction on cement), causing it to get stuck in a few laps with the learned model. This problem can be addressed by adding forward looking camera to predict future wheel-terrain interaction in future work. The overall success rates of the three models for all turns are shown in the first row in Tab. \ref{tab::success_rates}.
\begin{table}
\centering
\caption{Overall Success Rates of All Speeds and Turns}
\begin{tabular}{cccc}
\toprule
& Baseline & Ablation & \textbf{Learning}\\
\midrule
Seen & 52.4\% & 82.9\% & \textbf{86.9}\% \\
Unseen & 49.5\% & -- & \textbf{87.0}\% \\
\bottomrule
\end{tabular}
\label{tab::success_rates}
\vspace{-12pt}
\end{table}
\subsection{Navigation on Unseen Terrain}
To test that the learned model generalizes to different global planners $\Pi$ and also to unseen terrain, we further conduct an indoor experiment with a different track and global path ($\Pi_3$) on an unseen wooden floor (Fig. \ref{fig::living_room}). Note that $f^{+}_{\theta^*}$ has only seen training data from random exploration on the outdoor terrain (Fig. \ref{fig::backyard}). Since the unseen wooden floor is relatively more consistent and therefore easier to navigate than the outdoor unstructured terrain,\footnote{We speculate that the generalization would not be as good were the model trained indoors (on easy terrain) but applied outdoors.} we increase the navigation target speed to 2.4m/s - 2.8m/s, also with 0.1m/s intervals. The baseline and the learned model are applied with these five different target speeds, ten repetitions each. Fig. \ref{fig::all_data_unseen} shows the results from the 100 laps on the unseen terrain with a different global path. Similar to the results on seen terrain, the learned model produces more concentrated and also closer robot trajectories to the global path to be tracked. As shown in Fig. \ref{fig::failure_per_velocity_unseen} and \ref{fig::failure_per_turn_unseen}, the learned model also outperforms the baseline in terms of failure rate at all target speeds and in most turns (except Turn 2). The overall success rates of the baseline and learned model are shown in the second row in Tab. \ref{tab::success_rates}.
\section*{ACKNOWLEDGMENTS}
This work has taken place in the Learning Agents Research Group (LARG) and Autonomous Mobile Robotics Laboratory (AMRL) at UT Austin. LARG research is supported in part by NSF (CPS-1739964, IIS-1724157, NRI-1925082), ONR (N00014-18-2243), FLI (RFP2-000), ARO (W911NF-19-2-0333), DARPA, Lockheed Martin, GM, and Bosch.
AMRL research is supported in part by NSF (CAREER-2046955, IIS-1954778, SHF-2006404), ARO (W911NF-19-2-0333), DARPA (HR001120C0031), Amazon, JP Morgan, and Northrop Grumman Mission Systems.
Peter Stone serves as the Executive Director of Sony AI America and receives financial compensation for this work. The terms of this arrangement have been reviewed and approved by the University of Texas at Austin in accordance with its policy on objectivity in research.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:intro}
Recent observations of circumstellar discs by ALMA have revealed a rich variety of substructure. Features such as gaps and asymmetries
\citep{2015ApJ...808L...3A, 2016Sci...353.1519P, PhysRevLett.117.251101, 2016ApJ...820L..40A, 2016Natur.535..258C} in the emission provide
diagnostics for the physical processes that drive the evolution of the discs. In many gas-rich systems, these gap features are argued to indicate the
presence of a giant planet, either embedded in the disc \citep{2015MNRAS.453L..73D} or orbiting adjacent to it, although the details of gas-dust
interactions can have strong effects on the resulting morphologies
\citep{2018ApJ...866..110D}. In gas-poor/debris systems, a giant
planet perturber can also influence
the structure of the dust continuum emission. For example, a misaligned giant planet can produce nonaxisymmetric features such as warps
\citep{2001A&A...370..447A}, highly eccentric perturbers can produce structures through secular interactions
\citep{2014MNRAS.443.2541P, 2015MNRAS.448.3679P}, and mean motion resonances (MMRs) can open gaps
\citep{2015ApJ...798...83N, 2016ApJ...818..159T, 2018ApJ...857....3T}.
Collisions between planetesimals are thought to be the principal source of dust for debris discs \citep[see][]{2008ARA&A..46..339W}. Although some
amount of primordial dust is likely still present during the early stages of debris disc evolution, ongoing collisions between small bodies will augment
this and could be used to trace the underlying dynamical activity of the planetesimals. This process was directly detected when the putative object
Fomalhaut b was confirmed to be an expanding debris cloud, most likely from a planetesimal collision \citep{2020PNAS..117.9712G}.
\citet{2013ApJ...777L..31D} showed that collisional dust that is generated near the gap opened by a giant planet in the inner regions of a transition disc should
produce a distinct observable marker that could be used to infer the presence of the planet. In a followup study, it was found that morphological
differences in the dust emission could be used to determine whether the planet had a circular or eccentric orbit \citep{2016ApJ...820...29D}. Moreover,
substructure due to mean-motion resonances (MMRs) with a giant planet may hold additional clues that could be used to constrain the properties of
the planet. The width of a resonance is set by both the mass of the perturbing planet and the unperturbed eccentricity of the planetesimals,
which is in turn set by the eccentricity of the planet through secular forcing. Therefore, the planetesimal collision profile and potentially the structure of second generation
dust produced near MMRs may encode information about both the mass and eccentricity of the planet.
The dynamics governing the motion of bodies near MMRs is extremely nonlinear, as is determining what the collision rates between planetesimals
should look like in these regions. For a collection of bodies massive enough to experience the effects of gravitational focusing, a large eccentricity
dispersion tends to reduce the probability of collision, while enhancements in surface density tends to increase it. Due to conservation of the Jacobi
energy, MMRs simultaneously enhance the local eccentricity dispersion and also enhance the surface density adjacent to the resonance
\citep{2000Icar..143...45R, 2017ApJ...850..103B}. Unfortunately, collision detection in an N-body simulation is extremely computationally expensive.
So far, studies of planetesimal dynamics near MMRs in which the planetesimals are directly resolved have involved either collisionless test particles
\citep{2017ApJ...850..103B, 2016ApJ...818..159T, 2018ApJ...857....3T} or integration times that do not fully capture the dynamics of the resonances
\citep{2000Icar..143...45R, 2013ApJ...777L..31D}.
To further elucidate this subject, we use the tree based N-body code {\sc ChaNGa} \citep{2008IEEEpds...ChaNGa, 2015AphCom..2..1} to follow the
collisional evolution of a planetesimal disc under the gravitational influence of a Jupiter-sized body. Because particle positions are sorted into a tree
structure, neighbor finding and collision detection can be done quickly and efficiently. This considerably relaxes the constraints on resolution and
integration time. With this toolset, we explore the collision rate structure of a planetesimal disc in the vicinity of mean-motion resonances with a planet. In
particular, we would like to determine (1) what dynamics govern the collision rate profile near MMRs, (2) whether MMRs leave a detectable signature in the
collisionally-generated dust, and (3) whether these signatures can be used to determine or constrain the orbital properties of the perturbing planet.
This work is organized in the following way: In section \ref{sec:dynamics}, we provide an overview of the relevant dynamics that drive the evolution
of a planetesimal disc under the gravitational influence of an external perturber. In section \ref{sec:sims} we provide an overview of the N-body code
used and describe the initial conditions chosen for five simulations in which a perturbing giant planet is given various masses and eccentricities. Section
\ref{sec:results} presents the results of these simulations and we take an in-depth look at the collision rate profiles of the planetesimals near the MMRs.
In section \ref{sec:dust}, we extrapolate the collision rate down to smaller bodies and motivate a direct correspondence between the planetesimal collision profile
and the resulting dust profile. In section \ref{sec:constrain}, we use the resolved collisions to generate synthetic dust
emission profiles that would be detected with observing facilities like ALMA. Under our simplifying assumptions, we show that a characteristic bump
or dip feature appears in the dust emission near the interior 2:1 MMR, the presence of which depends on the mass and eccentricity of the perturbing
planet. If the 2:1 MMR can be identified (presumably, by identifying another prominent MMR, such as the 3:1, and measuring the spacing between the two), we discuss the
potential for using this feature to place constraints on the mass and eccentricity of the planet. In addition, we highlight the caveats of connecting the dust and planetesimal collision
profiles in such a simple way. Finally, we conclude in section \ref{sec:conclusions}.
\section{Overview of Relevant Dynamics} \label{sec:dynamics}
We begin by providing a description of the dynamical effects responsible for shaping the orbital distribution of the planetesimals. The purpose of this is twofold: (1) to motivate
the initial conditions used for the simulations described in section \ref{sec:ics} and (2) to justify the exclusion of certain physical effects from our simulations. Here, we focus solely on the
physics relevant for full-sized ($\sim$100 km) planeteismals and save a discussion of the effects on smaller bodies generated through collisions for section \ref{sec:dust}.
\subsection{Secular Forcing}\label{sec:sec_force}
The most direct and widespread effect that a perturbing giant planet will have on a planetesimal disc is through secular forcing of the planetesimals. This will
cause the complex eccentricities of the planetesimals to take on a time-independent
forced value, given by \citep{1999ApJ...527..918W} as
\begin{equation}\label{eq:eforced}
z_{f} = \frac{b^{2}_{3/2} (\alpha)}{b^{1}_{3/2} (\alpha)} e_{g} ~ \mathrm{exp} ~ i \varpi_{g}.
\end{equation}
\noindent Here, $\alpha = a_{g} / a$ where $a_{g}$ and $a$ are the semi-major axes of the giant planet and a planetesimal,
respectively. $e_{g}$ and $\varpi_{g}$ are the eccentricity and
longitude of pericenter of the giant planet, and $b^{j}_{s} (\alpha)$ is a
Laplace coefficient given by \citep{1999ssd..book.....M} (ch. 6, pg. 237, eq. 6.67) as
\begin{equation}\label{eq:lap}
b_{s}^{j}(\alpha) = \frac{1}{\pi} \int_{0}^{2 \pi} \frac{\cos \, j \psi \, d \psi}{\left( 1 - 2 \alpha \, \cos \psi + \alpha^2 \right)^{s}}.
\end{equation}
Without any nearby secular or mean motion-resonances, equation \ref{eq:eforced} will completely describe the eccentricities and longitude of
pericenter orientations of the planetesimals. Additional forces due to two-body scattering between planetesimals, along with
aerodynamic gas drag will add an additional free component to the complex eccentricity, which will be randomly oriented. The
magnitude of the free eccentricity describes how dynamically hot the planetesimal disc is and sets the random encounter speeds of
planetesimals. When the dynamical excitation of the disc is driven by gravitational stirring, the magnitude of the free eccentricity can
be described by a Rayleigh distribution \citep{1992Icar...96..107I}.
For the case of a Jupiter mass planet at 5.2 au perturbing a test particle at 3 au, the timescale for secular forcing is approximately 12,000 years. The details of this calculation
can be found in appendix \ref{sec:sec_forcing_timescale}.
\subsection{Mean-Motion Resonances}\label{sec:mmr}
In regions where there are commensurabilities between frequencies, Laplace-Langrange secular theory breaks down and bodies are subject to strong
perturbations. For the purposes of this study, we will ignore secular resonances, which generally occur on rather large timescales and will focus on
mean-motion resonances. A MMR occurs when the orbital period ratio between two bodies is sufficiently close to
\begin{equation}\label{eq:per_mmr}
\frac{P}{P'} = \frac{p + q}{p},
\end{equation}
\noindent where $p$ and $q$ are integers $>$ 0 and the unprimed and primed quantities correspond to the perturber and the body being perturbed,
respectively. In terms of these quantities, $P/P'$ corresponds to a $p+q:p$ resonance. If the perturber is much more massive than the other body and all of the
bodies lie in a near-Keplerian potential, the condition for MMR is set by
\begin{equation}\label{eq:a_mmr}
\frac{a}{a'} = \left( \frac{p}{p + q} \right)^{2/3}.
\end{equation}
If we further assume that the two bodies are orbiting in the same plane, the motion of the bodies near resonance is determined by the
behavior of a critical angle
\begin{equation}\label{eq:phi_crit}
\phi = (p + q) \lambda' - p \lambda - q \varpi,
\end{equation}
\noindent where $\lambda = \varpi + M$ is the mean longitude of a
body, with $M$ being the mean anomaly. For bodies in resonance, the critical angle will librate around an
equlibrium value, while this angle will circulate outside of resonance. For small eccentricities, this behavior is analogous to the motion of a
pendulum. Furthermore, variations in the critical angle are coupled to changes in the mean motion and semimajor axis \citep{1999ssd..book.....M}. An important
point to note, which we will revisit later, is that the variation frequency of this angle approaches zero near the edge of a resonance. The
width of a resonance can be defined by determining the largest variation in semimajor axis that permits librational, rather than circulatory motion of $\phi$.
Calculations for the widths of first and second order interior MMRs are shown in appendix \ref{sec:libration}, along with a timescale for the libration. For the most prominent interior
mean-motion resonances, including the 2:1 and 3:1, the timescale associated with these oscillations driven by a Jupiter mass planet is $\sim$ 1,000 - 2,000 years.
\subsection{Collisions Between Planetesimals}\label{sec:colleq}
A simple analytic model for the collision rate of a planetesimal population is given by \citet{1969edo..book.....S} as
\begin{equation}\label{eq:saf}
n \sigma v = n \pi s^{2} \left( 1 + 2 G m / s v^{2} \right) v,
\end{equation}
\noindent where $n$ is the number density of the population, $s$ and $m$ are the radii and masses of the bodies and $v$ is their typical encounter velocity. The encounter velocity is often described in terms of the rms eccentricity $\left<e^{2}\right>^{1/2}$ and inclination $\left<i^{2}\right>^{1/2}$ of the population (e.g. \citet{1993ARA&A..31..129L}) as
\begin{equation}\label{eq:eccincvel}
v = \sqrt{\left< e^{2} \right> + \left< i^{2} \right>} v_{k},
\end{equation}
\noindent where $v_{k}$ is the local Keplerian velocity. The second term in equation \ref{eq:saf} can be thought of as an additional enhancement to the collision cross section due to
gravitational focusing. When the typical encounter velocity is small compared to the mutual escape velocity of the planetesimals, the collision cross section greatly exceeds the geometric
value. An important feature of equation \ref{eq:saf} is that for a fixed value of $n$, the collision rate exhibits a global minimum as a function of $v$. For small $v$, gravitational focusing
facilitates more collisions, while for large $v$ the encounter rate simply becomes so great that the collision rate again increases, even though gravitational focusing is mostly suppressed.
This point will become relevant in section \ref{sec:vary_ecc} when we examine the qualitative changes in the collision rate near mean-motion resonances.
Although equation \ref{eq:saf} works well to describe the collision rate for a homogeneous collection of planetesimals, some problems arise when regions of commensurabilities are
introduced. Namely, the resonances cause the orbits of planetesimals to precess, and an interface between secularly aligned and randomly oriented orbits arise on each side of the
resonance. Additionally, an interface between dynamically cold and hot planetesimals develop. These two effects cause the number density, collision cross section and encounter velocity
to rather abruptly vary across the boundaries of the resonance. For this reason, an N-body treatment in which collisions are directly resolved is necessary to understand how the collision
rate varies near the MMRs.
\subsection{Gas Drag on Planetesimals}\label{sec:pl_drag}
Over the course of many orbits, the residual gas from the primordial nebula can damp the eccentricities and inclinations of bodies. For a planetesimal-sized body, gas drag operates
in the Stokes regime and the timescale for aerodynamic forces to significantly alter its relative velocity is given by \citep{1976PThPh..56.1756A}
\begin{equation}\label{eq:ts_stokes}
t_{s} = \frac{2 m}{C_{D} \pi s^{2} \rho_{g} v_{g}},
\end{equation}
\noindent where $m$ and $s$ are the mass and radius of the planetesimal. $C_{D}$ is a drag coefficient which is of order unity, $
\rho_{g}$ is the local density of the gas and $v_{g}$ is the headwind velocity of the gas experienced by the planetesimal. At 3 au, the gaseous component of the solar nebula has a
density of $3 \times 10^{-11}$ g cm$^{-3}$ and the typical headwind experienced by a body on a Keplerian orbit is $\sim 5,000$ cm s$^{-1}$ \citep{1981PThPS..70...35H}. For a 100 km
body with a density of 2 g cm$^{-3}$ (assuming a mixture of ice and rock) the stopping timescale is about 1 Myr. This is much longer than the timescales associated with
secular forcing and libration due to mean-motion resonances, as discussed above, but shorter than the typical protoplanetary disk lifetime and potentially shorter the planetesimal
formation timescale. For this reason, we do not model the effects of gas drag in the simulations, although we consider its effects when constructing initial conditions, which is discussed in
the next section.
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{figures/xy.png}
\caption{The positions of the remaining planetesimals at the end of the e1m2 (left), e2m2 (center) and e3m2 (right) simulations. The red dot
indicates the position of the giant planet, and the dashed line points in the direction of the planet's longitude of perihelion. Non-axisymmetric gaps
are apparent near the locations of MMRs. At higher eccentricities, more resonances become visible. At the 2:1 MMR, gap features at $\theta$ = 0
and $\theta = \pi$ follow the giant planet in its orbit.\label{fig:xy}}
\end{center}
\end{figure*}
\section{Simulations} \label{sec:sims}
\subsection{Numerical Methods}\label{sec:methods}
To follow the dynamical and collisional evolution of a planetesimal disc, we use the highly parallel N-body code {\sc ChaNGa}
\footnote{A public version of {\sc ChaNGa} can be downloaded from \url{http://www-hpcc.astro.washington.edu/tools/ChaNGa.html}}.
This code, which is written in the {\sc CHARM++} parallel programming language, was originally designed for cosmology simulations
and has been shown to perform well on up to half a million processors \citep{2015AphCom..2..1}. Using similar methods to {\sc PKDGRAV}, which has been used for numerous
studies of planet formation \citep{2000Icar..143...45R, 2005ApJ...625..427L, 2013ApJ...777L..31D, 2015ApJ...806...23L}, {\sc ChaNGa} calculates
gravitational forces using a modified Barnes-Hut \citep{1986Natur.324..446B} tree with hexadecapole expansions of the moments and integrates the equations of motion using a
kick-drift-kick leapfrog scheme. All of the simulations we perform use a node opening criterion of $\theta_{BH}$ = 0.7. More information about the code can be found in
\citet{2008IEEEpds...ChaNGa}.
We have recently modified {\sc ChaNGa} to handle solid-body collisions between particles, by assigning them a fixed radius, rather than treating them as tracers of a fluid with a
characteristic softening length. We provide a brief summary of the collision model here, although a full description is provided in \citet{2019MNRAS.489.2159W}. This work is largely based
on the solid-body collision implementation in {\sc PKDGRAV}, which is detailed in \citet{1994MNRAS.269..493R} and \citet{2000Icar..143...45R}. Imminent collisions are found during the
drift phase of each time step by extrapolating the positions of bodies forward using the velocity calculated during the first kick. For each body, the nearest 64 neighbors are tested for an
imminent collision. If an imminent collision is detected, the two particles of mass $m_{1}$ and $m_{2}$ are merged together together to form one single larger body with the same density
and a mass of $m_{1} + m_{2}$. The resulting body is then imparted
with the center of mass position and velocity of the two
colliders. Because the resolution of a collision can result in
another imminent collision, these events must be handled one at a time, with the soonest collision being resolved first. For this reason, another collision search is run each time a collision
is resolved, which continues until there are no more imminent collisions during the current time step. During the course of the simulation, no debris is created by collisions. To model the
collisionally generated dust distribution, we use the statistics of
the resolved planetesimal collisions to build a dust profile. This process is discussed in detail in section \ref{sec:dust}.
\begin{table}
\begin{center}
\caption{Summary of Simulations Run}
\begin{tabular}{lllll} \hline \hline
Name & Mass of Planet & Eccentricity of Planet & & \\ \hline
e1m2 & 1.0 $M_{jup}$ & 0.5 $e_{jup}$ & & \\
e2m2 & 1.0 $M_{jup}$ & 1.0 $e_{jup}$ & & \\
e3m2 & 1.0 $M_{jup}$ & 2.0 $e_{jup}$ & & \\
e2m1 & 0.5 $M_{jup}$ & 1.0 $e_{jup}$ & & \\
e2m3 & 2.0 $M_{jup}$ & 1.0 $e_{jup}$ & & \\ \hline
\end{tabular}
\label{tab:sims}
\end{center}
\end{table}
\subsection{Initial Conditions}\label{sec:ics}
In total, five simulations are run, which are listed in table \ref{tab:sims}. The second is a ``nominal'' case, in which the perturbing planet's mass and eccentricity are set to that of Jupiter's.
In the other four cases, the mass or eccentricity is altered by a factor of two from the nominal value. In all cases, the perturbing giant is placed on a 5.2 au orbit around a 1
$M_{\odot}$ star. The planetesimal disc extends from 2.2 to 3.8 au, which covers the two most prominent mean-motion resonances with the giant planet, the 2:1 at 3.27 au and the 3:1 at
2.5 au. The planetesimal disc loosely follows a minimum-mass solar nebula surface density profile \citep{1981PThPS..70...35H}
\begin{equation}\label{eq:surf_den}
\Sigma = \Sigma_{0} r^{-\alpha},
\end{equation}
\noindent with $\alpha$ = 3/2 and $\Sigma_{0}$ = 10 g cm$^{-2}$. Planetesimals are given a bulk density of 2 g cm$^{-3}$, which corresponds to a mixture of ice and rock, and
are given a diameter of 300 km. This choice of parameters is meant to mimic the setup used in simulation B from \citet{2000Icar..143...45R}. One difference from the
aformentioned study is that we use a narrower annulus, which reduces the number of particles required from $10^6$ down to roughly 500,000. This allows us to use a finer base timestep
size, which, as we will discuss in section \ref{sec:results}, appears to be the reason why no features were seen in the collision profile near the resonances by \citet{2000Icar..143...45R}.
Another difference from \citet{2000Icar..143...45R} is that the dynamical effects of secular forcing by the giant planet, along with the effects of viscous stirring and gas drag on the
planetesimals, are built into the initial conditions. Although we do not model the effects of gas drag on the planetesimals during the simulation, we construct the initial conditions
such that the effects of gas drag and viscous stirring are in balance. The viscous stirring timescale of the planetesimal disc is much longer than our chosen integration time, and so the
dynamical excitation of the disc (excluding resonances) stays constant, even without the inclusion of damping forces from the gas. This is done by first calculating the equilibrium
eccentricity $e_{eq}$ due to viscous stirring and gas drag as a function of semimajor axis according to equation 12 of \citet{2002ApJ...581..666K}. The eccentricities of the bodies are
drawn randomly from a Rayleigh distribution with a mode of $e_{eq}$, while the inclinations are drawn from a similar distribution with a mode of $e_{eq}$/2 \citep{1993MNRAS.263..875I}.
The arguments of perihelion $\omega$, longitude of ascending nodes $\Omega$ and the mean anomalies $M$ of the bodies are drawn uniformly $\in [0, 2 \pi)$.
To account for the effects of secular forcing by the planet, the eccentricity vectors of the planetesimals are first decomposed into real and imaginary
components:
\begin{equation}\label{eq:kh}
z = (k, ih) = e \, exp(i \varpi)
\end{equation}
\noindent and a forced component is added to $h$ according to equation \ref{eq:eforced} (where we have set $\varpi_{g}$ = 0).
\subsection{Time Stepping Scheme}\label{sec:timestep}
For the purposes of the integrator, there are two relevant timescales in this system. The first is the orbital dynamical time $\sqrt{a^3/
G M_{\odot}}$. Through experimentation, we have found that using a base timestep size with {\sc ChaNGa} of 3\% of an orbital dynamical time
at the inner edge of the disc keeps the integration symplectic. Although doing so preserves orbital frequencies, the errors associated with
precession frequencies do not average out to zero. This is especially important given the effects of secular forcing by the planet. To mitigate
this, we further reduce the base timestep size by a factor of 4 and use $\Delta t$ = 0.0025 yr. Doing so prevents the longitude of perihelia of
planetesimals at the inner edge of the disc, where this artificial precession is most severe, from drifting by more than the intrinsic spread in $\varpi$ due to the
free eccentricity over the course of our integrations.
An additional timescale is set by the dynamical time of the planetesimals ($\sim 1/\sqrt{G \rho}$), which is about 45 minutes. This timescale must be resolved in order to properly
follow close gravitational encounters between these bodies. To resolve the base time step and the dynamical timescale of planetesimals simultaneously, we use a two-tiered time
stepping scheme, following\citep{2015ApJ...806...23L}. To start, all bodies are placed on the base time step. A first pass of collision detection is then run in which the radii of all bodies are
inflated by a factor of 2.5. Any bodies with imminent collisions predicted using the inflated radii are placed on a time step that is a factor of 16 smaller than the orbital time step. Although
this is still a factor of roughly 7 larger than the dynamical time of a planetesimal, we found no difference in the collision rate when using any smaller of a minimum time step size.The
purpose of the two-tiered scheme is to properly resolve the gravitational interactions between any bodies that undergo a close encounter. This prevents the coarser base time step from
reducing the effectiveness of gravitational focusing, while minimizing the additional computational expense.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/ae.png}
\caption{The semimajor axes and eccentricities of the remaining planetesimals are shown, with the locations of prominent resonances indicated
by the vertical dashed lines. Libration of the critical angle drives large variations in eccentricity, which produce spikes in the a-e plane. Between
the resonances, the nonzero eccentricity is due to secular forcing by the planet.\label{fig:ae}}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{figures/long_ph.png}
\caption{Shown here are the longitudes of pericenter and semimajor axes of the remaining planetesimals. The dashed lines again indicate the
locations of prominent MMRs. Close to the resonances, the critical angle librates and drives fast precession of the longitudes of pericenter of the
bodies. This overpowers the secular forcing by the planet and effectively randomizes the orientations of their orbits.\label{fig:long_ph}}
\end{center}
\end{figure*}
\section{Results} \label{sec:results}
All five simulations are evolved for 5,000 years, which is roughly 400 complete orbits of the perturbing planet. Because the effects of the mean
motion resonances are not built into the initial conditions, the simulations must be run long enough for the distribution of orbital elements to reach
equilibrium near the resonances before the collision rate is measured. To do so, we allow each simulation to run for 2,000 years before we begin recording any collision statistics.
This is comparable to the libration period, which was calculated in section \ref{sec:mmr} for the 3:1 and 2:1 resonance. Although one might worry that a single libration period may not be
long enough for the orbital structure due to the resonances to develop, we find that the shape of the semimajor axis-eccentricity distribution of the planetesimals reaches a steady state by
this time. This is largely due to the fact that there are enough bodies in or near the resonances to allow an ensemble average of phase space, rather than a time average. As an additional
confirmation, we find that the time evolution of the collision rate near the resonances maintains a steady state in all five simulations after $\sim$ 2000 years.
\subsection{Varying the Eccentricity} \label{sec:vary_ecc}
We begin by examining simulations e1m2, e2m2 and e3m2. The positions of the planetesimals in the x-y plane after 5,000 years of integration are
shown in figure \ref{fig:xy}. In all cases, the coordinate system is rotated so that the giant planet lies at $\theta = 0$. The longitude of perihelion of
the planet is shown by the dashed line. Resonances with the perturbing planet are visible as nonaxisymmetric gaps in this figure. Upon close inspection of a
series of simulation snapshots, these gaps appear to follow the planet in its orbit, rather than aligning themselves with the longitude of perihelion. A
similar substructure, which is most noticeable near the 2:1 MMR, reveals itself in \citet{2000Icar..143...45R} (see their figures 3c and 3f) and
\citet{2016ApJ...818..159T} (see their figure 3a). This structure is also present in the \citet{2017ApJ...850..103B} simulations, although it was not reported at the
time (see Appendix \ref{sec:boley_plot}). It is worth noting that both \citet{2000Icar..143...45R} and \citet{2017ApJ...850..103B} started with a completely cold planetesimal disc.
Thus, the presence of these features seems robust to the choice of initial conditions.
The effects of the resonances become much more apparent in semimajor axis-eccentricity space, which is shown in figure \ref{fig:ae}. In all cases,
the 3:1, 2:1 and 5:3 resonances are readily visible as ``spikes'' in the eccentricity that bend slightly inward (due to the conservation of the Jacobi energy). In the
e3m2 simulation, features also appear near the 5:2, 7:3 and 5:3 resonances. The absence of these finer features from the e1m2 simulation can be
explained by the fact that the strength of a resonance scales with $e^{q}$ \citep{1994PhyD...77..289M}.
Another important effect of the resonances is visible in figure \ref{fig:long_ph}, which shows the orientation of the longitude of perihelia of
planetesimals in the disc. Inside of the resonances, orbits of planetesimals quickly precess, and their orientations are effectively randomized. An
important point to note is that this strong precession effect quickly disappears beyond the boundaries of the resonance. This turns out to be
key to explaining the nonaxisymmetric structure seen in figure \ref{fig:xy}, which will be addressed in more detail below.
Next, we examine the statistics of collisions resolved in each of the simulations. The 3D positions and velocities of the two colliding bodies are
recorded to a table at the moment of impact. We derive the Keplerian orbital elements of a collision using these positions and velocities.
First, we examine the semimajor axis of the first collider, which is shown in figure \ref{fig:coll_hist_a}. During a collision, the ``first'' particle is defined
as the more massive of the two. However, nearly all of the collisions happen between the initial, equal-mass planetesimals. In this case, the
distinction is set by the collision search algorithm and is rather arbitrary. In all of the plots where we show collision statistics, we have verified that
using the ``first'' or ``second'' collider does not qualitatively change any of the features. We find that some of the features present in figure
\ref{fig:coll_hist_a} in this and subsequent figures are highly sensitive to the number of bins and the location of the bin edges. For this reason, we
construct a probability density function (PDF) of the collisions using a Kernel Density Estimate (KDE). We use the {\sc neighbors.KernelDensity} function from the {\sc sklearn}
\citep{scikit-learn} package to construct our KDEs, using a gaussian kernel with a FWHM of 0.02 au. The curves shown in figure \ref{fig:coll_hist_a} are normalized such that the
area underneath is equal to 1.
Near the stronger resonances, there are noticeable suppressions or enhancements to the local collision rate. This contrasts with the findings of
\citet{2000Icar..143...45R}, who simulated a similar setup and found no discernible features near the MMRs. We attribute the differences to a more
conservative timestepping criterion in our simulations, along with the inclusion of secular perturbations in our initial conditions. The most prominent features appear as a
local maximum in the collision rate near the 2:1 MMR and a local minimum near the 3:1 MMR. At higher forced eccentricities, features near the 5:2, 8:3 and 5:3
resonances are also visible, due to the steep sensitivity of higher order resonant perturbations to eccentricity \citep{1994PhyD...77..289M}.
Although the features in figure \ref{fig:coll_hist_a} near the 3:1 and 2:1 MMRs appear qualitatively different, they can be explained by one single dynamical process. As discussed in
section \ref{sec:colleq}, the collision rate between planetesimals depends on both the encounter rate and the strength of gravitational focusing. The encounter rate grows as the relative
velocity $v$ between the planetesimals increases, while gravitational focusing is most effective for bodies with small relative velocities. There is therefore an intermediate value of $v$ for
which the average collision rate $\left< \sigma v \right>$ is at a minimum, which we will refer to as $\left< \sigma v \right>_{0}$. The dynamical excitation induced by a mean motion
resonance can either increase or decrease the local average collision rate, depending on the unperturbed value of $\left< \sigma v \right>$ relative to $\left< \sigma v \right>_{0}$.
In figure \ref{fig:gf}, we show the effect that the 2:1 MMR and 3:1 MMR have on the local average collision rate. Each pair of points connected by a line represents the average
collision rate before and after the effects of the mean-motion resonances develop. Here, the relative velocity between bodies is calculated by measuring the eccentricity and inclination
dispersion near each resonance and using equation \ref{eq:eccincvel}. We assume that the eccentricity and inclination dispersions are coupled \citep{1993MNRAS.263..875I} and use the
eccentricity dispersion as a free parameter. Because both the strength of secular forcing and the effects of the resonant perturbations are different near the 3:1 and 2:1 MMRs, the
planetesimal population at each of these locations in the disk experiences a qualitatively different change in the collision rate. Near the 3:1 MMR, the unperturbed collision rate is well
above the minimum value $\left< \sigma v \right>_{0}$ and the dynamical heating introduced by the resonance acts to suppress it further. Near the 2:1 MMR, the collision rate starts out
much closer to the minimum value and the additional perturbations instead act to drive it higher.
This explains the relative drop in the collision rate near the 3:1 MMR and the relative increase near the 2:1 and 5:3 resonances seen in figure \ref{fig:coll_hist_a}. Although
this could potentially serve as a useful diagnostic of the planetesimal size in a disc (because the size and mass of the planetesimals sets the eccentricity dispersion at which
$\left< \sigma v \right>$ is minimized), a direct measurement of the semimajor axes of the colliding planetesimals is not possible. Furthermore, as we will demonstrate next, collisions due
to bodies in resonance do not have a significant effect on the final shape of the radial dust distribution that we predict.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/coll_hist_a.png}
\caption{A PDF of the collision rate in each disc as a function of semimajor axis, generated using a KDE with a Gaussian kernel with a FWHM of 0.02 au. In semimajor axis space,
prominent features appear near the 3:1 and 2:1 MMRs. Near the 3:1, the collision rate exhibits a local minimum, while an enhancement appears near the
2:1.\label{fig:coll_hist_a}}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=\textwidth]{figures/gf_plot.png}
\caption{The collision rate (given by equation \ref{eq:saf}) of bodies in the vicinity of the 3:1 (left) and 2:1 (right) MMRs, relative to the minimum value, as a function of the local
eccentricity dispersion. The pairs of points connected by lines show the values of the unperturbed (left-hand points in each subplot) and perturbed (right-hand points in each subplot) collision rates in
the e1m2, e2m2 and e3m2 simulations. The unperturbed collision rate, along with the amount of dynamical heating that the planetesimal population experiences determines whether
the MMR acts to suppress or enhance the collision rate.\label{fig:gf}}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/coll_hist_r.png}
\caption{Here, collisions are instead ordered by cylindrical distance from the star. Features near the 3:1 and 2:1 MMRs are still present, but appear qualitatively different than in
semimajor axis space. At low eccentricities, a dip appears around the center of the resonance. At the highest eccentricity, a bump is formed instead. The dashed lines shown the
collision profile with collisions between bodies in resonance removed. The bump and dip features remain qualitatively the same, which suggests that they are produced by bodies
outside of the libration width of the resonance.\label{fig:coll_hist_r}}
\end{center}
\end{figure}
\section{Where Does the Dust End Up?}\label{sec:dust}
To construct a radial dust profile from the collision statistics, we begin by making two assumptions: (1) any dust generated by collisions is strongly coupled to the gas and (2) any
subsequent spatial evolution of the collisional debris is insignificant. So long as both of these assumptions are true, we can use the radial locations of planetesimal collisions in the
plane of the disk to generate a dust profile for each simulation.
We first provide a justification for assumption (1), which is as follows: At 3 au in the protosolar nebula, the mean free path of a gas particle is $\sim$ 50 cm (assuming a composition
of pure hydrogen with a local volume density of $\rho_{g} = 3 \times 10^{-11}$ g cm$^{-3}$). This places any dust grains in the Epstein drag regime, with a stopping timescale given by
\begin{equation}\label{eq:ts_epstein}
t_{s} = \frac{\rho s}{\rho_{g} v_{th}},
\end{equation}
\noindent where $\rho$ is the bulk density of the dust grain, $s$ is its size, and $v_{th}$ is the local thermal velocity of the gas. At 3 au in the protosolar nebula, $v_{th} \sim 10^{5}$ cm
s$^{-1}$. Assuming a 1 mm dust grain with a $\rho$ = 2 g cm$^{-3}$, the stopping time is $\sim 10^{-4}$ yr. This is orders of magnitude smaller than the
orbital timescale; therefore, we conclude that collisionally generated dust grains will immediately couple to the gas.
Assumption (2) is less straightforward to justify and may not be reasonable in all cases. Although the spatial redistribution of the full-sized planetesimals and mm-sized dust grains
is insignificant, the radial drift forces experienced by intermediate-sized bodies is much greater. Collisions between planetesimals do not immediately convert the mass to dust grains.
Rather, the debris will undergo a collisional cascade process, which
gradually grinds the resulting bodies down to smaller sizes. A proper
treatment of the collisional cascade process is
quite complicated and is well beyond the scope of this work. However, one may be able to ignore this process if the peak radial drift timescale is longer than the timescale for collisions
between these bodies.
This occurs for bodies of the size at which $t_{s} \Omega$ = 1. In the Stokes drag regime at 3 au, this occurs for bodies of 1 meter in size. Following a similar
calculation in \ref{sec:pl_drag} with a 1 m body, we find that an object of this size will drift across the 2:1 MMR on a timescale of $\sim$ 300 years. Although we do not model collisions
between bodies of this size, a lower limit on the collision timescale for these objects can be obtained by extrapolating the planetesimal collision rate measured in our simulations down. For
the e2m2 simulation, we find that 125 collisions occur within the libration width of the 2:1 MMR from T=2,000 yr to T=5,000 yr. There are a total of 15,000 planetesimals within this region,
which means that a single planetesimal undergoes a collision every $\sim$ 370,000 years on average. If the entire planetesimal population in this region were converted to 1 meter bodies
with the same bulk density, the number density of bodies increases by a factor of $10^{15}$, while the geometric collision cross section decreases by a factor of $10^{10}$. The collision
timescale would therefore decrease to $\sim$ 1 year, which is 2 orders of magnitude smaller than the drift timescale. As shown above, radial drift in the Stokes drag regime is maximized
for bodies of 1 meter in size, which is only a factor of two larger than the mean free path of the gas particles. It is therefore not entirely clear if the Stokes drag law is appropriate here.
Because of this, we also repeat the above calculation for bodies that experience Epstein drag. With this assumption, we find that radial drift is maximized for a cm-sized body and both the
radial drift and collision timescale for bodies of that size decrease by a factor of 10 relative to the Stokes values. In both cases, collisions between maximally drifting bodies occur on a
timescale two orders of magnitude shorter than the drift timescale. We therefore argue that it is reasonable and, indeed, best practice to start with such simple assumptions understand the
physics before additional complexity is later added. We caution, however, that the collision timescale estimates provided here should be interpreted as lower limits.
Operating under these assumptions \footnote{As discussed in \cite{2017ApJ...850..103B}, the local generation of second-generation dust is only one
consideration. The dust will also evolve spatially due to drag effects, but can also evolve in size or be re-accreted onto planetesimals. For simplicity,
we focus on local dust sourcing only while acknowledging that subsequent evolution could \textbf{also} affect the morphology.}, a map of the relative concentration
of second-generation dust can be constructed from the cylindrical distance at which the collisions occur. Here, cylindrical distance is defined as the separation between the central
star and the location of collision in the ($r, \theta$) plane. This is shown in figure \ref{fig:coll_hist_r}.
Similar to figure \ref{fig:coll_hist_a}, we use a KDE with a width of 0.02 au to assemble the collision locations into a radial distribution. Most
strikingly, the bump that was present near the 2:1 resonance is no longer visible in cylindrical distance space, although it suddenly appears again in
the e3m2 simulation. The local minimum that is visible near the 3:1 MMR also becomes a local maximum in the highest eccentricity case. To determine how the resonant
bodies actually contribute to the radial collision profile, we excluded collisions that fall between $2.495 < a < 2.505$ au (near the 3:1 resonance) and
between $3.2 < a < 3.35$ au (near the 2:1 resonance), which is shown by the dashed curve. As in figure \ref{fig:coll_hist_a}, the solid curves are normalized such that the area
underneath is equal to 1. For the dashed curves, the normalization factor is scaled according to the number of collisions in the entire sample, compared to the number of collisions in the
subsample, which excludes the collisions in resonance. Qualitatively, none of the bump or dip features present
are removed by making this exclusion. This suggests that the features near resonance seen in semimajor axis space become smeared out in
cylindrical distance space due to the large spread in eccentricity and orbital orientation.
This also suggests that the prominent bumps and dips seen in figure \ref{fig:coll_hist_r} are mainly produced by bodies outside of the libration width of the resonance, although
the resonant population does make some of the features more pronounced. To further understand this, we consider the nonaxisymmetric structure
seen in figure \ref{fig:xy}. In figure \ref{fig:coll_polar_e}, we compare the radial collision profiles to the radial structure seen in the e1m2, e2m2 and
e3m2 simulations. In the lowest forced eccentricity case (e1m2), the pileups near the edge of the 2:1 MMR line up with the boundaries of the dip
feature seen in the radial collision profile. The locations of the edges of the resonances are calculated using equations \ref{eq:res_fo} and
\ref{eq:res_so}. The pileups not as pronounced near the other resonances. As discussed previously, the planetesimal density enhancements and gaps
seen in polar coordinates follow the position of the giant planet in its orbit, which is shown by the vertical line. For the e2m2 case, the density
enhancements take up a much larger radial width, which smears out the corresponding collisional dust profile. In the highest forced eccentricity case
(e3m2), the density enhancements occur well within an annulus containing the resonant region, completely altering the radial peak-valley dust
morphology.
Although the nonaxisymmetric pileup of bodies near the edges of resonances has been seen previously
\citep{2000Icar..143...45R, 2016ApJ...818..159T}, we could not find a satisfactory explanation for why it happens, which we will try to provide here. As
mentioned in section \ref{sec:dynamics}, the circulation frequency of the critical angle slows as one approaches the edge of the resonance from the
outside. Following the pendulum analogy, this is equivalent to
approaching the point where the pendulum becomes suspended in the
at top dead center.
When the critical angle stays relatively stationary during an encounter, the net torque will drive the longitude of pericenter towards the point of
conjunction \citep{1976ARA&A..14..215P}. This is the basic mechanism that causes isolated resonances to be stable. Inside of resonance, the critical
angle librates about an equilibrium value and this configuration is maintained. Just outside of resonance, however, the critical angle slowly drifts away, with the
drift direction changing sign across the location of exact resonance. The net result of this is that bodies just outside of the resonance experience a torque from
the planet that temporarily drives their pericenters toward the current orbital position of the planet. After the encounter ends, differential rotation of the disc
slowly undoes the configuration. As the planet again passes by, adjacent planetesimals near the resonance edges temporarily align
themselves in this configuration. This explains why the concentration of planetesimals appears to `follow' the planet. Connecting this back to the radial collision profiles shown in
figure \ref{fig:coll_hist_r}, these bodies at the edges of the resonances must be responsible for the bumps and dips morphology seen because excluding collisions between bodies inside
of the MMRs does not qualitatively alter the profiles.
\begin{figure*}
\includegraphics[width=\textwidth]{figures/coll_polar_e.png}
\caption{The collision profiles in figure \ref{fig:coll_hist_r} are shown alongside the positions of the planetesimals, in polar coordinates. The dashed and solid lines show the pericenter
and apocenter, respectively, of bodies at each edge of the 3:1 and 2:1 MMR. In these figures, the longitude of pericenter of the planet lies at $\theta = 0$ and the present position of the
planet is indicated by the vertical line. For the 2:1 resonance, a bump, rather than a dip feature appears when the inner edge or the resonance's apocenter and outer edge's pericenter
distance cross. The collision profile near the 3:1 resonance does not appear to follow this same qualitative behavior.
\label{fig:coll_polar_e}}
\end{figure*}
\begin{figure*}
\includegraphics[width=\textwidth]{figures/coll_polar_m.png}
\caption{Similar to figure \ref{fig:coll_polar_e}, except the eccentricity of the planet is kept constant and the mass is varied.
This has the effect of changing the width of the resonances, without altering the relative apo or peri distances of
bodies near the edges. Except for the highest mass case, the apocenter and pericenter distances of the inner and outer edges of the 2:1
resonance are too close together to produce much of a dip or a bump feature.\label{fig:coll_polar_m}}
\end{figure*}
\citet{2016ApJ...818..159T} provide a different explanation for this phenomenon, suggesting that the nonaxisymmetric structure is the product
of the path that a resonant test particle takes in a frame co-rotating with the planet (see figure 8.4 of \citet{1999ssd..book.....M}). In other words, the gap structure is argued to be
due to the interfaces between the low eccentricity, non-resonant and high eccentricity, resonant regions of the disc. This explanation,
however, does not hold for a large collection of planetesimals, especially when there is no forced eccentricity. In such a case, the orbits of
planetesimals near resonance would be randomly aligned, and the axisymmetric structure should become washed out. Furthermore, when the perturbing
planet is on a circular orbit, this structure is still present (see figure 3 of \citet{2016ApJ...818..159T} and Appendix \ref{sec:boley_plot}). The only
connection between the structure of the gaps in the disc and the path of a resonant particle in the corotating frame is that both phenomena produce the
same symmetries, which depend on the particular MMR considered.
With no secular forcing, this dynamical phenomena should produce an underdensity in the radial surface density profile near the center of the
resonance and an overdensity at each edge in cylindrical distance space. When a forced eccentricity is introduced, the radial location of the
resonance edges vary over the course of an orbit. The maximum and minimum radial distance that the resonance edges occupy is indicated by the
solid and dashed lines in figure \ref{fig:coll_polar_e}. If the the aphelion distance of the inner edge becomes close to the perihelion distance of
the outer edge, the overdense regions on each side of the resonance meet and we should expect to see the central dip feature in the collision profile
disappear. This is exactly what appears to be happening in the middle panel of figure \ref{fig:coll_polar_e}. As the forced eccentricity is increased
further, a region forms at the center of the resonance where planetesimals on both sides of the resonance spend time (although not simultaneously).
When this occurs, a bump feature forms in the collision profile near the center of the resonance. This is apparent in the bottom panel of figure
\ref{fig:coll_polar_e}.
Although the same phenomenon appears to be happening near the 3:1 MMR, the pileups at the edges of the resonance are not as well defined. We
attribute this to the more complex symmetry, compared to the 2:1 MMR. In addition, a number of other nearby resonances, including the 7:2 (at
2.25 au), the 10:3 (at 2.33 au), the 8:3 (at 2.71 au) and the 5:2 (at 2.82 au) also likely are contributing to the collision profile in cylindrical distance
space near the 3:1 MMR. The 3:1 MMR is also much narrower than the 2:1, which means that any dip or bump features produced by it would
require much higher resolution observations. For these reasons, we will focus on the 2:1 MMR as the main diagnostic indicator.
\subsection{Varying the Mass}
The dynamical effects of varying the mass of the planet are somewhat simpler, in that doing so does not affect the forced eccentricity of the
planetesimal disc. Instead, only the width of the resonances changes. The width of a first order resonance scales with $m$ (see equation
\ref{eq:res_so}), while the leading order terms in the resonant part of the disturbing function, which set the strength of the resonance, also scale as
$m$. For the 2:1 MMR, the dynamics near the resonance are equally sensitive to changes in eccentricity and mass.
We show the polar structure of the e2m1, e2m2 and e2m3 simulations in figure \ref{fig:coll_polar_m} alongside the radial collision profile. In all three
cases, the eccentricity of the perturbing planet is set to $e_{jup}$. Changes in the apocenter and pericenter distances of the edges of the
resonances are entirely due to changes in the resonance width. For the e2m1 and e2m2 simulations, the inner apocenter and outer pericenter
distances near the 2:1 MMR are quite similar and no strong features appear in the collision profile near this region. For the e2m3 case, the edges of
the resonances are sufficiently separated to allow what appears to be the beginning of gap to form near the center of the 2:1 resonance in the collision profile.
\subsection{Observability of the Dust}
While the collisional dust profiles show radial amplitude variations, we still need to assess whether those variations could be observable. To proceed,
we create a sky model for simulated ALMA observations as follows. First, we use the radially-averaged collision profile from each simulation (see figures
\ref{fig:coll_polar_e} and \ref{fig:coll_polar_m}) as the template for an azimuthally symmetric disc. We then scale the size of each profile by a factor of
ten, which places the perturbing planet at 52 au. This scaling is permitted by the dynamics (both the resonance widths and the forced eccentricities
are scale-free) and makes the disc comparable in size to the many discs that have now been observed by ALMA \citep{2018ApJ...869L..42H}. The
angular size scale is then set by envisaging a face-on disc at a distance of 100 pc.
The sky model intensity (flux per pixel) is produced by interpolating the radial collision frequency onto a 2D Cartesian grid with cell widths of
approximately 2 mas. In doing so, we choose to make the intensity proportional to the collision frequency (i.e., the dust) because it is the most
straightforward case. Many different parameterizations are possible, depending on optical depth, grain size, disc temperature, etc., but we will avoid
such complications here, as they would affect the overall brightness profile, but not the presence of any gaps or rings. We normalize the intensity by
setting the total disc flux density to be 100 mJy at about 350 GHz.
The simulated observations are performed using the {\sc simobserve} task in {\sc CASA} \citep{2007ASPC..376..127M}. The disc is given an
$\rm RA = 11^{\rm h} 01^{\rm m} 52^{\rm s}$ and a $\delta = -34^\circ 42\arcmin 17\arcsec$ (i.e., the J2000 coordinates for TW Hydrae) and is
``observed'' through transit on March 20, 2020. We observe the disc using configurations 8, 9, and 10 at 350 GHz (cycle 6 antenna file), spending six
hours on-source in each configuration\footnote{We have imagined that the TAC really likes us.}. The visibilities are corrupted with thermal noise,
setting the precipitable water vapor to 1.5 mm. They are then combined, imaged, and cleaned
using the task {\tt tclean}. Imaging uses Briggs weighting with a robustness parameter of -1.
The results for the nominal simulation (e2m2) are shown in figure \ref{fig:alma_sim_obs}. The 3:1 and 2:1 resonances can be easily identified in the
cleaned image and in the corresponding radial profile. Moreover, the features are qualitatively similar to the bright rings and dark gaps seen in actual
disc profiles \citep{2015ApJ...808L...3A}.
In figure \ref{fig:alma_profiles}, we show the
azumithally-averaged radial profiles constructed from the simulated cleaned image from all five
simulations. In all cases, a bump or a dip feature is clearly visible at the location of the 3:1 and 2:1 MMRs, indicated by the dashed vertical lines.
As mentioned previously, the feature at the 2:1 MMR acts as the actual diagnostic indicator, while the 3:1 resonance (along with the gap that would presumably
open at the location of the planet) is mainly useful for
determining where the 2:1 MMR is actually located.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{alma/skymodel_e2m2.png}
\includegraphics[width=0.5\textwidth]{alma/xy_e2m2.png}
\caption{Simulated ALMA observations for the nominal (e2m2) case, operating under the assumption that the dust profile closely traces the planetesimal collision profile. Top:
The sky model based on the radial collision distribution. The size of the disc has been scaled by a factor of 10 and placed at a distance of 100 pc. Bottom: The cleaned image using
combined observations in configurations 8, 9, and 10. The circle in the bottom left indicates the simulated beam size. Gaps in the dust due to the 3:1 and 2:1 resonances are visible.
\label{fig:alma_sim_obs}}
\end{center}
\end{figure}
\begin{figure*}
\includegraphics[width=\textwidth]{figures/alma_profiles.png}
\caption{The azimuthally-averaged radial profile based on the simulated cleaned images from all five simulations. The eccentricity of the perturbing planet increases from left to right,
while the mass of the planet increases from bottom to top. The vertical dashed lines indicate the locations of the 3:1 and 2:1 resonances. As in figure \ref{fig:alma_sim_obs}, the size
scale of the disc has been expanded by a factor of 10 and then placed at a distance of 100 pc.\label{fig:alma_profiles}}
\end{figure*}
\section{Constraining the Mass and Eccentricity of the Planet}\label{sec:constrain}
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/bump_dip_diag.png}
\caption{The presence of a dip or bump in the collision profile (and therefore the dust profile) near the 2:1 MMR can be used to constrain the mass and eccentricity
of the perturbing planet. The color scale indicates the relative strength of the features, while the dashed lines
indicate boundaries in parameter space where a dip or a bump will be produced at the resonance. Combinations of mass and eccentricity that fall
between the dashed lines correspond to an inner apocenter and outer pericenter separation at the edges of the resonance that is smaller than the
resonance width and therefore will not create any significant feature in the collision profile.\label{fig:bump_dip_diag}}
\end{center}
\end{figure}
The simple bump vs dip structure that we expect to reveal itself in the dust emission from colliding planetesimals in near-resonance with a giant
planet could potentially be used to place constraints on the mass and eccentricity of the planet. Near the 2:1 MMR, the central dip feature is only
produced when bodies at the edges of the resonance stay sufficiently separated in cylindrical distance over the course of an orbit. This is achieved
when either the resonance width is large or the forced eccentricity is small. The bump feature, on the other hand, is produced when there is a
sufficiently large amount of overlap between the apocenters and pericenters of the inner and outer edges of the resonance, respectively. This is
achieved when the resonance is narrow and the forced eccentricity is large.
In figure \ref{fig:bump_dip_diag}, we show the constraints that the presence of a bump or a dip or dip at the 2:1 MMR places on possible values of $(m_{g}, e_{g})$. The dashed
lines indicate regions of parameter space where qualitatively different features in the collision profile are expected to form. Above the upper dashed line, the
resonance is wide while the radial excursion distance of planetesimals at the edge of the resonance is small, which gives rise to a dip in the collision profile.
Below the lower dashed line, the resonance is narrow, while the radial excursion distance is large, allowing the apocenter and pericenter distances of the edges
of the resonance to overlap and produce a bump in the collision profile. Between the dashed lines, the absolute value of separation between the apocenter and
pericenter distances of the inner and outer edges of the resonance is less than the resonance width. In this region of parameter space, no significant features in
the collision profile near the 2:1 MMR are expected to form.
Because the strength\footnote{More specifically, the strength of the resonant term in the disturbing potential} of the 2:1 resonance scales with $e_{g}$ and $m_{g}$, a sufficiently low
mass and low eccentricity planet will not create enough of a perturbation in the planetesimal disc to produce a detectable bump or dip feature. The colored contours in figure
\ref{fig:bump_dip_diag} indicate the relative strength of the resonance as these quantities are varied. Although the resonance strength is equally sensitive to changes in mass and
eccentricity, the peri-apocenter overlap of the resonance edges is more sensitive to changes in eccentricity. This is because changes to eccentricity affect both the resonance width and
the forced eccentricity. As we showed in figure \ref{fig:coll_polar_m}, changing $m_{g}$ by a factor of 4 produced very minimal changes to the resulting dust profile.
In order to actually identify the 2:1 MMR in the dust emission, at least one other resonance must be visible. Assuming the gravitational field in the
disc is near-Keplerian, the distance ratios between two features can be used to determine period ratios, which can be used to confirm whether the
features seen are indeed resonances. Although the 3:1 MMR does not appear to follow the simple bump vs dip dichotomy described above, its
presence should be marginally detectable in all of the cases shown above.
We would again like to emphasize that the dust emission profiles in figure \ref{fig:alma_profiles} are constructed upon the assumption that the radial dust structure closely traces the
collision profile of the planetesimals. For this assumption to be valid, the resulting dust must remain locally confined, which can be achieved if the grains are reaccreted onto
planetesimals \citep{2015SciA....1E0109J}. It is also necessary that the collisional cascade from which the dust grains form plays before radial drift has a chance to move debris away
from the resonances. Although we showed with a back-of-the-envelope calculation (see section \ref{sec:dust}) that this is plausible for a typical protoplanetary disk with properties similar
to the protosolar nebula, a more thorough treatment of collisional grinding and the subsequent evolution of the debris is necessary to be able to interpret the dust emission profiles
presented here as anything more than upper limits.
\section{Summary and Conclusions}\label{sec:conclusions}
In this work, we have shown that mean-motion resonances with a perturbing planet produce significant local variations in the collision rate of a
planetesimal disc. In contrast to \citet{2000Icar..143...45R}, we find that the more prominent interior MMRs, including the 2:1, 3:1 and 5:3, all
produce structure in the collision profile as a function of semimajor axis. Furthermore, we find that a series of distinctly different features appear
when collisions are ordered by cylindrical distance from the central star. The morphology of these features are tied to dynamical behavior of bodies near the edges of the
resonances, where the circulation frequency of the critical angle approaches infinity. Particularly near the 2:1 MMR, we find that a bump or dip in the collision profile forms when collisions
are ordered by cylindrical distance. The presence of one of these two features depends on the peri- and apocenter distances of the edges of the resonance, relative to the libration width.
Because these quantities depend on the mass and eccentricity of the perturbing planet, these properties of the planet are actually encoded in the collision profile.
Near the interior 2:1 MMR, we find that a distinct bump or dip feature is generated in the collision profile, depending on the properties of the perturbing
planet. The presence of one of these two structures can used to constrain the mass and eccentricity of the planet. For a high mass, low eccentricity
planet, a dip will form because the edges of the resonance, where many collisions occur, stay sufficiently separated. If the planet
has a low enough mass (which shrinks the size of the resonance) or is sufficiently eccentric (which decreases the separation between the apocenter
and pericenter distances of the edges of the resonance), a bump will instead form. We tested this hypothesis for five different combinations of the
planet's mass and eccentricity and found that a distinct bump or dip signature is produced as long as the planet properties are sufficiently far from
the dividing line in parameter space. This diagnostic is more useful for massive, eccentric planets, because the strength of the resonant
perturbations scale linearly with both of these quantities.
Although the planetesimal collision profile would not be directly observable in a planet-forming disc, the dust from the resulting collisional cascade could potentially be used to trace
it. So long as the dust grains are well-coupled to the gas and radial drift does not have a significant effect on the morphology of the dust, the bump or dip features seen in the collision
profile will also appear in the dust profile. Assuming that this is the case, we have generated simulated observations of a protoplanetary disk with ALMA and showed that the bump or dip
features seen in the collision profiles should be detectable through the dust emission. From a dust emission profile with sufficiently strong radial features, one could constrain the
properties of a perturbing planet on an exterior orbit in the following way: (1) identify the locations of two MMRs in the disc (2) measure the spacing between them to verify that one of the
features is indeed associated with the 2:1 MMR (3) determine whether an under- or overdensity in the radial dust profile exists near this location (4) use the presence of one of these two
features to determine the region of the eccentricity-mass plane in which the planet lies.
In line with \citet{2017ApJ...850..103B}, the results of this work suggest that, in planet forming discs, the observed dust structure should not be
interpreted as simply the result of primordial dust that is perturbed by the planets and gas. Instead, some morphological features could be the product
of collisional dynamics between larger bodies. Within the confines of the simulations presented here, we have shown that this second-generation dust
can give rise to significant features in the dust emission profile. For this reason, we emphasize that planet-forming discs should be thought of as
existing along an evolutionary continuum, with the youngest discs being dominated by primordial dust and gas and the most evolved discs being
comprised of mainly collisionally generated dust.
The results presented here appear to be broadly consistent with the planetesimal dynamics seen by \citet{2016ApJ...818..159T}. In their case, a dip
feature can be seen in the azimuthally averaged surface density profile, although they did not test a high enough eccentricity planet in any of their
simulations to produce a bump. Another thing to note is that the gap morphology seen by \citet{2016ApJ...818..159T} was markedly different for the
2:1 exterior MMR. Whether this would alter the radial collision profile for the exterior, rather than the interior 2:1 resonance, is not immediately clear.
On one hand, the solid surface density diminishes with distance ($\sim r^{-3/2}$ for the MMSN \citep{1981PThPS..70...35H}), which would weaken the signal from locally
generated collisional dust. On the other hand, exterior MMRs are quite effective at capturing inward drifting bodies \citep{1985Icar...62...16W}, which could locally enhance the
surface density well beyond the MMSN value. A full treatment of the collisional cascade process and the radial drift of debris is a subject that we leave for future work.
Compared to using the corotating gap opened by the planet as a diagnostic \citep{2013ApJ...777L..31D, 2016ApJ...820...29D}, measuring
variations in the dust emission near MMRs is much more subtle and requires much higher spatial resolution and sensitivity. As we have shown, the
bump vs dip feature near the 2:1 MMR is marginally detectable with ALMA for the nearest protoplanetary discs, so long as the dust profile closely matches the collisional profile of
the planetesimals. Another complicating factor is that
the inner $\sim$ 10 au of most protoplanetary discs are optically thick in the sub-mm. The solution to this problem is to instead observe at longer
wavelengths where the inner disc is optically thin. This comes at the expense of much poorer resolution. However, future radio facilities like the NG-
VLA are expected to achieve sub-au resolution for nearby planet-forming discs \citep{2018ASPC..517..147R}. In the more near-term, techniques
like Gaussian process fitting present a promising way to recover substructure at sub-beam resolution \citep{2020arXiv200507709J}.
\section*{Acknowledgements}
This work was facilitated through the use of advanced computational, storage and networking infrastructure provided by the Hyak supercomputer system at the
University of Washington. We acknowledge the people of the Dkhw’Duw’Absh, the Duwamish Tribe, the Muckleshoot Tribe, and other tribes on whose traditional
lands we have performed this work.
\textit{Software:} Astropy \citep{2013A&A...558A..33A}, {\sc CASA} \citep{2007ASPC..376..127M}, {\sc ChaNGa} \citep{2008IEEEpds...ChaNGa, 2015AphCom..2..1}, Matplotlib
\citep{2007CSE.....9...90H}, NumPy \citep{2011CSE....13b..22V}, Pandas \citep{mckinney-proc-scipy-2010}, {\sc PYNBODY} \citep{2013ascl.soft05002P}, {\sc sklearn} \citep{scikit-learn}
\section*{Data Availability}
The data presented in this article are available on our GitHub repository, at \url{https://github.com/spencerw/planetesimal_coll_paper}.
\bibliographystyle{mnras}
|
\section{The \texttt{PSRL-POMDP} Algorithm}
\label{sec: posterior sampling algorithm}
We propose a general Posterior Sampling Reinforcement Learning for POMDPs (\texttt{PSRL-POMDP}) algorithm (Algorithm~\ref{alg: posterior sampling}) for both the finite-parameter and the general case. The algorithm maintains a joint distribution on the unknown parameter $\theta_*$ as well as the state $s_t$. \texttt{PSRL-POMDP} takes the prior distributions $h$ and $f$ as input. At time $t$, the agent computes the posterior distribution $f_t(\cdot)$ on the unknown parameter $\theta_*$ as well as the posterior conditional probability mass function (pmf) $h_t(\cdot;\theta)$ on the state $s_t$ for $\theta \in \Theta$. Upon taking action $a_t$ and observing $o_{t+1}$, the posterior distribution at time $t+1$ can be updated by applying the Bayes' rule as\footnote{When the parameter set is finite, $\int_{\theta}$ should be replaced with $\sum_\theta$.}
\begin{align}
\label{eq: update rule of ft and ht}
f_{t+1}(\theta) &= \frac{\sum_{s, s'}\eta(o_{t+1}|s') \theta(s'|s, a_t)h_t(s;\theta)f_t(\theta)}{\int_{\theta} \sum_{s,s'}\eta(o_{t+1}|s') \theta(s'|s, a_t)h_t(s;\theta)f_t(\theta)d\theta}, \nonumber \\
h_{t+1}(\cdot ;\theta) &= \tau (h_t(\cdot;\theta), a_t, o_{t+1};\theta),
\end{align}
with the initial condition
\begin{align}
\label{eq: initial update of ht and ft}
f_1(\theta) &= \frac{\sum_s \eta(o_1|s)h(s;\theta)f(\theta)}{\int_{\theta} \sum_s \eta(o_1|s)h(s;\theta)f(\theta)d\theta}, \nonumber \\
h_1(s;\theta) &= \frac{\eta(o_1|s)h(s;\theta)}{\sum_s\eta(o_1|s)h(s;\theta)}.
\end{align}
Recall that $\tau (h_t(\cdot;\theta), a_t, o_{t+1};\theta)$ is a compact notation for \eqref{eq: update rule of ht}.
In the special case of perfect observation at time $t$, $h_{t}(s;\theta) = \mathbbm{1}(s_{t}=s)$ for all $\theta \in \Theta$ and $s \in {\mathcal{S}}$. Moreover, the update rule of $f_{t+1}$ reduces to that of fully observable MDPs (see Eq. (4) of \cite{ouyang2017learning}) in the special case of perfect observation at time $t$ and $t+1$.
Let $n_{t}(s, a) = \sum_{\tau=1}^{t-1} \mathbbm{1}(s_\tau=s, a_\tau=a)$ be the number of visits to state-action $(s, a)$ by time $t$. The number of visits $n_{t}$ plays an important role in learning for MDPs \citep{jaksch2010near,ouyang2017learning} and is one of the two criteria to determine the length of the episodes in the \texttt{TSDE} algorithm for MDPs \citep{ouyang2017learning}. However, in POMDPs, $n_{t}$ is not ${\mathcal{F}}_{{(t-1)}+}$-measurable since the states are not observable. Instead, let $\tilde n_{t}(s, a) := \field{E}[n_{t}(s, a) | {\mathcal{F}}_{{(t-1)}+}]$, and define the \textit{pseudo-count} $\tilde m_t$ as follows:
\begin{definition}
$(\tilde m_t)_{t=1}^T$ is a pseudo-count if it is a non-decreasing, integer-valued sequence of random variables such that $\tilde m_t$ is ${\mathcal{F}}_{{(t-1)}+}$-measurable, $\tilde m_t(s, a) \geq \ceil{\tilde n_t(s, a)}$, and $\tilde m_t(s, a) \leq t$ for all $t \leq T+1$.
\end{definition}
An example of such a sequence is simply $\tilde m_t(s, a) = t$ for all state-action pair $(s, a) \in {\mathcal{S}} \times {\mathcal{A}}$. This is used in Section~\ref{sec: finite}. Another example is $\tilde m_t(s, a) := \max \{\tilde m_{t-1}(s, a), \ceil{\tilde n_t(s, a)}\}$ with $\tilde m_0(s, a) = 0$ for all $(s, a) \in {\mathcal{S}} \times {\mathcal{A}}$ which is used in Section~\ref{sec: general}. Here $\ceil{\tilde n_t(s, a)}$ is the smallest integer that is greater than or equal to $\tilde n_t(s, a)$. By definition, $\tilde m_t$ is integer-valued and non-decreasing which is essential to bound the number of episodes in the algorithm for the general case (see Lemma~\ref{lem: number of episodes}).
\begin{algorithm}
\caption{\textsc{\texttt{PSRL-POMDP}}}
\label{alg: posterior sampling}
\begin{algorithmic}[1]
\REQUIRE prior distributions $f(\cdot), h(\cdot;\cdot)$ \\
\textbf{Initialization: }$t \gets 1, t_1 \gets 0$\\
Observe $o_1$ and compute $f_1, h_1$ according to \eqref{eq: initial update of ht and ft} \\
\FOR{{\normalfont episodes} $k=1, 2, \cdots$}
\STATE $T_{k-1} \gets t - t_k$\\
\STATE $t_k \gets t$\\
\STATE Generate $\theta_k \sim f_{t_k}(\cdot)$ and compute $\pi_k(\cdot) = \pi^*(\cdot;\theta_k)$ from \eqref{eq: the optimal policy}\\
\WHILE{$t \leq \texttt{SCHED}(t_k, T_{k-1})$ and $\tilde m_{t}(s, a) \leq 2 \tilde m_{t_k}(s, a)$ for all $(s, a) \in {\mathcal{S}} \times {\mathcal{A}}$}
\STATE Choose action $a_t = \pi_k(h_t(\cdot;\theta_k))$ and observe $o_{t+1}$\\
\STATE Update $f_{t+1}, h_{t+1}$ according to \eqref{eq: update rule of ft and ht}\\
\STATE $t \gets t+1$
\ENDWHILE
\ENDFOR
\end{algorithmic}
\end{algorithm}
Similar to the \texttt{TSDE} algorithm for fully observable MDPs, \texttt{PSRL-POMDP} algorithm proceeds in episodes. In the beginning of episode $k$, POMDP $\theta_k$ is sampled from the posterior distribution $f_{t_k}$ where $t_k$ denotes the start time of episode $k$. The optimal policy $\pi^*(\cdot; \theta_k)$ is then computed and used during the episode. Note that the input of the policy is $h_t(\cdot;\theta_k)$. The intuition behind such a choice (as opposed to the belief $b_t(\cdot) := \int_{\theta} h_t(\cdot;\theta)f_t(\theta)d\theta$) is that during episode $k$, the agent treats $\theta_k$ to be the true POMDP and adopts the optimal policy with respect to it. Consequently, the input to the policy should also be the conditional belief with respect to the sampled $\theta_k$.
A key factor in designing posterior sampling based algorithms is the design of episodes. Let $T_k$ denote the length of episode $k$. In \texttt{PSRL-POMDP}, a new episode starts if either $t > \texttt{SCHED}(t_k, T_{k-1})$ or $\tilde m_{t}(s, a) > 2\tilde m_{t_k}(s, a)$. In the finite parameter case (Section~\ref{sec: finite}), we consider $\texttt{SCHED}(t_k, T_{k-1}) = 2t_k$ and $\tilde m_t(s, a) = t$. With these choices, the two criteria coincide and ensure that the start time and the length of the episodes are deterministic. In Section~\ref{sec: general}, we use $\texttt{SCHED}(t_k, T_{k-1}) = t_k + T_{k-1}$ and $\tilde m_t(s, a) := \max \{\tilde m_{t-1}(s, a), \ceil{\tilde n_t(s, a)}\}$. This guarantees that $T_k \leq T_{k-1} + 1$ and $\tilde m_t(s, a) \leq 2\tilde m_{t_k}(s, a)$. These criteria are previously introduced in the \texttt{TSDE} algorithm \citep{ouyang2017learning} except that \texttt{TSDE} uses the true count $n_t$ rather than $\tilde m_t$.
\section{Regret Decomposition}
\label{app: aux}
\begin{lemma}
\label{lem: decomposition}
$R_T$ can be decomposed as $R_T = H\field{E_{\theta_*}}[K_T] + R_1 + R_2 + R_3$, where
\begin{align*}
R_1 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}T_k\Big[J(\theta_k) - J(\theta_*)\Big]}, \\
R_2 &:= H\field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sbr{\sum_{s'}\abr{\theta_*(s'|s_t, a_t) - \theta_k(s'|s_t, a_t)} + \sum_s \abr{h_t(s;\theta_*) - h_t(s;\theta_k)}}}, \\
R_3 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[c(h_t(\cdot;\theta_*), a_t) - c(h_t(\cdot;\theta_k), a_t)\Big]}.
\end{align*}
\end{lemma}
\begin{proof}
First, note that $\field{E_{\theta_*}}[C(s_t, a_t)| {\mathcal{F}}_{t+}] = c(h_t(\cdot; \theta_*), a_t)$ for any $t \geq 1$. Thus, we can write:
\begin{align*}
R_T &= \field{E_{\theta_*}}\Big[\sum_{t=1}^T\Big[C(s_t, a_t) - J(\theta_*)\Big]\Big] = \field{E_{\theta_*}}\Big[\sum_{t=1}^T\Big[c(h_t(\cdot; \theta_*), a_t) - J(\theta_*)\Big]\Big].
\end{align*}
During episode $k$, by the Bellman equation for the sampled POMDP $\theta_k$ and that $a_t = \pi^*(h_t(\cdot;\theta_k);\theta_k)$, we can write:
\begin{align*}
&c(h_t(\cdot;\theta_k), a_t) - J(\theta_k) = v(h_t(\cdot;\theta_k);\theta_k) - \sum_{o}P(o|h_t(\cdot;\theta_k), a_t;\theta_k)v(h';\theta_k),
\end{align*}
where $h' = \tau(h_t(\cdot;\theta_k), a_t, o;\theta_k)$.
Using this equation, we proceed by decomposing the regret as
\begin{align*}
R_T &= \field{E_{\theta_*}}\Big[\sum_{t=1}^T\Big[c(h_t(\cdot;\theta_*), a_t) - J(\theta_*)\Big]\Big] \\
&= \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[c(h_t(\cdot;\theta_*), a_t) - J(\theta_*)\Big]\Big] \\
&= \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\underbrace{\sum_{t=t_k}^{t_{k+1}-1}\Big[v(h_t(\cdot;\theta_k); \theta_k) - v(h_{t+1}(\cdot;\theta_k); \theta_k)\Big]}_{\text{telescopic sum}}\Big] + \underbrace{\field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}T_k\Big[J(\theta_k) - J(\theta_*)\Big]}}_{=: R_1} \\
&\qquad + \underbrace{\field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[v(h_{t+1}(\cdot;\theta_k); \theta_k) - \sum_{o \in O}P(o | h_t(\cdot;\theta_k), a_t; \theta_k)v(h'; \theta_k)\Big]\Big]}_{=: R_2'} \\
&\qquad + \underbrace{\field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[c(h_t(\cdot;\theta_*), a_t) - c(h_t(\cdot;\theta_k), a_t)\Big]\Big]}_{=: R_3}
\end{align*}
where $K_T$ is the number of episodes upto time $T$, $t_k$ is the start time of episode $k$ (we let $t_k = T+1$ for all $k > K_T$). The telescopic sum is equal to $v(h_{t_k}(\cdot;\theta_k); \theta_k) - v(h_{t_{k+1}}(\cdot;\theta_k); \theta_k) \leq H$. Thus, the first term on the right hand side is upper bounded by $H\field{E_{\theta_*}}[K_T]$. Suffices to show that $R_2' \leq R_2$. Throughout the proof, we change the order of expectation and summation at several points. A rigorous proof for why this is allowed in the case that $K_T$ and $t_k$ are random variables is presented in the proof of Lemma~\ref{lem: r3}.
We proceed by bounding the term $R_2'$. Recall that $h' = \tau(h_t(\cdot;\theta_k), a_t, o; \theta_k)$ and $h_{t+1}(\cdot; \theta_k) = \tau(h_t(\cdot;\theta_k), a_t, o_{t+1}; \theta_k)$.
Conditioned on ${\mathcal{F}}_{t}, \theta_*, \theta_k$, the only random variable in $h_{t+1}(\cdot;\theta_k)$ is $o_{t+1}$ ($a_t=\pi^*(h_t(\cdot;\theta_k); \theta_k)$ is measurable with respect to the sigma algebra generated by ${\mathcal{F}}_t, \theta_k$). Therefore,
\begin{align}
\label{eq: app expected v}
&\field{E_{\theta_*}}\Big[v(h_{t+1}(\cdot;\theta_k); \theta_k) | {\mathcal{F}}_{t}, \theta_k\Big] = \sum_{o \in O} v(h'; \theta_k) \field{P_{\theta_*}}(o_{t+1}=o | {\mathcal{F}}_{t}, \theta_k).
\end{align}
We claim that $\field{P_{\theta_*}}(o_{t+1}=o | {\mathcal{F}}_{t}, \theta_k) = P(o | h_t(\cdot;\theta_*), a_t; \theta_*)$: by the total law of probability and that $\field{P_{\theta_*}} (o_{t+1}=o | s_{t+1}=s', {\mathcal{F}}_{t}, \theta_k) = \eta(o|s')$, we can write
\begin{align*}
&\field{P_{\theta_*}}(o_{t+1}=o | {\mathcal{F}}_{t}, \theta_k) = \sum_{s'} \eta(o|s') \field{P_{\theta_*}}(s_{t+1}=s' | {\mathcal{F}}_{t}, \theta_k).
\end{align*}
Note that
\begin{align*}
\field{P_{\theta_*}}(s_{t+1}=s' | {\mathcal{F}}_{t}, \theta_k) &= \sum_{s} \field{P_{\theta_*}}(s_{t+1}=s' | s_t=s, {\mathcal{F}}_t, a_t, \theta_k)\field{P_{\theta_*}}(s_t=s | {\mathcal{F}}_t, \theta_k) \\
&=\sum_{s} \theta_*(s' | s, a_t) \field{P_{\theta_*}}(s_t=s | {\mathcal{F}}_t).
\end{align*}
Thus,
\begin{align}
\label{eq: app p equal to p}
\field{P_{\theta_*}}(o_{t+1}=o | {\mathcal{F}}_{t}, \theta_k) &= \sum_{s, s'} \eta(o | s') \theta_*(s' | s, a_t) h_t(s; \theta_*) = P(o | h_t(\cdot;\theta_*), a_t; \theta_*).
\end{align}
Combining \eqref{eq: app p equal to p} with \eqref{eq: app expected v} and substituting into $R_2'$, we get
\begin{align*}
R_2' &= \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[\sum_{o \in O}\Big(P(o | h_t(\cdot; \theta_*), a_t; \theta_*) - P(o | h_t(\cdot;\theta_k), a_t; \theta_k)\Big)v(h'; \theta_k)\Big]\Big].
\end{align*}
Recall that for any $\theta \in \Theta$, $P(o | h_t(\cdot;\theta), a_t; \theta) = \sum_{s'} \eta(o | s') \sum_{s} \theta(s' | s, a_t) h_t(s; \theta)$. Thus,
\begin{align}
\label{eq: app r1bar in terms of h minus h}
R_2' &= \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sum_{o, s'}v(h'; \theta_k) \eta(o | s')\sum_{s} \theta_*(s' | s, a_t) h_t(s; \theta_*)\Big] \nonumber \\
&\qquad- \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sum_{o, s'}v(h'; \theta_k) \eta(o | s')\sum_{s} \theta_k(s' | s, a_t) h_t(s; \theta_*)\Big] \nonumber \\
&\qquad + \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sum_{o, s'}v(h'; \theta_k) \eta(o | s')\sum_{s}\theta_k(s' | s, a_t) \big(h_t(s; \theta_*) -h_t(s; \theta_k)\big)\Big].
\end{align}
For the first term, note that conditioned on ${\mathcal{F}}_{t}, \theta_*$, the distribution of $s_t$ is $h_t(\cdot;\theta_*)$ by the definition of $h_t$. Furthermore, $a_t$ is measurable with respect to the sigma algebra generated by ${\mathcal{F}}_t, \theta_k$ since $a_t = \pi^*(h_t(\cdot;\theta_k); \theta_k)$. Thus, we have
\begin{align}
\label{eq: app removing ht theta star}
&\field{E_{\theta_*}}\Big[v(h'; \theta_k)\sum_{s} \theta_*(s' | s, a_t) h_t(s; \theta_*)\Big|{\mathcal{F}}_{t}, \theta_k\Big] = v(h'; \theta_k)\field{E_{\theta_*}}\Big[\theta_*(s'|s_t, a_t)\Big|{\mathcal{F}}_{t}, \theta_k\Big].
\end{align}
Similarly, for the second term on the right hand side of \eqref{eq: app r1bar in terms of h minus h}, we have
\begin{align}
\label{eq: app removing ht theta k}
&\field{E_{\theta_*}}\Big[v(h'; \theta_k)\sum_{s} \theta_k(s' | s, a_t) h_t(s; \theta_*)\Big|{\mathcal{F}}_{t}, \theta_k \Big] = v(h'; \theta_k)\field{E_{\theta_*}}\Big[\theta_k(s'|s_t, a_t)\Big|{\mathcal{F}}_{t}, \theta_k\Big].
\end{align}
Replacing \eqref{eq: app removing ht theta star}, \eqref{eq: app removing ht theta k} into \eqref{eq: app r1bar in terms of h minus h} and using the tower property of conditional expectation, we get
\begin{align}
\label{eq: app intermediate bound of r1bar}
R_2' &= \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[\sum_{s'}\sum_{o} v(h'; \theta_k) \eta(o | s')\Big(\theta_*(s' | s_t, a_t) - \theta_k(s'|s_t, a_t)\Big)\Big]\Big] \nonumber \\
&\qquad+ \field{E_{\theta_*}}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[\sum_{s'}\sum_{o}v(h'; \theta_k) \eta(o | s')\sum_{s} \theta_k(s' | s, a_t) \big(h_t(s; \theta_*) -h_t(s; \theta_k)\big)\Big]\Big].
\end{align}
Since $\sup_{b \in \Delta_{\mathcal{S}}}v(b, \theta_k) \leq H$ and $\sum_o \eta(o|s') = 1$, the inner summation for the first term on the right hand side of \eqref{eq: app intermediate bound of r1bar} can be bounded as
\begin{align}
\label{eq: app common lem 1 tmp1}
&\sum_{o \in O}v(h'; \theta_k) \eta(o | s')\Big(\theta_*(s' | s_t, a_t) - \theta_k(s'|s_t, a_t)\Big) \leq H\Big|\theta_*(s' | s_t, a_t) - \theta_k(s'|s_t, a_t)\Big|.
\end{align}
Using $\sup_{b \in \Delta_{\mathcal{S}}}v(b, \theta_k) \leq H$, $\sum_o \eta(o|s') = 1$ and $\sum_{s'}\theta_k(s'|s, a_t) = 1$, the second term on the right hand side of \eqref{eq: app intermediate bound of r1bar} can be bounded as
\begin{align}
\label{eq: app common lem 1 tmp2}
&\sum_{s'}\sum_{o \in O}v(h'; \theta_k) \eta(o | s')\sum_{s} \theta_k(s' | s, a_t)\big|h_t(s;\theta_*) - h_t(s;\theta_k)\big| \leq H\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_k)\big| \nonumber \\
\end{align}
Substituting \eqref{eq: app common lem 1 tmp1} and \eqref{eq: app common lem 1 tmp2} into \eqref{eq: app intermediate bound of r1bar} proves that $R_2' \leq R_2$.
\end{proof}
\section{Proofs of Section~\ref{sec: finite}}
\label{app: finite}
\subsection{Proof of Lemma~\ref{lem: 1-ft}}
\textbf{Lemma} (restatement of Lemma~\ref{lem: 1-ft})\textbf{.} Suppose Assumption~\ref{ass: kl divergence} holds. Then, there exist constants $\alpha > 1$ and $\beta > 0$ such that
\begin{align*}
\field{E}[1 - f_t(\theta_*) | \theta_*] \leq \alpha \exp(-\beta t).
\end{align*}
\begin{proof}
Let $\tau_t$ be the trajectory $\{a_1, o_1, \cdots, a_{t-1}, o_{t-1}, o_t\}$ and define the likelihood function
\begin{align}
{\mathcal{L}}(\tau_t | \theta) &:= \prime(\tau_t | \theta) = \prime(o_1|\theta)\prod_{\tau=2}^t \prime(o_{\tau}|o_{1:\tau-1}, a_{1:\tau-1}|\theta) = \prime(o_1|\theta)\prod_{\tau=2}^t \nu_{\theta}^{o_{1:\tau-1}, a_{1:\tau-1}}(o_\tau)
\end{align}
Note that $\prime(o_1|\theta)$ is independent of $\theta$, thus
\begin{align*}
\frac{{\mathcal{L}}(\tau_t | \theta)}{{\mathcal{L}}(\tau_t | \gamma)} = \prod_{\tau=2}^t\frac{\nu_{\theta}^{o_{1:\tau-1}, a_{1:\tau-1}}(o_\tau)}{\nu_{\gamma}^{o_{1:\tau-1}, a_{1:\tau-1}}(o_\tau)}
\end{align*}
Recall that $f_t(\cdot)$ is the posterior associated with the likelihood given by
\begin{align*}
f_t(\theta) = \frac{{\mathcal{L}}(\tau_t | \theta) f(\theta)}{\sum_{\gamma \in \Theta}{\mathcal{L}}(\tau_t | \gamma)f(\gamma)}.
\end{align*}
We now proceed to lower bound $f_t(\theta_*)$. We can write
\begin{align*}
f_t(\theta_*) &= \frac{{\mathcal{L}}(\tau_t | \theta_*) f(\theta_*)}{\sum_{\theta}{\mathcal{L}}(\tau_t | \theta)f(\theta)} = \frac{1}{1 + \sum_{\theta \neq \theta_*}\frac{f(\theta)}{f(\theta_*)} \frac{{\mathcal{L}}(\tau_t | \theta)}{{\mathcal{L}}(\tau_t | \theta_*)}} \\
&= \frac{1}{1 + \sum_{\theta \neq \theta_*}\frac{f(\theta)}{f(\theta_*)} \exp(-\sum_{\tau=1}^t\log\Lambda_\tau^\theta)},
\end{align*}
where we define $\Lambda_1^\theta := 1$ and for $\tau \geq 2$,
\begin{align*}
\Lambda_\tau^\theta := \frac{\nu_{\theta_*}^{o_{1:\tau-1}, a_{1:\tau-1}}(o_\tau)}{\nu_{\theta}^{o_{1:\tau-1}, a_{1:\tau-1}}(o_\tau)}
\end{align*}
Note that without loss of generality, we can assume that the denominator in the definition of $\Lambda_\tau^\theta$ is positive (otherwise, ${\mathcal{L}}(\tau_t | \theta) = 0$ and can be excluded from the denominator of $ f_t(\theta_*)$) and thus $\Lambda_\tau^\theta$ is well-defined.
Denote by $Z_t^\theta := \sum_{\tau=1}^t\log\Lambda_\tau^\theta$ and decompose it as $Z_t^\theta = M_t^\theta + A_t^\theta$ where
\begin{align*}
M_t^\theta &:= \sum_{\tau=1}^t\Big(\log\Lambda_\tau^\theta - \field{E}\Big[\log\Lambda_\tau^\theta \big| {\mathcal{F}}_{\tau-1}, \theta_*\Big]\Big), \\
A_t^\theta &:= \sum_{\tau=1}^t \field{E}\Big[\log\Lambda_\tau^\theta\big| {\mathcal{F}}_{\tau-1}, \theta_*\Big].
\end{align*}
Note that the terms inside the first summation constitute a martingale difference sequence with respect to the filtration $({\mathcal{F}}_\tau)_{\tau \geq 1}$ and conditional probability $\prime(\cdot|\theta_*)$. Each term is bounded as $|\log\Lambda_\tau^\theta - \field{E}[\log\Lambda_\tau^\theta | {\mathcal{F}}_{\tau-1}, \theta_*]| \leq d$ for some $d > 0$.
The second term, $A_t^\theta$ can be lower bounded using Assumption~\ref{ass: kl divergence} as follows
\begin{align*}
\field{E}\Big[\log\Lambda_\tau^\theta\big| {\mathcal{F}}_{\tau-1}, \theta_*\Big] &= \field{E}\bigg[\field{E}\Big[\log\Lambda_\tau^\theta\big| {\mathcal{F}}_{\tau-1}, a_{\tau-1}, \theta_*\Big]\Big| {\mathcal{F}}_{\tau-1}, \theta_*\bigg] \\
&= \field{E}\Big[{\mathcal{K}}(\nu_{\theta_*}^{o_{1:\tau-1}, a_{1:\tau-1}} \| \nu_{\theta}^{o_{1:\tau-1}, a_{1:\tau-1}})\big|{\mathcal{F}}_{\tau-1}, \theta_*\Big] \geq \epsilon
\end{align*}
Summing over $\tau$ implies that
\begin{align}
\label{eq: atdelta bound}
A_t^\theta \geq \epsilon t.
\end{align}
To bound $M_t^\theta$, let $0 < \delta < \epsilon$, and apply Azuma's inequality to obtain
\begin{align*}
\prime\Big(|M_t^\theta| \geq \delta t \big| \theta_*\Big) \leq 2\exp(-\frac{\delta^2t}{2d^2}).
\end{align*}
Union bound over all $\theta \neq \theta_*$ implies that the event $B_t^\delta := \cap_{\theta \neq \theta_*}\{|M_t^\theta| \leq \delta t\}$ happens with probability at least $1 - 2(|\Theta|-1)\exp(-\frac{\delta^2t}{2d^2})$. If $B_t^\delta$ holds, then $-M_t^\theta \leq \delta t$ for all $\theta \neq \theta_*$. Combining this with \eqref{eq: atdelta bound} implies that $\exp(-M_t^\theta - A_t^\theta) \leq \exp(\delta t - \epsilon t)$.
Therefore,
\begin{align*}
\field{E}[f_t(\theta_*) | \theta_*] &= \field{E}\Bigg[\frac{1}{1 + \sum_{\theta \neq \theta_*}\frac{f(\theta)}{f(\theta_*)} \exp(-M_t^\theta - A_t^\theta)}\bigg| \theta_* \Bigg] \\
&\geq \field{E}\Bigg[\frac{\mathbbm{1}(B_\delta^t)}{1 + \sum_{\theta \neq \theta_*}\frac{f(\theta)}{f(\theta_*)} \exp(\delta t - \epsilon t)} \bigg| \theta_* \Bigg] \\
&= \frac{\prime(B_\delta^t | \theta_*)}{1 + \frac{1 - f(\theta_*)}{f(\theta_*)} \exp(\delta t - \epsilon t)} \\
&\geq \frac{1 - 2(|\Theta| - 1)\exp(-\frac{\delta^2t}{2d^2})}{1 + \frac{1 - f(\theta_*)}{f(\theta_*)} \exp(\delta t - \epsilon t)}.
\end{align*}
Now, by choosing $\delta = \epsilon/2$, and constants $\alpha = 2\max \{\max_{\theta \in \Theta}\frac{1 - f(\theta)}{f(\theta)}, 2(|\Theta|-1)\}$, and $\beta = \min \{\frac{\epsilon}{2}, \frac{\epsilon^2}{8d^2}\}$, we have
\begin{align*}
\field{E}[1 - f_t(\theta_*) | \theta_*] &\leq 1 - \frac{1 - 2(|\Theta| - 1)\exp(-\frac{\delta^2t}{2d^2})}{1 + \frac{1 - f(\theta_*)}{f(\theta_*)} \exp(\delta t - \epsilon t)} \\
&= \frac{\frac{1 - f(\theta_*)}{f(\theta_*)} \exp(\delta t - \epsilon t) + 2(|\Theta| - 1)\exp(-\frac{\delta^2t}{2d^2})}{1 + \frac{1 - f(\theta_*)}{f(\theta_*)} \exp(\delta t - \epsilon t)} \\
&\leq \frac{1 - f(\theta_*)}{f(\theta_*)} \exp(\delta t - \epsilon t) + 2(|\Theta| - 1)\exp(-\frac{\delta^2t}{2d^2}) \\
&= \frac{1 - f(\theta_*)}{f(\theta_*)} \exp(-\frac{\epsilon t}{2}) + 2(|\Theta| - 1)\exp(-\frac{\epsilon^2t}{8d^2}) \\
&\leq \alpha \exp(-\beta t).
\end{align*}
\end{proof}
\section{Proofs of Section~\ref{sec: general}}
\label{app: general}
\subsection{Full Upper Bound on the Expected Regret of Theorem~\ref{thm: regret bound}}
\label{app: full upper bound}
The exact expression for the upper bound of the expected regret in Theorem~\ref{thm: regret bound} is
\begin{align*}
&\field{E}[R_T] = H\field{E}[K_T] + \field{E}[R_1] + \field{E}[R_2] + \field{E}[R_3] \\
&\leq (1 + H)\field{E}[K_T] + 12HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3} \\
&\quad+ (H+1)K_1 \field{E}\Big[\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}}\Big] + 2 + H \\
&\leq (1 + H)\sqrt{2T(1 + |{\mathcal{S}}||{\mathcal{A}}| \log (T+1))} + 12HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3} \\
&\quad+ 7(H+1)K_1\sqrt{2T}(1 + |{\mathcal{S}}||{\mathcal{A}}|\log (T+1)) \log\sqrt{2T} + 2 + H.
\end{align*}
\subsection{Finite-parameter Case Satisfies Assumptions~\ref{ass: concentrating belief} and \ref{ass: concentration}}
In this section, we show that Assumptions~\ref{ass: concentrating belief} and \ref{ass: concentration} are satisfied for the finite-parameter case i.e., $|\Theta| < \infty$ as long as the \texttt{PSRL-POMDP} generates a deterministic schedule. As an instance, a deterministic schedule can be generated by choosing $\tilde m_t(s, a) = t$ for all state-action pairs $(s, a)$ and running Algorithm~\ref{alg: posterior sampling} with either $\texttt{SCHED}(t_k, T_{k-1}) = 2t_k$ or $\texttt{SCHED}(t_k, T_{k-1}) = t_k + T_{k-1}$.
\begin{lemma}
\label{lem: assumptions satisfied for finite-parameter}
Assume $|\Theta| < \infty$. If Algorithm~ref{alg: posterior sampling} generates a deterministic schedule, then Assumptions~\ref{ass: concentrating belief} and \ref{ass: concentration} are satisfied.
\end{lemma}
\begin{proof}
Observe that the left hand side of \eqref{eq: ass concentrating belief} is zero if $\theta_{k(t)} = \theta_*$, and is upper bounded by $2$ if $\theta_{k(t)} \neq \theta_*$. Thus, we can write
\begin{align*}
\field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_{k(t)})\big|\Big | \theta_*\Big] \leq 2\prime(\theta_{k(t)} \neq \theta_* | \theta_*) = 2\field{E}\sbr{1 - f_{t_{k(t)}}(\theta_*) | \theta_*} \leq \alpha \exp(-\beta t_{k(t)}),
\end{align*}
which obviously satisfies Assumption~\ref{ass: concentrating belief} by choosing a large enough constant $K_1$. Here, the last equality is by Lemma~\ref{lem: 1-ft} and that the start time of episode $k(t)$ is deterministic.
To see why Assumption~\ref{ass: concentration} is satisfied, let $\hat \theta_t$ be the Maximum a Posteriori (MAP) estimator, i.e., $\hat \theta_t = \argmax_{\theta \in \Theta} f_t(\theta)$. Then, the left hand side of \eqref{eq: ass concentration} is equal to zero if $\hat \theta_t = \theta_*$. Note that this happens with high probability with the following argument:
\begin{align*}
\prime(\hat \theta_t \neq \theta_* | \theta_*) \leq \prime\rbr{f_t(\theta_*) \leq 0.5 | \theta_*} = \prime(1 - f_t(\theta_*) \geq 0.5 | \theta_*) \leq 2\field{E}[1 - f_t(\theta_*) | \theta_*] \leq 2\alpha \exp(-\beta t).
\end{align*}
Here the first inequality is by the fact that if $f_t(\theta_*) > 0.5$, then the MAP estimator would choose $\hat \theta_t = \theta_*$. The second inequality is by applying Markov inequality and the last inequality is by Lemma~\ref{lem: 1-ft}. Note that $\tilde m_t(s, a) \leq t$ by definition. We claim that Assumption~\ref{ass: concentration} is satisfied by choosing $K_2 = 2\sqrt{(-1/\beta)\log (\delta/2\alpha)}$. To see this, note that $2\alpha \exp(-\beta t) \leq \delta$ for $t \geq (-1/\beta)\log (\delta/2\alpha)$. In this case, \eqref{eq: ass concentration} automatically holds since with probability at least $1 - \delta$ the left hand side is zero. For $t < (-1/\beta)\log (\delta/2\alpha)$, note that the left hand side of \eqref{eq: ass concentration} can be at most 2. Therefore, $K_2$ can be found by solving $2 \leq K_2/\sqrt{(-1/\beta)\log (\delta/2\alpha)}$.
\end{proof}
\subsection{Auxiliary Lemmas for Section~\ref{sec: general}}
\begin{lemma}
\label{lem: j minus jstar}[Lemma 3 in \cite{ouyang2017learning}]
The term $\field{E}[R_1]$ can be bounded as $\field{E}[R_1] \leq \field{E}[K_T]$.
\end{lemma}
\begin{proof}
\begin{align*}
\field{E}[R_1] &= \field{E}\Big[\sum_{k=1}^{K_T}T_k\Big[J(\theta_k) - J(\theta_*)\Big]\Big] = \field{E}\Big[\sum_{k=1}^{\infty}\mathbbm{1}(t_k \leq T)T_kJ(\theta_k)\Big] - T\field{E}[J(\theta_*)].
\end{align*}
By monotone convergence theorem and the fact that $J(\theta_k) \geq 0$ and $T_k \leq T_{k-1} + 1$ (the first criterion in determining the episode length in Algorithm~\ref{alg: posterior sampling}), the first term can be bounded as
\begin{align*}
&\field{E}\Big[\sum_{k=1}^{\infty}\mathbbm{1}(t_k \leq T)T_kJ(\theta_k)\Big] = \sum_{k=1}^{\infty}\field{E}\Big[\mathbbm{1}(t_k \leq T)T_kJ(\theta_k)\Big] \\
&\leq \sum_{k=1}^{\infty}\field{E}\Big[\mathbbm{1}(t_k \leq T)(T_{k-1}+1)J(\theta_k)\Big].
\end{align*}
Note that $\mathbbm{1}(t_k \leq T)(T_{k-1} + 1)$ is ${\mathcal{F}}_{t_k}$-measurable. Thus, by the property of posterior sampling (Lemma~\ref{lem: property of posterior sampling}), $\field{E}[\mathbbm{1}(t_k \leq T)(T_{k-1}+1)J(\theta_k)] = \field{E}[\mathbbm{1}(t_k \leq T)(T_{k-1}+1)J(\theta_*)]$. Therefore,
\begin{align*}
\field{E}[R_1] &\leq \field{E}\Big[\sum_{k=1}^{\infty}\mathbbm{1}(t_k \leq T)(T_{k-1}+1)J(\theta_*)\Big] - T\field{E}[J(\theta_*)] \\
&= \field{E}\Big[J(\theta_*)(K_T + \sum_{k=1}^{K_T}T_{k-1})\Big] - T\field{E}[J(\theta_*)] \\
&= \field{E}[J(\theta_*)K_T] + \field{E}\Big[J(\theta_*)(\sum_{k=1}^{K_T}T_{k-1} - T)\Big]\leq \field{E}[K_T],
\end{align*}
where the last inequality is by the fact that $\sum_{k=1}^{K_T}T_{k-1} - T \leq 0$ and $0 \leq J(\theta_*) \leq 1$.
\end{proof}
\begin{lemma}
\label{lem: r3}
The term $\field{E}[R_3]$ can be bounded as
\begin{align*}
\field{E}[R_3] \leq K_1\field{E}\Big[\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}}\Big] + 1,
\end{align*}
where $K_1:= K_1(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ is the constant in Assumption~\ref{ass: concentrating belief}.
\end{lemma}
\begin{proof}
Recall that
\begin{align*}
\field{E}[R_3] &= \field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[c(h_t(\cdot;\theta_*), a_t) - c(h_t(\cdot;\theta_k), a_t)\Big]\Big].
\end{align*}
Let $k(t)$ be a random variable denoting the episode number at time $t$, i.e., $t_{k(t)} \leq t < t_{k(t) + 1}$ for all $t \leq T$.
By the definition of $c$, we can write
\begin{align*}
\field{E}[R_3] &= \field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sum_{s}C(s, a_t)\Big[h_t(s;\theta_*) - h_t(s;\theta_k)\Big]\Big] \\
&=\field{E}\Big[\sum_{t=1}^T \sum_{s}C(s, a_t)\Big[h_t(s;\theta_*) - h_t(s;\theta_{k(t)})\Big]\Big] \\
&\leq \sum_{t=1}^T \field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_{k(t)})\big|\Big] \\
&= \field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_k)\big|\Big]\Big],
\end{align*}
where
the inequality is by $0 \leq C(s, a_t) \leq 1$. Let $K_1:= K_1(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ be the constant in Assumption~\ref{ass: concentrating belief} and define event $E_1$ as the successful event of Assumption~\ref{ass: concentrating belief} where $\field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_k)\big|\Big] \leq \frac{K_1}{\sqrt{t_k}}$ happens. We can write
\begin{align*}
&\field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_k)\big|\Big] \\
&= \field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_k)\big|\Big] (\mathbbm{1}(E_1) + \mathbbm{1}(E_1^c)) \\
&\leq \frac{K_1}{\sqrt{t_k}} + 2\mathbbm{1}(E_1^c).
\end{align*}
Recall that by Assumption~\ref{ass: concentrating belief}, $\prime(E_1^c) \leq \delta$. Therefore,
\begin{align*}
\field{E}[R_3] \leq K_1\field{E}\Big[\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}}\Big] + 2T\delta.
\end{align*}
Choosing $\delta = \min(1/(2T), 1/(2HT))$ completes the proof.
\end{proof}
\begin{lemma}
\label{lem: r2bar}
The term $\bar R_2$ can be bounded as
\begin{align*}
\bar R_2 \leq H + 12HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3},
\end{align*}
where $K_2 := K_2(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ in Assumption~\ref{ass: concentration}.
\end{lemma}
\begin{proof}
Recall that
\begin{align}
\label{eq: app nice bound of r1bar}
\bar R_2 = H\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sum_{s'}\Big|\theta_*(s' | s_t, a_t) - \theta_k(s'|s_t, a_t)\Big|\Big].
\end{align}
We proceed by bounding the inner term of the above equation. For notational simplicity, define $z := (s, a)$ and $z_t := (s_t, a_t)$. Let $\hat \theta_{t_k}$ be the estimator in Assumption~\ref{ass: concentration} and define the confidence set $B_k$ as
\begin{align*}
B_k &:= \Big\{\theta \in \Theta_H : \sum_{s' \in {\mathcal{S}}} \Big|\theta(s'|z) - \hat \theta_k(s'|z)\Big| \leq \frac{K_2}{\sqrt{\max\{1, {\tilde m}_{t_k}(z)\}}}, \forall z \in {\mathcal{S}} \times {\mathcal{A}}\Big\},
\end{align*}
where $K_2 := K_2(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ is the constant in Assumption~\ref{ass: concentration}. Note that $B_k$ reduces to the confidence set used in \cite{jaksch2010near,ouyang2017learning} in the case of perfect observation by choosing $\tilde m_t(s, a) = n_t(s, a)$. By triangle inequality, the inner term in \eqref{eq: app nice bound of r1bar} can be bounded by
\begin{align*}
&\sum_{s'}\Big|\theta_*(s' | z_t) - \theta_k(s'|z_t)\Big| \\
&\leq \sum_{s'}\Big|\theta_*(s' | z_t) - \hat \theta_{t_k}(s'|z_t)\Big| + \sum_{s'}\Big|\theta_k(s' | z_t) - \hat \theta_{t_k}(s'|z_t)\Big| \\
&\leq 2\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big) + \frac{2K_2}{\sqrt{\max\{1, {\tilde m}_{t_k}(z_t)\}}}.
\end{align*}
Substituting this into \eqref{eq: app nice bound of r1bar} implies
\begin{align}
\label{eq: app two terms r1bar}
\bar R_2 &\leq 2H\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big)\Big] + 2H\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\frac{K_2}{\sqrt{\max\{1, {\tilde m}_{t_k}(z_t)\}}}\Big].
\end{align}
We need to bound these two terms separately.
\paragraph{Bounding the first term.} For the first term we can write:
\begin{align*}
\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big)\Big] &= \field{E}\Big[\sum_{k=1}^{K_T}T_k\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big)\Big] \\
&\leq T\field{E}\Big[\sum_{k=1}^{K_T}\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big)\Big] \\
&\leq T\sum_{k=1}^{T}\field{E}\Big[\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big)\Big],
\end{align*}
where the last inequality is by the fact that $K_T \leq T$. Now, observe that since $B_k$ is ${\mathcal{F}}_{t_k}$-measurable, Lemma \ref{lem: property of posterior sampling} implies that $\field{E}[\mathbbm{1}(\theta_k \notin B_k)] = \field{E}[\mathbbm{1}(\theta_* \notin B_k)]$. Moreover, by Assumption~\ref{ass: concentration}, $\field{E}[\mathbbm{1}(\theta_* \notin B_k)] = \prime(\theta_* \notin B_k) \leq \delta$. By choosing $\delta = \frac{1}{4T^2}$, we get
\begin{align}
\label{eq: app bound on the first term}
\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\big(\mathbbm{1}(\theta_* \notin B_k) + \mathbbm{1}(\theta_k \notin B_k)\big)\Big] \leq \frac{1}{2}.
\end{align}
\paragraph{Bounding the second term.} To bound the second term of \eqref{eq: app two terms r1bar}, observe that by the second criterion of the algorithm in choosing the episode length, we have $2\tilde m_{t_k}(z_t) \geq \tilde m_{t}(z_t)$. Thus,
\begin{align}
\label{eq: app nt and alpha}
&\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\frac{K_2}{\sqrt{\max\{1, {\tilde m}_{t_k}(z_t)\}}}\Big] \leq \field{E}\Big[\sum_{t=1}^{T}\frac{\sqrt 2K_2}{\sqrt{\max\{1, {\tilde m}_{t}(z_t)\}}}\Big] \nonumber \\
&= \sum_{t=1}^{T} \sum_{z}\field{E}\Big[\frac{\sqrt 2K_2 \mathbbm{1}(z_t=z)}{\sqrt{\max\{1, {\tilde m}_{t}(z)\}}}\Big] \nonumber \\
&= \sum_{t=1}^{T} \sum_{z}\field{E}\Big[\frac{\sqrt 2K_2 \mathbbm{1}(z_t=z)}{\sqrt{\max\{1, {\tilde m}_{t}(z)\}}}\mathbbm{1}\big({\tilde m}_{t}(z) \geq \alpha n_t(z)\big)\Big] \nonumber \\
&\qquad\qquad+ \sum_{t=1}^{T} \sum_{z}\field{E}\Big[\frac{\sqrt 2K_2 \mathbbm{1}(z_t=z)}{\sqrt{\max\{1, {\tilde m}_{t}(z)\}}}\mathbbm{1}\big({\tilde m}_{t}(z) < \alpha n_t(z)\big)\Big] \nonumber \\
&\leq \sum_{t=1}^{T} \sum_{z}\field{E}\Big[\frac{\sqrt 2K_2 \mathbbm{1}(z_t=z)}{\sqrt{\max\{1, \alpha{n}_{t}(z)\}}}\Big] + \sum_{t=1}^{T} \sum_{z}\field{E}\Big[\sqrt 2 K_2 \mathbbm{1}\big({\tilde m}_{t}(z) < \alpha n_t(z)\big)\Big].
\end{align}
Lemma~\ref{lem: pseudo count vs true count general case} implies that $\field{E}\Big[\mathbbm{1}\big({\tilde m}_{t}(z) < \alpha n_t(z)\big)\Big] = \prime({\tilde m}_{t}(z) < \alpha n_t(z)) \leq \alpha$. Thus, the second term in \eqref{eq: app nt and alpha} can be bounded by $\sqrt 2K_2|{\mathcal{S}}||{\mathcal{A}}|T\alpha$. To bound the first term of \eqref{eq: app nt and alpha}, we can write:
\begin{align*}
&\sum_{t=1}^{T} \sum_{z}\field{E}\Big[\frac{\sqrt 2K_2 \mathbbm{1}(z_t=z)}{\sqrt{\max\{1, \alpha{n}_{t}(z)\}}}\Big] \\
&\leq \sqrt \frac{2}{\alpha}K_2 \field{E}\Big[\sum_{z}\sum_{t=1}^{T}\frac{\mathbbm{1}(z_t=z)}{\sqrt{\max\{1, {n}_{t}(z)\}}}\Big].
\end{align*}
Observe that whenever $z_t=z$, $n_t(z)$ increases by $1$. Since, $n_t(z)$ is the number of visits to $z$ by time $t-1$ (including $t-1$ and excluding $t$), the denominator will be $1$ for the first two times that $z_t=z$. Therefore, the term inside the expectation can be bounded by
\begin{align*}
\sum_{z}\sum_{t=1}^{T}\frac{\mathbbm{1}(z_t=z)}{\sqrt{\max\{1, {n}_{t}(z)\}}} &= \sum_{z} \field{E}\Big[\mathbbm{1}(n_{T+1}(z)>0) + \sum_{j=1}^{n_{T+1}(z)-1}\frac{1}{\sqrt j}\Big] \\
&\leq \sum_{z} \field{E}\Big[\mathbbm{1}(n_{T+1}(z)>0) + 2\sqrt{n_{T+1}(z)}\Big]\\
& \leq 3\sum_z\sqrt{n_{T+1}(z)}.
\end{align*}
Since $\sum_z n_{T+1}(z) = T$, Cauchy Schwartz inequality implies
\begin{align*}
3\sum_z\sqrt{n_{T+1}(z)} \leq 3\sqrt{|{\mathcal{S}}||{\mathcal{A}}|\sum_{z}n_{T+1}(z)} = 3\sqrt{|{\mathcal{S}}||{\mathcal{A}}|T}.
\end{align*}
Therefore, the first term of \eqref{eq: app nt and alpha} can be bounded by
\begin{align*}
\sum_{t=1}^{T} \sum_{z}\field{E}\Big[\frac{\sqrt 2K_2 \mathbbm{1}(z_t=z)}{\sqrt{\max\{1, \alpha{n}_{t}(z)\}}}\Big] \leq 3K_2 \sqrt{\frac{2|{\mathcal{S}}||{\mathcal{A}}|T}{\alpha}}.
\end{align*}
Substituting this bound in \eqref{eq: app nt and alpha} along with the bound on the second term of \eqref{eq: app nt and alpha}, we obtain
\begin{align*}
&\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\frac{K_2}{\sqrt{\max\{1, {\tilde m}_{t_k}(z_t)\}}}\Big] \leq 3K_2 \sqrt{\frac{2|{\mathcal{S}}||{\mathcal{A}}|T}{\alpha}} + \sqrt 2K_2|{\mathcal{S}}||{\mathcal{A}}|T\alpha.
\end{align*}
$\alpha = (3/2)^{2/3}(|{\mathcal{S}}||{\mathcal{A}}|T)^{-1/3}$ minimizes the upper bound, and thus
\begin{align}
\label{eq: app bound on the second term}
\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}&\frac{K_2}{\sqrt{\max\{1, {\tilde m}_{t_k}(z_t)\}}}\Big] \leq 6K_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3}.
\end{align}
By substituting \eqref{eq: app bound on the first term} and \eqref{eq: app bound on the second term} into \eqref{eq: app two terms r1bar}, we get
\begin{align*}
\bar R_2 \leq H + 12HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3}.
\end{align*}
\end{proof}
\begin{lemma}
\label{lem: number of episodes}
The following inequalities hold:
\begin{enumerate}
\item The number of episodes $K_T$ can be bounded as $K_T \leq \sqrt{2T(1 + |{\mathcal{S}}||{\mathcal{A}}| \log (T+1))} = \ensuremath{\tilde{\mathcal{O}}}(\sqrt{|{\mathcal{S}}||{\mathcal{A}}|T})$.
\item The following inequality holds: $\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}} \leq 7\sqrt{2T}(1 + |{\mathcal{S}}||{\mathcal{A}}|\log (T+1)) \log\sqrt{2T} = \ensuremath{\tilde{\mathcal{O}}}(|{\mathcal{S}}||{\mathcal{A}}|\sqrt{T})$.
\end{enumerate}
\end{lemma}
\begin{proof}
We first provide an intuition why these results should be true. Note that the length of the episodes is determined by two criteria. The first criterion triggers when $T_k = T_{k-1} + 1$ and the second criterion triggers when the pseudo counts doubles for a state-action pair compared to the beginning of the episode. Intuitively speaking, the second criterion should only happen logarithmically, while the first criterion occurs more frequently. This means that one could just consider the first criterion for an intuitive argument. Thus, if we ignore the second criterion, we get $T_k = \ensuremath{\mathcal{O}}(k)$, $K_T = \ensuremath{\mathcal{O}}(\sqrt{T})$, and $t_k = \ensuremath{\mathcal{O}}(k^2)$ which implies $\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}} = \ensuremath{\mathcal{O}}(K_T) = \ensuremath{\mathcal{O}}(\sqrt T)$. The rigorous proof is stated in the following.
1. Define macro episodes with start times $t_{m_i}$ given by $t_{m_1} = t_1$ and $t_{m_i} :=$
\begin{align*}
\min\{t_k > t_{m_{i-1}} : \tilde m_{t_k}(s, a) > 2\tilde m_{t_{k-1}}(s, a) \text{ for some } (s, a)\}.
\end{align*}
Note that a new macro episode starts when the second criterion of episode length in Algorithm~\ref{alg: posterior sampling} triggers. Let $M_T$ be the random variable denoting the number of macro episodes by time $T$ and define $m_{M_T+1} = K_{T}+1$.
Let $\tilde T_i$ denote the length of macro episode $i$. Note that $\tilde T_i = \sum_{k=m_i}^{m_{i+1}-1}T_k$. Moreover, from the definition of macro episodes, we know that all the episodes in a macro episode except the last one are triggered by the first criterion, i.e., $T_k = T_{k-1}+1$ for all $m_i \leq k \leq m_{i+1}-2$. This implies that
\begin{align*}
&\tilde T_i = \sum_{k=m_i}^{m_{i+1}-1}T_k = T_{m_{i+1}-1} + \sum_{j=1}^{m_{i+1} - m_i - 1} (T_{m_{i}-1} + j) \\
&\geq 1 + \sum_{j=1}^{m_{i+1} - m_i - 1}(1 + j) = \frac{(m_{i+1} - m_i)(m_{i+1} - m_i+1)}{2}.
\end{align*}
This implies that $m_{i+1} - m_i \leq \sqrt{2\tilde T_i}$. Now, we can write:
\begin{align}
\label{eq: app kt and mt}
&K_T = m_{M_T+1} - 1 = \sum_{i=1}^{M_T}(m_{i+1} - m_i) \nonumber \\
&\leq \sum_{i=1}^{M_T}\sqrt{2 \tilde T_i} \leq \sqrt{2M_T\sum_{i}\tilde T_i} = \sqrt{2M_TT},
\end{align}
where the last inequality is by Cauchy-Schwartz.
Now, it suffices to show that $M_T \leq 1 + |{\mathcal{S}}||{\mathcal{A}}| \log (T+1)$. Let ${\mathcal{T}}_{s, a}$ be the start times at which the second criterion is triggered at state-action pair $(s, a)$, i.e.,
\begin{align*}
{\mathcal{T}}_{s, a} := \{t_k \leq T : \tilde m_{t_k}(s, a) > 2 \tilde m_{t_{k-1}}(s, a)\}.
\end{align*}
We claim that $|{\mathcal{T}}_{s, a}| \leq \log(\tilde m_{T+1}(s, a))$. To prove this claim, assume by contradiction that $|{\mathcal{T}}_{s, a}| \geq \log (\tilde m_{T+1}(s, a))+1$, then
\begin{align*}
&\tilde m_{t_{K_T}}(s, a) \geq \prod_{t_k \leq T, \tilde m_{t_{k-1}}(s, a) \geq 1}\frac{\tilde m_{t_{k}}(s, a)}{\tilde m_{t_{k-1}}(s, a)} \\
&\geq \prod_{t_k \in {\mathcal{T}}_{s, a}, \tilde m_{t_{k-1}}(s, a) \geq 1}\frac{\tilde m_{t_{k}}(s, a)}{\tilde m_{t_{k-1}}(s, a)} \\
&> \prod_{t_k \in {\mathcal{T}}_{s, a}, \tilde m_{t_{k-1}}(s, a) \geq 1} 2 = 2^{|{\mathcal{T}}_{s, a}|-1}\geq \tilde m_{T+1}(s, a),
\end{align*}
which is a contradiction. The second inequality is by the fact that $\tilde m_t(s, a)$ is non-decreasing, and the third inequality is by the definition of ${\mathcal{T}}_{s, a}$.
Therefore,
\begin{align}
\label{eq: app bound on mt}
&M_T \leq 1 + \sum_{s, a}|{\mathcal{T}}_{s, a}| \leq 1 + \sum_{s, a}\log(\tilde m_{T+1}(s, a)) \nonumber \\
&\leq 1 + |{\mathcal{S}}||{\mathcal{A}}| \log(\sum_{s, a}\tilde m_{T+1}(s, a)/|{\mathcal{S}}||{\mathcal{A}}|) \nonumber \\
&= 1 + |{\mathcal{S}}||{\mathcal{A}}|\log (T+1),
\end{align}
where the third inequality is due to the concavity of $\log$ and the last inequality is by the fact that $\tilde m_{T+1}(s, a) \leq T+1$.
2. First, we claim that $T_k \leq \sqrt{2T}$ for all $k \leq K_T$. To see this, assume by contradiction that $T_{k^*} > \sqrt{2T}$ for some $k^* \leq K_T$. By the first stopping criterion, we can conclude that $T_{k^*-1} > \sqrt{2T}-1$, $T_{k^*-2} > \sqrt{2T}-2$, \dots, $T_1 > \max\{\sqrt{2T} - k^* + 1, 0\}$ since the episode length can increase at most by one compared to the previous one. Note that $k^* \geq \sqrt{2T}-1$, because otherwise $T_1 > 2$ which is not feasible since $T_1 \leq T_0+1 = 2$.
Thus, $\sum_{k=1}^{k^*}T_k > 0.5\sqrt{2T}(\sqrt{2T}+1) > T$ which is a contradiction.
We now proceed to lower bound $t_k$. By the definition of macro episodes in part (1), during a macro episode length of the episodes except the last one are determined by the first criterion, i.e., for macro episode $i$, one can write $T_k = T_{k-1}+1$ for $m_i \leq k \leq m_{i+1}-2$. Hence, for $m_i \leq k \leq m_{i+1}-2$,
\begin{align*}
t_{k+1} &= t_k + T_k = t_k + T_{m_i-1} + k - (m_i-1) \\
&\geq t_k + k - m_i + 1.
\end{align*}
Recursive substitution of $t_k$ implies that $t_{k} \geq t_{m_i} + 0.5(k-m_i)(k-m_i+1)$ for $m_i \leq k \leq m_{i+1}-1$. Thus,
\begin{align}
\label{eq: app sum of Tk divided by tk}
&\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}} \leq \sqrt{2T} \sum_{i=1}^{M_T}\sum_{k=m_i}^{m_{i+1}-1}\frac{1}{\sqrt{t_k}} \nonumber \\
&\leq \sqrt{2T} \sum_{i=1}^{M_T}\sum_{k=m_i}^{m_{i+1}-1}\frac{1}{\sqrt{t_{m_i} + 0.5(k-m_i)(k-m_i+1)}}.
\end{align}
The denominator of the summands at $k=m_i$ is equal to $\sqrt{t_{m_i}}$. For other values of $k$ it can be lower bounded by $0.5(k-m_i)^2$. Thus,
\begin{align*}
&\sum_{i=1}^{M_T}\sum_{k=m_i}^{m_{i+1}-1}\frac{1}{\sqrt{t_{m_i} + 0.5(k-m_i)(k-m_i+1)}} \\
&\leq \sum_{i=1}^{M_T} \frac{1}{\sqrt{t_{m_i}}} + \sum_{i=1}^{M_T}\sum_{k=m_i+1}^{m_{i+1}-1} \frac{\sqrt 2}{k-m_i} \\
&\leq M_T + \sum_{i=1}^{M_T}\sum_{j=1}^{m_{i+1}-m_i-1} \frac{\sqrt 2}{j} \\
&\leq M_T + \sqrt 2 (M_T + \sum_{i=1}^{M_T}\log (m_{i+1}-m_i)) \\
&\leq M_T(1 + \sqrt 2) + \sqrt 2M_T \log(\frac{1}{M_T}\sum_{i=1}^{M_T}(m_{i+1}-m_i)) \\
&\leq M_T(1 + \sqrt 2) + \sqrt 2M_T \log\sqrt{2T} \\
&\leq 7M_T\log\sqrt{2T},
\end{align*}
where the second inequality is by $t_{m_i} \geq 1$, the third inequality is by the fact that $\sum_{j=1}^K 1/j \leq 1 + \int_1^K dx/x = 1 + \log K$, the forth inequality is by concavity of $\log$ and the fifth inequality is by the fact that $\sum_{i=1}^{M_T}(m_{i+1}-m_i) = m_{M_T+1} - 1 = K_T$ and $K_T/M_T \leq \sqrt{2T/M_T} \leq \sqrt{2T}$ (see \eqref{eq: app kt and mt}). Substituting this bound into \eqref{eq: app sum of Tk divided by tk} and using the upper bound on $M_T$ \eqref{eq: app bound on mt}, we can write
\begin{align*}
\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}} &\leq \sqrt{2T} \Big(7M_T\log\sqrt{2T}\Big) \\
&\leq 7\sqrt{2T}(1 + |{\mathcal{S}}||{\mathcal{A}}|\log (T+1)) \log\sqrt{2T}.
\end{align*}
\end{proof}
\section{Other Proofs}
\subsection{Proof of Lemma~\ref{lem: optimal policy}}
\textbf{Lemma} (restatement of Lemma~\ref{lem: optimal policy})\textbf{.} Suppose Assumption~\ref{ass: bellman equation} holds. Then, the policy $\pi^*(\cdot, \theta):\Delta_{\mathcal{S}} \to {\mathcal{A}}$ given by
\begin{equation}
\label{eq: the optimal policy}
\pi^*(b; \theta) := \argmin_{a \in {\mathcal{A}}} \{c(b, a) + \sum_{o \in O}P(o | b, a; \theta)v(b'; \theta)\}
\end{equation}
is the optimal policy with $J_{\pi^*}(h; \theta) = J(\theta)$ for all $h \in \Delta_{\mathcal{S}}$.
\begin{proof}
We prove that for any policy $\pi$, $J_\pi(h, \theta) \geq J_{\pi^*}(h, \theta) = J(\theta)$ for all $h \in \Delta_{\mathcal{S}}$. Let $\pi:\Delta_{\mathcal{S}} \to {\mathcal{A}}$ be an arbitrary policy. We can write
\begin{align*}
&J_\pi(h, \theta) = \limsup_{T \to \infty} \frac{1}{T} \sum_{t=1}^T \field{E}[C(s_t, \pi(h_t)) | s_1 \sim h] \\
&= \limsup_{T \to \infty} \frac{1}{T} \sum_{t=1}^T \field{E}\Big[\field{E}[C(s_t, \pi(h_t)) | {\mathcal{F}}_t, s_1 \sim h] \big| s_1 \sim h\Big] \\
&=\limsup_{T \to \infty} \frac{1}{T} \sum_{t=1}^T \field{E}[c(h_t, \pi(h_t)) | s_1 \sim h] \\
&\geq\limsup_{T \to \infty} \frac{1}{T} \sum_{t=1}^T \field{E}[J(\theta) + v(h_t, \theta) - v(h_{t+1}, \theta) | s_1 \sim h] \\
&= J(\theta),
\end{align*}
with equality attained by $\pi^*$ completing the proof.
\end{proof}
\section{Finite-Parameter Case ($|\Theta| < \infty$)}
\label{sec: finite}
In this section, we consider $\Theta \subseteq \Theta_H$ such that $|\Theta| < \infty$. When $\Theta$ is finite, the posterior distribution concentrates on the true parameter exponentially fast if the transition kernels are separated enough (see Lemma~\ref{lem: 1-ft}). This allows us to achieve a regret bound of $\ensuremath{\mathcal{O}}(H\log T)$. Let $o_{1:t}, a_{1:t}$ be shorthand for the history of observations $o_1, \cdots, o_t$ and the history of actions $a_1, \cdots, a_t$, respectively. Let $\nu_{\theta}^{o_{1:t}, a_{1:t}}(o)$ be the probability of observing $o$ at time $t+1$ if the action history is $a_{1:t}$, the observation history is $o_{1:t}$, and the underlying transition kernel is $\theta$, i.e.,
\begin{align*}
\nu_{\theta}^{o_{1:t}, a_{1:t}}(o) := \prime(o_{t+1}=o | o_{1:t}, a_{1:t}, \theta_*=\theta).
\end{align*}
The distance between $\nu_{\theta}^{o_{1:t}, a_{1:t}}$ and $\nu_{\gamma}^{o_{1:t}, a_{1:t}}$ is defined by Kullback Leibler (KL-) divergence as follows. For a fixed state-action pair $(s, a)$ and any $\theta, \gamma \in \Theta$, denote by ${\mathcal{K}}(\nu_{\theta}^{o_{1:t}, a_{1:t}} \| \nu_{\gamma}^{o_{1:t}, a_{1:t}})$, the Kullback Leibler (KL-) divergence between the probability distributions $\nu_{\theta}^{o_{1:t}, a_{1:t}}$ and $\nu_{\gamma}^{o_{1:t}, a_{1:t}}$ is given by
\begin{align*}
{\mathcal{K}}(\nu_{\theta}^{o_{1:t}, a_{1:t}} \| \nu_{\gamma}^{o_{1:t}, a_{1:t}}) := \sum_{o} \nu_{\theta}^{o_{1:t}, a_{1:t}}(o) \log\frac{\nu_{\theta}^{o_{1:t}, a_{1:t}}(o)}{\nu_{\gamma}^{o_{1:t}, a_{1:t}}(o)}.
\end{align*}
It can be shown that ${\mathcal{K}}(\nu_{\theta}^{o_{1:t}, a_{1:t}} \| \nu_{\gamma}^{o_{1:t}, a_{1:t}}) \geq 0$ and that equality holds if and only if $\nu_{\theta}^{o_{1:t}, a_{1:t}} = \nu_{\gamma}^{o_{1:t}, a_{1:t}}$. Thus, KL-divergence can be thought of as a measure of divergence of $\nu_{\gamma}^{o_{1:t}, a_{1:t}}$ from $\nu_{\theta}^{o_{1:t}, a_{1:t}}$. In this section, we need to assume that the transition kernels in $\Theta$ are distant enough in the following sense.
\begin{assumption}
\label{ass: kl divergence}
For any time step $t$, any history of observations $o_{1:t}$ and actions $a_{1:t}$, and any two transition kernels $\theta, \gamma \in \Theta$, there exists a positive constant $\epsilon > 0$ such that ${\mathcal{K}}(\nu_{\theta}^{o_{1:t}, a_{1:t}} \| \nu_{\gamma}^{o_{1:t}, a_{1:t}}) \geq \epsilon$.
\end{assumption}
This assumption is similar to the one used in \cite{kim2017thompson}.
\begin{theorem}
\label{thm: finite parameter regret}
Suppose Assumptions~\ref{ass: bellman equation} and \ref{ass: kl divergence} hold. Then, the regret bound of Algorithm~\ref{alg: posterior sampling} with $\texttt{SCHED}(t_k, T_{k-1}) = 2t_k$ and $\tilde m_t(s, a) = t$ for all state-action pairs $(s, a)$ is bounded as
\begin{align*}
R_T \leq H\log T + \frac{4(H+1)}{(e^{-\beta}-1)^2},
\end{align*}
where $\beta > 0$ is a universal constant defined in Lemma~\ref{lem: 1-ft}.
\end{theorem}
Observe that with $\texttt{SCHED}(t_k, T_{k-1}) = 2t_k$ and $\tilde m_t(s, a) = t$, the two stopping criteria in Algorithm~\ref{alg: posterior sampling} coincide and ensure that $T_k = 2T_{k-1}$ with $T_0 = 1$. In other words, the length of episodes grows exponentially as $T_k = 2^k$.
\subsection{Proof of Theorem~\ref{thm: finite parameter regret}}
In this section, proof of Theorem~\ref{thm: finite parameter regret} is provided. A key factor in achieving $\ensuremath{\mathcal{O}}(H\log T)$ regret bound in the case of finite parameters is that the posterior distribution $f_t(\cdot)$ concentrates on the true $\theta_*$ exponentially fast.
\begin{lemma}
\label{lem: 1-ft}
Suppose Assumption~\ref{ass: kl divergence} holds. Then, there exist constants $\alpha > 1$ and $\beta > 0$ such that
\begin{align*}
\field{E}[1 - f_t(\theta_*) | \theta_*] \leq \alpha \exp(-\beta t).
\end{align*}
\end{lemma}
Equipped with this lemma, we are now ready to prove Theorem \ref{thm: finite parameter regret}.
\begin{proof}
Note that the regret $R_T$ can be decomposed as $R_T = H\field{E_{\theta_*}}[K_T] + R_1 + R_2 + R_3$, where
\begin{align*}
R_1 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}T_k\Big[J(\theta_k) - J(\theta_*)\Big]}, \\
R_2 &:= H\field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sbr{\sum_{s'}\abr{\theta_*(s'|s_t, a_t) - \theta_k(s'|s_t, a_t)} + \sum_s \abr{h_t(s;\theta_*) - h_t(s;\theta_k)}}}, \\
R_3 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[c(h_t(\cdot;\theta_*), a_t) - c(h_t(\cdot;\theta_k), a_t)\Big]}.
\end{align*}
Note that the start time and length of episodes in Algorithm~\ref{alg: posterior sampling} are deterministic with the choice of $\texttt{SCHED}$ and $\tilde m_t$ in the statement of the theorem, i.e., $t_k$, $T_k$ and hence $K_T$ are deterministic. Note that if $\theta_k = \theta_*$, then $R_1 = R_2 = R_3 = 0$. Moreover, we have that $J(\theta_k) - J(\theta_*) \leq 1$, $\sum_{s'}\abr{\theta_*(s'|s_t, a_t) - \theta_k(s'|s_t, a_t)} \leq 2$, $\sum_s \abr{h_t(s;\theta_*) - h_t(s;\theta_k)} \leq 2$, and $c(h_t(\cdot;\theta_*), a_t) - c(h_t(\cdot;\theta_k), a_t) \leq 1$. Therefore,
\begin{align*}
R_1 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}T_k\mathbbm{1}(\theta_k \neq \theta_*)} = \sum_{k=1}^{K_T}T_k \field{P_{\theta_*}}(\theta_k \neq \theta_*), \\
R_2 &:= 4H\field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\mathbbm{1}(\theta_k \neq \theta_*)} = 4H\sum_{k=1}^{K_T}T_k \field{P_{\theta_*}}(\theta_k \neq \theta_*), \\
R_3 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\mathbbm{1}(\theta_k \neq \theta_*)} = \sum_{k=1}^{K_T}T_k \field{P_{\theta_*}}(\theta_k \neq \theta_*).
\end{align*}
Note that $\field{P_{\theta_*}}(\theta_k \neq \theta_*) = \field{E_{\theta_*}}[1 - f_{t_k}(\theta_*)] \leq \alpha \exp(-\beta t_k)$ by Lemma~\ref{lem: 1-ft}. Combining all these bounds, we can write
\begin{align*}
R_T \leq HK_T + (4H+2)\alpha \sum_{k=1}^{K_T}T_k \exp(-\beta t_k).
\end{align*}
With the episode schedule provided in the statement of the theorem, it is easy to check that $K_T = O(\log T)$.
Let $n = 2^{K_T}$ and write
\begin{align*}
\sum_{k=1}^{K_T}T_k \exp(-\beta t_k) = \sum_{k=1}^{K_T}2^k e^{-\beta (2^k-1)} \leq \sum_{j=2}^{n}je^{-\beta (j-1)} = \frac{d}{dx} \frac{x^{n+1}-1}{x-1} \Big|_{x = e^{-\beta}}-1.
\end{align*}
The last equality is by geometric series. Simplifying the derivative yields
\begin{align*}
\frac{d}{dx} \frac{x^{n+1}-1}{x-1} \Big|_{x = e^{-\beta}} &= \frac{nx^{n+1} - (n+1)x^n + 1}{(x-1)^2}\Big|_{x = e^{-\beta}} \leq \frac{nx^{n} - (n+1)x^n + 1}{(x-1)^2}\Big|_{x = e^{-\beta}} \\
&= \frac{-x^n + 1}{(x-1)^2}\Big|_{x = e^{-\beta}} \leq \frac{2}{(e^{-\beta}-1)^2}.
\end{align*}
Substituting these values implies $R_T \leq H\log T + \frac{4(H+1)}{(e^{-\beta}-1)^2}$.
\end{proof}
\section{General Case ($|\Theta| = \infty$)}
\label{sec: general}
We now consider the general case, where the parameter set is infinite, and in particular, $\Theta = \Theta_H$, an uncountable set.
We make the following two technical assumptions on the belief and the transition kernel.
\begin{assumption}
\label{ass: concentrating belief}
Denote by $k(t)$ the episode at time $t$. The true conditional belief $h_t(\cdot;\theta_*)$ and the approximate conditional belief $h_t(\cdot;\theta_{k(t)})$ satisfy
\begin{equation}
\label{eq: ass concentrating belief}
\field{E}\Big[\sum_{s}\big|h_t(s;\theta_*) - h_t(s;\theta_{k(t)})\big|\Big] \leq \frac{K_1(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)}{\sqrt{t_{k(t)}}},
\end{equation}
with probability at least $1 - \delta$, for any $\delta \in (0, 1)$. Here $K_1(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ is a constant that is polynomial in its input parameters and $\iota$ hides the logarithmic dependency on $|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, T, \delta$.
\end{assumption}
Assumption~\ref{ass: concentrating belief} states that the gap between conditional posterior function for the sampled POMDP $\theta_k$ and the true POMDP $\theta_*$ decreases with episodes as better approximation of the true POMDP is available.
There has been recent work on computation of approximate information states as required in Assumption \ref{ass: concentrating belief} \citep{subramanian2020approximate}.
\begin{assumption}
\label{ass: concentration}
There exists an ${\mathcal{F}}_{t}$-measurable estimator $\hat \theta_t : {\mathcal{S}} \times {\mathcal{A}} \to \Delta_{\mathcal{S}}$ such that
\begin{align}
\label{eq: ass concentration}
\sum_{s'}|\theta_*(s'|s, a) - \hat \theta_t(s' | s, a)| \leq \frac{K_2(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)}{\sqrt{\max\{1, {\tilde m}_{t}(s, a)\}}}
\end{align}
with probability at least $1 - \delta$, for any $\delta \in (0, 1)$, uniformly for all $t=1, 2, 3, \cdots, T$, where $K_2(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ is a constant that is polynomial in its input parameters and $\iota$ hides the logarithmic dependency on $|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, T, \delta$.
\end{assumption}
There has been extensive work on estimation of transition dynamics of MDPs, e.g., \citep{grunewalder2012modelling}. Two examples where Assumptions~\ref{ass: concentrating belief} and \ref{ass: concentration} hold are:
\begin{itemize}
\item \textbf{Perfect observation.} In the case of perfect observation, where $h_t(s;\theta) = \mathbbm{1}(s_t=s)$, Assumption~\ref{ass: concentrating belief} is clearly satisfied. Moreover, with perfect observation, one can choose $\tilde{m}_{t}(s, a) = n_{t}(s, a)$ and select $\hat \theta_k(s'|s, a) = \frac{n_{t}(s, a, s')}{n_{t}(s, a)}$ to satisfy Assumption~\ref{ass: concentration} \citep{jaksch2010near,ouyang2017learning}. Here $n_{t}(s, a, s')$ denotes the number of visits to $s, a$ such that the next state is $s'$ before time $t$.
\item \textbf{Finite-parameter case.} In the finite-parameter case with the choice of $\tilde m_t(s, a) = t$ for all state-action pairs $(s, a)$ and $\texttt{SCHED}(t_k, T_{k-1}) = t_k + T_{k-1}$ or $\texttt{SCHED}(t_k, T_{k=1}) = 2t_k$, both of the assumptions are satisfied (see Lemma~\ref{lem: assumptions satisfied for finite-parameter} for details). Note that in this case a more refined analysis is performed in Section~\ref{sec: finite} to achieve $\ensuremath{\mathcal{O}}(H\log T)$ regret bound.
\end{itemize}
Now, we state the main result of this section.
\begin{theorem}
\label{thm: regret bound}
Under Assumptions~\ref{ass: bellman equation}, \ref{ass: concentrating belief} and \ref{ass: concentration}, running \texttt{PSRL-POMDP} algorithm with $\texttt{SCHED}(t_k, T_{k-1}) = t_k + T_{k-1}$ yields $\field{E}[R_T] \leq \ensuremath{\tilde{\mathcal{O}}}(HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3})$, where $K_2 := K_2(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ in Assumption~\ref{ass: concentration}.
\end{theorem}
The exact constants are known (see proof and Appendix~\ref{app: full upper bound}) though we have hidden the dependence above.
\subsection{Proof Sketch of Theorem~\ref{thm: regret bound}}
We provide the proof sketch of Theorem~\ref{thm: regret bound} here.
A key property of posterior sampling is that conditioned on the information at time $t$, the sampled $\theta_t$ and the true $\theta_*$ have the same distribution \citep{osband2013more,russo2014learning}. Since the episode start time $t_k$ is a stopping time with respect to the filtration $({\mathcal{F}}_t)_{t \geq 1}$, we use a stopping time version of this property:
\begin{lemma}[Lemma 2 in \cite{ouyang2017learning}]
\label{lem: property of posterior sampling}
For any measurable function $g$ and any ${\mathcal{F}}_{t_k}$-measurable random variable $X$, we have $\field{E}[g(\theta_k, X)] = \field{E}[g(\theta_*, X)]$.
\end{lemma}
Introducing the pseudo count $\tilde m_t(s, a)$ in the algorithm requires a novel analysis to achieve a low regret bound. The following key lemma states that the pseudo count $\tilde m_{t}$ cannot be too smaller than the true count $n_t$.
\begin{lemma}
\label{lem: pseudo count vs true count general case}
Fix a state-action pair $(s, a) \in {\mathcal{S}} \times {\mathcal{A}}$. For any pseudo count $\tilde m_t$ and any $\alpha \in [0, 1]$,
\begin{equation}
\prime\big(\tilde m_{t}(s, a) < \alpha n_{t}(s, a)\big) \leq \alpha.
\end{equation}
\end{lemma}
\begin{proof}
We show that $\prime\big(\tilde n_{t}(s, a) < \alpha n_{t}(s, a)\big) \leq \alpha$. Since by definition $\tilde m_t(s, a) \geq \tilde n_t(s, a)$, the claim of the lemma follows. For any $\alpha \in [0, 1]$,
\begin{equation}
\tilde n_{t}(s, a) \mathbbm{1}\big(\alpha n_{t}(s, a) > \tilde n_{t}(s, a)\big) \leq \alpha n_{t}(s, a).
\end{equation}
By taking conditional expectation with respect to ${\mathcal{F}}_{{(t-1)}+}$ from both sides and the fact that $\field{E}[n_{t}(s, a)|{\mathcal{F}}_{{(t-1)}+}] = \tilde n_{t}(s, a)$, we have
\begin{equation}
\label{eq: ntilde indicator smaller than ntilde}
\tilde n_{t}(s, a) \field{E}\Big[\mathbbm{1}\big(\alpha n_{t}(s, a) > \tilde n_{t}(s, a)\big) \Big| {\mathcal{F}}_{{(t-1)}+} \Big] \leq \alpha \tilde n_{t}(s, a).
\end{equation}
We claim that
\begin{equation}
\label{eq: indicator less than alpha}
\field{E}\Big[\mathbbm{1}\big(\alpha n_{t}(s, a) > \tilde n_{t}(s, a)\big) \Big| {\mathcal{F}}_{{(t-1)}+}\Big] \leq \alpha,~~\text{a.s.}
\end{equation}
If this claim is true, taking another expectation from both sides completes the proof.
To prove the claim, let $\Omega_0, \Omega_+$ be the subsets of the sample space where $\tilde n_t(s, a) = 0$ and $\tilde n_t(s, a) > 0$, respectively. We consider these two cases separately: (a) on $\Omega_+$ one can divide both sides of \eqref{eq: ntilde indicator smaller than ntilde} by $\tilde n_t(s, a)$ and reach \eqref{eq: indicator less than alpha}; (b) note that by definition $\tilde n_t(s, a) = 0$ on $\Omega_0$. Thus, $ n_t(s, a)\mathbbm{1}(\Omega_0) = 0$ almost surely (this is because $\field{E}[n_t(s, a)\mathbbm{1}(\Omega_0)] = \field{E}[\field{E}[n_t(s, a)\mathbbm{1}(\Omega_0)|{\mathcal{F}}_{(t-1)+}]] = \field{E}[\tilde n_t(s, a)\mathbbm{1}(\Omega_0)] = 0$). Therefore,
\begin{align*}
\mathbbm{1}(\Omega_0)\mathbbm{1}\big(\alpha n_{t}(s, a) > \tilde n_{t}(s, a)\big) = 0,\quad\text{a.s.,}
\end{align*}
which implies
\begin{equation*}
\mathbbm{1}(\Omega_0)\field{E}\Big[\mathbbm{1}\big(\alpha n_{t}(s, a) > \tilde n_{t}(s, a)\big) \Big| {\mathcal{F}}_{{(t-1)}+}\Big] = 0,\quad\text{a.s.,}
\end{equation*}
which means on $\Omega_0$, the left hand side of \eqref{eq: indicator less than alpha} is indeed zero, almost surely, proving the claim.
\end{proof}
The parameter $\alpha$ will be tuned later to balance two terms and achieve $\ensuremath{\tilde{\mathcal{O}}}(T^{2/3})$ regret bound (see Lemma~\ref{lem: r2bar}). We are now ready to provide the proof sketch of Theorem~\ref{thm: regret bound}.
By Lemma~\ref{lem: decomposition}, $R_T$ can be decomposed as $R_T = H\field{E_{\theta_*}}[K_T] + R_1 + R_2 + R_3$, where
\begin{align*}
R_1 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}T_k\Big[J(\theta_k) - J(\theta_*)\Big]}, \\
R_2 &:= H\field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sbr{\sum_{s'}\abr{\theta_*(s'|s_t, a_t) - \theta_k(s'|s_t, a_t)} + \sum_s \abr{h_t(s;\theta_*) - h_t(s;\theta_k)}}}, \\
R_3 &:= \field{E_{\theta_*}}\sbr{\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\Big[c(h_t(\cdot;\theta_*), a_t) - c(h_t(\cdot;\theta_k), a_t)\Big]}.
\end{align*}
It follows from the first stopping criterion that $T_k \leq T_{k-1} + 1$. Using this along with the property of posterior sampling (Lemma~\ref{lem: property of posterior sampling}) proves that $\field{E}[R_1] \leq \field{E}[K_T]$ (see Lemma~\ref{lem: j minus jstar} for details). $\field{E}[R_3]$ is bounded by $K_1\field{E}\big[\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}}\big] + 1$ where $K_1:= K_1(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ is the constant in Assumption~\ref{ass: concentrating belief} (see Lemma~\ref{lem: r3}). To bound $\field{E}[R_2]$, we use Assumption~\ref{ass: concentrating belief} and follow the proof steps of Lemma~\ref{lem: r3} to conclude that
\begin{equation*}
\field{E}[R_2] \leq \bar R_2 + HK_1 \field{E}\Big[\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}}\Big] + 1,
\end{equation*}
where
\begin{equation*}
\bar R_2 := H\field{E}\Big[\sum_{k=1}^{K_T}\sum_{t=t_k}^{t_{k+1}-1}\sum_{s'}\Big|\theta_*(s' | s_t, a_t) - \theta_k(s'|s_t, a_t)\Big|\Big].
\end{equation*}
$\bar R_2$ is the dominating term in the final $\ensuremath{\tilde{\mathcal{O}}}(T^{2/3})$ regret bound and can be bounded by $H + 12HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3}$ where $K_2 := K_2(|{\mathcal{S}}|, |{\mathcal{A}}|, |O|, \iota)$ is the constant in Assumption~\ref{ass: concentration}. The detailed proof can be found in Lemma \ref{lem: r2bar}. However, we sketch the main steps of the proof here. By Assumption~\ref{ass: concentration}, one can show that
\begin{equation*}
\bar R_2 \leq \ensuremath{\tilde{\mathcal{O}}} \Big(\field{E}\Big[\sum_{t=1}^{T}\frac{HK_2}{\sqrt{\max\{1, {\tilde m}_{t}(s_t, a_t)\}}}\Big]\Big).
\end{equation*}
Now, let $E_2$ be the event that $\tilde m_t(s, a) \geq \alpha n_t(s, a)$ for all $s, a$. Note that by Lemma~\ref{lem: pseudo count vs true count general case} and union bound, $\prime(E_2^c) \leq |{\mathcal{S}}||{\mathcal{A}}|\alpha$. Thus,
\begin{align*}
&\bar R_2 \leq \ensuremath{\tilde{\mathcal{O}}} \Big(\field{E}\Big[\sum_{t=1}^{T}\frac{HK_2}{\sqrt{\max\{1, {\tilde m}_{t}(s_t, a_t)\}}}\big(\mathbbm{1}(E_2) + \mathbbm{1}(E_2^c)\big)\Big]\Big) \\
&\leq \ensuremath{\tilde{\mathcal{O}}} \Big( H\field{E}\Big[\sum_{t=1}^{T}\frac{K_2}{\sqrt{\alpha\max\{1, {n}_{t}(s_t, a_t)\}}}\Big] + HK_2 |{\mathcal{S}}||{\mathcal{A}}|T\alpha\Big)
\end{align*}
Algebraic manipulation of the inner summation yields $\bar R_2 \leq \ensuremath{\tilde{\mathcal{O}}} \Big( HK_2\sqrt \frac{|{\mathcal{S}}||{\mathcal{A}}|T}{\alpha} + HK_2 |{\mathcal{S}}||{\mathcal{A}}|T\alpha\Big)$.
Optimizing over $\alpha$ implies $\bar R_2 = \ensuremath{\tilde{\mathcal{O}}}(HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3})$.
Substituting upper bounds for $\field{E}[R_1], \field{E}[R_2]$ and $\field{E}[R_3]$, we get
\begin{align*}
\field{E}[R_T] &= H\field{E}[K_T] + \field{E}[R_1] + \field{E}[R_2] + \field{E}[R_3] \\
&\leq (1 + H)\field{E}[K_T] + 12HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3} + (H+1)K_1 \field{E}\Big[\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}}\Big] + 2 + H.
\end{align*}
From Lemma~\ref{lem: number of episodes}, we know that $\field{E}[K_T] = \ensuremath{\tilde{\mathcal{O}}}(\sqrt{|{\mathcal{S}}||{\mathcal{A}}|T})$ and $\sum_{k=1}^{K_T}\frac{T_k}{\sqrt{t_k}} = \ensuremath{\tilde{\mathcal{O}}}(|{\mathcal{S}}||{\mathcal{A}}|\sqrt{T})$. Therefore, $\field{E}[R_T] \leq \ensuremath{\tilde{\mathcal{O}}}(HK_2(|{\mathcal{S}}||{\mathcal{A}}|T)^{2/3}).$
\section{Introduction}
\label{sec: introduction}
Reinforcement learning (RL) considers the sequential decision-making problem of an agent in an unknown environment with the goal of minimizing the total cost. The agent faces a fundamental \textit{exploration-exploitation} trade-off: should it \textit{exploit} the available information to minimize the cost or should it \textit{explore} the environment to gather more information for future decisions? Maintaining a proper balance between exploration and exploitation is a fundamental challenge in RL and is measured with the notion of cumulative regret: the difference between the cumulative cost of the learning algorithm and that of the best policy.
The problem of balancing exploration and exploitation in RL has been successfully addressed for MDPs and algorithms with near optimal regret bounds known \citep{bartlett2009regal,jaksch2010near,ouyang2017learning,azar2017minimax,fruit2018efficient,jin2018q,abbasi2019exploration,zhang2019regret,zanette2019tighter,hao2020provably,wei2020model,wei2021learning}.
MDPs assume that the state is perfectly observable by the agent and the only uncertainty is about the underlying dynamics of the environment. However, in many real-world scenarios such as robotics, healthcare and finance, the state is not fully observed by the agent, and only a partial observation is available. These scenarios are modeled by Partially Observable Markov Decision Processes (POMDPs). In addition to the uncertainty in the environment dynamics, the agent has to deal with the uncertainty about the underlying state. It is well known \citep{kumar2015stochastic} that introducing an information or belief state (a posterior distribution over the states given the history of observations and actions) allows the POMDP to be recast as an MDP over the belief state space. The resulting algorithm requires a posterior update of the belief state which needs the transition and observation model to be fully known. This presents a significant difficulty when the model parameters are unknown. Thus, managing the exploration-exploitation trade-off for POMDPs is a significant challenge and to the best of our knowledge, no online RL algorithm with sub-linear regret is known.
In this paper, we consider infinite-horizon average-cost POMDPs with finite states, actions and observations. The underlying state transition dynamics is unknown, though we assume the observation kernel to be known. We propose a Posterior Sampling Reinforcement Learning algorithm (\texttt{PSRL-POMDP}) and prove that it achieves a Bayesian expected regret bound of $\ensuremath{\mathcal{O}}(\log T)$ in the finite (transition kernel) parameter set case where $T$ is the time horizon. We then show that in the general (continuous parameter set) case, it achieves $\ensuremath{\tilde{\mathcal{O}}}(T^{2/3})$ under some technical assumptions.
The \texttt{PSRL-POMDP} algorithm is a natural extension of the \texttt{TSDE} algorithm for MDPs \citep{ouyang2017learning} with two main differences. First, in addition to the posterior distribution on the environment dynamics, the algorithm maintains a posterior distribution on the underlying state. Second, since the state is not fully observable, the agent cannot keep track of the number of visits to state-action pairs, a quantity that is crucial in the design of algorithms for tabular MDPs. Instead, we introduce a notion of \textit{pseudo count} and carefully handle its relation with the true counts to obtain sub-linear regret. To the best of our knowledge, \texttt{PSRL-POMDP} is the first online RL algorithm for POMDPs with sub-linear regret.
\section{Conclusions}
In this paper, we have presented one of the first online reinforcement learning algorithms for POMDPs. Solving POMDPs is a hard problem. Designing an efficient learning algorithm that achieves sublinear regret is even harder. We show that the proposed \texttt{PSRL-POMDP} algorithm achieves a Bayesian regret bound of $\ensuremath{\mathcal{O}}(\log T)$ when the parameter is finite. When the parameter set may be uncountable, we showed a $\ensuremath{\tilde{\mathcal{O}}} (T^{2/3})$ regret bound under two technical assumptions on the belief state approximation and transition kernel estimation.
There has been recent work that does approximate belief state computation, as well as estimates transition dynamics of continuous MDPs, and in future work, we will try to incorporate such estimators. We also assume that the observation kernel is known. Note that without it, it is very challenging to design online learning algorithms for POMDPs. Posterior sampling-based algorithms in general are known to have superior numerical performance as compared to OFU-based algorithms for bandits and MDPs. In future work, we will also do an experimental investigation of the proposed algorithm. An impediment is that available POMDP solvers mostly provide approximate solutions which would lead to linear regret. In the future, we will also try to improve the regret for the general case to $\ensuremath{\tilde{\mathcal{O}}}(\sqrt{T})$.
\bibliographystyle{plainnat}
\section{Preliminaries}
\label{sec: preliminaries}
An infinite-horizon average-cost Partially Observable Markov Decision Process (POMDP) can be specified by $({\mathcal{S}}, {\mathcal{A}}, \theta, C, O, \eta)$ where ${\mathcal{S}}$ is the state space, ${\mathcal{A}}$ is the action space, $C: {\mathcal{S}} \times {\mathcal{A}} \to [0, 1]$ is the cost function, and $O$ is the set of observations. Here $\eta: {\mathcal{S}} \to \Delta_O$ is the observation kernel, and $\theta: {\mathcal{S}} \times {\mathcal{A}} \to \Delta_{\mathcal{S}}$ is the transition kernel such that $\eta(o | s) = \prime(o_t=o | s_t=s)$ and $\theta(s'|s, a) = \prime(s_{t+1}=s'|s_t=s, a_t=a)$ where $o_t \in O$, $s_t \in {\mathcal{S}}$ and $a_t \in {\mathcal{A}}$ are the observation, state and action at time $t=1, 2, 3, \cdots$. Here, for a finite set ${\mathcal{X}}$, $\Delta_{\mathcal{X}}$ is the set of all probability distributions on ${\mathcal{X}}$. We assume that the state space, the action space and the observations are finite with size $|{\mathcal{S}}|, |{\mathcal{A}}|, |O|$, respectively.
Let ${\mathcal{F}}_t$ be the information available at time $t$ (prior to action $a_t$), i.e., the sigma algebra generated by the history of actions and observations $a_1, o_1, \cdots, a_{t-1}, o_{t-1}, o_t$ and let ${\mathcal{F}}_{t+}$ be the information after choosing action $a_t$. Unlike MDPs, the state is not observable by the agent and the optimal policy cannot be a function of the state. Instead, the agent maintains a belief $h_t(\cdot;\theta) \in \Delta_{\mathcal{S}}$ given by $h_t(s ; \theta) := \prime(s_t=s | {\mathcal{F}}_t;\theta)$, as a sufficient statistic for the history of observations and actions. Here we use the notation $h_t(\cdot ; \theta)$ to explicitly show the dependency of the belief on $\theta$. After taking action $a_t$ and observing $o_{t+1}$, the belief $h_t$ can be updated as
\begin{equation}
\label{eq: update rule of ht}
h_{t+1}(s';\theta) =\frac{\sum_{s}\eta(o_{t+1}|s') \theta(s'|s, a_t)h_t(s;\theta)}{\sum_{s'}\sum_{s}\eta(o_{t+1}|s') \theta(s'|s, a_t)h_t(s;\theta)}.
\end{equation}
This update rule is compactly denoted by $h_{t+1}(\cdot ; \theta) = \tau (h_t(\cdot;\theta), a_t, o_{t+1}; \theta)$, with the initial condition
\begin{equation*}
h_1(s;\theta) = \frac{\eta(o_1|s)h(s;\theta)}{\sum_s\eta(o_1|s)h(s;\theta)},
\end{equation*}
where $h(\cdot;\theta)$ is the distribution of the initial state $s_1$ (denoted by $s_1 \sim h$). A deterministic stationary policy $\pi : \Delta_{\mathcal{S}} \to {\mathcal{A}}$ maps a belief to an action. The long-term average cost of a policy $\pi$ can be defined as
\begin{equation}
J_\pi(h; \theta) := \limsup_{T \to \infty} \frac{1}{T} \sum_{t=1}^T \field{E}\Big[C\Big(s_t, \pi\big(h_t(\cdot;\theta)\big)\Big) \Big].
\end{equation}
Let $J(h, \theta) := \inf_\pi J_\pi(h, \theta)$ be the optimal long-term average cost that in general may depend on the initial state distribution $h$, though we will assume it is independent of the initial distribution $h$ (and thus denoted by $J(\theta)$), and the following Bellman equation holds:
\begin{assumption}[Bellman optimality equation]
\label{ass: bellman equation}
There exist $J(\theta) \in \field{R}$ and a bounded function $v(\cdot; \theta):\Delta_{\mathcal{S}} \to \field{R}$ such that for all $b \in \Delta_{\mathcal{S}}$
\begin{equation}
\label{eq: bellman equation}
J(\theta) + v(b; \theta) = \min_{a \in {\mathcal{A}}} \{c(b, a) + \sum_{o \in O}P(o | b, a; \theta)v(b'; \theta)\},
\end{equation}
\end{assumption}
where $v$ is called the relative value function, $b' = \tau(b, a, o; \theta)$ is the updated belief, $c(b, a) := \sum_s C(s, a)b(s)$ is the expected cost, and $P(o|b, a; \theta)$ is the probability of observing $o$ in the next step, conditioned on the current belief $b$ and action $a$, i.e.,
\begin{equation}
P(o | b, a; \theta) = \sum_{s' \in {\mathcal{S}}}\sum_{s \in {\mathcal{S}}} \eta(o | s') \theta(s' | s, a) b(s).
\end{equation}
Various conditions are known under which Assumption \ref{ass: bellman equation} holds, e.g., when the MDP is weakly communicating \citep{bertsekas2017dynamic}.
Note that if Assumption~\ref{ass: bellman equation} holds, the policy $\pi^*$ that minimizes the right hand side of \eqref{eq: bellman equation} is the optimal policy. More precisely,
\begin{lemma}
\label{lem: optimal policy}
Suppose Assumption~\ref{ass: bellman equation} holds. Then, the policy $\pi^*(\cdot, \theta):\Delta_{\mathcal{S}} \to {\mathcal{A}}$ given by
\begin{equation}
\label{eq: the optimal policy}
\pi^*(b; \theta) := \argmin_{a \in {\mathcal{A}}} \{c(b, a) + \sum_{o \in O}P(o | b, a; \theta)v(b'; \theta)\}
\end{equation}
is the optimal policy with $J_{\pi^*}(h; \theta) = J(\theta)$ for all $h \in \Delta_{\mathcal{S}}$.
\end{lemma}
Note that if $v$ satisfies the Bellman equation, so does $v$ plus any constant. Therefore, without loss of generality, and since $v$ is bounded, we can assume that $\inf_{b \in \Delta_{\mathcal{S}}}v(b; \theta) = 0$ and define the span of a POMDP as $\spn(\theta) := \sup_{b \in \Delta_{\mathcal{S}}}v(b; \theta)$. Let $\Theta_H$ be the class of POMDPs that satisfy Assumption~\ref{ass: bellman equation} and have $\spn(\theta) \leq H$ for all $\theta \in \Theta_H$. In Section~\ref{sec: finite}, we consider a finite subset $\Theta \subseteq \Theta_H$ of POMDPs. In Section~\ref{sec: general}, the general class $\Theta = \Theta_H$ is considered.
\paragraph{The learning protocol.} We consider the problem of an agent interacting with an unknown randomly generated POMDP $\theta_*$, where $\theta_* \in \Theta$ is randomly generated according to the probability distribution $f(\cdot)$.\footnote{In Section~\ref{sec: finite}, $f(\cdot)$ should be viewed as a probability mass function.} After the initial generation of $\theta_*$, it remains fixed, but unknown to the agent. The agent interacts with the POMDP $\theta_*$ in $T$ steps. Initially, the agent starts from state $s_1$ that is randomly generated according to the conditional probability mass function $h(\cdot;\theta_*)$. At time $t=1, 2, 3, \cdots, T$, the agent observes $o_t \sim \eta(\cdot|s_t)$, takes action $a_t$ and suffers cost of $C(s_t, a_t)$. The environment, then determines the next state $s_{t+1}$ which is randomly drawn from the probability distribution $\theta_*(\cdot | s_t, a_t)$. Note that although the cost function $C$ is assumed to be known, the agent cannot observe the value of $C(s_t, a_t)$ since the state $s_t$ is unknown to the agent. The goal of the agent is to minimize the expected cumulative regret defined as
\begin{equation}
R_T := \field{E_{\theta_*}}\Big[\sum_{t=1}^T\Big[C(s_t, a_t) - J(\theta_*)\Big]\Big],
\end{equation}
where the expectation is with respect to the prior distribution $h(\cdot;\theta_*)$ for $s_1$, the randomness in the state transitions, and the randomness in the algorithm. Here, $\field{E_{\theta_*}}[\cdot]$ is a shorthand for $\field{E}[\cdot|\theta_*]$. In Section~\ref{sec: finite}, a regret bound is provided on $R_T$, however, Section~\ref{sec: general} considers $\field{E}[R_T]$ (also called Bayesian regret) as the performance measure for the learning algorithm. We note that the Bayesian regret is widely considered in the MDP literature \citep{osband2013more,gopalan2015thompson,ouyang2017learning,ouyang2017learningbased}.
\subsection{Related Literature}
\label{sec: related literature}
We review the related literature in two main domains: efficient exploration for MDPs, and learning in POMDPs.
\noindent\textbf{Efficient exploration in MDPs.} To balance the exploration and exploitation, two general techniques are used in the basic tabular MDPs: \textit{optimism in the face of uncertainty (OFU)}, and \textit{posterior sampling}. Under the OFU technique, the agent constructs a confidence set around the system parameters, selects an optimistic parameter associated with the minimum cost from the confidence set, and takes actions with respect to the optimistic parameter. This principle is widely used in the literature to achieve optimal regret bounds
\citep{bartlett2009regal,jaksch2010near,azar2017minimax,fruit2018efficient,jin2018q,zhang2019regret,zanette2019tighter,wei2020model}.
An alternative technique to encourage exploration is posterior sampling \citep{thompson1933likelihood}. In this approach, the agent maintains a posterior distribution over the system parameters, samples a parameter from the posterior distribution, and takes action with respect to the sampled parameter \citep{strens2000bayesian,osband2013more,fonteneau2013optimistic,gopalan2015thompson,ouyang2017learning,jafarnia2021online}. In particular, \citep{ouyang2017learning} proposes \texttt{TSDE}, a posterior sampling-based algorithm for the infinite-horizon average-cost MDPs.
Extending these results to the continuous state MDPs has been recently addressed with general function approximation \citep{osband2014model,dong2020root,ayoub2020model,wang2020reinforcement}, or in the special cases of linear function approximation \citep{abbasi2019politex,abbasi2019exploration,jin2020provably,hao2020provably,wei2021learning,wang2021provably}, and Linear Quadratic Regulators \citep{ouyang2017learningbased,dean2018regret,cohen2019learning,mania2019certainty,simchowitz2020naive,lale2020explore}. In general, POMDPs can be formulated as continuous state MDPs by considering the belief as the state. However, computing the belief requires the knowledge of the model parameters and thus unobserved in the RL setting. Hence, learning algorithms for continuous state MDPs cannot be directly applied to POMDPs.
\noindent\textbf{Learning in POMDPs.} To the best of our knowledge, the only existing work with sub-linear regret in POMDPs is \cite{azizzadenesheli2017experimental}. However, their definition of regret is not with respect to the optimal policy, but with respect to the best memoryless policy (a policy that maps the current observation to an action). With our natural definition of regret, their algorithm suffers linear regret. Other learning algorithms for POMDPs either consider linear dynamics \citep{lale2020logarithmic,tsiamis2020online} or do not consider regret \citep{shani2005model,ross2007bayes,poupart2008model,cai2009learning,liu2011infinite,liu2013online,doshi2013bayesian,katt2018bayesian,azizzadenesheli2018policy} and are not directly comparable to our setting. |
\section{Introduction}
\textcolor{black}{\emph{Introduction.}}---
In recent years periodically driven isolated systems emerge as an exciting field of research, giving justice to the fact that
driven systems exhibit intriguing properties as compared to their
equilibrium counterparts \cite{shirley65,dunlap86,grifoni98}.
The quantum systems are studied extensively in this context theoretically \cite{goldman14,eckardt17atomic,bukov2015universal} as well as experimentally \cite{Wang453,Experiment2013,Experiment2016,fleury2016floquet,Experiment2017}; for example, dynamical localization \cite{kayanuma08,nag14,nag15}, many-body localization \cite{d13many,d14long,ponte15periodically,ponte15,lazarides15fate,zhang16}, quantum phase transitions \cite{eckardt05superfluid,zenesini09}, Floquet topological insulator \cite{oka09photovoltaic,kitagawa11transport,lindner11floquet,rudner13anomalous,vega19,seshadri19,nag19,nag20a,nag20b}, Floquet topological superconductor \cite{ghosh21a,ghosh21b},
Floquet time crystals \cite{else16floquet,khemani16phase,zhang17observation,yao17discrete}, higher harmonic generation \cite{faisal97,nag17,ikeda18,neufeld2019floquet} are
remarkable nonequilibrium phenomena. Consequently the heating happens to be very crucial factor as far as the stability of
the driven systems is concerned \cite{Bilitewski15,Reitter17,Boulier19}.
The consensus so far is that the driven quantum many-body systems heat up to an infinite-temperature state \cite{moessner2017equilibration,luitz17,d14long,Seetharam18} with some exceptions \cite{Prosen98,halder18}.
However, it has been shown that heating can be suppressed for integrable systems due to infinite number of constants of motion, as manifested through the non-equilibrium steady states \cite{russomanno12periodic,nag14,gritsev2017integrable}. On the other hand,
many-body localized systems prevent heating for their effective local integrals of motion in the presence of interaction and disorder \cite{d13many,d14long,ponte15periodically,ponte15,lazarides15fate}. The high frequency driving is another alternative route to prohibit the heating in the long-lived prethermal region, that grows exponentially with frequency, before heating up at the infinite temperature state
~\cite{choudhury14stability,bukov15prethermal,citro15dynamical,Mori15,chandran16interaction,canovi16stroboscopic,mori2016rigorous,lellouch17parametric,weidinger17floquet,abanin2017effective,else17prethermal,zeng17prethermal,Peronaci18,rajak2018stability,Mori18,Howell19,rajak2019characterizations,abanin2017rigorous}.
Interestingly, the quasistationary prethermal state is concomitantly described by an effective
static Hamiltonian, obtained using the Floquet-Magnus expansion, in the high-frequency regime \cite{Mori15,canovi16stroboscopic,mori2016rigorous,weidinger17floquet,abanin2017effective,else17prethermal,zeng17prethermal,Peronaci18}.
Here arises a very relevant question whether the classical systems exhibit such interesting intermediate prethermal plateau. Recently, using generic many-body systems of classical chaos theory \cite{rajak2018stability,rajak2019characterizations} and periodically driven classical spin chains \cite{Howell19,Mori18}, the classical systems are also found to demonstrate the Floquet prethermalization.
Similar to the quantum case, Floquet-Magnus expansion leads to an effective static classical Hamiltonian describing the prethermal phase where heating is exponentially suppressed \cite{Howell19,Mori18,rajak2019characterizations,torre2020statistical,Hodson21}.
The prethermal phase is further characterised by
generalized Gibbs ensemble (GGE) causing
hydrodynamic behavior to emerge in the above phase \cite{takato16,mori2018thermalization,deutsch2018eigenstate}.
The framework of fluctuating hydrodynamics
becomes a convenient tool to investigate the equilibrium transport in classical non-linear systems
\cite{das2014role,Mendl13,Mendl14,spohn2014nonlinear,mendl2015current,kundu16,dhar2019transport,doyon2019generalized,das2020nonlinear}. The integrable (non-integrable) classical systems typically admit ballistic (non-ballistic) transport \cite{spohn2014nonlinear,das14pre,bastianello2018generalized,doyon18,spohn2020ballistic}. The theory of fluctuating hydrodynamics is also employed to understand the transport in non-linear Fermi-Pasta-Ulam-Tsingou like systems \cite{Mendl13,das14pre}.
Given the above background, we would like to investigate the non-equilibrium dynamics of fluctuation
correlation of the kinetic energy
in a model of interacting classical kicked rotors as a probe to the hydrodynamic behavior of the problem.
The motivation behind choosing such model is that
in the limit of large number of particles per site,
a Bose-Hubbard model can be mapped to the above model \cite{rajak2019characterizations}. More importantly, kicked rotor systems can be realized in experiments using Josephson junctions with Bose-Einstein condensates \cite{cataliotti2001josephson}. The time-periodic delta function kicks can be implemented by varying the potential depth and width controlling the intensity of laser light \cite{cheneau2012light,goldman2014periodically}.
The main questions that we pose in this work are as follows:
How does a typical fluctuation
behave in quasi-stationary prethermal states, as well as in the regime where kinetic energy grows in an unbounded
manner \cite{rajak2019characterizations}? Provided the notion of the GGE in the dynamic prethermal regime, does diffusive transport as seen for the case of static Hamiltonian \cite{das2020nonlinear} persist? Moreover, our questions are very pertinent experimentally
where Floquet prethermalization has been realized in optical lattice platforms \cite{messer18,rubio20}.
Given the fact that the time-dynamics of the kinetic energy for the above system can be divided into three different temporal regimes depending on the nature of its growth \cite{rajak2018stability,rajak2019characterizations}, in this work, while numerically investigating the propagation of fluctuation (\ref{eq_correlation}) through the system as a function of time,
we show that these dynamical regimes are characterized by distinct space-time behavior of the kinetic energy fluctuation correlation (see Fig.~\ref{fig:schematic}).
Following the initial transient, the system enters into the prethermal regime, characterized by almost constant kinetic energy with exponentially suppressed heating,
where the fluctuation spreads over space diffusively as a function of time (see Fig.~\ref{fig:PTcorr}). The spatio-temporal correlation becomes
Gaussian whose variance $W$ increases linearly with time.
Once the system starts absorbing energy from the drive, the
fluctuation becomes exponentially localized around the site of disturbance and
temporally frozen referring to the constant nature of $W$ with time (see Fig.~\ref{fig:IL}). We refer
this intermediate window as a crossover region that connects the spatially and temporally quasi-localized behavior of correlations at long time (see Fig.~\ref{fig:k1p2}) with the
prethermal phase. In that quasi-localized phase, $W$ decays to vanishingly small values while
the kinetic energy of
the system grows linearly with time. Therefore, the kinetic energy localization (diffusion) corresponds to the diffusion (localization) of fluctuation correlation. We qualitatively understand the underlying energy absorption mechanism in prethermal phase
based on the hydrodynamic description.
Our study considering fluctuation correlation of kinetic energy as a two-point observable, reveals new insight to the dynamic phases that are not accounted by the one-point observables. Moreover,
dynamic features such as quasi-localization and localization of spatio-temporal correlations do not have any static analogue.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Fig1.pdf}
\caption{Schematic shows the
evolution of relative width of spatio-temporal correlations (\ref{eq_correlation}), indicated by the \textcolor{black}{log of} variance $W$ (left axis), and the \textcolor{black}{log of} average
kinetic energy per rotor $E_{\rm kin}(n)=(1/N)\sum^N_{j=1} \langle p^2_j(n)/2\rangle$ (right axis, brown solid line),
as a function of number of kicks $n$ for two different values of driving parameters $K_1$ and $K_2$. The kinetic energy is quasi-conserved ($E_{\rm kin}(n) \sim n^0$, brown solid line) and the spread of spatio-temporal correlations is self-similar diffusive
($W \sim n$, green solid line) in the prethermal phase \textcolor{black}{(see Fig.~\ref{fig:PTcorr})}. The intermediate crossover regime is associated with non-diffusive growth of kinetic energy ($E_{\rm kin}(n) \not\sim n$, brown solid line) and temporally frozen correlations but exponentially localized in
space ($W \sim n^{0}$, blue solid line), \textcolor{black}{see Fig.~\ref{fig:IL}}. The diffusive growth of
kinetic energy ($E_{\rm kin}(n) \sim n^1$, brown solid line) corresponds to heating regime where
fluctuations tend to be almost $\delta$-correlated in space and time ($W \to 0$, with $n \to \infty$, red solid line), \textcolor{black}{see Fig.~\ref{fig:k1p2}}. Interestingly, the lifetime of prethermal (heating) regime decrease (increases) with increasing driving parameter $K_1 > K_2$. The different regimes of $W$ are connected
by black (grey) dotted line for $K_1$ ($K_2$). \textcolor{black}{ We note that this plot is a cartoon representation while the exact numerical analysis for the evolution of the kinetic energy is given in Fig.~$2$ of Ref.~\cite{rajak2019characterizations}}.}
\label{fig:schematic}
\end{figure}
\textcolor{black}{\emph{Model and correlation fluctuation.}}---
We consider a generic non-equilibrium many-body system of classical chaos theory as given by
\cite{kaneko89diffusion,konishi90diffusion,falcioni91ergodic,chirikov1993theory,chirikov97arnold,mulansky11strong,rajak2020stability}.
\eqa{H&=\sum_{j=1}^{N}\left[\frac{p_{j}^{2}}{2} -\kappa \sum_{n=-\infty}^{+\infty} \delta(t-n \tau) V\left(r_j\right) \right]
,\label{eq_hamiltonian}}
where stretched variable $r_j = \phi_{j}-\phi_{j+1}$ and $V(r_j) = 1+\cos(r_j)$. Here $\phi_j$, $j=1,\cdots,N$, are the angles of the rotors
and $p_j$ are the corresponding angular momenta. The parameter $\kappa$ denotes the interaction as well as kick strength,
and $\tau$ is the time period of delta kicks. The system described in Eq.~(\ref{eq_hamiltonian}) can have infinite energy
density due to the unbounded nature of kinetic energy.
We note that the total angular momentum of the
system is an exact constant of motion, \textcolor{black}{because the Hamiltonian in Eq.~(\ref{eq_hamiltonian}) is invariant under a global translation $\phi_j\rightarrow\phi_j+\alpha$, $\alpha$ being an arbitrary real number. Moreover, the Hamiltonian has discrete time translation symmetry $H(t)=H(t+\tau)$.}
Using classical Hamilton's equations of motion, one can get the discrete maps of $\phi_j$ and $p_j$ between $n$-th and $(n+1)$-th kicks:
\eqa{
p_j(n+1) &= p_j(n) - \kappa \left( V'(r_{j-1}) -V'(r_j) \right)\nonumber\\
\phi_j(n+1) &= \phi_j(n) + p_j(n+1) \tau.
\label{eq_motion}
}
Here $V'$ describes derivative of $V$ with respect to $r_j$
evaluated after $n$-th kick.
We consider periodic boundary conditions $\phi_{N+i} = \phi_i$.
From Eq.~(\ref{eq_motion}), it can be noticed that the dynamics of the system is determined by only one dimenisonless
parameter, $K=\kappa\tau$ that we use for all our further calculations ~\cite{rajak2018stability,rajak2019characterizations}.
We compute here the spatio-temporal correlation of kinetic energy fluctuations, defined by
\eqa{ C(i,j,t,t_w) = \frac{1}{4} \left[ \langle{p_i^2(t)}{p_j^2(t_w)} \rangle - \langle {p_i^2(t)} \rangle \langle {p_j^2(t_w)} \rangle \right],
\label{eq_correlation}}
where $i$ and $j$, respectively, represent the positions of $i$-th and $j$-th rotors; $t$ ($t_w$) represents an arbitrary final time (initial waiting time). \textcolor{black}{We always consider $t>t_w$ throughout the paper.}
The symbol $\langle..\rangle$ denotes the average over the initial conditions where $\phi_j(0)$ are chosen from a uniform distribution $\in [-\pi,\pi ]$, and
the corresponding momenta, $p_j(0)=0$ for $j=1,\cdots,N$.
The spatio-temporal correlation captures how a typical small perturbation \textcolor{black}{applied at time $t_w$}
spreads in space $x \equiv i-j$ (with translation symmetry)
and time $t$ through the system. \textcolor{black}{
We refer to the correlator in
Eq.~(\ref{eq_correlation}) as $C(x,t)$ while investigating below.
The system shows an exponentially long prethermal state where the kinetic energy
becomes almost constant, and eventually heats up after a crossover regime when the kinetic energy grows linearly with time~\cite{rajak2018stability,rajak2019characterizations}.
We have further analyzed different temporal regimes investigating the behavior of spatio-temporal correlation.
The prethermal state can be characterized appropriately by a time-averaged Hamiltonian (see Eq.~(\ref{eq_eff_ham})).
Therefore, although, the driven system breaks continuous time-translation symmetry, it preserves an effective
time-translation symmetry inside the prethermal regime due to quasi-conservation of Floquet Hamiltonian at high frequencies. Thus, the spatio-temporal correlator becomes a function
of space $x \equiv i-j$ and time $t \equiv t-t_w$ inside the prethermal regime. However, for other two regimes where the
total energy is not a constant of motion, the correlator generally becomes a function of both $t$ and $t_w$ in addition to $x$.
}
\textcolor{black}{We have summarized our main result of spatio-temporal correlation and its connection with the evolution of kinetic
energy schematically in Fig.~\ref{fig:schematic}. In this context, we consider the width of the spatio-temporal correlations, i.e.,
variance \eqa{W =\frac{\sum_{x=-N/2}^{N/2} x^2 C(x,t)}{\sum_{x=-N/2}^{N/2} C(x,t)},} to characterize different dynamical regimes.
In order to measure the relative width of the fluctuations, the appropriate normalization of the distribution as described by the denominator is crucial. In the prethermal regime, the denominator is independent of time due to conservation of energy, while in the other regimes, the denominator is time dependent and normalizes the distribution at all times.
We have schematically drawn the evolution of $W$ in Fig.~\ref{fig:schematic} by acquiring detailed knowledge about spatio-temporal correlation
in different phases as discussed below.
}
The observables are averaged over $10^{5}$ to $10^{6}$ initial conditions. Otherwise specifically mentioned
in our simulations, we fix $\tau=1$ and $N=2048$. This makes $t=n$, and we use these terms interchangeably.
The lifetime of the prethermal state for such systems (see Eq.~(\ref{eq_hamiltonian})) increases exponentially in $1/K$ \cite{rajak2019characterizations}.
For small values of $K$, the prethermal state persists for astronomically large time, thus making the numerical calculation
extremely costly to probe all the dynamical phases by varying time. In order to circumvent this problem, we choose to tune $K$ such that
the lifetime of the prethermal state can be substantially minimized and we can investigate the phase where fluctuations get localized within our numerical facilities.
However, provided the distinct nature of these regimes, our findings would remain unaltered if one addresses them by varying time only.
In our numerical calculations, we choose both $t,~t_w$ within the same phase.
\textcolor{black}{ We would like to emphasize the choice of $t$ and $t_w$ such that the fluctuation correlation of kinetic energy can behave distinctly in different regimes.
There can be some
quantitive but no qualitative changes in the correlator for $t$ and $t_w$ chosen from same regimes while quantitive changes are observed for $t$ and $t_w$ chosen from different regimes. In order to give an idea about the choice of $t$ and $t_w$, we exemplify a situation with $K=0.3$ where $10^1<t,~t_w<10^3$ for prethermal regime, $6\times10^3<t,~t_w<10^4$ for the crossover regime and $2\times10^5<t,~t_w<10^7$ for the heating regime
\cite{rajak2019characterizations}. As discussed above the temporal width of various regimes
vary with $K$ and hence $t$ and $t_w$ are needed to be appropriately chosen within the same regimes.
}
\textcolor{black}{\emph{Results.}}---
We first focus on the spreading of fluctuation (\ref{eq_correlation}) in
the prethermal phase that is denoted by the green solid lines in Fig. \ref{fig:schematic}.
The prethermal phase can be described by a
GGE with the total energy as a quasi-conserved quantity
\cite{rajak2019characterizations}.
In terms of the inverse frequency Floquet-Magnus expansion, the lowest order term of the Floquet Hamiltonian is
the average Hamiltonian that governs the prethermal state at high frequency, given by
\eqa{H^{*}=\frac{1}{\tau} \int_{0}^{\tau} H(t) d t =\sum_{j=1}^{N}\left[\frac{p_{j}^{2}}{2}-\frac{\kappa}{\tau} \left(1+\cos \left(\phi_{j}-\phi_{j+1}\right)\right)\right].
\label{eq_eff_ham}}
Employing the notion of GGE, the
composite probability distributions can be written as
\eqa{ P^*(\{p_j,r_j\}) =\frac{1}{Z^*} \prod_{j=1}^{N} e^{- \big(\frac{p_j^2}{2} - \frac{\kappa V(r_j)}{\tau }\big)/T^* },
\label{eq_distribution}}
where $Z^*$ is the partition function for the GGE and
$T^*$ is the temperature associated with prethermal phase.
Given the particular choice of the initial conditions here, the prethermal temperature is found to be $T^* = 0.938363 \frac{K}{\tau^2}$ \cite{rajak2019characterizations}; for more details see Appendix \ref{GGE_appendix}.
Moreover, this description of the GGE does not depend on the number of rotors $N$, thus indicating the thermodynamic stability of this phase.
We associate the prethermal phase with the diffusive spatio-temporal spread of kinetic energy correlation as shown in
Fig.~\ref{fig:PTcorr} (a). A relevant renormalization
of $x$ and $y$-axes with time yields the following
scaling form of the correlation:
$C(x,t) \sim A_K t^{-1/2} f\big(xt^{-1/2} \big)$,
with $f(y) = e^{-y^2/2D}/\sqrt{2\pi D}$, as depicted in Fig.~\ref{fig:PTcorr} (b); for more \textcolor{black}{detailed discussions} see Appendix \ref{diffusion_appendix}. We thus find that the correlation at different space time collapse together.
Here, $A_K$ denotes the amplitude of the Gaussian distribution respectively for a given value of $K$. The diffusion constant $D$ is a measure of variance $W$, being weakly dependent on the parameter $K$, grows linearly
with time $W \sim t$.
One can observe that the fluctuation spreads
in a way such that the area under the correlation curves keep their area constant i.e. the sum rule $\sum_x C(x,t)$ is approximately independent of time in the prethermal phase.
The sum rule determines $A_K = \sum_{x=-N/2}^{N/2} C(x,t) \approx \sum_{x=-N/2}^{N/2} C(x,0) = \delta_{x,0}\left(\langle p^4 \rangle - \langle p^2 \rangle^2 \right)/4= 0.4402K^2/\tau^4 \delta_{x,0}$, by considering the fact that the energy absorption is exponentially suppressed in the prethermal regime \cite{rajak2019characterizations}; for more details see Appendix \ref{GGE_appendix}. This supports our numerical result of quadratic growth of $A_K$ as shown in the inset of
Fig.~\ref{fig:PTcorr} (b). The apparent $27\%$ mismatch
in the prefactor of $A_K$ with the numerically value might be due to the fact that exponentially slow variation of the kinetic energy in the prethermal phase is not taken into account theoretically.
The energy correlations of static rotor system at high temperature platform exhibit diffusive transport \cite{lepri2016thermal,dhar2008heat}. This is in resemblance with the present case of Floquet prethermalization at high frequency. Owing to the quasi-validity of equipartition theorem in the GGE picture \cite{rajak2018stability,rajak2019characterizations},
the kinetic and potential energy behave in an identical fashion. As a result, the correlation of total energy qualitatively follows the correlation of kinetic energy in the prethermal regime.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Fig2_a.pdf}
\includegraphics[width=\linewidth]{Fig2_b.pdf}
\caption{(a) The space-time spreading of kinetic energy correlation (\ref{eq_correlation}) for $K=0.14$ in the prethermal regime.
(b) The diffusive Gaussian scaling of correlation is observed with appropriate renormalization:
$C(x,t) =A_K t^{-1/2} f(xt^{-1/2}),~ f(y) = e^{-y^2/2D}/\sqrt{2\pi D}$ with parameters $D\sim 0.727$ and $A_K \sim 0.0026$. Inset shows the quadratic variation of
the total area under the curve $\sum_{x=1}^N C(x,t)$ with $K$. \textcolor{black}{ The parameters are $N=2048$, and $t_w = 64$}.
\label{fig:PTcorr}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{Fig3_a.pdf}
\includegraphics[width=1.0\linewidth]{Fig3_b.pdf}
\caption{The space-time spreading of kinetic energy correlation (\ref{eq_correlation}) for (a) $K=0.45$, \textcolor{black}{$t_w = 142$} and
$K=0.7$, \textcolor{black}{$t_w = 64$} (b), respectively, in the intermediate crossover regime. The correlation distribution is fitted with stretched exponential $C(x,t) \sim \frac{A_K}{2 \alpha ^{-1/\beta } \Gamma \left(1+\frac{1}{\beta }\right)} e^{-\alpha |x|^\beta}$ with $\alpha \sim 0.36~[1.05], \beta \sim 0.63~[0.65], A_K \sim 5.16~[41.51]$ for (a) [(b)]. \textcolor{black}{The system size is $N=2048$ for both cases.}
\label{fig:IL}
\end{figure}
We now investigate the spatio-temporal evolution of correlation (\ref{eq_correlation}) in the intermediate crossover regime, designated by the blue solid line in Fig. \ref{fig:schematic}, that lies
between the prethermal and the heating region of
kinetic energy. The system starts to absorb energy from the drive through many-body resonance channels causing the
kinetic energy to grow in sub-diffusive followed by super-diffusive manner \cite{rajak2018stability,rajak2019characterizations}.
However, the probability of the occurrence of such resonances decreases exponentially with $1/K$ in the high frequency limit.
We find that the oscillators are maximally correlated with each other at $x=0$ and falls rapidly to zero in two sides $x$ as shown in Fig.~\ref{fig:IL} (a) and (b), for $K=0.45$ and $0.7$, respectively. To be precise, correlation
decays stretched exponentially in short distances:
$C(x,t) \sim A_K e^{-\alpha |x|^\beta}$ while it falls exponentially (i.e., more rapidly than stretched exponential) in long distances: $C(x,t) \sim e^{-\gamma |x|} $; for more details see Appendix \ref{crossover_appendix}. Here, $\beta$ and $\gamma$ weakly depend on $K$ referring to the fact that driving parameter $K$ can in general control the spatial spread of fluctuation.
These profiles do not change with time within the crossover region
referring to the fact that variance of the spatial correlation distribution remains constant with time $W \sim t^0 $. This allows
us to differentiate it from the diffusive transport that occurs in the prethermal phase.
However, with increasing time, one can observe that long distance correlation becomes more noisy
leaving the spatial structures qualitatively unaltered.
At the end, we discuss the time zone where the average kinetic energy shows unbounded chaotic diffusion, as denoted by the red solid line in Fig.~\ref{fig:schematic}, resulting in
the effective temperature to increase linearly with time~\cite{rajak2018stability, rajak2019characterizations}. The correlation of the kinetic energy is fully localized in space and temporally frozen as shown in Fig.~\ref{fig:k1p2}). To be precise, the correlation
is nearly a $\delta$-function centered around $x=0$ i.e., fluctuation gets localized at the site of disturbance for all time. In this regime, the system shows fully chaotic behavior in the phase space and the angles of the rotors
become statistically uncorrelated both in space and time.
It is noteworthy that
there is no description of average Hamilton exist here as the inverse frequency Floquet-Magnus expansion does not converge \cite{Mori18,Howell19}.
In contrast to the prethermal phase, the amplitude of correlation peak in the crossover and heating regime increases as $K^\eta$ with $\eta>2$. On the other hand, the variance in the heating regime becomes decreasing function of $n$, precisely, $W \to 0$ for $n \to \infty$ that is markedly different from the behavior of $W$ in remaining two
earlier regimes. Finally, we stress that our findings in this heating regime do not suffer from finite size effect suggesting the
thermodynamic stability of this phase.
\textcolor{black}{It is noteworthy that $\langle{p_i^2(t)}{p_j^2(t_w)} \rangle$ and $\langle{p_i^2(t)}\rangle \langle{p_j^2(t_w)} \rangle$ in the kinetic energy fluctuation both individually depend on time $t$ and the initial waiting time $t_w$. On the contrary, the connected part i.e., $C(i,j,t,t_w)$ as a whole does not depend on $t$ instead depends on $t_w$ such that the peak height of $C(i,j,t,t_w) $-profile increases with $t_w$. This can be physically understood as an initial value problem in terms of $p_{i,j}(t_w)$ for the rotors that subsequently uncorrelated in the heating region.
The peak value of the correlator is also found to be dependent on the coupling parameter $K$.
In this region, the time-independent
correlator effectively freezes into three discrete spatial points i.e., at the site of disturbance with $i=j$ and the remaining two adjacent sites with $i=j\pm 1$.}
\textcolor{black}{We shall now exploit the assumption of statistically uncorrelated (both in space and time) nature of the rotors to shed light on this intriguing behavior analytically. In terms of the stretched variable $r_j(n)=\phi_{j+1}(n)-\phi_j(n)$, the assumption leads to the following mathematical form
\begin{eqnarray}
&&\langle\sin(r_i(n))\sin(r_j(m))\rangle=\frac{1}{2}~\delta_{n,m}\delta_{i,j},\nonumber \\
&&\langle\sin(r_i(n))\sin(r_j(n') \sin(r_k(m))\sin(r_l(m'))\rangle=\nonumber \\
&&\frac{1}{4}~(\delta_{n,n'} \delta_{m,m'} \delta_{i,j} \delta_{k,l} + \delta_{n,m} \delta_{n',m'} \delta_{i,k} \delta_{j,l} \nonumber \\
&&+ \delta_{n,m'} \delta_{n',m} \delta_{i,l} \delta_{j,k})
+ \frac{3}{8} \delta_{n,n'} \delta_{m,m'} \delta_{m,n} \delta_{i,j} \delta_{k,l} \delta_{i,k}
\label{approx}
\end{eqnarray}
where the average is carried over different initial conditions, $i,j,k,l$ represent the position of the rotor and $n,n',m,m'$ denote the various times $t$'s in terms of the number of kicks.
We note that $r_j(n)$ becomes uniform random variable in the heating region.
The momentum of $i$-th rotor at any time
$t$ can be formulated from the equation of motion (Eq.~\ref{eq_motion}).
The first term of the kinetic energy fluctuation correlator with $t>t_w\gg1$ can be calculated in the heating regime as follows \begin{align}
&\langle p_i^2(t)p_j^2(t_w)\rangle
=K^4\Big(2~ t_w^2 \delta_{i,j} + \frac{1}{2}~ t_w^2 \delta_{i,j+1} + \frac{1}{2} t_w^2 \delta_{i,j-1} \nonumber \\
& + \frac{3}{4} t_w \delta_{i,j} + \frac{3}{8}~ t_w \delta_{i,j+1} + \frac{3}{8}~ t_w \delta_{i,j-1}+
tt_w \delta_{i,i} \delta_{j,j}
\Big).
\label{heating_term1}
\end{align}
The last term reminds us the diffusive growth of kinetic energy $\langle p_i^2(t)\rangle=K^2 t$ that is obtained in the heating region. Therefore, the kinetic energy fluctuation correlation (Eq.~(\ref{eq_correlation})) takes the following form
\begin{align}
&C(i,j,t,t_w)
=K^4\Big( \frac{t_w^2}{2} \delta_{i,j} + \frac{3t_w}{16} \delta_{i,j}+
\frac{t_w^2}{8} \delta_{i,j+1} + \frac{3t_w}{32} \delta_{i,j+1}\nonumber \\
&+ \frac{t_w^2}{8} \delta_{i,j-1} + \frac{3t_w}{32} \delta_{i,j-1}\Big )\nonumber \\
&= K^4\Big( \frac{t_w^2}{2} \delta_{i,j} +
\frac{t_w^2}{8} \delta_{i,j+1} + \frac{t_w^2}{8} \delta_{i,j-1} \Big ) + O(t_w)
\label{heating_corr}
\end{align}
}
\textcolor{black}{We note that for $t_w\gg 1$, the $O(t^2_w)$ terms dominate over $O(t_w)$ term.
This clearly suggests that the assumption of uncorrelated nature of rotors is able to mimic the numerical outcome convincingly in the heating region i.e., $C(x,t)\ne 0$ for $x=0$ and $\pm 1$. In Fig.~\ref{fig:k1p2}, we compare the numerical results with the prediction of Eq.~\ref{heating_corr} and find a match within $6\%$ error.
The above result is derived considering the assumption that rotors are always uncorrelated under driving.
As a result, the independent rotor approximation works better to explain the numerical outcomes
for higher values of $K$ as the driven system enters into the heating region quite early.
Most importantly, the analytical form in Eq.~(\ref{heating_corr}) correctly captures the value of the correlator at the adjacent sites of the disturbance $i=j\pm 1$
drops to $\frac{1}{4}$ of the peak value at $i=j$, as observed in Fig.~\ref{fig:k1p2}.
The non-zero value of correlation in the adjacent sites might be the effect of nearest-neighbor interactions of the system.
In addition, the correlator is independent of final time $t$, whereas it depends on $t_w$, when the disturbance is applied on the system.
It indicates that the correlator depends on the value
of momentum at $t=t_w$, from where we start measuring the correlator, but correlator does not spread further with time since the rotors are effectively uncorrelated in this regime.
The detailed analytical derivation is presented in Appendix \ref{heating_analysis}.
}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Fig4.pdf}
\caption{The correlations (\ref{eq_correlation}) become
fully localized such as nearly a $\delta$-function in the heating region for $K=3$, \textcolor{black}{with $t_w=142$, $N=2048$. The blue circles denote the correlation values as a function of initial wait time and the kick strength given by Eq.~\ref{heating_corr}. } }
\label{fig:k1p2}
\end{figure}
\textcolor{black}{We here stress that Fig.~\ref{fig:schematic}, being a cartoon representation, can compactly demonstrate the results shown in Figs.~\ref{fig:PTcorr}, \ref{fig:IL}, and \ref{fig:k1p2} as well as shed light on the physical understanding. We superimpose
the time evolution of the variance $W$ of the correlation $C(x,t)$, obtained from analyzing the correlation profile in Figs.~\ref{fig:PTcorr}, \ref{fig:IL}, and \ref{fig:k1p2}, with the kinetic energy $E_{\rm kin}(n)$ that has already been reported \cite{rajak2019characterizations}. In the prethermal regime, $E_{\rm kin}(n) \sim n^0$ and the self-similar Gaussian scaling of $C(x,t)\sim f( x t^{-1/2})$ suggests $ W \sim n $ (see Fig.~\ref{fig:PTcorr}(b)). In the crossover region, $E_{\rm kin} \not \sim n$ and the temporally frozen stretched exponential
$C(x,t)\sim f(x,t^0)$ yields $W \sim n^0$ (see Fig.~\ref{fig:IL}). In the heating region, $E_{\rm kin} \sim n$, fluctuations become
nearly $\delta$-correlated but frozen in time $C(x,t)\sim f(x,t^0)\delta_{x,l} $ with $l \in [0,\pm 1]$, leads to the following situation $W \to 0$ as $n\to \infty$ (see Fig.~\ref{fig:k1p2}).
Therefore, the disjoint nature between the spatial and temporal structures of the correlation in crossover and heating regions causes distinct features in $W$ as compared to its combined spatio-temporal nature for prethermal region.
}
\textcolor{black}{\emph{Discussions.}}---
This is clearly noted in our study that the spatio-temporal correlation provides
a deep insight to characterize different dynamical phases.
The phase matching between adjacent rotors, captured by stretched variables $r_j(n)=\phi_{j+1}(n)-\phi_j(n)$, play very important
role in determining the nature of spreading of the fluctuations.
The time-evolution
of the stretched variable, following the equations of motion (\ref{eq_motion}), is given as $r_j(n+1)=r_j(n)+(p_{j+1}(n+1)-p_j(n+1))\tau$. Upon satisfying the resonance condition $p_j-p_{j+1}=2 \pi m/\tau$ with $m$ as an integer number,
the stretched variable rotates by $2\pi$ angle
between two subsequent kicks. When all the rotors go through these resonances, their relative phase matching is lost and the eventually the coupled rotor system turned into an array
of uncoupled independent rotor. At this stage,
the system absorbs energy from the drive at a constant rate
in an indefinite manner. This is precisely the case for heating up regime where fluctuation correlations do not spread in time and space. On the other hand, the resonances are considered to be extremely rare events
in the prethermal regime suggesting the fact that
relative phase matching between
adjacent rotors allows the fluctuation to propagate diffusively throughout the system in time.
The notion of the time independent average Hamilton in the prethermal region might be related to the fact that
all the rotors rotate with a common collective phase and eventually controlled by an underlying synchronization phenomena \cite{nag17,khasseh19}.
Coming to the {phenomenological} mathematical description,
the exponential suppressed heating and the validity of the constant sum-rule in the prethermal phase suggest a hydrodynamic diffusion picture (with diffusion constant $D$) for the fluctuation \cite{spohn14,spohnKPZ,dhar2021revisiting,Norman20}:
\eqa{ \partial_t u(x,t) = \partial_x \bigg( \frac{D}{2}~ \partial_x u(x,t)
+ B ~ \zeta (x,t) \bigg )
\label{eq_diffusion}}
where $u(x,t) = \frac{1}{2} \left(p^2(x,t) - \langle p^2(x,t) \rangle\right)$ such that $\langle p^2(x,t) \rangle \neq \langle p^2(x, t') \rangle$. The conservative noise $\zeta(x,t)$ of strength $B$
is delta correlated in space and time $\langle \zeta(x,t) \zeta(x',t')\rangle = \delta_{xx'} \delta(t-t')$.
In the high-frequency limit, the equilibrium fluctuation dissipation relation can be extended to long-lived prethermal regime: $B^2 \sim D { T^*}^2$; for more details see Appendix \ref{hydrodynamic_appendix}. Given the plausible assumption that the noise part $B$ of the fluctuating current increases with increasing $K$, one can understand the diffusion process in a phenomenological way. The diffusion constant $D$ is then considered to be independent of $K$ while prethermal temperature is determined by $K$, as observed in GGE picture.
Moreover, the self-similar Gaussian nature of fluctuation correlation in the prethermal phase can be understood from the solution of Eq.~\ref{eq_diffusion} such that $ C(x,t) \sim \langle u^2(0,\tau_i) \rangle e^{-x^2/(2Dt)} /\sqrt{2\pi D t}$. In the other limit inside the
infinite temperature heating regime, the non-diffusive transport leaves the $u(x,t)$ to be $\delta$ correlated in space while almost frozen in time. Now there is an extended crossover region, connecting the prethermal state to heating regime, where the diffusion equation does not take such simple form causing the system to exhibit an amalgamated
behavior.
\textcolor{black}{Before we conclude, we would like to re-emphasize that unlike the equilibrium case $t_w$ plays crucial role for the non-equilibrium case. To be more precise, only relevant time variable is the relative time $t \equiv t-t_w$ for equilibrium case respecting time translation symmetry. This is not true in the present case and the correlation is expected to exhibit complex structure for arbitrarily chosen $t$ and $t_w$ with $t>t_w$.
Thanks to the GGE description of the prethermal region, we can consider the relative time $t \equiv t-t_w$ as the appropriate variable.
One can think of the average Hamiltonian embeds an effective time translation symmetry in the prethermal phase.
However, the effective GGE description fails for crossover and heating region, resulting in the fact that the two time instants $t$
and $t_w$ are equally important there.
The effective time translation symmetry is no longer valid in the above two regions.
This is what we clearly see for the heating regime where the peak value of the correlator depends on $t_w$. }
\textcolor{black}{\emph{Conclusions.}}---
In conclusion, our study demonstrates that a typical two-point observable such as,
the fluctuation correlation of kinetic energy, can be scrutinized to probe different dynamic phases of classical many-body kicked rotor system \cite{computation}. It is indeed counter intuitive that the average kinetic energy per rotor and their spatio-temporal correlations, being derived from the former quantity, yet yield opposite behavior.
The GGE description of prethermal phase obeys diffusive transport where spatio-temporal correlation follows
self similar Gaussian profile. During this diffusion, the correlation curves keep their area constant
due to the quasi-conservation of the kinetic energy. In the long time limit where kinetic energy grows diffusively, fluctuation interestingly becomes frozen in space and time.
\textcolor{black}{We have also calculated the behavior of kinetic energy fluctuation analytically using the statistically uncorrelated nature of the angles of the rotors inside the heating regime and find a good match with the numerical results.}
There exists an extended crossover region where kinetic energy increases in a complicated way exhibiting both sub-diffusive to super-diffusive nature. The fluctuation interestingly shows a rapidly (slowly) decaying short (long) range stretched (regular) exponential localization. In this case the fluctuation does not have any time dynamics. Therefore, correlated phenomena in prethermal phase gradually assembles to completely uncorrelated heat up phase through a crossover region. These non-trivial phases of matter are the consequences of the driving and do not have any static analogue. Provided the understanding on long range quantum systems \cite{Lerose19,Saha19,Tomasi19,garg19,modak20a,modak20b}, it would be interesting to study the fate of the above phases along with the crossover region in long-range classical systems.
\textcolor{black}{Furthermore, various intriguing nature of correlators can be observed for
$t$ and $t_w$ chosen from different phases that is beyond the scope of the present study.}
The microscopic understanding of hydrodynamic picture and fluctuation-dissipation relation in dynamic systems are yet to be extensively analyzed in future.
{\it{Acknowledgement} }:
We thank Andrea Gambassi and Emanule Dalla Torre for reading the manuscript and their constructive comments. A.R. acknowledges UGC, India for start-up research grant F. 30-509/2020(BSR). AK would like to thank the computational facilities Mario at ICTS and Ulysses at SISSA.
|
\section{Introduction}
The electron-positron pair production has been studied theoretically~\cite{2010The,2011Interference,2005Worldline,2006Worldline,2002Quantum,2010Schwinger} and experimentally~\cite{1985Observation,1997Search,1997Positron,1996Observation,2010Collision,2012Breit,2009Relativistic,2015The,Augustin_2014} by many authors since Sauter~\cite{Sauter1931} considered pair creation in a static electric field. In 1951~\cite{1951On}, the pair creation rate in a constant field was calculated by Schwinger and he pointed out the electric field strength for creating observable pairs, which is $E_c={10}^{16}$ \rm{V/cm}, corresponding to laser intensity of ${10}^{29}\rm{W/cm^2}$ respectively. With the rapid developments of laser technology and the chirped-pulse amplification (CPA) technology ~\cite{1985Compression,1988Amplification}, laser intensity could reach ${10}^{24}\rm{W/cm^2}$ at the present~\cite{Eli}, but still far more smaller than the critical value of Schwinger effect.
However, there is no experiment that directly achieves the conversion from energy to matter, while it is full of hope in the future.
In terms of experiment, Cowan~\cite{1985Observation} observed positive and negative electron beams in the heavy ion collision experiment in 1986.
After a decade, Ahmad~\cite{1997Search} produced a single energy positron beam through heavy ion collisions. In these experiments, however, the nuclear reaction caused by the collision of high-energy relativistic ions and the structure change of the high-Z nucleus are much greater than the effect of the Coulomb fields` superposition of two nuclei.
In general, most positrons came from a nuclear reaction rather than a direct vacuum breakdown. Also, in 1997, Burke et al.~\cite{1997Positron} used the 46.6 GeV electron beam generated by Stanford Linear Accelerator (SLAC) to collide with a ${10}^{18}W/cm^2$ laser beam, gained pair productions.
In this experiment, the non-linear Compton scattering of laser photons and electrons generates high-energy gamma photons~\cite{1996Observation} , and these high-energy gamma photons continually interact with the laser to generate positron and electron pairs, called the Breit-Wheeler process~\cite{2010Collision,2012Breit} .
In 2009, Chen~\cite{2009Relativistic} irradiated a gold target with an ultra-strong laser of ${10}^{20}W/cm^2$, and detected a positron beam of ${10}^{10}$ per sphere behind the gold target. In this experiment, laser photons and nuclear scattering generated high-energy gamma photons and these high-energy photons interact with high-Z nuclei to generate a large number of positrons and electrons, respectively, named the Bethe-Heitler process~\cite{2015The,Augustin_2014} .
In terms of theory, many theoretical ideas have been adopted for solving the non-perturbative and non-equilibrium process in pair production, such as Wentzel-Kramers-Brillouin approximation (WKB)~\cite{2010The,2011Interference} , World-line instanton technique ~\cite{2005Worldline,2006Worldline} , solving Vlasov equation ~\cite{2002Quantum} , Wigner function formalism~\cite{2010Schwinger} and so on. In this article, we discuss the pair creation problem in asymmetric potential well with computational quantum field theory (CQFT)~\cite{1999Numerical,2005Creation,2006Timing} which has had a great achievement not only with pair creation, also with Zitterbewegung~\cite{2004Relativistic} , Klein paradox~\cite{2004Klein,article} , relativistic localization problem~\cite{2004Relativistic} and so on.
Recently more interesting works were performed, which include one of ours for the effective interaction time mechanism to improve the pair numbers significantly \cite{2019Wang}. In this paper, furthermore, we study the pair creation problem in asymmetric Sauter potential well. We examine it by using the fields which include both the static asymmetric Sauter potential and the alternating field. It is found that the width of the potential well plays an important role in this electron-positron production process.
The paper is organized as follows. In Sec.\ref{the} we discuss the theoretical framework for computational quantum field theory. In Sec.\ref{multi} we study the momentum spectrum and the location distribution of the created electrons. In Sec.\ref{evolution} we compare the evolution of the number of particles created in different types of asymmetric potential well. Finally, in Sec.\ref{sum} a brief summary of the work is given.
\section{The theoretical framework for CQFT}\label{the}
As for electron and positron we use Dirac equation, we use the atomic units (a.u.) as $\hbar=m_0=e=1$ in this whole paper, as for fine structure constant $\alpha=1/c=1/137.036$, and consider a one-dimensional system along the $z$ direction for the simplicity of our calculation and the model. Here~\cite{1961An,Campbell:2017hsr}
\begin{equation}
i\partial\hat{\psi}\left({z},t\right)/\partial{t}=\lbrack{c\alpha_z{\hat{p}}_z+\beta c^2+V(z,t)}\rbrack\hat{\psi}\left(z,t\right),
\end{equation}
$\alpha_{z}$ denotes as $z$ component of the Dirac matrix, $\beta$ denotes unit Dirac matrix, and we only focus on a single spin thanks to there is no magnetic field in our one-dimensional system. In results, the four-component spinor wave function becomes two components and Dirac matrix $\alpha_{z}$ and $\beta$ are replaced by the Pauli matrix $\sigma_1$ and $\sigma_3$, where:
\begin{equation}
\sigma_1=\left(\begin{matrix}0 & 1\\1 & 0\\\end{matrix}\right), \ \ \ \ \ \ \ \ \ \ \sigma_3=\left(\begin{matrix}1&0\\0&-1\\\end{matrix}\right).
\end{equation}
The Hamiltonian is
\begin{equation}
{H}=c\sigma_1{\hat{p}}_z+\sigma_3 c^2+V(z,t),
\end{equation}
whereby $V\left({z},t\right)$ is the classical external scalar potential along the $z$ direction. We can express $\hat{\psi}\left(z,t\right)$ as the combine of creation and annihilation operators as follows:
\begin{equation}
\sum_p \hat{b}_p(t) W_p(z)+\sum_{n} \hat{d}_n(t) W_n(z) = \sum_p \hat{b}_p W_p(z,t)+\sum_n \hat{d}_n W_n(z,t),
\end{equation}
where $ \sum_{p(n)}\ $ stands for the summation over all states with positive and negative energy, and $ W_{p(n)}\left({z},t\right)=\langle{z}|p\left(n\right)(t)\rangle $ is the solution of the Dirac equation for the initial condition $W_{p(n)}\left({z},t=0\right)=W_{p(n)}\left({z}\right)$, where $W_{p(n)}\left({z}\right)$ is the energy eigen-function of the field-free Dirac equation. We also can express the fermion operators as:
\begin{equation}
{\hat{b}}_p(t)=\sum_{{p^\prime}}{{\hat{b}}_{p^\prime}U_{pp^\prime}\left({z},t\right)+\sum_{n^\prime}\ {\hat{d}}_{n^\prime}^\dag U_{pn^\prime}\left({z},t\right),\ \ \ \ \ \ \ \ \ \ }
\end{equation}
\begin{equation}
{\hat{d}}_n^\dag(t)=\sum_{{p^\prime}}{{\hat{b}}_{p^\prime}U_{np^\prime}\left({z},t\right)+\sum_{n^\prime}\ {\hat{d}}_{n^\prime}^\dag U_{nn^\prime}\left({z},t\right),\ \ \ \ \ \ \ \ \ \ }
\end{equation}
\begin{equation}
{\hat{b}}_p^\dag(t)=\sum_{{p^\prime}}{{\hat{b}}_{p^\prime}^\dag U_{pp^\prime}^\ast\left({z},t\right)+\sum_{n^\prime}\ {\hat{d}}_{n^\prime} U_{pn^\prime}^\ast\left({z},t\right),\ \ \ \ \ \ \ \ \ \ }
\end{equation}
\begin{equation}
{\hat{d}}_p(t)=\sum_{{p^\prime}}{{\hat{b}}_{p^\prime}^\dag U_{np^\prime}^\ast\left({z},t\right)+\sum_{n^\prime}\ {\hat{d}}_{n^\prime} U_{nn^\prime}^\ast\left({z},t\right),\ \ \ \ \ \ \ \ \ \ }
\end{equation}
where $U_{pp^\prime}\left(t\right)=\left\langle p\left|\hat{U}\left(t\right)\right|{p^\prime}\right\rangle,\ U_{pn^\prime}\left(t\right)=\left\langle p\left|\hat{U}\left(t\right)\right|{n^\prime}\right\rangle,\ U_{nn^\prime}\left(t\right)=\left\langle n\left|\hat{U}\left(t\right)\right|{n^\prime}\right\rangle,\ U_{np^\prime}\left(t\right)=\left\langle n\left|\hat{U}\left(t\right)\right|{p^\prime}\right\rangle.$
The time evolution operator of the field
$ \hat{U}\left(t\right)\equiv\hat{T}\exp{(-i\int_{0}^{t}{Hd\tau)}}$, $\hat{T} $
denotes time-order operator and due to the operators at different times may not be commutatived,
and it helps to sort the operators at different times so that the operator at the early time is classified to the right side of the operator at the late time.
The electronic portion of the field operator is defined as $ \hat{\psi}_{e}^{+}\left({z},t\right)\equiv\sum_{p}{{\hat{b}}_p\left(t\right)W_p({z})} $
so that the created electrons` spatial number density can be written as:
\begin{equation}
\rho_{e}\left({z},t\right)=\langle \text{vac}|{\hat{\psi}}_{e}^{+\dag}\left({z},t\right){\hat{\psi}}_{e}^+\left({z},t\right)|\text{vac}\rangle.
\end{equation}
The anticommutator relations $ \left\{{\hat{b}}_p,{\hat{b}}_{p^\prime}^\dag\right\}=\delta_{pp^\prime}$ and $\left\{{\hat{d}}_n,{\hat{d}}_{n^\prime}^\dag\right\}=\delta_{nn^\prime}$, the number density of the electrons can be rewritten as:
\begin{equation}
\rho_{e}\left({z},t\right)=\sum_{{n}}{\mid\sum_{{p}}{U_{pn}\left(t\right)W_p\left({z}\right)\mid^2,\ }}
\label{density}
\end{equation}
where $U_{pn}\left(t\right)$ can be computed with the split operator numerical technique~\cite{1999Numerical}. By integrating the Eq.\eqref{density}, we can obtain the total number of created electrons as:
\begin{equation}
\label{eq2}
N\left(t\right)=\int\rho_{e}\left({z},t\right)d{z}=\sum_{{p}}\sum_{{n}}{\left|U_{pn}\right|^2.\ }
\end{equation}
\section{The Multi-photon process}\label{multi}
We set the amplitude of potential well $V_1$ and alternating field $V_2$ both equal to $V_1=V_2=2c^2-10000$, and our total field is:
\begin{equation}
V\left(z,t\right)=V_1S\left(z\right)f\left(t\right)+V_2\sin{\left(\omega t\right)}S\left(z\right)\theta(t;t_0,t_0+t_1),
\end{equation}
where
\begin{equation}
S\left(z\right)=\{[\tanh{\left(z-{D/2}\right)}/{W}_1]-[\tanh{\left(z+{D/2}\right)}/{W}_2]\}/2
\label{shape},
\end{equation}
and
\begin{equation}
f\left(t\right)=\sin[\pi{t}/2{t}_0]\theta(t;0,t_0)+\theta(t;t_0,t_0+t_1)+\cos[\pi({t-{t}_0-{t}_1})/2{t}_0]\theta(t;t_0+t_1,2t_0+t_1).
\label{time}
\end{equation}
Eq.\eqref{time} describes how the processes of our potential well to turn on and off, and we bring a step function $\theta\left(t;t_1,t_2\right)$ on. As demonstrated by the total field function, the oscillating field doesn't work during $(0,t_0)$ in order to reduce the trigger effect of the pair production, so we set $t_0=5/c^2$ for the well establishment of the potential well. Then we turn on the oscillating field during $(t_0,t_0+t_1)$, and $t_1=20\pi/c^2$. The total time for our calculation is $T=2t_0+t_1=(10+20\pi)c^2$ and divided into $N_t=10000$ time intervals. We have also divided the calculation space into $N_z=2048$ grid points. When we set these two parameters larger, we will get more precise results in our simulation experiment, meanwhile, the experiment time also gets longer. Therefore, they should be controlled within a reasonable range.
For the Eq.\eqref{shape}, $D$ is the width of the potential well, and $W_1,\ W_2$ are widths of the electric field at the edge. We set the $D=10\lambda_e$, where the $\lambda_e$ is the Compton wavelength. $W_1=0.3\lambda_e$ and keep stable, $W_2$ has different values, which is from $0.075\lambda_e$ to $1.5\lambda_e$, exactly quater to five times of the $W_1$, respectively. The frequency of the oscillating field is $\omega=2.1c^2$.
\begin{figure}[htbp]
\suppressfloats
\begin{center}
\includegraphics[width=\textwidth]{Fig1}
\end{center}
\caption{Momentum spectra of the pair creation for one symmetric combined field(red), and two asymmetric combined field with different widths of the left side(black and blue spectrum). The frequency of the oscillating field is $\omega=2.1c^2$, other parameters are $N_z=2048$,\ $N_t=10000$,\ $V_1=V_2=2c^2-10000$,\ $D=10\lambda_e$, and $W$ is given in the figure.}
\label{fig:1}
\end{figure}
The FIG.\ref{fig:1} is the momentum spectrum of the pair production in three different shapes of our potential well. As we compare the black line which represents to the asymmetric potential well with $W_1=0.3\lambda_e,\ W_2=0.15\lambda_e$, and the red line which represents to the symmetric well, we can easily notice that the $N_p>0$ part has very clear changes compared with the $N_p<0$ part. The reason of this effect is that the electric field gets stronger at the left edge of our potential well as the $W_2$ gets smaller, and the electrons with positive momentum get more and more increase. On the contrary, blue line which represents to the asymmetric potential well with $W_1=0.3\lambda_e,\ W_2=0.6\lambda_e$, has clear decline in its momentum distribution compared with the symmetric potential well (red line). All this three spectra have one very common feature, which is when $N_p$ close to 0 the change is not very clear, but as $N_p>>0$ there has a marked differences between these three momentum spectra. We will discuss this in our location distribution part.
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.48\textwidth}
\centering
\includegraphics[width=1\textwidth]{Fig2a}
\end{minipage}
\begin{minipage}[t]{0.48\textwidth}
\centering
\includegraphics[width=1\textwidth]{Fig2b}
\end{minipage}
\caption{The location distribution of the created electrons, (a) is the comparison of asymmetric potential well with $W_1=0.3\lambda_e,\ W_2=0.15\lambda_e$ (black), and symmetric well with $W_1=W_2=0.3\lambda_e$ (red), (b) is the comparison of asymmetric potential well with $W_1=0.3\lambda_e,\ W_2=0.6\lambda_e$ (black), and symmetric well with $W_1=W_2=0.3\lambda_e$ (red).}
\label{fig:2}
\end{figure}
FIG.\ref{fig:2} is the location distribution of the created electrons along the $z$ direction. In FIG.\ref{fig:2} (a), compare with the symmetric potential well, the asymmetric potential well with $W_1=0.3\lambda_e,\ W_2=0.15\lambda_e$ (black line in a) shows there are more electrons in our space. Around $z=0$ area, particle number density is little bit higher in the right side than the left side or we can say the location distribution is asymmetric when $z$ close to 0, it's very logical with our conclusions on the top. But there is also an equal amount of increase on the both side along the $\pm {z}$ direction. We can explain this phenomenon with the Fourier transformation, small-scale spatial distribution is related to large momentum and the large-scale spatial distribution is related to small momentum. As in FIG.\ref{fig:1}, the changing of width in left side has more effects to the large momentum electrons so that small-scale spatial distribution has obvious asymmetric values. The large-scale spatial distribution increase in equal amount or we can say they are symmetric, because of our asymmetric potential well has few effects to the small momentum electrons as in FIG.\ref{fig:1}. On the contrary, in FIG.\ref{fig:2} (b), there is a clear decrease on particle number density when we make the $W_2$ two times wider than the $W_1$.
Thanks to the left side of our potential well keep changing while the right side is being stable, our spectra have optimized when width gets narrower. So in our next step, we will only discuss $W_2=0.5W_1=0.15\lambda_e$ shaped one, numerically, the detailed information about the momentum spectrum of the pair production in symmetric potential well is discussed in the reference \cite{2013Tang}. We calculate the eigenvalues of the bound states in the well with~\cite{Campbell:2017hsr}:
\begin{equation}
cp_2\cot{\left(p_2D\right)}=\frac{EV_1}{cp_1}-cp_1,
\end{equation}
whereby $p_1=\sqrt{c^2-E^2/c^2} ,\ p_2=\sqrt{(E+V_1)^2/c^2-c^2},\ \ V_1={2c}^2-10000,\ D=10\lambda_e$. This equation is symmetric in $\pm E$, so in our asymmetric well would have some uncertainties occurred.
\begin{figure}[htbp]
\centering
\begin{minipage}[t]{0.51\textwidth}
\centering
\includegraphics[width=1\textwidth]{Fig3a}
\end{minipage}
\begin{minipage}[t]{0.48\textwidth}
\centering
\includegraphics[width=1.\textwidth]{Fig3b}
\end{minipage}
\caption{Momentum spectrum of the asymmetric fields, the numerical grids are same as FIG.\ref{fig:1}}\label{fig:3b}
\end{figure}
We get the results as follow : $E_1=-0.4247c^2\ ,\ E_2=-0.3069c^2\ ,\ E_3=-0.1361c^2\ \ ,\ E_4=0.0680c^2\ \ ,\ E_5=0.2919c^2\ \ ,\ E_6=0.5260c^2\ \ ,\ E_7=0.7618c^2\ \ ,\ E_8=0.9978c^2$.
The relationship between peak energy in the FIG.\ref{fig:1} and the eigenvalue of the bound states in the symmetric potential well is given in the reference~\cite{2013Tang}, which is $E_{pi}=E_i+n\omega$.
We verify this relationship whether it works for our asymmetric potential well. In the FIG.\ref{fig:3b} (a), it shows the one-photon process, which means the electron jumps from the well in its bound states through absorbing one photon with frequency of $\omega=2.1c^2$. We test it with calculation of its energy,
\begin{equation}
E=\sqrt{c^2p^2+c^4},
\label{eq2}
\end{equation}
whereby $p=2N_p\pi/L$, where $L$=2.0 is the length of the numerical grid and the $N_p$ is the pick value we have marked in FIG.\ref{fig:3b}. Take an example of $N_p=58$, the eigenvalue $E_1=-0.4247c^2$ and we calculate the energy which correspond to the $N_p=58$ with Eq.\eqref{eq2}, and we get $E_{p1}=1.6637c^2$, the relationship is roughly $E_{pi}-E_i=1.6637c^2-\left(-0.4247c^2\right)=2.0884c^2\approx\omega=2.1c^2$.
We take FIG.\ref{fig:3b} (c) as an example of our calculation for multi-photon effect which is estimated that electrons in the bound states absorb three photons and become free particles. According to the values of each peak, we know $N_p=254,\ 260,\ 268,\ 279,\ 291,\ 302,\ 315,\ 324$ precisely , and each of them stand for $i=1,\ 2,\ 3,\ 4,\ 5,\ 6,\ 7,\ 8.$ And we calculate the energy of the electrons at every single peak, they can be written as $E_{N_p=254}=5.9083c^2,\ {\ E}_{N_p=260}=6.0449c^2,\ E_{N_p=268}=6.2248c^2,\ E_{N_p=279}=6.4739c^2,\ E_{N_p=291}=6.7458c^2,\ E_{N_p=302}=6.9953c^2,\ E_{N_p=315}=7.2904c^2,\ E_{N_p=324}=7.4948c^2.$
\
\begin{center}
TABLE1: The value of $E_{pi}$ and $E_i$ at the peaks
\vskip 1pt
\centering
\begin{tabular}{ |c|c|c|c|c|c|c|c|c| }
\hline
$i$ & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8\\
\hline
$E_i$ & $-0.4247c^2$ & $-0.3069c^2$ & $-0.1361c^2$ & $0.0680c^2$ & $0.2919c^2$ & $0.5260c^2$ & $0.7618c^2$ & $0.9978c^2$\\
\hline
$N_p$ & 254 & 260 & 268 & 279 & 291 & 302 & 315 & 324\\
\hline
$E_{pi}$ & $5.9083c^2$ & $6.0449c^2$ & $6.2248c^2$ & $6.4739c^2$ & $6.7458c^2$ & $6.9953c^2$ & $7.2904c^2$ & $7.4948c^2$\\
\hline
\end{tabular}
\end{center}
\
With very simple calculation we verify that $E_{pi}=E_i+3\omega$, also indicate our estimation is fully correct.
The changing of the momentum spectrum as we turnover the widths at the two side of the field is predictable. We have just compared two groups of data, which are $W_1=0.3\lambda_e,W_2=0.6\lambda_e$ and $W_1=0.6\lambda_e,W_2=0.3\lambda_e$ respectively. Obviously, the momentum distribution of positive and negative $N_p$ has transferred its corresponding values as we turnover the widths at each edge.
\section{Evolution of the electron number}\label{evolution}
In this section we study the pair production in subcritical asymmetric potential well with a subcritical oscillating field, but with different widths of electric fields as we have considered in the last section. We set $V_1=V_2={2c}^2-10000$, and $\omega=2.1c^2$.
\begin{figure}[htbp]
\suppressfloats
\begin{center}
\includegraphics[width=1\textwidth]{Fig4}
\end{center}
\caption{The total number $N(t)$ for the created electrons in the different well with different widths given very clearly in the figure. The parameters are $\omega=2.1c^2,\ N_z=2048,\ N_t=10000,\ V_1=V_2=2c^2-10000,\ D=10\lambda_e,\ L=2.0.$ }\label{fig:5}
\end{figure}
In FIG.\ref{fig:5}. All the curves have only one difference, other parameters are the same. The right width of the electric field at the edge of the potential well is stable($W_1=0.3\lambda_e$), and we keep changing the left side from 0.25 time to 5 times of the right side which are namely $W_2=0.075\lambda_e,0.15\lambda_e,0.3\lambda_e,0.6\lambda_e,0.9\lambda_e,1.2\lambda_e,1.5\lambda_e$. The curve $c$ represents to the symmetric potential well, other six curves represent to the asymmetric potential well.
We prefer to take curve $c$ as a comparison with other six curves. As shown in FIG.\ref{fig:5}, when we keep the width of right side of the potential well stable, make the left side wider, the rate of the electron creation would be slower and slower over time, as in curves $d$, $e$, $f$ and $g$, and close to one very specific value as in curve $g$, which approximately equals to the value of a one-sided potential well. The shape of this potential-well can be written as follow:
\begin{equation}
V\left(z,t\right)=V_1S\left(z\right)f\left(t\right)+V_2\sin{\left(\omega t\right)}S\left(z\right)\theta(t;t_0,t_0+t_1),
\end{equation}
where $S\left(z\right)=({1+\tanh[{z/W}]})/2$, it is shown that when one side of the potential well is enough wider, and it doesn't have any effect to the whole system, which we call a critical value of the electric field that can hardly effect to the increasing of the pair production.
We realize that compared with the curve $c$, the yields of the pair production has increased in curve $a$ and curve $b$, as the curve $a$ represents to the yields of the electrons in the potential well with $W_1=0.3\lambda_e,$ $W_2=0.075\lambda_e$ and the curve $b$ represents to the yields of the electrons in the potential well with $W_1=0.3\lambda_e,$ $W_2=0.15\lambda_e$. Surely it is logical due to the intensity of the electric field would be stronger as the width gets narrower. As shown in the FIG.\ref{fig:5}, increase of the yields is evident from $W_2=0.3\lambda_e$ to $W_2=0.15\lambda_e$, and it has very weak effect as we make the $W_2$ narrower, $W_2=0.075\lambda_e$, precisely. That is approximately the maximum value of the yields in our asymmetric potential well.
\section{Summary}\label{sum}
We have compared the momentum spectrum, location distribution and the yields of the pair production in different shape of potential well, in other word, symmetric and asymmetric potential well. As for the momentum spectrum in symmetric case, it had obtained a simple law which is $E_{pi}=E_i+n\omega,$ we find this relationship is also useful for the asymmetric potential well after calculation. It means electrons in the bound states created in the well, escape from their positions after absorbing one or more photons. The momentum distribution of the electrons with positive momentum gets obvious increase than the negative momentum electrons when the left side of potential well becomes narrower. On the contrary, the momentum distribution of the electron with positive momentum gets obvious decrease than the negative momentum electrons when the left side of potential well becomes wider, as expected. The reason for this effects is explicated by the location distribution as shown in FIG.\ref{fig:2}.
For the yields of the pair production, the growth of $N(t)$ gets slower increase when we make one of the width of the potential well wider, because of the electric intensity gets weaker correspondingly, and close to the value of one-sided potential well. The growth of $N(t)$ increases much significantly as the width gets narrower. In general, because of the multi-photon effects, the pair production always keeps increasing in any type of our asymmetric potential well.
\begin{acknowledgments}
\noindent
We would like to thank Binbing Wu and Suo Tang for useful discussions. The work was supported by the National Natural Science Foundation of China (NSFC) under Grants No. 11875007 , No. 11935008 and No. 11965020. The numerical simulation was carried out at CTP office in Xinjiang University.
\end{acknowledgments}
|
\section{Introduction}
The development of chirped pulse amplification \cite{strickland1985compression}
has enabled an array of exotic physics \cite{faure2004laser,liu1997laser,di2012extremely},
including atmospheric filamentation \cite{braun1995self}.
Filamentation in turn produces interesting
and useful effects,
from supercontinuum generation \cite{brodeur1999ultrafast,kandidov2003self},
broadband THz radiation \cite{amico2008forward,andreeva2016ultrabroad}, and RF pulses,
which have recently been explored in great detail
\cite{forestier2010radiofrequency,englesbe2018gas,
janicek2020length,mitrofanov2021coherently,englesbe2021ultrabroadband,
mitrofanov2021polarization}.
We have discovered that a novel combination
of a plasma wake field and a surface wave
is responsible for this filament RF.
Femtosecond lasers are becoming
important spectroscopic tools
(both in the lab \cite{kampfrath2013resonant}
and for remote sensing \cite{labutin2016femtosecond}),
and the added production of pulsed RF
promises to expand the utility of these systems
\cite{park2016perspective,doll2013adiabatic,capmany2007microwave},
and to provide insight
into the physics of filamentation.
During typical $\lambda=$ 800 nm filamentation
Kerr self-focusing and plasma defocusing are in rough
balance, resulting a plasma column with a density of
$n_e \sim 10^{22}$ m${}^{-3}$
\cite{couairon2007femtosecond,chin2010femtosecond}.
This electron density corresponds to a plasma
frequency of $\omega_{pl} \simeq 6 \times 10^{12}$ rad/s,
which is closely linked to
the generation of THz radiation.
In the single color THz theory of \cite{amico2008forward}
the freshly ionized electrons receive a ponderomotive push
from the laser pulse,
thus exciting a coherent longitudinal
current $I_z$ in the plasma (which is quickly
damped by the high collision frequency).
This short current pulse translates at $c$ behind the
laser pulse, thereby producing a conical shell
of radially polarized THz radiation.
The RF resembles the THz radiation in some ways,
and differs critically in others.
It has a similar conical shell spatial profile and radial polarization,
which indicates that a GHz scale longitudinal current pulse
is also generated within the plasma.
However it also has a well defined broadband peak in the 5-15 GHz range
(and is not the tail of the THz radiation
\cite{englesbe2018gas,englesbe2021ultrabroadband,mitrofanov2021coherently})
and the RF amplitude grows strongly with decreasing pressure
\cite{englesbe2018gas,mitrofanov2021coherently},
in contrast to the THz which falls off sharply \cite{rodriguez2010scaling}.
A distinct physical mechanism thus must be responsible for the RF.
Exploratory Particle In Cell (PIC) simulations
\cite{birdsall2004plasma,peterkin2002virtual,
taflove2005computational,yee1966numerical,
boris1970relativistic,villasenor1992rigorous,
berenger1994perfectly}
(see section $1$ in \cite{supplementary_ref_1} for more details)
provided the key insights that explain the RF generation.
They revealed that a hot shell of electrons expands off the plasma surface
into the surrounding atmosphere over roughly 50 ps
(as anticipated by \cite{zhou2011measurement}),
and that the corresponding current density $J_r$ depends strongly
on the electron-neutral collision frequency.
Subsequent simulations demonstrated that this Plasma
Wake field (PW) excites a Surface Wave (SW)
on the outer boundary of
the plasma (namely, a long wavelength
Surface Plasmon Polariton (SPP)\cite{pitarke2006theory,maier2006terahertz},
with a speed approaching $c$).
Finally the SW was found to effectively detach from the end of the plasma,
with most of the energy being then converted into a broadband RF pulse.
The PWSW numerical model is split into 4 stages,
as length scales spanning 7 orders of magnitude need to be resolved.
First the initial electron distribution function
$\mathcal{N}_e$ is determined.
In a typical $\lambda = 800$ nm experiment
a linearly polarized, single color, 40 mJ
pulse is focused into a quartz cell with a ${f}/60$ mirror,
which sets the beam waist $\omega_0$ to 30 $\mu$m
(10 mJ and ${f}/15$ focusing are common for $\lambda =$ 3.9 $\mu$m).
The resulting Keldysh parameter
$\gamma = \omega_{laser} \sqrt{2 m_e I_{pot}}/(q_e E)$
is close to 1, so we use the general ionization rate $\scriptstyle\mathcal{W}$
from \cite{popruzhenko2008strong},
which is supported by experimental data \cite{sharma2018measurements}
near atmospheric pressure
(other ionization mechanisms will likely be needed
for 10 $\mu$m lasers \cite{tochitsky2019megafilament}).
A laser intensity of $I \sim 5\times10^{17}$ W/m${}^2$
generates enough plasma at 1 atm for intensity clamping \cite{couairon2007femtosecond}
to occur (with $\scriptstyle\mathcal{W}$ $\sim$ $10^{12}$ s${}^{-1}$),
which leads to an estimated plasma length $L_{pl}$ of $25$ cm,
and a radius $r_{pl}$ of $\sim$ $1$ mm
(see section 2 of \cite{supplementary_ref_1} for more details).
At 1 Torr $I$ increases to $\sim 2\times10^{18}$ W/m${}^2$,
which ionizes much of the O${}_2$, and
the plasma is reduced to $L_{pl} \simeq 13$ cm
and $r_{pl} \simeq 0.5$ mm.
In turn, the lower energy 3.9 $\mu$m laser plasmas
are shorter, with $L_{pl}$ ranging from about $1$ to $2$ cm.
These estimates match well (within 25\%) with photographic
measurements of the plasma taken
at different pressures (available in
section 2 of \cite{supplementary_ref_1}).
Simulations of filamentation with these parameters have also been
performed using a Unidirectional Pulse Propagation Equations (UPPE)
solver \cite{kolesik2004nonlinear}.
These agree with the overall dimensions of the
plasma column, but also indicate that there is rich fine scale
structure throughout the plasma
(as expected as we are in the multifilamentary regime
\cite{berge2004multiple,ettoumi2015laser,ettoumi2015spin}).
For this letter we simplify and do not include
fine scale plasma perturbations.
The plasma column is approximated as having
a constant density central core
due to intensity clamping,
and then a 100 $\mu$m radial ramp down to
zero density at the outer radius $r_{pl}$.
The velocity distribution is determined by our PIC code:
given $E(\vec{x},t)$ the air is ionized
at the $\scriptstyle\mathcal{W}$ rate \cite{popruzhenko2008strong}
and the new electrons
are accelerated through the remainder of the pulse
\cite{boris1970relativistic}
(code to perform these calculations is included
in \cite{supplementary_ref_1}).
In general the velocity distribution
is influenced by both the details of strong field
ionization near $\gamma = 1$ (e.g. \cite{corkum1993plasma}),
and by the deformation of the laser pulse
during filamentation.
For this work we further simplify and assume that
the electrons are ionized with zero initial velocity and
are then accelerated by the remainder
of a Gaussian pulse (with $\hat{x}$ polarization
and propagating in $+z$).
As a whole the initial $\mathcal{N}_e$ is highly non-Maxwellian,
and at 100 Torr has a peak kinetic energy $K_{tail} \simeq$ 5 eV,
and an average kinetic energy $K_{avg} \simeq 0.6$ eV,
while these increase at 1 Torr to $K_{tail} \simeq 16$ eV and $K_{avg} \simeq 2$ eV.
For 3.9 $\mu$m lasers the kinetic energy is $\sim 25$ times larger,
as the laser intensities are
comparable and the energy scales as $\lambda^2$.
We next consider the evolution of the plasma column.
We construct thin transverse slices of plasma
given $\mathcal{N}_e$,
with periodic boundary conditions used in the
longitudinal direction $\hat{z}$
(which is valid to leading order as the electron velocities
are a small fraction of $c$)
and simulate the radial evolution with our PIC code.
The Debye length is quite small: $\lambda_{Debye} \simeq 10$ nm,
so we use an energy conserving method \cite{pointon2008second}
to calculate the Lorentz force.
The electron-neutral elastic collision frequency
$\nu_{eN}$ depends on the cross sections for O${}_2$ and N${}_2$,
which for our energies is roughly 10 \AA${}^2$ \cite{itikawa2006cross}.
In turn the electron-ion momentum-transfer collision frequency is
given by $\nu_{ei} = 7.7{\times}10^{-12} n_e \ln ( \Lambda_C ) / K_{eV}^{3/2} $,
where $\Lambda_C = 6 \pi n_e \lambda_{Debye}^3$ \cite{thorne2017modern}.
The resulting radial current density
$J_r$ and electron density $n_e$
are then recorded as functions of radius and time
(more detail can be found in section 3 of \cite{supplementary_ref_1}).
These are well resolved with a mesh resolution
of $\Delta x = \Delta y =$ 2 $\mu$m,
and a macro-particle weight of
${\sim}10$ on the outer edge of the plasma.
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{log_ne_v5.pdf}
\caption{Log plots of the electron density $n_e(r,t)$
showing the plasma wake variation at pressures
of 100 (a), 10 (b), and 1 Torr (c)
for a $\lambda = 800$ nm laser pulse.
The plasma outer radius has been set to $r_{pl}=0.5$ mm
for all 3 pressures for comparison purposes.}
\label{log_ne_plots}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{jr_19_20_combo_v3.pdf}
\caption{Log plots of the radial current density $J_r(r,t)$
at 1 Torr (800 nm source),
as it evolves in a transverse slice of plasma.
The initial density at the edge of the plasma has been given
a step function profile, with a value of $n_e = 10^{19}$ m${}^{-3}$
in the top plot and $n_e = 10^{20}$ m${}^{-3}$ on the bottom.
The profiles of the plasma wakes outside the column ($r > r_{pl} = 0.5$ mm)
are quite similar, while at the surface
SPPs are excited at roughly
${{\sim}}20$ and ${{\sim}}60$ GHz respectively, along with
higher frequency volume plasmons in the interior.
}
\label{log_jr_plots}
\end{figure}
The electron number density for $\lambda=800$ nm simulations of the
PW at 100, 10, and 1 Torr can be seen in Fig.~\ref{log_ne_plots}.
The outer edge of the plasma at $t=0$ has a
simplified step function profile,
with $n_e = 10^{20}$ m${}^{-3}$
at radius $r_{pl} = 0.5$ mm.
Thus, in addition to the PW being launched off of the edge of this plasma,
a coherent radial plasma frequency oscillation of ${\sim} 90$ GHz
is stimulated in the interior \cite{dawson1959nonlinear} and a SPP
at ${\sim} 63$ GHz is excited on the surface
\cite{pitarke2006theory,maier2006terahertz,sakai2018wave}.
The PW that expands out into the neutral atmosphere ($r > r_{pl}$)
is insensitive to the density
of the outer edge of the plasma (see Fig.~\ref{log_jr_plots}),
as opposed to the excited surface and volume plasmons.
Analytic approximations provide a useful complement to the simulations.
Consider the late time electron density at a small
distance $r_{\Delta}$ off the surface of the plasma at radius $r_{pl}$.
Using Gauss's law and approximating $\ln( ( r_{pl} + r_{\Delta} ) / r_{pl} )$
as $r_{\Delta} / r_{pl}$ for $r_{\Delta} \ll r_{pl}$
we find the that the number of electrons per unit
length $n_L$ that escape to $r_{\Delta}$ scales as:
\begin{equation}
n_L(r_{\Delta}) \simeq \frac{2 \pi \epsilon_0 K_{eV} }{q_e} \frac{r_{pl}}{r_{\Delta}},
\label{nL_eq}
\end{equation}
where $K_{eV} \simeq K_{tail}$ is the kinetic energy
of the hot electrons that lead the escape.
The radial electrostatic field $E_{r,stat}$ that corresponds to $n_L$ has a
simple form: $E_{r,stat}(r_{\Delta}) = K_{eV} / r_{\Delta}$.
At $\lambda=800$ nm and 1 Torr
$K_{tail} \simeq 16$ eV and $r_{\Delta} \simeq$ 60 $\mu$m,
giving $E_{r,stat} \simeq 2.7 \times 10^5$ V/m.
This allows for an estimate of the PW
evolution timescale:
an electron with energy $K_{tail} = 16$ eV
in an electric field of this magnitude
follows a parabolic trajectory over 100 ps.
This compares well with the spread
of electrons seen in Fig.~\ref{log_ne_plots}.
The magnitude of the radial current density can likewise be approximated as:
\begin{equation}
J_r \simeq v_{eff}(t) \frac{q_e n_L}{2 \pi r_{pl} r_{\Delta} }; v_{eff}(t) = \min \begin{cases} v_{tail} \\ v_{diff}(t) \end{cases}
\label{Jr_eq}
\end{equation}
where the effective velocity $v_{eff}(t)$ is the minimum of the tail
velocity and the Fickian diffusion speed
$v_{diff}(t) = (K_{tail} /(m_e \nu t))^{1/2}$.
At 1 Torr this gives $J_r \simeq 1\times10^5$ A/m${}^{2}$.
Finally we note that caution is needed for
small $r_{\Delta}$ values: Eq.s \eqref{nL_eq} and \eqref{Jr_eq}
only hold up to densities that are comparable to the original plasma
edge density.
We next switch from these transverse PIC simulations
to a continuum FDTD-Drude model in a 2D axisymmetric coordinate system
\cite{taflove2005computational} that spans the length of the plasma.
The Drude model:
$\partial_t \vec{J}_{pl} + \nu_{pl} \vec{J}_{pl} = \epsilon_0 \omega_{pl}^2 \vec{E}$
is integrated into the code
via an auxiliary differential equation \cite{okoniewski1997simple}.
A mixed resolution of $\Delta r =$ 2 $\mu$m and $\Delta z =$ 50 $\mu$m
suffices to resolve the SWs.
The PW current profile $J_r$ is driven across
the surface of the plasma column
in the longitudinal direction at the speed of light.
In this work only the external radial currents $J_{r > r_{pl}}$
are used to drive the SW and subsequent RF
(see section 4 of \cite{supplementary_ref_1} for more detail).
The higher frequency volume plasmons can also
excite waves in neighboring lower density plasma, but we
suspect that including fine scale plasma structure
will act as surface roughness \cite{kretschmann1972angular}
and cause these to continuously detach.
As expected the PW current
excites a broadband surface wave with similar frequency content.
The frequency $\omega$ of the wave is considerably lower than
the limit SPP frequency
and it thus travels at velocity
$c ((2 \omega^2 - \omega_{pl}^2)/(\omega^2 - \omega_{pl}^2))^{1/2}$
which approaches the speed of light.
The resulting phase-matched copropagation leads to
a steady growth in the surface wave intensity
(see also \cite{smith1953visible,dahan2020resonant}).
The SW is well approximated by the Sommerfeld-Goubau
\cite{sommerfeld1899ueber,goubau1950surface}
solution for SPPs on a cylinder of finite conductivity.
The external radial component of the electric field $E_r$ has the form:
\begin{equation}
E_{r}( r, z, t ) = -\frac{\pi r_{pl} E_0 }{2 r_{outer}} e^{i(\omega t - h z)} H^{(1)}_1 (r \sqrt{k^2 - h^2} ) ,
\label{SG_SW}
\end{equation}
for waves of frequency $\omega$ propagating in the $+z$ direction,
with an amplitude of $E_0$ at the surface.
The complex $h$ wavenumber describes the SW wavelength and
attenuation length scale,
$k$ is the free space wavenumber $k=\omega / c$,
$H^{(1)}_1$ is a Hankel
function of the first kind with order 1,
and $r_{outer} = 1/|\sqrt{k^2 - h^2}|$.
For radii $r_{pl} < r < r_{outer}$
the Hankel function with complex argument
is approximately $H^{(1)}_1 \simeq -2 r_{outer}/(\pi r)$,
and for larger radii it falls off exponentially
(as is typical for SPPs).
Recent work \cite{jaisson2014simple,mendoncca2019electromagnetic}
has simplified the calculation of $h$
through the use of the Lambert $W$ function.
Given a typical plasma conductivity $\sigma$ a
complex $\xi$ can be defined:
\begin{equation}
\xi = -\frac{k e^{-2 \gamma_{euler}}}{2} \sqrt{ \frac{ \epsilon_0 \omega }{ 2 \sigma } } r_{pl} (1+i),
\end{equation}
(where $\gamma_{euler}$ is the Euler-Mascheroni constant),
and $h$ is then given by:
\begin{equation}
h = \sqrt{ k^2 + \frac{4 \xi e^{-2 \gamma_{euler}}}{r_{pl}^2 W_1(\xi)} },
\end{equation}
where the first branch of $W$ is used.
At 1 Torr the plasma has a radius $r_{pl} \simeq 0.5$ mm
and conductivity
$\sigma \simeq 2 \times 10^3$ S/m, and thus for
a 10 GHz SW we find $h \simeq 210 + 3.2i$.
The dispersive SW thus travels at about $0.988 c$,
attenuates by a factor of $1/e$ every $30$ cm
(if not being actively excited),
and has a $1/r$ transverse envelope out to
$r_{outer} \simeq 2.4$ cm,
all of which compares well with simulations.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{SW_3_9_micron_v2.pdf} \\
\caption{Log plot of electric field strength shortly after the detachment of a $\lambda=3.9$ $\mu$m, 10 Torr surface wave from a 2 cm plasma column, as
simulated by the axisymmetric FDTD-Drude model.}
\label{SW_39}
\end{figure}
We next consider the excitation of the surface wave
by the plasma wake currents.
An isolated Hertzian dipole with uniform current $I$
along an antenna of length $L$ will
radiate electromagnetic waves at a wavelength $\lambda_{Dipole}$
with power:
$P_{Dipole} = (\pi\eta_0/3) I^2 L^2 / \lambda^2_{Dipole}$.
In our case we have a radial current distribution
of approximate length $r_{\Delta}$
that is normal to a conducting cylinder of radius $r_{pl}$.
When there is a separation of length scales
$r_{\Delta} \ll r_{pl} \ll \lambda_{SW}$
and an overall radial current of $I_r$ then the
power that is primarily radiated into SWs is:
\begin{equation}
P_{SW} \simeq \frac{\pi \eta_0}{60} I_r^2 \frac{r_{\Delta}^2}{r_{pl}^{5/3} \lambda^{1/3}_{SW}},
\label{power_SW}
\end{equation}
(where we have simplified with $\lambda_{SW} = 2\pi/\operatorname{\mathbb{R}e}(h)$).
We note that the dimensionless term
$r_{\Delta}^2/(r_{pl}^{5/3} \lambda^{1/3}_{SW})$
(compare to $L^2 / \lambda^2_{Dipole}$ for the Hertzian dipole),
has different asymptotic scalings in other length scale regimes.
Equation \eqref{power_SW} applies both to a stationary
radial antenna and the $\pm z$ wavetrains it emits, and to our case,
where the antenna translates at $c$ with its comoving SW pulse.
The total effective current $I_r(z)$
is found by integrating
the current density $J_r$ from 0 to $z$,
and absent dissipation the amplitude $E_0$
of the SW in \eqref{SG_SW} would grow linearly (for constant $J_r$).
However there is dissipation due to the finite
conductivity of the plasma,
and the attenuation coefficient $\alpha = \operatorname{\mathbb{I}m}(h)$
is similar in magnitude to $1/L_{pl}$.
The SW thus grows linearly initially
and then saturates after several $1/\alpha$, and the effective
total radial current after propagating a distance $z$ is:
\begin{equation}
I_r(z) = 2 \pi r_{pl} J_r \frac{1 - e^{-\alpha z}}{\alpha}.
\label{Ir_integrated}
\end{equation}
We express the amplitude of the final SW in terms of
its longitudinal current, as $I_z = 2 \pi r_{pl} E_0 / \eta_0$.
Integrating \eqref{SG_SW} to find the SW peak power
and setting this equal to \eqref{power_SW} we find:
\begin{equation}
I_z(z) = \frac{2 \pi^2 r_{\Delta} J_r (1-e^{-\alpha z})}{\alpha(30 \ln (r_{outer}/r_{pl}))^{1/2}}
\left( \frac{r_{pl}}{\lambda_{SW}} \right)^{1/6}.
\label{final_Iz}
\end{equation}
For the $\lambda=800$ nm laser this gives
$I_z \simeq 0.6$ A at the end of the
plasma at 1 Torr, a current of $I_z \simeq 0.2$ A at 10 Torr,
and $I_z \simeq 0.05$ A at 100 Torr, which are close to the
measured currents in the corresponding simulations.
When the surface wave reaches the end of plasma the majority
of the energy detaches and propagates forward as the RF pulse,
with the reflected wave decaying as it travels back down the plasma.
Fig.~\ref{SW_39} shows an example of this,
where the SW has just
detached from a 2 cm plasma column produced by a
$\lambda=$ 3.9 $\mu$m laser source at 10 Torr.
In this way the system resembles a surface wave
end-fire antenna \cite{andersen1967radiation,sporer201724}.
The second half of the primarily one cycle RF pulse forms as the
SW separates from the plasma, with some variation in the profile
dependent on the abruptness of the detachment.
The recorded $I_z(z,t)$ currents from the SW simulation
are used in a ${\sim}1$ m${}^2$
laboratory scale axisymmetric FDTD model,
with resolution $\Delta r = \Delta z = 0.5$ mm.
Plots of the resulting Fourier transformed RF field strengths
at 10 Torr for
both $\lambda = 800$ nm and $\lambda = 3.9$ $\mu$m
simulations can be seen in Fig.~\ref{Far_RF_1}
(see section 5 of \cite{supplementary_ref_1} for more detail).
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{fft_800_39_v2.pdf}
\caption{Plots of simulated electric field strength as a function of frequency
and angle for RF pulses from a 40 mJ, f${}_N=60$, 800 nm laser source
(left, \cite{englesbe2018gas}) and a 12 mJ, f${}_N=15$, 3.9 $\mu$m source
(right, \cite{mitrofanov2021coherently}), with both cases run at 10 Torr.
The numerical profiles and amplitudes
closely match the experimental results.
}
\label{Far_RF_1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.5\textwidth]{combo_polar_v4.pdf}
\caption{Polar plots of the detector peak to peak Voltage $V_{pp}$
for the $\lambda=800$ nm laser source
as a function of angle and pressure as
measured in the lab (left), and as calculated
by both the simulation framework (right, triangles),
and by the associated analytic approximations
(right, $\times$ ticks) that culminate in \eqref{THz_eqn}.}
\label{Vpp_compare}
\end{figure}
To directly compare with lab results the simulated
electric fields are converted to peak
to peak voltages $V_{pp}$ via a frequency
dependent calibration factor
$C_{E/V}(f_{GHz}) \simeq 50 + 10 f_{GHz}$
(which captures the horn antenna response \cite{englesbe2018gas}).
We adapt an equation for energy spectral density
radiated per unit solid angle derived
for THz radiation \cite{amico2008forward}:
\begin{equation}
\frac{d^2 U}{d\omega d\Omega} = \frac{|I_z(\omega)|^2 \sin^2 \theta }{4 \pi \epsilon_0 c (1-\cos \theta)^2}
\sin^2 \left( \frac{L_{pl} \omega}{2c} (1-\cos \theta) \right),
\label{THz_eqn}
\end{equation}
to construct voltage profiles for our analytic approximations.
The average amplitude of $I_z$ \eqref{final_Iz}
over the length of the plasma is applied to a ${\sim}10$ GHz pulse,
which is Fourier transformed for use with \eqref{THz_eqn}
(along with the calibration factor), and then
converted back to an effective $V_{pp}$
(code that implements all of these
analytic approximations is available in \cite{supplementary_ref_1}).
In general simulations with varying pressures,
focal geometries, and
laser pulse energies and wavelengths
all produce RF that nicely matches recent
experimental results
\cite{englesbe2018gas,janicek2020length,
mitrofanov2021coherently,englesbe2021ultrabroadband}.
The magnitude of the simulated RF is a bit lower than the
lab data for the runs shown in Fig.~\ref{Vpp_compare},
but is within the errors for both the
simulations and in the experimental data.
Subtle changes in Ti:Sapphire laser alignment
generates $\pm 50\%$ amplitude variations at a given pressure
(with the spatial and frequency content being much more stable).
It is suspected that the slight laser pulse variations
are amplified during filamentation,
leading to fine scale changes in the plasma radius,
density, and electron velocities.
Plausible variations in the same parameters lead to
comparable changes in the simulated RF amplitude as well.
We have demonstrated that a plasma wake field that
excites a surface wave is responsible
for the generation of filament RF.
The rich internal dynamics of the plasma
merits further research: we expect that
stochastic higher frequency RF
may inform on the fine scale structure of the plasma and the
nonlinear optics that generate it.
The ionization model needs to be expanded to include more physical effects
(e.g. \cite{tochitsky2019megafilament,corkum1993plasma,amini2019symphony}),
and strong field ionization can be further explored with noble gas
filament RF.
The plasma can also be driven into the
strongly coupled regime with harder focusing,
and we intend to model resulting RF
with the use of \cite{baalrud2013effective,donko2009molecular}.
Finally we note that it may be possible
to recreate the core mechanism
-- an antenna that copropagates with the wave that it excites --
in other settings and thereby
drive strong coherent radiation.
\begin{acknowledgments}
\vspace{4mm}
The authors thank Edward Ruden, Serge Kalmykov and Charles Armstrong
for useful discussions,
and the Air Force Office of Scientific Research (AFOSR) for support
via Laboratory Tasks No. FA9550-18RDCOR0009 and No. FA9550-19RDCOR027.
This work was supported in part by high-performance computer time
and resources from the DoD High Performance Computing Modernization Program.
Approved for public release; distribution is unlimited. Public Affairs
release approval No. AFRL-2020-0489.
\end{acknowledgments}
|
\section{Introduction} \label{sec_intro}
While the Norma cluster (ACO\,3627) was identified and catalogued by \citet{Abell_89}, this cluster's massive nature and dynamical role in the Great Attractor (GA) region were not realised until the mid 1990s \citep{Kraan_96,Boehringer_96, Tamura_98}. The Norma cluster has since been identified by \citet{Woudt_08} as the richest cluster in the GA region. Studies in this region are affected by high levels of Galactic extinction and star-crowding due to its location at low Galactic latitudes. Some large scale structures are therefore expected to still remain hidden within this zone of obscuration \citep{Sorce_17, Courtois_19a} as evidenced by the recent discovery of the Vela supercluster \citep{Kraan_2017a}. Currently, there is significant improvement in our understanding of the large scale structures residing in the zone of avoidance through observations at near-infrared wavelengths where the effect of Galactic extinction significantly reduces \citep[e.g.][]{Macri_19}. Using telescopes with superior resolution such as the IRSF and NTT minimises the effects of high stellar density (star-crowding) in this region \citep[see e.g.][]{Said_016}. In addition, 21-cm radio observations \citep{Said_2016, Staveley_16, Kraan_18, Schroder_19} have revealed the abundance and massive nature of galaxies in this region. Such {\scshape{Hi}} observations are not affected by Galactic extinction but may suffer from radio contamination from neighbouring radio sources.
The Local Group has a well known peculiar velocity ($v_{_{\text{LG}}} = 627 \pm 22$~km\;s$^{-1}$) with respect to the CMB rest frame in the direction $\ell = 276^\circ \pm 3^\circ$, $b = 30^\circ \pm 3^\circ$ \citep{Kogut_93} --- which is in the direction of Shapley \citep{Shapley_30, Raychaudhury_89, Scaramella_89} and Vela \citep{Kraan_2017a} superclusters. %
Despite more than two decades of study, the fractional contribution of the GA and Shapley supercluster on the Local Group's peculiar velocity
remains a topic of debate \citep[see e.g. ][]{Smith_00, Lucey_05, Erdogdu_06, Kocevski_06, Lavaux_10, Carrick_15}.
In the analysis by \citet{Lavaux_10}, they recovered the majority of the amplitude of the Local Group's peculiar velocity within 120$h^{-1}$\;Mpc but the direction of motion did not agree with the expected direction of the CMB dipole --- they suggested convergence may lie as far as 200$h^{-1}$\;Mpc which is supported by recent measurements using the 2M++ sample \citep{Carrick_15}.
While some authors found small contributions from the Shapley Supercluster (e.g., \citealt{Bardelli_00, Bolejko_08}), significant contributions from distant structures have been supported by e.g. \citet{Kocevski_06, Springob_14, Springob_16, Hoffman_17, Qin_2018, Tully_19, Said_20}. %
Although the zone of avoidance or zone of obscuration has received less attention, structures hidden within or behind this region may play a role in the general flow and may significantly contribute to the Local Group's motion.
An example is the recent discovery of an extended supercluster of galaxies at $cz \sim 18 000$\;km\;s$^{-1}$, located at Galactic longitudes and latitudes $(\ell, b) \sim (272^\circ, 0^\circ)$ \citep{Kraan_2017a}. This previously hidden Vela supercluster was later kinematically confirmed by \citet{Courtois_19a}. %
Using Cosmicflows-3 which is a compilation of distances for $\sim$\,18000 galaxies \citep{Tully_16}, \citet{Courtois_17} found dominant features (repulsive and attractive), the major basin of attraction corresponding to an extension of the Shapley supercluster. This is in agreement with the previous findings based on the 6dF Galaxy Peculiar Velocity Survey
where positive line-of-sight peculiar velocities were estimated, suggesting substantial mass overdensities originating at distances further away than the Great Attractor \citep{Springob_16}. %
There is currently increasing support for flow towards more distant structures such as Vela and Shapley superclusters \citep[see e.g. ][]{Kocevski_06, Springob_14, Staveley_16, Hoffman_17, Tully_19}. %
Located at \mbox{($\ell$, $b$, $v_{\rm{hel}}$)\,$=$\,(325$^\circ$, --7$^\circ$, \,$4871\pm 54$\,km\;s$^{-1}$)} --- see for example, \citet{Woudt_08}, the Norma cluster lies close to the anticipated centre of the GA \citep{Kolatt_95}. As such, a reliable estimate of the redshift-independent distance (and hence the peculiar velocity) of the Norma cluster is crucial to understanding the GA flow. Our previous measurements \citep{Mutabazi_14} show the Norma cluster to have a small peculiar velocity, consistent with zero (i.e.\ $v_{_{\rm{pec}}}$\,$=$\,$-$\,72\,$\pm$\,170\,km\;s$^{-1}$).
In order to improve the Fundamental Plane (FP) distance of Norma, we have made new $J$ and $K_s$-band photometric measurements using near-infrared (NIR) images from the 1.4\,m Japanese InfraRed Survey Facility (IRSF) at the Sutherland site of the South African Astronomical Observatory. We improve on our previous results by studying the effect of Galactic extinction using the $J-K_s$ colour for these new data. In addition, we have extended/improved our previous $K_s$-band FP analysis from $N$\,$=$\,31 to a total sample of 41 Norma cluster early-type galaxies (ETGs) selected from within $\frac{2}{3}R_{\rm A}$ where $R_{\rm A}$ is the Norma cluster's Abell radius.
In this paper, we therefore present (1) distance and peculiar velocity for the Norma cluster using the $J$ and $K_s$-band FP \citep{Djorgovski_87, Dressler_87_fp} analysis for 31 ETGs whose NIR images were obtained using IRSF (2) a combined $K_s$-band FP analysis for a total of 41 ETGs whose NIR images were obtained using the IRSF and NTT telescopes \citep[see][for telescope details]{Mutabazi_14}. For ETGs that were observed using both the IRSF and NTT, the NTT data was given priority due to its superior resolution and good seeing conditions.
Note however that both datasets have high quality, highly resolved data (NTT and IRSF have pixel scales of 0.29 and 0.45 arcsec per pixel, respectively) which are very well suited for this study.
This paper is structured as follows: Sample selection and observations for our FP sample is discussed in $\S$\;\!\ref{sample_data} while the photometric data and analysis are presented and discussed in $\S$\;\!\ref{data_analysis}. A re-calibration of the NIR reddening maps is presented in $\S$\;\!\ref{extinc_recalib}. %
In $\S$\;\!\ref{offset}, we present the measured zero-point offset using the FP relation. %
The distance and peculiar velocity measurements are presented in $\S$\;\!\ref{dist_vpec_fp}. The discussion and conclusions of our findings are presented under $\S$\;\!\ref{sec_discuss}.
Unless stated, we have adopted standard cosmology with $\Omega_\text{m}$\,$=$\,$0.27$, \mbox{$\Omega_{\Lambda}$\,$=$\,$0.73$} and \mbox{$\text H_0$\,$=$\,$70.5\; \rm{km} \; \rm s^{-1} \!\;\rm{Mpc}^{-1}$} \citep{Hinshaw_09}. While measuring the distance of the Norma cluster relative to the Coma cluster, we used \mbox{$z_{_{\rm{CMB}}} = 0.02400 \pm 0.00016$} for the Coma cluster \citep{Hudson_04}.
\section{Observations and data reduction} \label{sample_data}
\subsection{Sample selection}
The ETGs used in the Norma cluster sample were such that (1) 2096\,km\;s$^{-1}$~$<v_{\rm hel}<$~7646\,km\;s$^{-1}$ where $v_{\rm hel}$ is the
galaxy's heliocentric velocity (2) lie within $\frac{2}{3}$ of the Abell radius,
and (3) we had successfully measured the central velocity dispersion using the 2dF spectrograph.
The above selection criteria resulted in a total of 31 ETGs with $J$ and $K_s$-band images from the IRSF telescope. Ten of these 31 ETGs with IRSF images are unique to the IRSF while 21 ETGs have NIR images from both the IRSF and NTT.
For the Coma cluster sample, galaxies selected were those typed as E or E/S0 or S0 by \citet{Dressler_80}, were identified and confirmed as Coma cluster members based on their redshifts, and had a central velocity dispersion in the Sloan Digital Sky Survey Data Release~8 \citep{Aihara_11}. In addition, only those which we successfully measured the FP photometry parameters from the 2MASS Atlas images were selected. This resulted in a Coma cluster sample of 121 ETGs.
\subsection{Observations with the IRSF}
The IRSF telescope is equipped with SIRIUS, a three-colour band camera with three 1024$\times$1024 pixels HgCdTe infrared detectors and is capable of simultaneous imaging in the $J$, $H$ and $K_s$ passbands \citep{Nagashima_99}.
Our IRSF observations were carried out using a total integration time of 600\,s split over 20 short exposures of 30\,s each. Standard NIR data reduction procedures (such as dark subtraction, flat-fielding, sky subtraction) were applied before combining the 20 dither frames into one science image. Note that both the NTT and IRSF telescopes are suitable for the high stellar density GA region since they are equipped
with the SOFI and SIRIUS imaging instruments which have a low pixel scale, that is, 0.29\;arcsec per pixel and 0.45\;arcsec per pixel, respectively. It is also important to note that the combined $K_s$-band data (IRSF+NTT) is the largest cluster sample ($N$\,$=$\,$41$) used for peculiar velocity studies in the GA region to date.
To ensure that all our photometry data sets are on the same photometric system, the astrometric and photometric calibrations were performed using the 2MASS Point Source Catalogue \citep{Skrutskie_06}. During the photometric calibration process, we applied a small correction due to the slight difference between the 2MASS, NTT and IRSF filters (see e.g., \citealt{Carpenter_01}). %
\section{Photometry data analysis} \label{data_analysis}
In order to apply the FP relation, measurements of the effective radii and mean effective surface brightnesses are required. These were measured through galaxy surface brightness profile fitting.
\subsection{Galaxy surface brightness profile fitting} \label{ellipses}
For accurate photometry, stars in each galaxy field were subtracted after which a reliable estimate of the sky background was calculated using an annulus as described in \citet{Mutabazi_14}. We performed the photometric analysis (for both Norma and Coma galaxies) using the \textsc{Ellipse}\ \citep{Jedrzejewski_87} task in \textsc{Iraf}. %
The resulting surface brightness profiles were fitted using a combination of two S\'{e}rsic functions. The best fit was extrapolated to infinity so as to estimate the total extrapolated magnitude. The effective radius was then measured through interpolation. To correct for the seeing (\textsc{Psf}) effects, the \textsc{Galfit}\ algorithm \citep{Peng_10} was used with and without a \textsc{Psf}\ convolution.
Figure~\ref{extrapolate:singDsersics_Nfit1d} is an example of the fitted profiles for WKK\,6250, one of the Norma cluster's galaxies from the IRSF data. The top panel is the $J$-band while the bottom panel represents the $K_s$-band.
The fit data were restricted to within the radius ranging from twice the seeing to where the galaxy flux averages to 1$\sigma_s$ above the sky background.
These are represented by the two small magenta arrows. The black solid line is the best fit which is a combination of the two S\'{e}rsic component fits represented by the red and blue dashed lines. The inset shows the residuals.
\begin{figure}[htb!]
\centering
\begin{tabular}{c}
\includegraphics[width=0.45\textwidth]{WKK6250_J.eps} \\
~\\~\\
\includegraphics[width=0.45\textwidth]{WKK6250_K.eps}
\end{tabular}
\caption{\label{extrapolate:singDsersics_Nfit1d} Double S\'{e}rsic component fits to the galaxy surface brightness profile for WKK\,6250 in the $J$ (top panel) and $K_s$-band (bottom panel). The red and blue dashed lines represent the individual S\'{e}rsic components while the red solid line represents the combined fit (best fit to the galaxy surface brightness profile). The small magenta arrows represent twice the seeing (FWHM) and the radius where the galaxy surface brightness reaches the sky background noise level. Only the data within the range indicated by these two magenta arrows were used to fit the galaxy surface brightness profile.
}
\end{figure}
\subsection{Photometry comparisons} \label{systematics}
As a photometric reliability check, we took advantage of the high quality, high resolution images provided by both the NTT (pixel scale of 0.29 arcsec per pixel) and IRSF (pixel scale of 0.45 arcsec per pixel) telescopes. We compared the $K_s$-band total extrapolated magnitudes for an overlapping sample of 21 ETGs (ETGs which are common to both the IRSF and NTT data). The comparison shows excellent agreement as seen in Figure \ref{phot-comp-ntt-irsf}, i.e., an average difference in total extrapolated magnitudes of $0.000\pm 0.021$\,mag although with a relatively large scatter of 0.095 mag. %
\begin{figure}[htb!]
\centering
\includegraphics[width=0.43\textwidth]{Mag_IRSFF+NTT.eps}
\caption{\label{phot-comp-ntt-irsf}Comparison between the total extrapolated magnitudes for the 21 ETGs observed using both the NTT and IRSF telescopes. The red solid line shows a one-to-one relation. Both the average magnitude and rms were computed based on this line.}
\end{figure}
This scatter is dominated by WKK\,6204 with a very uncertain IRSF $K_s$-band magnitude. When this galaxy (WKK\,6204) is excluded from the comparison between the IRSF and NTT magnitudes, the mean difference is $0.0125\pm 0.0180$ mag and the rms reduces from 0.095 to 0.080. However, the effect on the IRSF Ks-band Fundamental Plane analysis is not very significant -- the rms from the FP fit decreases slightly from 0.0778 to 0.0776 when WKK\,6204 is included and excluded, respectively. The source of the large uncertainty on the IRSF $K_s$-band magnitude for this galaxy (WKK\,6204) is not very clear since the error/uncertainty in the IRSF $J$-band magnitude is significantly lower than the IRSF $K_s$-band magnitude for this same galaxy (0.07 mag in $J$ and 0.28 mag in $K_s$-band). It should however be noted that there were sky gradients in the observed field for this galaxy using IRSF as compared to NTT possibly due to a superior resolution and better seeing conditions associated with NTT. These sky gradients are most likely due to a nearby bright galaxy in the field of WKK\,6204. Therefore, the large uncertainty on the IRSF magnitude for WKK\,6204 could be due to the error associated with the measured sky background in the $K_s$-band as a result of this nearby galaxy. The effect may be more pronounced in the $K_s$ image since the nearby galaxy is brighter in the $K_s$ than in the $J$-band. %
Note that the central velocity dispersions used in our analysis were taken from the Sloan Digital Sky Survey Data Release 8 \citep{Aihara_11} for the Coma cluster sample. For the Norma cluster sample, data using the fibre spectroscopy on the 2dF facility \citep{Lewis_02} on the 3.9\,m Anglo-Australian Telescope were used.
\subsection{Coma cluster sample -- Photometric analysis}
For the Coma cluster, 2MASS Atlas images were used in the photometric analysis. For photometric quality check, we have compared our measurements and the 2MASS~XSC for the 121 Coma ETGs. The mean differences \mbox{(2MASS -- This work)} between our measured total extrapolated magnitudes and the values taken from the 2MASS~XSC \citep{Skrutskie_06} are $-$\,0.003\,$\pm$\,0.014~mag and $-$\,0.015\,$\pm$\,0.009~mag in the $J$ and $K_s$-band, respectively. The photometric results and the corrected FP variables for the Coma sample ($N$\,$=$\,121) are presented in
Table~\ref{Coma_JKs_phot}. %
\subsection{Galactic extinction: re-calibration using NIR colours} \label{extinc_recalib}
As Norma is at Galactic latitude, $b=-7$, the Galactic extinction is relatively large. Note that for the Coma cluster, the effect of Galactic extinction is very small, i.e., $\sim$\,0.007~mag and 0.003~mag in the $J$ and $K_s$-band, respectively. %
The \citet{Schlegel_98} NIR reddening maps have been found to over-estimate the Galactic extinction especially at low Galactic latitudes \citep{Bonifacio_00, Yasuda_07, Schroder_07, Schlafly_11}. A re-calibration of the NIR reddening maps based on the \mbox{$J-K_s$} colour for the Norma and Coma cluster ETGs used in our FP analysis is presented here. This closely follows the re-calibration by \citet{Schroder_07}.
The extinction in the $J$ and $K_s$ bands, assuming the \citet{Fitzpatrick_99} reddening law, with $R_{\rm v}=3.1$ is given by
\begin{equation}
A_J = 0.937 E(B-V), \qquad A_{K_s} = 0.382 E(B-V), \label{ext_eq2}
\end{equation}
where $E(B-V)$ is the colour excess or selective extinction. %
The redshift ($k$-correction) and extinction-corrected $(J-K_s)$ colour is given by
\begin{equation}
(J-K_s)^o = (J-K_s) - 0.555 \, E(B-V) + 4.0\,z; \label{extinc_factor1}
\end{equation}
where, $ (J-K_s)$ is the measured NIR colour with no corrections applied while $(J-K_s)^o$ is the NIR colour, corrected for both redshift ($k$-correction, see $\S$\;\!\ref{k_correct}: Equation~\ref{k-corr}) and Galactic extinction effects. The colour excess, $E(B-V)$ values, were obtained using the DIRBE/IRAS reddening maps \citep{Schlegel_98}. %
Figure~\ref{extinc_JKs} shows the least-squares fit to the Norma sample (top panel) and Coma sample (bottom panel). The $J$ and $K_s$ aperture magnitudes were measured using aperture radii of 5 arcsec. The best fit which is represented by the red dashed line is
{\small{
\begin{eqnarray}
\text{Norma:} ~ (J-K_s)^o = -0.283\pm0.087 A_{K_s} + 0.984\pm0.037, \label{eq1_norma}\\
\text{Coma:~} ~ (J-K_s)^o = -0.096\pm0.056 A_{K_s} + 1.050\pm0.067. \label{eq2_coma}
\end{eqnarray}
}}
\begin{figure}[htb!]
\centering
\begin{tabular}{c}
\includegraphics[width=0.43\textwidth]{Norma_extinc2.eps} \\
~\\
\includegraphics[width=0.43\textwidth]{Coma_extinc2.eps}
\end{tabular}
\caption[$J$ and $K_s$-band photometry for the Coma sample]{\label{extinc_JKs}Measuring the extinction correction factor using the $(J-K_s)$ colour for the Norma (top panel) and the Coma (bottom panel) samples. The y-axis is the $(J-K_s)$ colour which has been corrected for both Galactic extinction and redshift effects. %
}
\end{figure}
Equations~\ref{eq1_norma} and \ref{eq2_coma} can be expressed in the form
\begin{equation}
(J-K_s)^o = aA_{K_s} + b \, ; \label{extinc_factor2}
\end{equation}
where $A_{K_s}$ is the extinction in the $K_s$-band which is given by Equation~\ref{ext_eq2}, $a$ and $b$ are fit parameters which were measured through simple least-squares regression. %
The reddening law for the DIRBE/IRAS reddening maps \citep{Schlegel_98} following the derivation in \citet{Schroder_07} can be written as:
\begin{equation}
J - K_s = \left(J - K_s\right)^o + \left[ \frac{E(J-K_s)}{A_{K_s}} \right] A_{K_s} \label{jtm_1}
\end{equation}
For the true extinction-corrected colour, $\left(\widetilde{J - K_s}\right)^o$, and true $K_s$-band extinction, $\tilde{A}_{K_s}$, we have
\begin{equation}
J - K_s = \left(\widetilde{J - K_s}\right)^o + \left[ \frac{E(J-K_s)}{A_{K_s}} \right] \tilde{A}_{K_s}.\label{jtm_2}
\end{equation}
Combining Equations~\ref{extinc_factor2} and \ref{jtm_1} yields
\begin{equation}
J - K_s = a A_{K_s} + b + \left[ \frac{E(J-K_s)}{A_{K_s}} \right] A_{K_s}. \label{jtm_3}
\end{equation}
Equating Equations~\ref{jtm_2} and \ref{jtm_3} leads to
\begin{equation}
\left(\widetilde{J - K_s}\right)^o = aA_{K_s} + b + \left( A_{K_s} - \tilde{A}_{K_s} \right) \left( \frac{E(J-K_s)}{A_{K_s}} \right)\cdot \label{jtm_4}
\end{equation}
The trend in Equations~\ref{eq1_norma} and \ref{eq2_coma}
implies a multiplicative factor, $f$, between the applied and true extinction, that is,
\begin{equation}
\tilde{A}_{K_s} = f A_{K_s}. \label{jtm_5}
\end{equation}
Substituting Equation~\ref{jtm_5} into \ref{jtm_4} results in
\begin{equation}
\left(\widetilde{J - K_s}\right)^o = a + \left[\left(1-f\right) \frac{E(J-K_s)}{A_{K_s}}\right] A_{K_s} + b. \label{jtm_6}
\end{equation}
Note that the true colour, $\left(\widetilde{J - K_s}\right)^o$, is independent of extinction, $A_{K_s}$. This implies that
\begin{equation}
\left(1-f\right) \frac{E(J-K_s)}{A_{K_s}} = 0. \label{jtm_7}
\end{equation}
Equation~\ref{jtm_7} implies that
\begin{equation}
f = 1 + a \left[\frac{E(J-K_s)}{A_{K_s}} \right]^{-1}. \label{extinc_factor_final}
\end{equation}
From the measured slope (Equations~\ref{eq1_norma} and \ref{eq2_coma}) and the selective extinction given by Equation
\ref{ext_eq2}, the extinction correction factors computed from Equation~\ref{extinc_factor_final} are \mbox{$0.81\pm0.06$} and \mbox{$0.93\pm0.05$} for the Norma and Coma samples, respectively. The challenge in our calibration is that, our sample covers a very small area on the sky and hence a very small range of extinction values (see the top and bottom panels of Figure~\ref{extinc_JKs}). It is however worth the effort given our high quality (high resolution) photometric data. After applying these extinction correction factors (0.81 for Norma and 0.93 for Coma ETGs), the mean differences between the \citet{Schlafly_11} $K_s$-band extinction and our values are relatively small, that is, 0.0081 mag (rms\,=\,0.0008 mag) and 0.0003 mag (rms\,=\,0.000045 mag) for the Norma and Coma ETGs, respectively. The extinction correction factor of \mbox{$0.81\pm0.06$} implies a range of 0.75\,--\,0.87 which translates into a difference of $\sim$\,43\;km\;s$^{-1}$ in Norma's distance.
\subsection{Corrected effective surface brightnesses} \label{k_correct}
The mean effective surface brightness ($\langle \mu_e \rangle$) was computed using:
\begin{equation}
\langle \mu_e \rangle = m_{\rm{tot}} + 2.5 \log \left( 2\pi r^2_e \right) - A_{\lambda} + k_{\lambda} - 10 \log \left( 1+z \right),
\end{equation}
where $m_{\text{tot}}$ is the total extrapolated magnitude corrected for the star-subtraction and sky background effects. $A_{\lambda}$ is the Galactic extinction, $k_{\lambda}$ is the $k$-correction, and
$10 \log \left( 1 + z \right)$ is a cosmological dimming correction term ($\lambda$ denotes the filter/band -- i.e., $J$, and $K_s$). The effective radii for both the Norma and Coma cluster ETGs were determined by measuring the radius of a circle containing half the total flux. These were thereafter corrected for seeing effects (measured using \textsc{Galfit}). The applied $k$-correction was adopted from \citet{Glass_99}, i.e., %
\begin{equation}
k_{J} = 0.7z, \qquad k_{K_s} = 3.3z. \label{k-corr}
\end{equation}
We have compared our adopted $k$-corrections with those of \citet{Chilingarian_10}. The mean differences (for the $K_s$-band) are 0.022 mag (rms\,=\,0.004 mag) and 0.039 mag (rms\,=\,0.005 mag) for the Norma and Coma samples, respectively. Using the \citeauthor{Chilingarian_10} $k$-correction values lowers the distance of the Norma cluster by $\sim$\,67\;km\;s$^{-1}$. %
Table~\ref{Norma_JKs_phot} shows the results for the Norma cluster sample where the final fully corrected FP variables ($r_e$, $\langle \mu_e \rangle$, and $\log \sigma$) are presented (also see Table~\ref{Coma_JKs_phot} for the Coma sample). %
Note that for the Norma cluster sample, $m_{\rm{tot}}$ was first corrected for star-subtraction effects (which are --\,0.0143 and {\mbox{--\,0.0135}} mag in the $J$ and $K_s$-band, respectively). An additional small correction (--\,0.011 and \mbox{--\,0.008} mag in the $J$ and $K_s$-band, respectively) due to sky background variation was also applied. This correction was measured (in our analysis) through simulations conducted using our IRSF $J$ and $K_s$-band images.
\section{Norma Distance: FP zero-point offset} \label{offset}
\subsection{Construction of the Fundamental Plane data sets} \label{fp_data_sets}
The FP relation
requires the effective radius, $r_e$, mean effective surface brightness, $\langle \mu_e \rangle$, and the central velocity dispersion, $\sigma$, that is,
\begin{equation}
\log r_e = a \log \sigma + b \langle\mu_e\rangle + c, \label{fp_eq}
\end{equation}
where $a$ and $b$ are the FP slopes and $c$ is the intercept. %
In Table~\ref{Norma_JKs_phot}, we present the FP data for the Norma ETGs ($N$\,$=$\,31).
Presented in this table are the central velocity dispersions and the photometric results obtained using the IRSF NIR $J$ and $K_s$-band images. The Galactic extinction correction applied to $\langle \mu_e \rangle$ is from the re-calibration presented in Section~\ref{extinc_recalib}, i.e., \mbox{$A_J = 0.937\,f\,E(B-V)$} and \mbox{$A_{K_s} = 0.382\,f\, E(B-V)$} where $f=0.81\pm0.06$ for the Norma cluster sample and $f=0.93\pm0.05$ for the Coma cluster calibration sample.
\begin{table*
\centering
\caption[Norma cluster $J$ and $K_s$ photometry results]{\label{Norma_JKs_phot} Norma cluster $J$ and $K_s$-band photometry results from f\mbox{}itting and extrapolating the galaxy surface brightness prof\mbox{}iles. NIR images taken using the IRSF telescope were used for the photometry analysis. The total extrapolated magnitudes presented here have been corrected for only, the star-subtraction and sky background effects. The $J$ and $K_s$-band mean ef\mbox{}fective surface brightness in $\rm{mag}\;\rm{arcsec}^{-2}$ have been corrected for the Galactic extinction, redshift, and the cosmological dimming ef\mbox{}fects. The Galactic extinction values are from the re-calibration presented under Section~\ref{extinc_recalib}, where, \mbox{$A_J = 0.937\,f\,E(B-V)$} and \mbox{$A_{K_s} = 0.382\,f\,E(B-V)$}.
Both the aperture correction and run offset have been applied to the central velocity dispersion values (in dex) presented here. The FP variables, that is, effective radii, mean effective surface brightnesses and the central velocity dispersions presented here have been fully corrected. These were used in the final FP fitting and analysis.}
\begin{tabular}{c@{\hskip 0.15in} r@{\hskip 0.15in} r@{\hskip 0.15in} r@{\hskip 0.15in} r@{\hskip 0.15in} c@{\hskip 0.15in} c@{\hskip 0.15in} c@{\hskip 0.15in} c@{\hskip 0.15in} c} \\ \hline\hline
\multicolumn{1}{c@{\hskip 0.15in} }{Identification} & \multicolumn{2}{c@{\hskip 0.15in} }{Total magnitude} & \multicolumn{2}{c@{\hskip 0.15in} }{Effective radius} & \multicolumn{2}{c@{\hskip 0.15in} }{Mean eff. surf. brightness} & \multicolumn{2}{c@{\hskip 0.15in} }{Extinction} & \multicolumn{1}{c}{$\log \sigma$} \\
& \multicolumn{1}{c@{\hskip 0.15in} }{$J$} & \multicolumn{1}{c@{\hskip 0.15in} }{$K_s$} & \multicolumn{1}{c@{\hskip 0.15in} }{$J$} & \multicolumn{1}{c@{\hskip 0.15in} }{$K_s$} & \multicolumn{1}{c@{\hskip 0.15in} }{$J$} & \multicolumn{1}{c@{\hskip 0.15in} }{$K_s$} & \multicolumn{1}{c@{\hskip 0.15in} }{$A_J$} & \multicolumn{1}{c@{\hskip 0.15in} }{$A_{K_s}$} & \\
\multicolumn{1}{c@{\hskip 0.15in} }{(1)} & \multicolumn{1}{c@{\hskip 0.15in} }{(2)} & \multicolumn{1}{c@{\hskip 0.15in}}{(3)} & \multicolumn{1}{c@{\hskip 0.15in} }{(4)} & \multicolumn{1}{c@{\hskip 0.15in} }{(5)} & \multicolumn{1}{c@{\hskip 0.15in} }{(6)} & \multicolumn{1}{c@{\hskip 0.15in} }{(7)} & \multicolumn{1}{c@{\hskip 0.15in} }{(8)} & \multicolumn{1}{c@{\hskip 0.15in} }{(9)} & \multicolumn{1}{c}{(10)} \\ \hline
WKK6019 & 10.87$\pm$0.06 &9.82$\pm$0.03 & 4.56$\pm$0.50 &4.22$\pm$0.23 & 15.98$\pm$0.24 &14.81$\pm$0.12 & 0.172 &0.070 &2.409$\pm$0.010 \\
WKK6047 & 12.81$\pm$0.06 &11.83$\pm$0.03 & 3.07$\pm$0.22 &2.92$\pm$0.13 & 17.09$\pm$0.16 &16.04$\pm$0.10 & 0.173 &0.070 &2.010$\pm$0.015 \\
WKK6075 & 12.95$\pm$0.06 &11.92$\pm$0.03 & 2.32$\pm$0.27 &2.14$\pm$0.06 & 16.56$\pm$0.26 &15.38$\pm$0.07 & 0.181 &0.074 &2.099$\pm$0.017 \\
WKK6116 & 10.50$\pm$0.06 &9.47$\pm$0.04 & 7.48$\pm$1.00 &7.59$\pm$0.70 & 16.71$\pm$0.30 &15.71$\pm$0.21 & 0.163 &0.066 &2.344$\pm$0.012 \\
WKK6148 & 12.67$\pm$0.05 &11.69$\pm$0.03 & 3.84$\pm$0.18 &3.33$\pm$0.14 & 17.39$\pm$0.11 &16.13$\pm$0.10 & 0.155 &0.063 &1.998$\pm$0.019 \\
WKK6180 & 11.16$\pm$0.06 &10.13$\pm$0.03 & 6.84$\pm$0.69 &7.18$\pm$0.37 & 17.12$\pm$0.23 &16.24$\pm$0.12 & 0.159 &0.065 &2.308$\pm$0.011 \\
WKK6183 & 11.13$\pm$0.06 &10.16$\pm$0.03 & 5.32$\pm$0.82 &5.58$\pm$0.42 & 16.68$\pm$0.34 &15.66$\pm$0.17 & 0.155 &0.063 &2.377$\pm$0.012 \\
WKK6193 & 11.10$\pm$0.07 &10.00$\pm$0.03 & 6.48$\pm$0.68 &6.60$\pm$0.28 & 16.97$\pm$0.24 &15.99$\pm$0.10 & 0.159 &0.065 &2.226$\pm$0.010 \\
WKK6204 & 10.35$\pm$0.07 &9.63$\pm$0.28 & 6.87$\pm$1.12 &5.22$\pm$3.16 & 16.37$\pm$0.36 &14.98$\pm$1.34 & 0.160 &0.065 &2.499$\pm$0.009 \\
WKK6207 & 11.05$\pm$0.07 &9.90$\pm$0.03 & 4.38$\pm$0.83 &6.39$\pm$0.52 & 16.21$\pm$0.42 &15.78$\pm$0.18 & 0.152 &0.062 &2.241$\pm$0.012 \\
WKK6221 & 11.31$\pm$0.06 &10.77$\pm$0.04 & 9.14$\pm$1.05 &7.47$\pm$0.38 & 17.89$\pm$0.26 &16.62$\pm$0.11 & 0.164 &0.067 &2.038$\pm$0.017 \\
WKK6229 & 12.56$\pm$0.07 &11.59$\pm$0.03 & 2.30$\pm$0.25 &1.97$\pm$0.09 & 16.19$\pm$0.24 &14.88$\pm$0.11 & 0.156 &0.063 &2.210$\pm$0.015 \\
WKK6233 & 12.73$\pm$0.07 &11.91$\pm$0.04 & 2.69$\pm$0.22 &2.14$\pm$0.07 & 16.66$\pm$0.19 &15.26$\pm$0.08 & 0.152 &0.062 &2.243$\pm$0.014 \\
WKK6235 & 11.91$\pm$0.06 &10.96$\pm$0.03 & 4.73$\pm$0.42 &4.39$\pm$0.21 & 17.07$\pm$0.20 &15.99$\pm$0.11 & 0.170 &0.069 &2.141$\pm$0.018 \\
WKK6242 & 11.82$\pm$0.07 &10.90$\pm$0.03 & 2.74$\pm$0.24 &2.16$\pm$0.15 & 15.86$\pm$0.20 &14.41$\pm$0.15 & 0.151 &0.061 &2.424$\pm$0.011 \\
WKK6250 & 11.56$\pm$0.05 &10.55$\pm$0.02 & 3.58$\pm$0.31 &3.41$\pm$0.13 & 16.15$\pm$0.20 &15.01$\pm$0.08 & 0.159 &0.065 &2.324$\pm$0.010 \\
WKK6252 & 13.00$\pm$0.06 &12.18$\pm$0.03 & 2.80$\pm$0.24 &2.44$\pm$0.16 & 17.04$\pm$0.19 &15.88$\pm$0.14 & 0.153 &0.063 &2.073$\pm$0.012 \\
WKK6269 & 8.97$\pm$0.06 &8.14$\pm$0.03 & 20.40$\pm$2.47&17.57$\pm$1.03 & 17.44$\pm$0.27 &16.22$\pm$0.13 & 0.148 &0.060 &2.579$\pm$0.011 \\
WKK6275 & 11.40$\pm$0.06 &10.22$\pm$0.03 & 7.74$\pm$0.78 &8.39$\pm$0.60 & 17.62$\pm$0.23 &16.68$\pm$0.16 & 0.149 &0.061 &2.165$\pm$0.014 \\
WKK6282 & 12.17$\pm$0.06 &11.22$\pm$0.03 & 3.05$\pm$0.33 &2.60$\pm$0.17 & 16.42$\pm$0.24 &15.14$\pm$0.15 & 0.147 &0.060 &2.274$\pm$0.012 \\
WKK6297 & 12.69$\pm$0.05 &11.78$\pm$0.02 & 4.38$\pm$0.25 &4.01$\pm$0.18 & 17.72$\pm$0.14 &16.71$\pm$0.10 & 0.149 &0.061 &1.844$\pm$0.026 \\
WKK6305 & 10.23$\pm$0.05 &9.24$\pm$0.11 & 6.77$\pm$0.79 &7.93$\pm$1.49 & 16.25$\pm$0.26 &15.45$\pm$0.42 & 0.198 &0.081 &2.327$\pm$0.009 \\
WKK6318 & 9.61$\pm$0.06 &8.73$\pm$0.03 & 15.84$\pm$1.76&13.75$\pm$0.79 & 17.49$\pm$0.25 &16.35$\pm$0.13 & 0.180 &0.073 &2.354$\pm$0.013 \\
WKK6342 & 12.08$\pm$0.05 &11.06$\pm$0.02 & 2.59$\pm$0.13 &3.05$\pm$0.15 & 15.98$\pm$0.12 &15.11$\pm$0.11 & 0.152 &0.062 &2.326$\pm$0.009 \\
WKK6360 & 10.92$\pm$0.06 &10.02$\pm$0.02 & 4.19$\pm$0.66 &3.60$\pm$0.15 & 15.91$\pm$0.35 &14.61$\pm$0.10 & 0.154 &0.063 &2.505$\pm$0.009 \\
WKK6383 & 11.94$\pm$0.06 &10.78$\pm$0.03 & 4.04$\pm$0.26 &4.59$\pm$0.22 & 16.77$\pm$0.15 &15.88$\pm$0.11 & 0.176 &0.072 &2.195$\pm$0.012 \\
WKK6402 & 11.56$\pm$0.05 &10.62$\pm$0.02 & 3.70$\pm$0.29 &3.23$\pm$0.15 & 16.20$\pm$0.18 &14.98$\pm$0.10 & 0.190 &0.077 &2.238$\pm$0.011 \\
WKK6429 & 12.30$\pm$0.04 &11.63$\pm$0.04 & 5.04$\pm$0.37 &5.56$\pm$0.17 & 17.73$\pm$0.16 &16.91$\pm$0.09 & 0.159 &0.065 &2.084$\pm$0.012 \\
WKK6431 & 11.74$\pm$0.05 &10.78$\pm$0.03 & 3.56$\pm$0.24 &3.18$\pm$0.19 & 16.28$\pm$0.16 &15.17$\pm$0.13 & 0.184 &0.075 &2.286$\pm$0.010 \\
WKK6459 & 11.37$\pm$0.06 &10.45$\pm$0.04 & 6.18$\pm$0.85 &6.25$\pm$0.21 & 17.09$\pm$0.31 &16.13$\pm$0.08 & 0.200 &0.081 &2.322$\pm$0.010 \\
WKK6477 & 12.67$\pm$0.05 &11.73$\pm$0.04 & 3.02$\pm$0.32 &2.68$\pm$0.13 & 16.90$\pm$0.23 &15.76$\pm$0.11 & 0.179 &0.073 &2.106$\pm$0.019 \\
\hline
\end{tabular}
\end{table*}
\subsection
Fundamental Plane -- Fit results}\label{JKs_fittings}
The FP was fitted using the inverse least-squares fit. The top and bottom
panels of Figure~\ref{JKs:fig_FP} show the FP projection for the $J$- and $K_s$-band, respectively. The black open circles represent the Coma cluster galaxies while the red filled circles are the Norma cluster galaxies shifted to the Coma distance. This was done assuming the Coma cluster has a zero peculiar velocity. %
\begin{figure}[htb!]
\centering
\begin{tabular}{c}
\includegraphics[width=0.45\textwidth]{FP_IRSF-Jband.eps} \\
~\\~\\
\includegraphics[width=0.45\textwidth]{FP_IRSF-Ksband.eps}
\end{tabular}
\caption{\label{JKs:fig_FP}The $J$ and $K_s$-band FP based on 31 ETGs from the IRSF data (for the Norma cluster) and 121 ETGs from 2MASS (for the Coma calibration sample). The top (a) and bottom (b) panels
represent the $J$- and $K_s$-band FP projections, respectively. The red filled circles represent the Norma cluster ETGs after shifting them to the Coma cluster's distance. The black open circles represent the Coma cluster galaxies.}
\end{figure}
The combined $K_s$-band FP analysis was performed using the combined IRSF and NTT data. In combining the two data sets, the latter were given preference due to the higher resolution imaging. The FP projection obtained from the combined $K_s$-band analysis is shown in Figure~\ref{Ks:fig_FP}. The black open circles represent the Coma cluster galaxies while the red filled circles represent the Norma cluster galaxies after shifting them to the Coma cluster's distance.
A summary of results for the IRSF $J$ and $K_s$-band as well as the combined $K_s$-band FP fit parameters is presented in Table~\ref{JKs_FP_distance}. The derived values for the FP parameters $a$, and $b$ are comparable with other previous FP studies in the near-infrared \citep[see e.g.,][]{LaBarbera_10, Magoulas_12}. %
\begin{figure}[htb!]
\centering
\begin{tabular}{c}
\includegraphics[width=0.45\textwidth]{FP_IRSFF+NTT.eps}
\end{tabular}
\caption[Combined $K_s$-band FP]{\label{Ks:fig_FP}Same as Figure~\ref{JKs:fig_FP} but for the combined $K_s$-band FP projection.
}
\end{figure}
\begin{table} \tiny
\centering
\caption[$J$- and $K_s$-band FP fit parameters]{\label{JKs_FP_distance}The $J$- and $K_s$-band FP fit parameters ($a$ and $b$) as well as the zero-point offset.
The first column indicates the source of the data used while
column~5 is the FP zero-point offset. Columns 6 and 7 are the scatter and the corresponding percentage uncertainty in the distance to the individual Norma cluster galaxies, respectively. Column 8 indicates the number of Norma cluster galaxies used in our analysis; this is 121 for the Coma cluster.}
\begin{tabular}{l@{\hskip 0.075in}c@{\hskip 0.075in}c@{\hskip 0.075in}c@{\hskip 0.075in}c@{\hskip 0.075in}c@{\hskip 0.075in}c@{\hskip 0.075in}l} \hline\hline
Sample & Filter & $a$ & $b$ & $\Delta c$ & rms & \% err & N \\
& & & & & & & \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) \\ \hline
NTT & $K_s$ & $1.465\pm0.059$ & $0.326\pm0.020$ & $0.154\pm0.014$ & 0.083 & 28 & 31$*$ \\
IRSF & $J$ & $1.356\pm0.058$ & $0.317\pm0.018$ & $0.149\pm0.016$ & 0.079 & 25 & 31 \\
IRSF & $K_s$ & $1.470\pm0.053$ & $0.357\pm0.019$ & $0.145\pm0.014$ & 0.078 & 26 & 31 \\
NTT+IRSF & $K_s$ & $1.490\pm0.053$ & $0.325\pm0.017$ & $0.163\pm0.013$ & 0.080 & 27 & 41 \\
\hline
\end{tabular}
~\\$*$ --- All data in this row were taken from \citet{Mutabazi_14}.
\end{table}
\section{Norma cluster: distance and peculiar velocity} \label{dist_vpec_fp}
Table~\ref{JKs_FP_distance} shows the different zero-point offsets measured using the $J$ and $K_s$-band FP analysis. These are in excellent agreement with each other. The best measured value (least scatter/error) is from the combined $K_s$-band FP analysis using 41 ETGs within the Norma cluster. We adopted this value to measure the distance and peculiar velocity for the Norma cluster. This implies that
\begin{equation}
\log {D_A \left(\rm{Coma}\right)} - \log{D_A\left(\rm{Norma}\right)} =
0.163\,\pm\,0.013,
\label{finedist_here}
\end{equation}
where $D_A (\rm{Coma})$ and $D_A (\rm{Norma})$ are the angular diameter distances for the Coma and Norma clusters, respectively. The angular diameter distances were measured following \citet{Wright_06}. The distance (Hubble redshift, $z_{_{\rm H}}$) and the peculiar velocity for the Norma cluster were derived using the same method described in \citet{Mutabazi_14}. That is, the peculiar velocity for the Norma cluster was computed using
\begin{equation}
v_{\text{pec}} = c\,z_{\text{pec}},
\end{equation}
where $c$ is the speed of light and
\begin{equation}
z_{\text{pec}} = \frac{1+z_{_{\text{CMB}}}}{1+z_{_{\text{H}}}} - 1.
\end{equation}
The resulting Norma cluster's peculiar velocity after applying the homogeneous Malmquist bias correction \citep[see][for details]{Hudson_97} presented in Table~\ref{JKs_FP_distance_vpec} is $44\pm151$\;km\;s$^{-1}$ (distance of 4915\,$\pm$\,121\;km\;s$^{-1}$ or $d$\,$=$\,69.7\,$\pm$\,1.7~Mpc).
\begin{table} \tiny
\centering
\caption[Distance and peculiar velocity from the $J$- and $K_s$-band FP analysis]{\label{JKs_FP_distance_vpec}The $J$- and $K_s$-band FP offsets (column~3) and the corresponding measured peculiar velocity (column~5). The homogeneous Malmquist bias correction shown in column 6 has not been applied (this should be subtracted from the peculiar velocity presented in column~5). %
}
\begin{tabular}{l@{\hskip 0.1in}c@{\hskip 0.1in}c@{\hskip 0.1in}c@{\hskip 0.1in}r@{\hskip 0.1in}c@{\hskip 0.1in}l} \hline\hline
Sample & Filter & $\Delta c$ & $z_{_{\rm H}}$ & \multicolumn{1}{c@{\hskip 0.1in}}{v$_{\rm{pec}}$} & Malm. bias & N \\
& & [dex] & & [km\;s$^{-1}$] & [km\;s$^{-1}$] & \\
(1) & (2) & (3) & (4) & \multicolumn{1}{c@{\hskip 0.1in}}{(5)} & (6) & (7) \\ \hline
NTT & $K_s$ & $0.154\pm0.014$ & $0.01667\pm0.00055$ & $-43\pm170$ & 29 & 31$*$ \\
IRSF & $J$ & $0.149\pm0.016$ & $0.01684\pm0.00062$ & $-95\pm187$ & 33 & 31 \\
IRSF & $K_s$ & $0.145\pm0.014$ & $0.01704\pm0.00058$ & $-152\pm173$ & 30 & 31 \\
NTT+IRSF & $K_s$ & $0.163\pm0.013$ & $0.01632\pm0.00050$ & $+58\pm151$ & 14 & 41 \\
\hline
\end{tabular}
~\\$*$ --- All data in this row were taken from \citet{Mutabazi_14}.
\end{table}
\section{Discussion and conclusions} \label{sec_discuss}
We have measured a small peculiar velocity of the Norma cluster of $+44$\,$\pm$\,$151$\;km\;s$^{-1}$ which within errors, is consistent with zero and in excellent agreement with our previous measurement of $-72\pm170$\;km\;s$^{-1}$ \citep{Mutabazi_14}. Our analysis and results have demonstrated, that, despite the challenges of the large Galactic extinction and severe stellar contamination, distances using the FP and TF relations can be derived reliably for galaxies that lie relatively close to the Galactic plane.
We have carefully considered the possible sources of measurement and systematic errors. Results from simulations of ETGs showed that the ef\mbox{}fect of star-subtraction is very small, ranging from
$-0\!\mbox{\ensuremath{\;\mathrm{.}\!\!^\textrm{m}}} 011$ (NTT $K_s$-band data) to
$-0\!\mbox{\ensuremath{\;\mathrm{.}\!\!^\textrm{m}}} 014$ (IRSF $K_s$-band data), which offsets the measured peculiar velocity by $\sim$\,41\;km\;s$^{-1}$ to $\sim$\,52\;km\;s$^{-1}$, respectively. The systematic effect arising from possible gradients in the sky background
is very small, i.e., $\sim$\,$-0\!\mbox{\ensuremath{\;\mathrm{.}\!\!^\textrm{m}}} 008$, which corresponds to a change in Norma's measured peculiar velocity of $\sim$\,30\;km\;s$^{-1}$. %
Using the NIR $J$ and $K_s$-band images obtained using the IRSF telescope has also helped re-calibrate the \citet{Schlegel_98} extinction values as described in Section~\ref{extinc_recalib}. For the Norma cluster sample, the measured extinction correction factor is 0.81 for the \citeauthor{Schlegel_98} maps, which is in good agreement with the \citet{Schlafly_11} extinction values. The combined $K_s$-band FP sample resulted in the largest cluster sample ($N=41$) used in the peculiar velocity studies at relatively low Galactic latitudes to date, and hence the most precise distance and peculiar velocity of the Norma cluster measured to date.
Figure~\ref{GA_vpec_plot} shows the peculiar velocities for Clusters/Groups in the GA region. The peculiar velocity measurements were taken from the literature.
Shown are the velocity measurements from ENEARc \citep{Bernardi_02}, SMAC \citep{Hudson_04}, SFI++ \citep{Springob_07}, and our Norma cluster measurement (red filled circle). The dashed horizontal lines represent $\pm 3\sigma$ where $\sigma$\,$=$\,151\;km\;s$^{-1}$ is the error on the Norma cluster's peculiar velocity derived using the combined $K_s$-band FP analysis. $N_{\phi}$ is the angular separation between the cluster and Norma. %
The improved sample of 41 ETGs taken from within $\frac{2}{3}$ of the Norma cluster's Abell radius makes this the most precise measurement for the distance and peculiar velocity of the Norma cluster to date. %
\begin{figure}[h
\centering
\includegraphics[width=0.43\textwidth]{GA_vpec.eps}
\caption{\label{GA_vpec_plot}
Peculiar velocities for Clusters/Groups in the GA region. The Norma cluster is represented by the red filled circle. The dashed horizontal lines represent $\pm3\sigma$ where $\sigma$ is the error on the Norma cluster's peculiar velocity measured using the combined $K_s$-band FP analysis. $N_{\phi}$ is the angle on the sky between the cluster and Norma. %
}
\end{figure}
In the study by \citet{Carrick_15} the resisual bulk flow due to structures not covered by the 2M++ redshift sample has a velocity of $159\pm 23$~km\;s$^{-1}$ in the direction $\ell = 303^\circ \pm 11^\circ$, $b = 6^\circ \pm 11^\circ$. This is $\sim 25^\circ$ from Norma, and therefore
corresponds to $\sim 144\pm 21$~km\;s$^{-1}$. This is the external bulk flow contribution on the Norma peculiar velocity arising from structures beyond the volume covered by the 2M++ redshift catalogue \citep{Lavaux_2011}. Similar residual bulk flows have been reported with a bulk flow contribution on the Norma cluster of e.g., $136\pm39$~km\;s$^{-1}$ \citep{Turnbull_12}, $154\pm10$~km\;s$^{-1}$ \citep{Boruah_19}, $147\pm12$~km\;s$^{-1}$ \citep{Said_20}, and $161\pm11$~km\;s$^{-1}$ \citep{Said_20}. These are all significantly different from zero and are, within the error margin ($1\sigma$), consistent with our Norma's peculiar velocity of $44$\,$\pm$\,$151$\;km\;s$^{-1}$. Dedicated surveys such as Taipan \citep{Cunha_17, Taylor_20} and WALLABY \citep{Duffy_12, Koribalski_20} will help improve the accuracy of distance measurements and hence enable us infer more accurate peculiar velocities.
\acknowledgments
TM acknowledges financial support from the Square Kilometre Array South Africa (SKA~SA). Travel costs to Sutherland site of the South African Astronomical Observatory were funded by the multi-wavelength funding (Department of Astronomy, University of Cape Town).
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is \url{http://www.sdss3.org/}. SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University. This publication has made use of the NASA/IPAC ExtraGalactic Data base (NED), and also data products from the 2MASS, a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation.
\software
\textsc{Galfit}\ \citep{Peng_10}, \matplotlib\ \citep{Hunter_07}, \textsc{SCIPY}\ \citep{Virtanen_17}, \numpy\ \citep{Oliphant_06,
van_11}, \textsc{PGPLOT}\ \citep{Pearson_11}, \textsc{Ellipse}\ task under \textsc{Iraf}'s \textsc{Stsdas}\ package \citep{Jedrzejewski_87}.
\facilities{1.4m Infrared Survey Facility (IRSF) at South Africa Astronomical Observatory Sutherland site,
European Southern Observatory (ESO) 3.6m New Techology Telescope (NTT) at La Silla Observatory,
Australian National University (ANU) 3.9m Anglo-Australian Telescope (AAT) at Siding Spring Observatory (SSO).}
\bibliographystyle{aasjournal}
|
\section{Introduction}
\label{Sec:1}
Coulomb scattering based muon tomography was proposed in 2003~\cite{Borozdin2003aa}. Since then many research groups have demonstrated that the method can be successfully implemented in different fields to detect high-Z materials, e.g., nuclear safety~\cite{Clarkson:2014xfa, Poulson:2016fre, Mahon:2019aa} or security applications~\cite{Morris:2008aaa, Riggi:2018aaa, Harel:2019aaa}. In recent years several reviews have appeared providing excellent insights into last developments~\cite{Bonechi:2019ckl, Kaiser:2019aa, Vanini:2019aab, Procureur:2018aaa, Bonomi:2017uow}.
Coulomb scattering depends strongly on atomic number $Z$. It makes the recognition of high-Z materials (e.g. nuclear materials) in muon tomography or more generally, in atmospheric ray tomography (ART), easier than the recognition of low-Z materials. Thus ART applications on low-Z materials have merged slowly. Technological challenges are particularly complex in security applications, where low-Z materials should be identified within minute-scale time.
However, some dedicated research papers have appeared on the feasibility of the low-Z ART for security applications. Klimenko et al have studied whether electrons can be used for flux attenuation as an additional signature in muon tomography~\cite{Klimenko:2005aaa}. Other accompanying physical effects such as muonic X-rays and electromagnetic showers have also been investigated. The authors summarize that the inclusion of the latter requires additional detection systems. Cuellar et al reported results based on Geant4 simulations that the stopping effect of the low momentum part of the leptons spectrum could be used as a source of additional information to scattering tomography, especially for medium and low-Z materials~\cite{Cuellar:2009aa}. Later the effect has been demonstrated with simulated and experimental results~\cite{Blanpied:2015aaa}. The authors have proposed to exploit the ratio of stopping power to scattering, which enables to eliminate the sample thickness as an unknown variable. First, one identifies the material from the ratio, then the mean scattering angle and the known radiation length can be computed to calculate the thickness. The simulated results with materials of 1~m$^2$ for heavy metal plates or cubes of 1~m$^3$ for paper and nylon, as well as experimental results on the pallet-size units with low-Z materials, were presented with 30, 10 and 5 minute long measurements.
The detection of explosives and narcotics has been investigated at the TUMUTY facility based on the Geant4 simulations, where the material classification was done using the machine learning (ML), Support Vector Machine (SVM)~\cite{Yifan:2018aac}. The authors conclude that ART allows to discern narcotics and explosives from background and metals, but not be separated from each other (the various $20\times20\times20$~cm$^3$ size objects with 10 to 30~minute measuring time were studied). The experimental validations of the numeric results on an RPC-based tomography system concluded that the actual spatial resolution (0.6~mm) was lower than previously were simulated~\cite{Pan:2019aaa}. Four materials were tested: flour (substitute of narcotics), aluminium, steel, and lead. The overall discrimination rate with the SVM classifier could reach 70, 95, and 99\% with 1, 5, and 10 minutes measurement, accordingly.
We have found only a few examples of low-Z ART in non-security applications. Bikit et al demonstrated that low-Z materials can be imaged by the simultaneous detection of cosmic muons and photons created via bremsstrahlung inside the investigated bone and soft tissue sample~\cite{Bikit:2016abc, Mrdja:2016aaa}. Another example is inspection of nuclear waste encapsulated in concrete matrix. The investigations on the containers of nuclear waste showed that litre-size gas bubbles in concrete due to uranium dioxide oxidation can be discovered within days of exposure~\cite{Dobrowolska:2018abc}. Two recently papers reported that ART can detect reinforcement rebars in concrete. The first study is based on Monte Carlo simulations~\cite{Dobrowolska:2020aaa} and the second on experimental tests~\cite{Niederleithinger:2020wrf}.
Three main factors influence the ART performance: (i) the accuracy of particle trajectory estimation, (ii) ability to determine particle type and energy, (iii) the tomographic reconstruction methods deployed. The fourth, a system independent factor, is the measurement time available. In the early days of muon tomography, a geometry-based reconstruction approach, Point of Closest Approach (PoCa)~\cite{SCHULTZ2004687}, was introduced. PoCa is approximating multiple scattering processes with a single interaction point. It shows fairly good results in detection of high-Z materials submerged to low-Z materials. In conventional X-ray Computed Tomography (CT) analytical methods such as filtered back projection (FPB) have been historically widely used for its computational lightness. In CT the positions of source and detector are well known, but non-isotropic exposure and high Poisson noise in ART reduce its performance considerably. Iterative methods seem to be most suited, either algebraic~\cite{Liu:2019aab} or statistical approaches~\cite{Schultz:2007aaa, PESENTE:2009aaa}, where the best trajectory estimates are obtained using scattering and displacement data.
Due to the properties of the Coulomb multiple scattering, the performance of tomographic reconstruction is greatly improved if the energy of passing particle is estimated. For example, the measured angular distribution of passing particles allows to classify the particles according their estimated energy~\cite{Morris:2012aca}. The proposed momentum multi-group model method was experimentally tested and the reconstruction performance was quantitatively compared against the squared average, median angle squared and the average of squared angle using the receiver operating characteristic (ROC) curves~\cite{Perry:2014bad}. The results demonstrate clear superiority for multi-group approach. The CRIPT-project presented an another approach of momentum estimation employing massive iron sheets between their tracking detectors~\cite{ANGHEL:2015abs}.
The accurate reconstruction of particle tracks is especially important in low-Z ART and in applications having sub-cubic-meter volumes scanned. The resolution of scattering angle at mrad is required. The scattering density of muons is directly related to the atomic number $Z$ of material. The density is an order of magnitude lower for materials with low-Z (Ca or lighter elements, $Z \lesssim 20$) in comparison to high-Z (Pb or heavier elements, $Z \gtrsim 80$). For example, many experiments have applied the cylindrical drift chamber detectors in low-Z ART. The spatial resolution of small scale drift tubes can go down to 400~$\mu$m (FWHM) and angles of 2~mrad (FWHM)~\cite{Morris:2008aaa}. For large scale systems, a sub-mm position resolution perpendicular to the drift tube wire has been reported~\cite{Blanpied:2015aaa}. For an another type, resistive plate chambers (RPC), spatial resolution of 0.6~mm on lab-scale has been achieved~\cite{Pan:2019aaa}. The authors show no data on the spacing between the detector plates, so we cannot estimate the angular resolution.
On the sea level, the electrons and positrons constitute about 30-40\% of the total lepton flux and about a half of that is in practical usable energy range~\cite{Klimenko:2005aaa}. Due to low energy, this fraction of the lepton spectrum is strongly influenced from the environment around the detector system, for example, the building construction structures above the detector system. However, one can employ it as an additional source of information for the detection of small and medium sized objects, such as passenger luggage or similar scale objects.
In this paper we present new methods and a proof-of-concept tomography system developed specifically for the ART of low-Z materials. We introduce the Particle Track Filtering (PTF) and Multi-Modality Tomographic Reconstruction (MMTR) methods. PTF allows to classify the passing particle events (of charged leptons) into different groups so that the absorption and the scattering effect are more isolated. Some aspects of the PTF method are described in a patent application by some authors of the paper~\cite{Patent:2018aaa}.
In tomographic reconstruction we have developed a novel edge detection method to separate the logical volumes of scanned object. We demonstrated the effectiveness of the method in case of single composition material (e.g. water, aluminum, steel) and double composition material (e.g. explosive RDX in flesh-like material) objects. The MMTR approach enables to combine multiple reconstructions obtained from the different PTF-classified particle event groups and the scattering and transmission parameters from the track reconstruction.
We have composed the numerical models using the Geant4 software package~\cite{Allison:2006ve} to optimize the tomography system and the parameters of PTF and MMTR procedures. For the tracker system (the hodoscope) we chose plastic scintillating fibers arrays. With this tracker we were able to achieve a spatial resolution of 120~$\mu$m for muons and the 1~mrad angular resolution in the particle track reconstruction.
To validate our theoretical and numerical tests we built a physical proof-of-concept tomography system. The measurements on the prototype showed excellent agreement with the numerical simulations. We are able to increases significantly the discriminating power of ART for low-Z materials.
The structure of the paper is the following. We start with the description of PTF realized on plastic scintillator fiber arrays. In Section~\ref{Sec:3} we present the MMTR method. In both sections we validate the methods with numerical Geant4 models. In Section~\ref{Sec:4} we describe the general design principles of the physical proof-of-concept ART system. In Section~\ref{Sec:5} we give the technical description of the prototype and the first results of the measurements. In the last section we summarize the results and give some hints on the next steps we have considered.
\section{The particle track classification: the PTF method}
\label{Sec:2}
In this section we describe the main elements of the Particle Track Filtering (PTF) method. We have developed a virtual hodoscope system in Geant4 addressing the critical factors of the low-Z ART described above. Mainly, we address the accuracy of particle track reconstruction and the ability to classify the type and energy of particle event candidates. The hodoscope consists of the three position sensitive detector plates composed by plastic scintillator fibers. The detector plate has 4-layered structure: the two orthogonally placed double-layered fiber-mats, which give the $x$ and $y$ position of the particle hit (see Fig.~\ref{Fig:1}). The fiber diameter is 1.0~mm, the pitch in a single layer is 1.1~mm. The top layer is shifted by a half pitch, its position has been aligned according to the positions of the lower layer fibers. This ensures close to 100\% geometrical detection efficiency at the every angle of an incidence and high spatial resolution (for details, see Sec.~\ref{Sec:4}).
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_1.pdf}
\caption{A snapshot from the Geant4 model of the position-sensitive detector plate composed by the two orthogonally placed double-fiber-layers (gray cylinders). The red line denotes the trajectory of a passing muon. The greened fibers illustrate the propagation of scintillation light generated in the fibers by the muon passage.}\label{Fig:1}
\end{figure}
The hodoscope has three detector plates in order to estimate the energy and the type of the passing particle using the intrinsic scattering angle $\theta$ (in the hodoscope). Each detector plate has a well-defined thickness and material thus the probability distribution of the intrinsic scattering can be calculated. Though the typical $\theta$ of passing muons is in the order of mrad it can be measured in the hodoscope. As a result we can classify the reconstructed particle events (with some statistical probability) into separated categories according to their scattering angle $\theta$ in the hodoscope. Fig.~\ref{Fig:2} illustrates the intrinsic scattering $\theta$ in the hodoscope.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_2.pdf}
\caption{The intrinsic scattering angle $\theta$ we use to classify the particle events (passing muons/electrons) in the hodoscope. The hodoscope has the three detector plates (black bold lines numbered as 1, 2, 3). The arrows denote the reconstructed particle trajectory. The angle $\theta$ denotes the intrinsic scattering angle of the particle in the plate 2.}\label{Fig:2}
\end{figure}
Fig.~\ref{Fig:3} shows the simulation results of the filtering spectrum for the hodoscope having the distance 100~mm between the two adjacent plates. We used the CRY cosmic ray event generator to model the atmospheric ray flux consisting of muons and electrons at sea level~\cite{CRY-paper:2007aaa}. We fixed the spatial resolution of detector plates at 0.1~mm, which corresponds to the angular resolution of 1 mrad for particles approaching the hodoscope orthogonally. Considering the angular resolution of the hodoscope, the total spectral range presented in Fig.~\ref{Fig:3} can be divided into different number of groups; we call those PTF groups below. For example, a possible robust PTF classification schema is to classify the particle events to muons and electrons or the muons with low, medium and high momentum. In Fig.~\ref{Fig:3} we have separated the spectrum into the three PTF groups: F1 (dominated by muons), F2 (mixed muons and electrons) and F3 (dominated by electrons). This classification schema has been used in the figures of the reconstruction results in the next sections.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_3.pdf}
\caption{The distribution of atmospheric ray muons and electrons as a function of the intrinsic scattering angle $\theta$ in the hodoscope (from the Geant4 model with the CRY event generator). The distribution shows we can apply the intrinsic scattering angle $\theta$ as a discriminating parameter classifying the type and energy range of the hodoscope passing particle. The latter improves the tomographic reconstruction of scanned samples very significantly. The colored areas denote the muon and electron dominated values of $\theta$ (blue, yellow) and the mixed region (gray).}\label{Fig:3}
\end{figure}
\section{Tomographic reconstruction: the MMTR method}
\label{Sec:3}
\subsection{Reconstructing the volume of interest}
With the design of the hodoscopes as described above, we developed a model of the tomography system in Geant4 with the interrogation volume $1 \times 1 \times 2$~m$^3$. The motivation of the size comes from practical applications, e.g., it would be suitable for small luggage, postal parcels or human body. Fig.~\ref{Fig:4} shows the dimensions of the simulated tomography system. The system consists of the two horizontal and the two vertical hodoscopes forming an enclosed interrogation volume aka the volume of interest (VOI). When a muon/electron passes the entrance hodoscope it provides the entrance incidence angle, the intrinsic scattering angle for PTF and the entrance trajectory to the VOI with its entrance location. In the same way one can characterize an exiting particle from VOI.
The geometry of the tomography system maximises the usage of side flux: on the $y$-axis the side hodoscopes measure the low incidence angle flux and on the $x$-axis the elongated hodoscopes with the 2~m length increase the field of view in the central part of VOI.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_4.pdf}
\caption{A schema of the Geant4 model of the tomography system with the four hodoscopes (upper and lower horizontal and left and right vertical bold dark lines), the radiation source (CRY, gray area above) and an example of studied objects in the center of the system (dark purple). The numbers 1-6 denote the horizontal detector plates.}\label{Fig:4}
\end{figure}
In our Geant4 simulations the exposure time was estimated by the internal code of CRY, which calculates the elapsed time on the generation of selected number of particles (all types) per the defined generation area, $10 \times 10$~m$^2$ in our case. The tomography system was centred below the CRY source. The CRY generation area was placed 30~cm from the upper hodoscope.
In order to test PTF, we implement a tomographic reconstruction procedure based on the back projection. Thus we fill VOI with a voxel grid. In the back-projection, we use the $z$-planes as a reference point and calculate the corresponding $x$- and $y$-coordinates of the particle on a particular $z$-plane. This process is repeated for all $z$-planes, up to filling the VOI, see Fig.~\ref{Fig:5}. The $z$-plane is formed from the voxel grid point at the $z$-coordinate. The amount of the $z$-planes depends on the voxel size set by the user. As we focus on low-$Z$ materials and moderate volumes we approximate the trajectories of the particles with a straight line between the entering and the exiting locations of a particle (at the boundary of VOI). Then we allocate so called score value to each voxel when a trajectory passes through the voxel (see Fig.~\ref{Fig:6}).
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_5.pdf}
\caption{The $z$-planes (blue) filling the volume of interest (VOI) and the voxel grid (black). The number 3 refers to the closest upper detector plate to VOI and 4 refers to the closest lower detector plate to VOI (see the numbering of the plates in Fig.~\ref{Fig:4}).}\label{Fig:5}
\end{figure}
We can apply the back-projection procedure to the transmission and the scattering parameters. In the case of transmission based reconstruction, the count of particle trajectories in the every voxel can be used as the score value. In the case of scattering based reconstruction then one can choose the score value as the total scattering angle or the scattering angle per unit length of the trajectory. The total scattering angle is defined as the angle between vectors BC and DE, see Fig.~\ref{Fig:6}. Alternatively, other back-projection parameters are possible (see Sec.~\ref{Sec:5}).
Summing up all particle trajectories one forms the volume density map (VDM) of VOI. We build many VDMs, a dedicated VDM per PTF group. In the case of scattering parameter, we calculate the median or the mean scattering angle for each voxel forming the VDM in the scattering based reconstruction.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_6.pdf}
\caption{The real and reconstructed particle trajectories through VOI in the Geant4 simulated tomography system. The designating scores to the voxels that the reconstructed trajectory passes through shown.}\label{Fig:6}
\end{figure}
It is relevant that one has different options how to combine the detector hit data in the different plate detectors. Not all the reconstructed trajectory candidates are complete, i.e. with hits in every plate detector of the entrance and the exiting hodoscope. One can have different strategies how to deal with incomplete trajectories. For simplicity, we remove all the candidates that have been registered only in the entering hodoscope. If a candidate has hits in every plate of the entering hodoscope and at least a hit in a plate of the exiting hodoscope (detector plates 1, 2, 3 and 4 or 5 or 6, see Fig.~\ref{Fig:6}) it forms so-called incomplete trajectory candidate. To apply the scattering based backprojection, at least the two hits in the plates are needed in the entrance and the exiting hodoscopes. We consider the complete and incomplete candidates in the reconstruction.
After the considering trajectory candidates we apply PTF as described in Sec.~\ref{Sec:2}. We remind that the PTF procedure classifies the candidates into three or more PTF groups. After that we feed the groups to the optimised tomographic reconstruction procedure per group. For example, we can apply different back-projection parameters on different groups.
The third step of the tomographic reconstruction is the composing of VDM. To reduce the amount of possible false trajectories we define the upper and lower limit values for the total scattering angle. Thus so far our reconstruction results the number of VDMs depending on the set of the PTF classification groups, back projection parameters etc.
Fig.~\ref{Fig:7} and \ref{Fig:8} shows the effect of filtering parameters on the composing of VDM. The figures compare the two reconstructions of VOI (VDMs) with and without the hodoscope and VOI filtering using scattering and transmission parameter, accordingly. The figures show the VDMs at z=100~cm, where the plane crosses the interrogated object at the half height. The object is a cube of water with the side length of 10~cm centered in the middle of VOI. The estimated real measurement time is approximately 18 minutes, the voxel size $1 \times 1 \times 1$~cm$^3$. The upper panels present the reconstruction procedure with no filtering. The lower panel of Fig.~\ref{Fig:7} presents the reconstruction, where the PTF extracts the highest momentum fraction (F1, see Fig.~\ref{Fig:3}) dominated by muons. The lower panel of Fig.~\ref{Fig:8} shows the reconstruction results, where the PTF extracts the lowest momentum fraction (F3, see Fig.~\ref{Fig:3}) of muons and the most of the electrons. We obtained both the figures from the same Geant4 simulation.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_7.pdf}
\caption{The effect of hodoscope and VOI filtering. A cross-sectional Geant4 snapshots on horizontal plane ($z = -100$) of a cube of water ($10\times10\times10$~cm$^3$) without (upper) and with (lower) PTF. The hodoscope filtering range $0 < \theta < 2$~mrad.}\label{Fig:7}
\end{figure}
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_8.pdf}
\caption{The effect of hodoscope and VOI filtering. A cross-sectional Geant4 snapshots on horizontal plane ($z = -100$) of a cube of water ($10\times10\times10$~cm$^3$) without (upper) and with (lower) PTF. In the lower panel, the hodoscope filtering range is $10 < \theta < 30$~mrad.}\label{Fig:8}
\end{figure}
The figures demonstrate that despite the reduced flux in the lower panels the filtering can significantly increase the contrast of VDMs. In those examples we have separated the two physical effects: the absorption and scattering. Thus the different settings of the filtering parameters and ranges can potentially increase the discrimination capability for object detection and material classification.
\subsection{Object detection}
In this subsection we describe the algorithms we developed to extract object shapes from VDMs. A crucial property of the algorithms is that they have to operate well in high Poisson noise conditions. We also notice that due do proprietary information the methodology of the edge detection procedure is disclosed with some generality.
We start the object reconstruction on the 2D data. An algorithm enhances the edges of objects from VDM iteratively. After several iterations, the image sharpness is enhanced to a level, where we apply an image threshold to extract the binary image, the final reconstructed shape of the object. Fig.~\ref{Fig:9} shows a double-object example, a smaller inner cube of RDX-explosive with the side length of 10~cm surrounded with a larger cube of flesh-like material with the side length 30~cm. As described above we apply PTF. In this case the VDM represents the distribution of mean scattering angles. The spatial resolution of the plates in the Geant4 simulation was 0.1~mm. The estimated measurement time in the Geant simulation is 18~minutes.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_9.pdf}
\caption{Example of the double-object detection algorithm. From top to down, the original reconstructed image down to the detected objects through the iterative object detection procedure in case of double-object system: a cube ($10\times10\times10$~cm$^3$) of RDX explosive material surrounded by a larger cube of flesh ($30\times30\times30$~cm$^3$). The estimated measurement time in the Geant simulation is 18~minutes. The voxel size is $1\times1\times1$~cm$^3$, the spatial resolution of the detector plates 0.1~mm.}\label{Fig:9}
\end{figure}
After the 2D reconstruction, the results are fed to the 3D reconstruction procedure. To differentiate between the object candidates, we introduce a labelling step that gives a unique index to the connected volumes. This can be achieved having so-called Connected-Component Labelling (CCL), where the subsets of connected components are uniquely labelled using, e.g., Two Pass Algorithm [33]. Fig.~\ref{Fig:10} presents a half-cut of reconstructed and detected objects (for the same simulation as in Fig.~\ref{Fig:9}). The result indicates clearly that the VDMs of the two materials are distinguishable enough for the edge detection and the two different logical volumes have been established. The two different logical volumes are denoted: yellow for RDX, orange and black edges for flesh.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_10.pdf}
\caption{The object reconstruction and detection of a double object with spectroscopic filtering relying on the scattering density parameter. The double object: a cube of RDX explosive material ($10\times10\times10$~cm$^3$) centered in a cube of flesh ($30\times30\times30$~cm$^3$), the voxel size is $1\times1\times1$~cm$^3$. The object detection and separation to the two logical volumes are denoted by yellowish colors. The spatial resolution of the detector plate is 0.1~mm.}\label{Fig:10}
\end{figure}
We tested the same object configuration with a shorter, 2-minutes estimated, measurement time. Fig.~\ref{Fig:11} shows the results. One can clearly see the loss of contrast. Due to lower statistics, we increased the voxel size from 1~cm$^3$ to 9~cm$^3$. However, the edge detection algorithm is still able to detect and separate the two materials as different logical volumes.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_11.pdf}
\caption{The object reconstruction and detection of a double object with spectroscopic filtering relying on the scattering density parameter. The double object: a cube of RDX explosive material ($10\times10\times10$~cm$^3$) centered in a cube of flesh ($30\times30\times30$~cm$^3$), the voxel size is $3\times3\times3$~cm$^3$. The object detection and separation to the two logical volumes are denoted by yellowish colors. The estimated measurement time 2 minutes and the spatial resolution of the detector plate 0.1~mm.}\label{Fig:11}
\end{figure}
Once we establish and separate the logical volumes of different object we can apply special algorithms for the material or other classifications on the reconstructed volumes. Those will be the subject of future work.
\section{Design of the physical prototype}
\label{Sec:4}
\subsection{Technical description}
As described above we developed the PTF and MMTR methods based on the virtual Geant4 models. To prove (and improve) the methods we built a simplified physical prototype system. Fig.~\ref{Fig:12} shows the principal design of the physical prototype. The prototype has an upper hodoscope with three detector plates based on the above described fiber-mat technology. The lower hodoscope has been replaced with a single detector plate. This design means the system cannot measure the total scattering angle. However, one can estimate the total scattering angle by calculating the angle between the vectors BC and CD, see Fig.~\ref{Fig:6}. This angle depends on the position of the object in a vertical axis, thus one can do a comparison material classification experiment fixing objects into the same position in every measurement. An another shortcoming of the minimal design is that the detector plates cover only a small fraction of the total solid angle around VOI and only a limited fraction of the flux is available. The detector plate size is $25 \times 25$~cm$^2$, the distance between the two adjacent plated in the upper hodoscope is 7.5~cm. The volume of VOI is $25 \times 25 \times 25$~cm$^3$. Thus the field of view is limited to 64~degrees in $x$- and $y$-axis. The limited field of view results in low vertical reconstruction accuracy if no correction algorithm implemented.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_12.pdf}
\caption{The general spatial design with the detector plates and the structural elements. An example object (gray cubic) is in the middle of the detector between the upper hodoscope (the three horizontal plates) and the lower single detector plate (a plate below the cube). The fiber-mat couplings with SiPMs nor the DAQ-system are not shown.}\label{Fig:12}
\end{figure}
First, we studied the detection and reconstruction performance of the prototype in the Geant4 simulations. We describe the main detection performance parameters, detection efficiency and spatial resolution, in the following subsections.
\subsection{Particle tracking in the fiber-mat system}
In this section we explain how we build up a particle track candidate form hit candidates (lit fibers). First, we explain our notations and the coordinate system. For simplicity, we deal with the $x$- and $y$-directions separately and so below we show only a direction if sufficient. Fig.~\ref{Fig:13} shows the fiber schematics and the notations through the four plates of the prototype in a selected direction ($x$ or $y$). As described in Sec.~\ref{Sec:2} and shown in Fig.~\ref{Fig:1} a plate has the four fiber layers in total arranged to the two orthogonal bi-layers. We address the fiber layers as h$AB$ where $A$ denotes the plate and $B$ denotes the layer in the plate.
\begin{figure}[ht]
\centering\includegraphics[width=0.6\linewidth]{figs/Figure_13.pdf}
\caption{The labeling schema of the bi-layered scintillating fibers of the detector (showing the $y$-directional by-layer only).}\label{Fig:13}
\end{figure}
For example, h11 denotes the plate 1, the y-directional bi-layer and the first layer. h14 denotes the plate 1, the x-directional bi-layer and the second layer. For the z-coordinate of the hit candidate we define a virtual plane between the bi-layers. So the layers h11 and h12 are above the virtual $z$-plane and the layers h13 and h14 are below the $z$-plane. We define the component of $x$ or $y$ layers as the fiber-centre offsets from the $z$-plane. The following parameters describe the fiber: the core diameter, the width and the pitch of the cladding. The pitch is the distance between the adjacent fiber centres. As we know all the positions and other parameters of the fibers we can compute the centres of the fibers in the well-defined 'absolute' Cartesian coordinate space. We enumerate the fibers starting from 0 to $N$ in a layer, where $N$ is the max number of fibers in the layer. Thus detecting a fiber lit we can then convert it the 'absolute' coordinates.
To reconstruct a plate hit we handle the $x$- and $y$-directions separately. For example, Fig.~\ref{Fig:14} shows the reconstructed hit for the $y$-direction from the three lit fibers in the $y$ by-layer. We compute the hit coordinate finding the weighed arithmetic mean of fiber centres lit. In the figure, red denotes the fibers lit and the black cross is the hit coordinate reconstructed. The same calculation is performed for the $x$ by-layer. The red line is an example of particle track candidate reconstructed having some other hit candidates from the other plates.
\begin{figure}[ht]
\centering\includegraphics[width=0.75\linewidth]{figs/Figure_14.pdf}
\caption{The reconstruction of the hit candidate for the $y$-directional by-layer. The red denotes the lit fibers. The cross depicts the reconstructed hit candidate. The red line is an example of particle track candidate reconstructed having some hit candidates from the other plates.}\label{Fig:14}
\end{figure}
The $x$ and $y$ coordinates of a hit reconstructed (at least) in the two planes we can project the results to the $z$-plane. We use the linear extrapolation based on the reconstructed track candidate. Fig.~\ref{Fig:15} shows the extrapolation results (black circled crosses). Combining the component values from the both axes projected to same $z$ plane, we have reconstructed the 3-dimensional coordinate of a hit candidate.
\begin{figure}[ht]
\centering\includegraphics[width=0.75\linewidth]{figs/Figure_15.pdf}
\caption{The hit extrapolation to the plate-z. The red fibers represent fibers generating signal, the cross depicts the calculated particle hit coordinate, the circled cross indicates the extrapolated coordinate.}\label{Fig:15}
\end{figure}
\begin{figure}[ht]
\centering\includegraphics[width=0.75\linewidth]{figs/Figure_16.pdf}
\caption{An clustering example, the red fibers represent the fibers generating signal.}\label{Fig:16}
\end{figure}
For the track reconstruction we need a clustering algorithm to cluster the lit fibers from the weighted mean of fiber centres computed. Fig.~\ref{Fig:16} shows some cases of clustered lit fibers to illustrate the approach. Clustering algorithm works in the two stages. First, we cluster the lit fibers from the single layers. We form clusters of consecutively lit fibers. In the layer h11 the fibers are {1, 3, 4} and in the layer h12 the fibers {0, 2, 5}. Second, we merge the clusters from the different layers if they intersect component-wise: either the $x$ or $y$ coordinates depending on the axis are handled. Fig.~\ref{Fig:16} shows this as the dotted intersecting region. We compute the minimum extreme of a cluster by taking the leftmost fiber, computing its intersections in the fiber layer $z$-plane and taking the minimum value. The same logic is followed to calculate the maximum extreme value.
\subsection{Geant4 simulations for the prototype}
We tested the hit and track reconstruction algorithms described in the last sub-section in the Geant4 simulations of the prototype. To estimate the quality and effectiveness of the algorithms we collected the following information of the muon/electron hits in the virtual detector plates: (i) the hit coordinates and (ii) the number of fiber(s) providing the signal from the same event for the every detector plate and layer. We generated the simulation data with a statistically sufficient number of events having the CRY-based source described above. Then we calculated the distance differences between the Geant4 provided coordinate values (i) and the coordinates provided by our algorithms calculated for every $x$ and $y$ layer and all the simulated events. We calculated the distributions and the standard deviations, some results are presented in Table~\ref{Tab:1}. The numbers in the table are based on the 13 699 simulated hit events: 10 907 muon events and 2792 electron events. The distribution has its mean at zero and it corresponds to the Laplace distribution.
\begin{table}[h!]
\begin{tabular}{c|c c c}
&
\multicolumn{3}{c}{\textbf{Stdev} ($\mu$m)} \\
\cline{2-4}
\textbf{Detector plate} & \textbf{Muons} & \textbf{Electrons} & \textbf{All} \\
\hline
Plate 1 & 113 & 1234 & 567 \\
Plate 2 & 110 & 726 & 343 \\
Plate 3 & 153 & 840 & 1112 \\
Plate 4 & 109 & 2769 & 1261 \\
\hline
\textbf{Average} & \textbf{121} & \textbf{1392} & \textbf{821} \\
\end{tabular}
\caption{Standard deviations between the hit locations and reconstructed hit locations from the Geant4 simulations of the prototype tomography system.}
\label{Tab:1}
\end{table}
In the case of muons, the standard deviation varies slightly above 100~$\mu$m. The reason of this rather large value is in a limited set of extreme cases, which have 1-2 mm deviations from the actual hit coordinates. In the case of electrons, we expect larger deviations due to more scattering events of electrons in the detector plates, decreasing the performance of the track reconstruction algorithm. However, the number of strongly deflecting extreme event is considerably small. If we remove some tens of most deflecting events, the standard deviation falls down to a few hundreds of $\mu$m.
To test the performance of the tomographic reconstruction in the Geant4 simulations we had different cube shaped materials in VOI. We fixed the location of the object in the same position size and had the $6 \times 6 \times 6$~cm$^3$ cube shaped objects. We centered the object in the horizontal plane and located the objects vertically 20~mm from the lower detector plate of the upper hodoscope. Fig.~\ref{Fig:17} and \ref{Fig:18} show the reconstruction results of the three different materials. To create the VDMs we applied the mean scattering angle. The estimated real measurement time was 30~minutes.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_17.pdf}
\caption{The reconstructed logical volumes by the prototype tomography system. The scanned object (plexiglass, aluminium, iron) has the size $6\times6\times6$~cm$^3$. The voxel size $0.5\times0.5\times0.5$~cm$^3$ and the estimated measurement time 30~minutes. The scale of the coordinates is denoted in voxels (e.g. 10 corresponds to 5~cm).}\label{Fig:17}
\end{figure}
The horizontal cross-sections have been reconstructed quite accurately, for lead object it is overestimated about 10~mm, for Plexiglas 10~mm less than its actual size. Due to the limited angle tomography, reconstructed objects are vertically elongated about 2-3 times their actual size.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_18.pdf}
\caption{The horizontal cross-sections of reconstructed logical volumes at the half-height of the object applying (the same reconstruction schema as in Fig.~\ref{Fig:17}). The scanned object (plexiglass, aluminium, iron) has the size $6\times6\times6$~cm$^3$. The voxel size $0.5\times0.5\times0.5$~cm$^3$ and the estimated measurement time 30~minutes. The scale of the coordinates is expressed in voxels (e.g. 10 corresponds to 5~cm).}\label{Fig:18}
\end{figure}
\section{The physical prototype and the first measurements}
\label{Sec:5}
\subsection{Construction of the prototype}
Based on the design of the prototype tested with Geant4, we constructed the physical prototype to have a physical validation of the proposed detector systems, PTF and MMTR. As in the case of virtual model, the physical prototype (see Fig.~\ref{Fig:19}) has the upper hodoscope with three detector plates and the lower hodoscope with single detector plate. We have the detector layers with the 1~mm round single cladding scintillation fibers (Saint-Gobain BCF-12) assembled as shown in Fig.~\ref{Fig:1}. The active area of each detector plate is $247 \times 247$~mm$^2$. We spaced the detector plates in the upper hodoscope with 75~mm pitch, the distance between lower plate of the upper hodoscope and the lower hodoscope is 250~mm and thus the dimensions of the VOI are $247 \times 247 \times 250$ mm$^3$.
We installed the detector plates into a rigid inner frame that fixed the aluminium support frames of the detector plates. The external frame is used to fix the positions of the SiPM arrays and the DAQ boards. On top of the upper hodoscope and below the lower hodoscope, we assembled the two trigger detectors (scintillator plates).
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_19.pdf}
\caption{The prototype tomography system is encased into a light-tight box. The upper casing plate removed for the photograph.}\label{Fig:19}
\end{figure}
We assembled the detector plates using grooved alignment tables (see Fig.~\ref{Fig:20}). This guarantees the accurate and controlled positioning of fibers in the fiber mat. The alignment tables have been produced from 1~mm thick sheets of phenolic paper laminate (PFCP201) by milling the grooves with a pitch of 1.100~mm. We glued the fibers on the alignment table using UV-curing glue (Loctite AA 3311). We added TiO2 powder to the glue to reduce optical leakage between the fibers.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_20.pdf}
\caption{The assembly of a detector plate. The inset shows the magnified front image of the grooved surface of the support plate and the fiber positions in the fiber mat.}\label{Fig:20}
\end{figure}
We assembled the two separately produced double-layered fiber mats orthogonally to each other in order to form a 2D-sensitive four-layered detector plate. We supported the assembly with an aluminium frame that provides sufficient mechanical rigidity, flatness and fixation with the frame. At the corners of each PFCP201 plate we have the alignment holes allow to align the two orthogonal 2D fiber mats against each other and to align the four detector plates with respect to each. For the latter we used the alignment rods once the detector plates were mounted in the frame. The structure ensures a mechanical tolerance for relative plate positions below $\leq0.1$~mm. We estimate the mechanical tolerance of the fiber positioning on the fiber mat to be below 0.01~mm and the orthogonality of the two fiber mats below 0.001~degrees.
\subsection{Data Acquisition System}
We composed the data acquisition system of the prototype implementing eight CAEN DT5550W boards. We use Ketek (PA3325-WB-0808) and Hamamatsu (S13361-3050AE-08) SiPMs arrays to collect scintillation light from the fibres.
Two plastic scintillator plates trigger the data readout. The active area of a trigger is $30 \times 30$ cm$^2$, the upper one placed on top of the upper hodoscope and the second one placed under the lower hodoscope. The triggers have two $6 \times 6$~mm$^2$ SiPMs (Hamamatsu S13360-6050CS) using the SiPM Readout Front-End Board CAEN A1702. When a muon/electron passes the trigger system (both plastic scintillator plates) a trigger signal is created that initiates the readout of all the CAEN DT5550W boards.
\subsection{Data analysis: testing PTF and MMTR}
Having the physical prototype our first objective was to test the detection efficiency of the completed particle tracks. The efficiency is defined as a completed particle track candidate reconstructed if the triggered signals are collected from all the 8 bi-layer fiber mats. The average count rate from the triggering system is about 3~Hz indoor (the lab room has concrete walls and ceilings and one additional floor above). Considering the triggering detector area (0.09~m$^2$) and the field of view (66 degrees) of the double trigger plate telescope, the average count rate is in the expected range. We performed a 30-minute measurement with the empty VOI volume to estimate the efficiency of the completed tracks. In the test we recorded the 6594 triggered events, out of which the 3562 were completed tracks. It means the total efficiency for the completed particle tracks is around 50\%.
Fig.~\ref{Fig:21-22} shows the comparisons of the reconstruction having the different filtering conditions on the 2500 completed trajectories. The voxel size was $0.5 \times 0.5 \times 0.5$~cm$^3$, the reconstruction parameter was the mean scattering angle. Due to the limited field of view, the horizontal cross-sectional images are a little unclear, different from the simulation results of the full model of the tomography system. After we applied PTF, the contrast of the reconstructed images improved, especially in the case of light materials.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_21-22.pdf}
\caption{The horizontal cross-sections with the different filtering conditions for a cube of water and aluminium ($6\times6\times6$~cm$^3$): no hodoscope nor VOI filtering applied (left); no hodoscope filtering applied (middle); hodoscope and VOI filtering applied (right).}\label{Fig:21-22}
\end{figure}
Having the optimised PTF and MMTR parameters, we reconstruct the 3D images for the three different cases: a cubic of aluminium, Plexiglas and ammonium nitrate (in granular form) with the dimensions $6 \times 6 \times 6$~cm$^3$ (see Fig.~\ref{Fig:23} and \ref{Fig:24}. The reconstruction parameter for VDMs was the mean scattering angle.
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_23.pdf}
\caption{The detected logical volumes as a result of the reconstruction of a cube samples ($6\times6\times6$~cm$^3$) of plexiglass (upper left), aluminium (upper right) and steel (lower left). The samples were centred 2~cm below the lower detector plate of the upper hodoscope. The red region represents the logical volume for the detected object. The scale of the coordinates is in cm.}\label{Fig:23}
\end{figure}
\begin{figure}[ht]
\centering\includegraphics[width=1.0\linewidth]{figs/Figure_24.pdf}
\caption{The horizontal cross-sectional cuts of the same samples as in Figure~\ref{Fig:23}. The cut is at the half-height of the object and the reconstruction schema same as in Figure~\ref{Fig:23}. The scale of the coordinates is in cm.}\label{Fig:24}
\end{figure}
The reconstruction results are vertically elongated as we predicted from the Geant4 simulations and vary between 15-25~cm. Horizontally, the side length vary between 7-8~cm. The dimensions of the objects are overestimated, which is expected due to limited angle tomography. Theoretically, one can apply a special correction if needed.
\section{Discussions and conclusions}
In this paper, we discussed on the two new methods in muon tomography: PTF (the classification of particle candidates in the hodoscope) and MMTR (the creating of multi-VDMs following the classification schema of PTF, etc). Our main motivation is to improve object reconstruction and material classification for low-Z materials in the case of ART. We applied the methods on the hodoscope system based on plastic scintillating fibers. We proposed the detector technology to ensure the required detector performance, mainly the spatial accuracy, the sensitivity and the lightness of the detector system. First, we presented the results based on the Geant4 simulations.
Second, we validated the methods on the minimal table-top prototype tomography system. We tested the performance of PTF and MMTR on the prototype having different low-Z materials. We showed that applying the proposed methods we can successfully detect the low-Z objects in a 10–30-minute time scale. For example, we tested the objects composed by water, Plexiglas, ammonium nitrate (mimicking explosive material). Our Geant4 simulations indicate that one can reduce the exposure time to a few minutes range having the full-scale tomography system (see Sec.~\ref{Sec:3}) based on the same technology like the physical minimal prototype.
The particle tracking algorithm determining the particle hit location in a detector plate and its trajectory in the hodoscope provide the accuracy 120~$\mu$m for muons. We demonstrated that VDM can be analysed as a set of 2D images applying an edge detection algorithm on the VDM. It allows to detect the objects as separated logical volumes. The method works efficiently in the case of high Poisson noise, hence applicable for short measurement times.
When logical volumes reconstructed one can be interested in the material classification of volumes. The possible classification procedure can follow as a single procedure or having the multiple procedures, e.g., one per PTF group. For example, if one combines the classification based the filtered fractions F1 and F3 (see Fig.~\ref{Fig:1}), the discrimination ability of materials increases including both the scattering and transmission data simultaneously. Furthermore, establishing the reconstructed objects as discrete entities enables to apply material classification techniques based on machine learning.
One of our next steps is to develop the material classification algorithms on VDMs. Some potential methods to start the classification study would be machine learning based such as the Kullback-Leibler Divergence (KLD) or the Support vector Machine (SVM). Another future step will be the reconstruction of complex objects representing more realistic user cases. For example, we can implement some iterative reconstruction methods in order to increase the ability of generating VDMs for complex scanning objects.
The PTF and MMTR methods combined with the proposed detector system improve the reconstruction of VOI significantly. The combination makes the detection and potentially classification of low-Z composed materials possible in small-scale detector system. It opens up new routes for the commercialization of ART tomography. The developed methods and the detector system is interesting in the context of industrial security applications such as baggage scanning in airports and cargo and shipping container scanners for illicit materials and contraband.
\section*{Acknowledgements}
We would like to thank Andrea Giammanco for very helpful discussions and comments. This work was continuously and strongly supported by the startup company GScan OÜ. This work was supported by the Enterprise Estonia, the grant EU48693. AH thanks the Estonian Research Council for the grant PRG434 and the EC for the ERDF CoE program project TK133.
\bibliographystyle{elsarticle-num-names}
|
\section{Introduction}
The use of Triplet loss~\cite{schroff2015facenet} was popularized with the introduction of the FaceNet architecture which was aimed to face identification tasks. This loss allows to train a neural network, commonly a Convolutional Neural Network (CNN), to produce a vector representation of an image. The goal is that the neural network learns a mapping from face images to an euclidean space, it is desirable that distances among face image positions directly correspond to a measure of face similarity. Within this setting a desired outcome is that images of the same face will cluster together and images of different faces will be separated by a margin. Triplet loss has been applied into different types of images: objects~\cite{wang2014learning}, person re-identification~\cite{cheng2016person}, information retrieval~\cite{hoffer2015deep}.
Recently, Triplet loss has been proposed for speech tasks, for instance: the speaker verification task ~\cite{zhang2018text}, for speaker turn~\cite{bredin2017tristounet}, speech emotion classification~\cite{huang2018speech}, among other tasks. However, its applicability to forensic speaker identification has not been explored, particularly for the Spanish language. \emph{Forensic Speaker Identification} (FSI) focuses on gathering and quantifying the evidence that will be presented in a court. FSI addresses the question if a specific recording registers or not, speech produced by a specific person~\cite{rose2002forensic,morrison2019forensic}. The more basic scenario in FSI consists of two sample speech recordings, a reference sample and a questioned sample. For the reference sample, we always know the identity of the speaker. This certainty is guaranty by the chain of custody, this is because we know the conditions in which the recording was taken, including the identity of the speaker. On the other hand, for the questioned recording we are not sure about the identity of the person whose voice is in the recording. In a case that involves FSI, the identity of the speaker in the questioned recording is contested regarding the identity in the reference recording; one of the involved parts affirms that the voice in the reference and the questioned recordings are the same (\emph{same-speaker hypothesis}); while the other part affirms the contrary (\emph{different-speaker hypothesis}).
The goal in a FSI case, is not only about matching the two recordings by their \emph{similarity}, same-speaker hypothesis, FSI requires a stronger legal standard which makes also necessary to quantify the chances of the questioned sample to be associated to other speakers of the population, this address the different-speaker hypothesis. This measurement is known as \emph{typicality}. With these two measurements, similarity and typicality it is common to calculate the Likelihood Radio (LR)~\cite{good1991weight}. LR offers a quantifiable measurement that updates the odds of one of the hypothesis, this information should be taken into consideration in the context of the presence of other evidence regarding the case.
In this work, we propose to extract speech embeddings from speech spectrogram samples for reference and questioned speech recordings in order to quantify the LR. We start by presenting related work in section~\ref{sec:rw}. We continue by presenting the details of our neural model and the implementation of the triplet loss in section~\ref{sec:nm}. We present two sets of results: first, in section~\ref{exp1}, we measure the quality of the embeddings by proposing inner and outer speaker distance metrics, together with the use of the well established silhouette clustering metric. Second, in section~\ref{exp2}, we propose two ways to calculate the LR in terms of the speech embedding distances. Once we show that speech embeddings are an option to be used in FSI we discuss some ethical aspects to be considered in section~\ref{sec:ethics}. Finally, we summarise our main findings in section~\ref{sec:con}.
\section{Related work}
\label{sec:rw}
Forensic speech science has advanced in last three decades in which it has established different methodologies and techniques to face the speech identification problem~\cite{morrison2019forensic}. In the case of the methodology, there has been a paradigm shift towards empirically grounded methods~\cite{saks2005coming,morrison2009forensic}. This shift has been motivated by the requirements of admissibility of science evidence that had become a standard in some courts around the world~\cite{giannelli1980admissibility,champod2011scientific}. The main result of this shift has been the adoption of the Likelihood-Ratio (LR) as a mean to introduce the evidence in court. LR is formulated in the following manner:
\begin{equation}
LR=\frac{p(E|H_s)}{p(E|H_d)}
\end{equation}
where $E$ represents the evidence, in FSI this is the quantification of speech properties in the questioned recorded sample. $H_s$ correspond to the same-speaker hypothesis and $H_d$ to the different-speaker hypothesis. The numerator can be considered a similarity score while the denominator a typicality one. LR is not to be considered independent of the other facts of the case, on the contrary its meaning only depends on the strength or lack of the rest of the evidence and their compatibility with either of the hypotheses.
On the other hand, from the point of the techniques there has been several proposals which allow the quantification of $p(E|{H_s})$ and $p(E|{H_d})$. One approach that was extensively explored was the statistical analysis supposing a Gaussian distribution of speech features~\cite{rose2002forensic}. A common approach is to measure a specific phonetic and phonological speech properties (e.g., formants) in a specific context (e.g., a word \cite{rose1998forensic,rose1999long}). Motivated by the application of multi-variable statistics in the forensic field~\cite{aitken2004evaluation} new approaches were suggested together with kernel approaches to improve the statistical analysis of the speech evidence~\cite{rose2004linguistic,morrison2011comparison}.
Another proposed method is the use Gausian-Mixture-Model/Universal Background Model (GMM-UBM)~\cite{bimbot2004tutorial}. This methodology reflects more a generative machine learning approach which depends on data intensive algorithms, for this reason in these methodologies there is no necessity of measuring a specific voice property. The GMM-UBM model is parametric deppends on a dataset of recordings which is used to define the model parameters, in this case the parameters of the Gaussian mixtures. A GMM speaker-specific model is created to quantify the similarity term, and a UBM general model, based on the population of possible speakers, is generated to quantify the typicality term. With the advent of different machine learning techniques some other machine learning-based approaches had been proposed, such as using Support Vector Machines, boosting algorithms and Random Forest~\cite{univaso2015data,morrison2020statistical}. A fundamental piece to adopt ML techniques is its appropriateness to calculate the LR~\cite{1202302}. Recently, there has been proposals that exploits the discriminative power of Neuronal Networks and their capability of producing representations, some examples of these approaches are: DNN senone i-vectors~\cite{garcia2014supervised}, bottleneck features~\cite{yaman2012bottleneck} and x-vectors~\cite{snyder2018x}.
On the other hand, with the progress of self-supervised methods for training deep learning networks there has been advances in proposals for learning representations for embedding spaces. Contrastive methods have been proposed to compare a questioned sample to a set of samples of known speakers, this setting correspond to a setting of \emph{speaker verification} were there is access to samples of the speaker to identify, it is a matter of verifying if the two sets of samples match~\cite{heigold2016end,snyder2018x}. Of particular interest are the advances with the Triplet Loss, since it maps raw speech representation into an euclidean space~\cite{schroff2015facenet}. The use of these embeddings for the speaker identification task is trivial since the distance among embeddings can be used to determine which speaker is close to a questioned recording. Here it is important to notice, that speaker identification does not quantify typicality, but just similarity. This approach had been used in different scenarios: \cite{li2017deep} uses a Residual CNN and GRU that transforms a spectogram into a $512$ dimension vector, it proposes to use cosine similarity to guide the triplet loss and it evaluates accuracy and error to recognise the speaker. \cite{zhang2017end} presented a CNN Inception Resnet to generate a $128$ dimension embedding, it focused on $L2$ norm, it proposed validation and false accept rate to evaluate its system. \cite{mo2020self} modifies the triplet loss to make it more efficient, it also uses accuracy (top 1 and top 5) to evaluate its system. In all these three cases, the resulting performances were superior to previous approaches.
Of particular interest for our experimentation is the difference between female and male voices, since as it will be presented further down that our dataset has an imbalance among this type of speakers. According to the medical notion, the voice originates in the throat of the speaker specifically in the larynx. There is an understanding that the size of the larynx correlates with sexual characteristics which at the same time determine the sex (or pitch) of the voice as an acoustic event. In average, the male larynx is larger than the female larynx and is naturally inclined to produce a lower pitched speech. However this is not a rule since for men and women speech can overlap. In this regard, the sexual characteristics of the vocal tract, biologically determined, determine sex in speech, however the identification phase is of one of gender, since it is constituted by a subjective interpretation~\cite{azul2013voices,mcdermott2011distinguishing,ertam2019effective}. With this in mind our experimentation will be on perception which means on gender.
\section{Triplet loss and Neural Model}
\label{sec:nm}
Triplet loss compares an anchor input with two other inputs, a positive input which shares a property with the anchor, in our case it is the identity of the speaker, and with a negative input which does not share such property. The comparison is guided by the following formulation:
\begin{equation}
L(A,P,N) = max( D(A,P) - D(A,N) + m, 0)
\end{equation}
where $A$, $P$ and $N$ are vectors representing the anchor, the positive and negative inputs respectively. $D$ is distance metric and $m$ is a margin. In a ideal setting it is expected that the distance between $A$ and $P$ to be less than distance between $A$ and $N$ at least by a margin $m$, if that is the case the loss is zero when the network calculating the vectors is doing a good job. However, if this is not the case the loss will be positive and by using back propagation the weights of the model that produces the vectors from raw information will be adjusted. Figure~\ref{fig:tripletloss} shows the relation between the CNN model and the triplet loss, it is important to notice that the CNN blocks are the same neuronal network that transforms all inputs since the weights are shared. The Figure also illustrates the three cases in relation to distances among the embeddings and the margin. $L2$ norm is used as a distance metric.
\begin{figure}[t]
\centering
\fontsize{5}{5}\selectfont
\includegraphics[width=0.8\textwidth]{pictures/Tripletloss.png}
\caption{Triplet loss applied to three speech segments which are transformed into three vectors. Possible cases relation among A, P and N inputs, the third case is a zero loss case.}
\label{fig:tripletloss}
\end{figure}
A common arrangement in Triplet loss is to use the same neuronal model to produce the vector inputs from raw inputs. In this work we propose the use a CNN that will receive segments of the spectogram of speech recordings. Figure~\ref{fig:cnn} shows this arrangement. We propose a simple CNN composed by $5$ layers when possible otherwise $4$. Figure~\ref{fig:cnn} shows a diagram with the specific details of our model.
\begin{figure}[t]
\centering
\fontsize{5}{5}\selectfont
\includegraphics[width=0.8\textwidth]{pictures/CNNtripletloss.png}
\caption{Convolutional network}
\label{fig:cnn}
\end{figure}
\subsection{Preprocessing of speech signal}
The input of our CNN is a segment of a spectogram that represents a time slice ($t$ in milliseconds) and frequencies information up to $8.5 kHz$ (enough to characterise the human voice), using always $256$ bins for the frequencies. All our recordings are down sampled to $16$K and the spectogram is normalized and pre-amph. These specific choices are followed from typical pre-processing of speech signals. In order to obtain the spectogram we use a \emph{hann} window, with variable window ($w$ in milliseconds) and hop ($h$ in milliseconds) size. The parameters $t$, $w$ and $h$ allows us to generate an image patch (spectogram segment) with a variable width but a constant height $W\times 256$. This patch is the CNN's raw input which will be transformed it into a $1024$ dimension embedding for all our experiments.
\section{Dataset}
\label{sec:dataset}
In this work we use the Spanish Voxforge dataset, which is entirely based on the recordings from the Voxforge Project\footnote{\url{http://www.voxforge.org}}. The Voxforge Project is a non-profit initiative that aims to collect transcribed speech for use with Free and Open Source Speech Recognition Engines. We chose the Voxforge because the speakers read always the same prompt, a paragraph of \emph{El Quijote}, this eliminates overfitting by the content of what it is said. We expect that our models focus on properties of how things are said. The speakers donated a sample of their voice by registering to the project website, filling a form with relevant information about the speaker and reading some prompts directly through the speaker's computer microphone. Thanks to this mechanism, we can know the following about every speaker, which is relevant for forensic purposes and our experimentation:
\begin{itemize}
\item Username: It could be left in blank or it could be an alias
\item Gender: Male / Female
\item Age: $13-17=$ Youth / $18-64=$ Adult / $65$-or greater$=$ Senior)
\item Native Speaker?: Yes / No
\item Dialect: Country or Region
\end{itemize}
Since the beginning of the project in $2006$, several languages have been added by the community, making necessary to clarify that our corpus only contains Spanish recordings until the year $2016$. The original recordings were manually segmented into utterances. Table~\ref{tbl:nationality} shows the total number of male and female speakers and how they are classified into the corpus according to their nationality.
\begin{table}[ht]
\centering
\begin{tabular}{|l|c|c|}\hline
Country & Males & Females\\\hline
Argentina & $143$ & $31$ \\
Chile & $69$ & $3$ \\
Latin America & $148$ & $10$ \\
Mexico & $68$ & $8$ \\
Spain & $1240$ & $411$ \\
Unknown & $ 45$ & $4$ \\\hline
Total & $1713$ & $467$ \\\hline
\end{tabular}
\caption{Number of Speakers in the Voxforge Corpus and their Nationalities.}
\label{tbl:nationality}
\end{table}
The Spanish Voxforge dataset is composed by $21,692$ recordings which in average last $8.25$ seconds. They come in a $16kHz$, $16$ bit, mono format. The total duration of the whole recordings is approximately $50$ hours. For the experiments, we split the speakers in $80\%$ training, $10\%$ validation and $10\%$ testing.
\section{Experiments and results}
We performed two levels of experimentation. First we explored the speech embeddings by measuring quality of the speech embeddings and the cluster they generate on samples from the same spaker. Second we evaluated the capability of the speech embeddings to be used to calculate an LR, we propose two approaches: one is to use directly distance and questioned as a proxy for the LR, second is to use a ratio between the previous distance divided by the shortest distance to other speaker in the population.
\subsection{Quality of the embeddings}
\label{exp1}
To quantify the quality of the embeddings, we calculate three metrics: inner average distance for the same speaker samples ($IAD$), outer average distance between speaker and centroids of other speakers ($OAD$). We expect a IAD small which will signal that samples from the same speaker will land in the same region. For OAD, we expect a large number that will signal that the samples from a different speaker will be far away from other speakers. With these two metrics we proposed to calculate a distance ratio ($DR$) that will tell us the relation between a speaker and the rest of speakers, we will expect a small amount for $DR$ signaling good speech embeddings. We also calculated the mean silhouette coefficient ($MSC$) which ranges from $-1$ (worst result) to $1$ (best result). In this case, a number closer to $1$ will indicate that there is less confusion among the clusters of embeddings from the same speaker samples.
The first question we address is how long we have to train the network using triplet loss. For these experiments we set the parameters to $t=2000ms$, $w=100$ and $h=50$, we also set a margin of $2$ and we train the model during $1$, $2$ and $3$ days (exploratory experiments with shorter time showed that the minimum training time was a day). Table~\ref{tbl:exp1_time} reports the results. As can be seen $1$ day of training is enough to reach a good results:
\begin{table}[ht]
\centering
\begin{tabular}{|c||c|c|c|c|c|}\hline
Training days & IAD & IOD & DR & MSC \\\hline
1 & $4.77$ & $27.61$ & $\textbf{0.1730}$ & $\textbf{0.2475}$ \\
2 & $\textbf{4.61}$ & $26.62 $ & $0.1732 $ & $0.2326$ \\
3 & $4.85$ & $\textbf{27.88}$ & $0.1742$ & $0.2358$ \\\hline
\end{tabular}
\caption{Validation results for different training duration, one day gives the best results amd more time does not affect the behaviour of the network. }
\label{tbl:exp1_time}
\end{table}
The second question we address was how large has to be the CNN's input speech segment. For this we have to explore different parameters of the spectogram: $h$, $t$ while we fix $w$ to $100$ ms which is a common window size for speech signal processing. Varying $h$ allows us to control the amount of information that passes through in a segment, we explored $25$ms, $33$ms and $50$ms. On the other hand $t$ allows to control the amount of signal that the CNN will 'see', we explore $1$s, $1.5$s and $2$s values; we did not try a larger time since some all of recordings were at least $2$s, but not necessary longer. We set the margin to $2$ since in previous experiments had allows to identify a good compromise between $DR$ and $MSC$. Table~\ref{tbl:exp1_all} summarise the main findings for different combinations of these parameters, each model was trained during a day. As can be seen the more information in the patch the smaller the DR is (which is a good result). This points that a more informative patch is obtained by a larger segment with a small hope size.
Figure~\ref{fig:projection} shows the projection of embeddings. As it can be seen same speaker embeddings clusters, in this case these cluster represent one speaker's recording from which we extracted $20$ samples that became $20$ embeddings. The figure illustrates our best ($t=2000$ and $h=25$) and worst models ($t=1000$ and $h=50$). As it can be seen in both projections there is ordering of the speakers (clusters of speakers) but also shows some speakers which are close among themselves. It can been seen that our best model produces more organized positions while for the worst model there is some confusion among the clusters in the middle.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c||c|c|c|c|c|}\hline
$t$ (ms) & $h$ (ms) & Patch size & IAD & IOD & DR & MSC \\\hline
$1000$ & 25 & $40\times 256$ & $3.21$ & $22.17$ & $\textbf{0.1445}$ & $\textbf{0.1924}$ \\
$1000$ & 33 & $30\times 256$* & $3.60$ & $19.36$ & $0.1860$ & $0.1754$ \\
$1000$ & 50 & $20\times 256$* & $4.68$ & $23.21$ & $0.2018$ & $0.1506$ \\\hline
$1500$ & 25 & $60\times 256$ & $3.57 $ & $25.74$ & $\textbf{0.1386} $ & $ \textbf{0.2209}$ \\
$1500$ & 33 & $45\times 256$ & $3.42$ & $21.31$ & $0.1605$ & $0.2276$ \\
$1500$ & 50 & $30\times 256$* & $3.98$ & $23.46$ & $0.1695$ & $0.1721$ \\\hline
$2000$ & 25& $80\times 256$ & $4.69$ & $30.24$ & $0.1541$ & $\textbf{0.2766}$ \\
$2000$ & 33 & $60\times 256$ & $4.44$ & $29.70$ & $\textbf{0.1495}$ & $0.2487$ \\
$2000$ & 50 & $40\times 256$ & $4.77$ & $27.60$ & $0.2019$ & $0.2475$ \\\hline
\end{tabular}
\caption{Validation results for different parameters of speech segments. }
\label{tbl:exp1_all}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\textwidth]{pictures/modelo_25_100_2000.png}
\includegraphics[width=0.45\textwidth]{pictures/modelo_50_100_1000.png}
\caption{Projections of a recording for speakaer in validation set. First projection corresponds to our best model ($t=2000$, $h=25$), second to our worst model ($t=1000$, $h=25$) (same color same speaker, 218 speakers)}
\label{fig:projection}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|}\hline
Metric & Value\\\hline
IAD & $4.46$\\
IOD & $28.68$\\
DR & $0.1555$\\
MSC & $0.2618$\\\hline
\end{tabular}
\caption{Evaluation results with parameters \textit{t=2000}, \textit{h=25}, \textit{w=100} and \textit{m=2}. }
\label{tbl:exp1_final}
\end{table}
As presented in the section \ref{sec:dataset} there is an imbalance between female and male speakers, and between the nationalities of the speakers. To quantify the effect of these imbalances we performed more evaluations. In the case of gender we created three models: with only female recordings ($F$, $373$ training and $47$ validation speakers), with only male recordings ($M*$,$1370/171$) and training with a comparable amount of male recordings with female recordings ($M$, $373/47$). For this experiments we set the parameters to $t=2000ms$, $w=100ms$ and $h=50ms$. Table~\ref{tbl:gender} shows the main results on the corresponding validation dataset. As it can be seen there is an effect of gender miss-match, however this is not severe for the case of training with female speakers and using the model with male speakers. On the other direction, we can notice a severe drop in performance. As expected, the best setting is to train with the biggest amount of recordings of a gender and evaluating on that gender.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|||l|l|l|l|l|}\hline
Model & Evaluating & IAD & IOD & DR & MSC \\\hline
$F$ & $F$ & $3.74$ & $22.23$ & $0.1686$ & $0.2500$ \\
$F$ & $M$ & $4.68$ & $19.84$ & $0.2357$ & $0.2571$ \\\hline
$M$ & $M$ & $6.29$ & $33.17$ & $0.1896$ & $0.3517$ \\
$M$ & $F$ & $5.97$ & $33.12$ & $0.1822$ & $0.2381$ \\\hline
$M$* & $M$ & $3.71$ & $21.03 $ & $0.1765 $ & $0.3236 $ \\
$M$* & $F$ & $3.32$ & $22.87 $ & $0.1454 $ & $0.2073 $ \\\hline
\end{tabular}
\caption{Evaluations measure the effect of gender imbalance, $F$ (female) and $M$ (male) are comparable since they have the same number of training speakers; $M$* is not directly comparable since it relies in a larger amount of speakers.}
\label{tbl:gender}
\end{table}
Table~\ref{tbl:nationallity} shows our findings for the nationality. Given the amount of speakers, we decided to compare two types of speakers from two regions: Latinamerica ($L$) and Spain ($S$). We set two models, $L$ ($373$/$48$) and $S$ ($373$/$48$). As we can see there is not a notable difference given the nationality, considering the region Latinamerica packs more nationalities in the dataset we can not see an effect of the regional accents on the capabilities of triplet loss in producing good embeddings. Similarly to gender, the miss-match between training and evaluation datasets speakers produces drops in the performance, but not as severe as one might expect.
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|||l|l|l|l|l|}\hline
Model & Evaluating & IAD & IOD & DR & MSC \\\hline
$L$ & $L$ & $5.05$ & $21.73$ & $0.2325$ & $0.3410$ \\
$L$ & $S$ & $4.68$ & $25.17$ & $0.1903$ & $0.3401$ \\\hline
$S$ & $S$ & $4.41$ & $26.42$ & $0.1670$ & $0.3517$ \\
$S$ & $L$ & $4.73$ & $20.10$ & $0.2355$ & $0.2911$ \\\hline
\end{tabular}
\caption{Evaluations measure the effect of nationality imbalance, $L$ (Latinamerica) and $S$ (Spain)}
\label{tbl:nationallity}
\end{table}
\subsection{Speech embeddings for FSI}
\label{exp2}
In these experiments we aim to establish a way to calculate the LR based on the distances among embeddings. The more straight forward proposal is to use a normalized distance between the centroids of the reference and questioned sample embeddings, we call this approach distance based ($D$). The second proposal is to use a radio between the same distance $D$ and the distance to the closest speaker in the population in order to account for tipicallity, we call this approach distance radio ($DR$). The first proposal can be formalized as:
\begin{equation}
LR_{D} = \frac{D(q,r)}{N};
\end{equation}
while the second proposal as:
\begin{equation}
LR_{DR} = \frac{min(D(q,p))}{D(r,q)};\forall p\in Population
\end{equation}
Where $q$ is the centroid of the questioned embedding samples, $r$ the centroid of the reference and $p$ is a centroid of a speaker from the population and $N$ is a normalizing factor.
For the experimentation we used two Forensic Speech Identification scenarios: genuine (were the same-speaker hypothesis is true) and impostor (where the different-speaker hypothesis is true). Per speaker we randomly selected three recordings from where we sampled $15$ segments per recording as a reference. In the case of the \emph{genuine} scenario we selected a fourth recording as the questioned source of samples while for the impostor scenario we randomly selected an extra recording from a different speaker. Additionally, as our population we randomly selected $100$ different speakers from which we extracted the same amount of samples that our reference (i.e., $45$). It is important to remark that the recordings for these settings came from the validation split of the data.
With this considerations, we have have $218$ genuine cases and $218$ impostor cases. Figure~\ref{fig:exp2_freqdet} shows for both approaches the LR scores: distance based ($D$) and distance radio ($DR$). For the case of the distance as a proxy for LR ($D$), as expected genuine cases are concentrated with lower values with a mean in $0.31$ while impostor are located with a mean of $0.59$. The Equal Error Rate (EER) for the metric is of $0.1467$ with a sensitivity of $1.80$. For the case of distance radio ($DR$) we see that it corresponds better with the common interpretation of LR were the score indicates an update in the believe on the same speaker hypothesis. In this case, lower scores are associated to impostor cases with a mean of $0.64$ and larger score correspond to genuine cases with a mean of $1.15$. This approach has an EER $0.13$ and sensitivity index of $1.79$.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{pictures/DistributionsD.png}
\includegraphics[width=0.5\textwidth]{pictures/DistributionsDR.png}
\caption{Histogram of scores for positive (Genuine scores) and negative (Impostor scores) for both LR cases}
\label{fig:exp2_freqdet}
\end{figure}
\newpage
Figure~\ref{fig:exp2_comparison} shows the DET and ROC curves in a log scale for both scores. DET curves are common in the forensic field, in particular it compares the False Match Rate (FMR) with False Non-Match Rates (FNMR) of the system. For both DET curves we see that as the number of false positives grows it is harder to missmatch a case. A better system produces a curve located to the left bottom corner. Our experiments show that bot approaches $D$ and $DR$ share a lot of predictive power. It is important to take into account that DET curves are related to ROC curves, which are more common in the ML field. Within the ROC curves is common to measure Area Under the Curve (AUC) to compare two systems, in this case we can observe that the $DR$ approach has AUC of $0.9475$ while $D$ a AUC score of $0.9402$, this points to a slightly more robust discriminative power for the $DR$ metric.
\begin{figure}[ht]
\centering
\includegraphics[width=0.50\textwidth]{pictures/DETLog.png}
\includegraphics[width=0.50\textwidth]{pictures/ROCLog.png}
\caption{DET and ROC Log curves, for both proposals: distance (D) and distance ratio (DR).}
\label{fig:exp2_comparison}
\end{figure}
\section{Ethical discussion}
\label{sec:ethics}
In recent years, the topic of sharing personal information has become extremely relevant due to the intensive interaction between modern societies and technology, in particular technology that are able to take autonomous decisions. Nowadays, simple things like clicking a ``like'' button or typing some words to search in a search engine could trigger undesired advertisements or the risk to be scammed in creative ways~\cite{irani2011modeling}. As computer scientists, these type of observations raise the question of how datasets are share through the internet in order to contribute to the advances of science but at the same time, not to harm anyone in anyway, as it can be understood in the \emph{hippocratic oath for artificial intelligence practitioners}~\cite{etzioni2018hippocratic}.
In this work, the use of the Voxforge Spanish dataset raises some privacy concerns, since the donors accepted that their voices are part of a database destined to create language technologies however while maintaining anonymity. On the other hand, due to the type of information provided by each donor (mainly a personalised username), it could be easy to identify some of them and performing actions against them like identity fraud in certain speech systems~\cite{safavi2016fraud}. Fortunately, some of this concerns were contemplated by the developers of the Voxforge datasets which allowed and option to opt-out, however it could hardly be enforced for all distributed copies~\cite{zimmer2010but}. Following this concern our experiments do not rely on the \emph{username} field and it only releases the models, not the recordings. We also will honor any opt-out petition to the Voxforge dataset.
Other aspect of concern regarding this work is if the proposed system is ready to be deployed and used in a real court case. From our perspective, even though the experiments show an interesting and promising performance, it is clear that our models are not ready for deployment in court. In particular, the selection of the population has to be properly. In a real case, it must be transparent who will be included in the population to calculate typicality. Additionally, the results of our models have to be used in the context of other evidence, as previously mentioned the LR has the goal of updating the odds of a legal hypothesis, but it is not a definitive test of identity. So in this context, any attempt to use our models as a recognition system will be ill-advised.
\section{Conclusions}
\label{sec:con}
In this work we explored the use of the triplet loss function and a convolution neural network (CNN) for the forensic identification of speakers. We have shown that this setting offers an alternative to well established approaches. Within this setting, the goal is to train a convolutional network (CNN) to produce an embedding of a speech sample. Although previous research have shown that such embeddings can cluster speakers, in our experimentation we have shown that the distance among clusters can be used to approximate a Likehood Radio (LR) which is a common measurement to convey the plausibility of a legal forensic hypothesis.
Our experimentation points out that the CNN is able to produce a good embedding representation for long speech samples and finer resolution of the spectogram. We hypothesize this is because it contains more information of a speaker voice. In particular we have focused on Spanish from Latin America and Spain, we have shown that our approach is not affected by variant of Spanish. However, gender has a strong effect on the performance. So far our results points out that models have to be gender dependent, but more research has to be done on this point (see~\cite{ertam2019effective}).
Finally, our results also showed that triplet loss conveys a good discriminative power when the distances are used to approximate LR. We proposed two approaches for the calculation of LR, the first one is solely based on the distances among a questioned and reference samples ($D$). This approach can be considered independent of the population. The second approach uses the radio between the first approximation $D$ and the distance of the questioned sample to the closest speaker from the population ($DR$). We conclude that this second option provides a better alternative, since it conveys information in two aspects: first numerically it corresponds to other LR scores and second it has a better predictive power (AUC=$0.9458$), this makes it suitable for the forensic identification of speakers.
\section{Acknowledgements}
The authors thank CONACYT for the computer resources provided through the
INAOE Supercomputing Laboratory's Deep Learning Platform for Language
Technologies with the project \emph{Experimentos con voz, traducci\'on y clasificaci\'on de textos} (id. PAPTL 01-008). We also acknowledge Fernanda Hernandez and Sandra V\'azquez which were involved in early stages of the development of the code and experimentation.
\section*{Conflict of interest}
The authors declare that they have no conflict of interest.
\bibliographystyle{unsrt}
|
\section{Introduction}
Three-body recombination occurs in ultracold atomic gases as a result of a three-body collision, in which the atoms gain kinetic energy due to the formation of a two-body bound state. This process is vitally important in a number of physical and chemical contexts and has been recognized as one of the most important scattering observables\;\cite{PhysRevLett.83.1566,PhysRevLett.83.1751,PhysRevLett.94.213201,BRAATEN2006259,efimov_phisics}. As an exothermic reaction, TBR is one of the main loss mechanisms in systems with ultracold atoms, limiting the density and lifetime of a Bose-Einstein condensate\;\cite{PhysRevLett.79.337,PhysRevLett.85.728,PhysRevLett.91.123201}. Additionally, the recombination process is employed as a way to form weakly bound diatoms in ultracold degenerate Fermi gases\;\cite{PhysRevLett.91.240401,PhysRevLett.91.240402,PhysRevLett.91.250401}.
Particles with resonant s-wave interactions will exhibit the Efimov effect, i.e., an infinite sequence of universal bound states characterized by discrete scale invariance\;\cite{Efimov1971,Efimov1973}. The universal properties of the Efimov effect have been investigated both theoretically\;\cite{jiawangPRL2012,yujunwangPRl2012,Sorensen2012,Wumengshan2014,Huang2014Li6,Wumengshan2016,
Hafner2017,Hanhuili2018,Wenz2009Li6,PhysRevA.100.052702}
and experimentally\;\cite{Kraemer2006Cs,Zaccanti2009Cs,Berninger2011Cs,Huang2014Cs,
Lompe2010Li6,Williams2009Li6,Huckans2009Li6,Ottenstein2008Li6,Pollack2009Li7,Gross2009Li7,Gross2010Li7,
Kunitski2015He,
Barontini2009rb87k41,Bloom2013rb87k40,HuMingGuang2014rb87k40,Wacker2016rb87k3941,Kato2017rb87k4041,Maier2015rb87Li7,Pires2014cslia-,
Tung2014cslia-,Ulmanis2016cslia+,Ulmanis_Kuhnle2016csli,Johansen2017csli,PhysRevLett.125.243401} in ultracold atomic gases. Due to the novel characteristic of the Efimov effect in few-body physics, whether the Efimov effect is possible for p-wave or higher partial-wave interaction is fundamentally important\;\cite{PhysRevLett.97.023201,PhysRevA.86.012711}. Nisida\;\cite{PhysRevA.86.012710} noted that Efimov states cannot be realized in physical situations for non-s-wave interactions. The studies of Efremov \textit{et al.}\;\cite{PhysRevLett.111.113201} indicated that the effective potential is attractive and decreases as the third power of the interatomic distance for a heavy-heavy-light system with a p-wave resonant interaction when employing the Born-Oppenheimer approximation method. With the same method, Zhu and Tan\;\cite{PhysRevA.87.063629} provided a more general discussion of the universal properties for atoms near higher partial-wave Feshbach resonances and found that the effective potential behaved as $1/\rho^{2L+1}$ when the distance $\rho$ between two heavy atoms was large. These two works also demonstrated that the Efimov effect does not occur in few-body systems interacting via higher partial-wave resonant interactions.
For d-wave dimers, Gao\;\cite{PhysRevA.62.050702} predicted that the bound states always appear at a universal value of the s-wave scattering length of $a_s\approx0.956\,r_{\scriptscriptstyle\textsl{vdW}}$. $r_{\scriptscriptstyle\textsl{vdW}}=(2\mu_{2b}C_6)^{1/4}/2$ is the van der Waals length, which characterizes the length scale between the two neutral atoms interacting via the van der Waals interaction $-C_6/r^6$ with two-body reduced mass $\mu_{2b}$. For a system of three identical bosons that interact via a d-wave resonant interaction, the study of Wang \textit{et al.}\;\cite{PhysRevA.86.062511} showed that a universal three-body state associated with the d-wave dimer is formed at $a_s\approx1.09\,r_{\scriptscriptstyle\textsl{vdW}}$. Calculations by Juan \textit{et al.}\;\cite{PhysRevA.99.012701} showed that the TBR rate monotonically increases through the unitary point and is nearly a constant on the quasi-bound side. Despite this progress, universal properties of heteronuclear systems interacting via resonant higher partial-wave interactions are less well understood\;\cite{PhysRevLett.120.023401}.
Experimentally, great efforts have been devoted to studying strongly interacting atomic Bose gases with s-wave resonances\;\cite{Kraemer2006Cs,Zaccanti2009Cs,Berninger2011Cs,Huang2014Cs,
Pollack2009Li7,Gross2009Li7,Gross2010Li7,Kunitski2015He,Barontini2009rb87k41,Bloom2013rb87k40,HuMingGuang2014rb87k40,Wacker2016rb87k3941,Kato2017rb87k4041,Maier2015rb87Li7,Pires2014cslia-,
Tung2014cslia-,Ulmanis2016cslia+,Ulmanis_Kuhnle2016csli,Johansen2017csli,PhysRevLett.110.163202,PhysRevLett.123.233402,PhysRevLett.125.243401,PhysRevLett.111.125303,PhysRevX.6.021025,PhysRevA.89.021601}, but only a few studies have investigated the many-body properties of interacting fermions near a p-wave resonance\;\cite{PhysRevLett.90.053201,PhysRevA.70.030702,PhysRevA.71.045601,PhysRevA.88.012710,PhysRevA.98.020702,PhysRevLett.120.133401,PhysRevA.99.052704,PhysRevA.101.062702}.
For higher partial-wave resonances, challenges arise due to their short lifetimes and narrow resonance widths\;\cite{PhysRevLett.119.203402}. Recently, two broad d-wave resonances were observed via atom loss in a $^{85}$Rb\,-$^{87}$Rb mixture\;\cite{PhysRevLett.119.203402}. Very recently, a d-wave shape resonance and Feshbach resonance were observed in degenerate $^{41}$K and $^{39}$K gases, respectively\;\cite{Yao2019,PhysRevA.99.022701}. Moreover, both p-wave and d-wave Feshbach resonances were observed in the $^{133}$Cs\,-$^{6}$Li system\;\cite{Zhu2019}. This experimental progress has provided a platform to study the universal properties of few-body physics with d-wave resonant interactions.
In this paper, we investigate the TBR process in a heteronuclear system with d-wave resonant interactions. In real ultracold atomic systems, inter- and intraspecies interactions are generally not controlled independently. Thus, complications arise in heteronuclear three-body systems due to the two different scattering lengths. Near the interspecies Feshbach resonance, two identical atoms interact with each other over a small scattering length. Thus, we also focus on the more realistic case in which two heteronuclear atoms are in d-wave resonance while homonuclear atoms interact over a finite scattering length. As a result, ultracold gases of heteronuclear systems are expected to show rich few-body physics compared to the homonuclear case.
The TBR rates are obtained using quantum calculations based on a combination of the slow variable discretization (SVD) method, traditional hyperspherical adiabatic method and
$R$-matrix propagation method\;\cite{RevModPhys.68.1015,PhysRevA.65.042725,
PhysRevA.78.062701,PhysRevA.84.052721,Tolstikhin_1996,BALUJA1982299}.
Following the method of Ref\;\cite{PhysRevA.84.052721}, first, we solve the Schr$\mathrm{\ddot{o}}$dinger equation with the hyperradius divided into two regimes.
At short distances, the SVD method is employed to overcome the numerical difficulties at sharp nonadiabatic avoided crossings, and at large distances, the traditional adiabatic hyperspherical method is utilized to avoid the large memory and central processing unit (CPU) time needed in SVD.
Second, by propagating the $\underline{\mathcal{R}}$ matrix from short distances to large distances, we can obtain scattering properties through the $\underline{\mathcal{S}}$ matrix by matching the $\underline{\mathcal{R}}$ matrix with asymptotic functions and boundary conditions. The Lennard-Jones potential, which has been shown to be an excellent model potential, is utilized to mimic the interactions between atoms\;\cite{yujunwangPRl2012,Ulmanis2016cslia+,PhysRevA.86.062511,jiawangPRL2012, PhysRevA.90.022106}.
This paper is organized as follows:
In Sec. II, our calculation method and all necessary formulas for calculations are presented.
In Sec. III, we discuss the results and emphasize the significant role of intraspecies interactions in heteronuclear systems.
We then provide a brief summary.
Atomic units are applied throughout the paper unless stated otherwise.
\section{Theoretical formalism}
This numerical study focuses on the heteronuclear system with total angular momentum $J=0$. We use $m_i$ (i=1,2,3) to represent the mass of three atoms and use $r_{ij}$ to represent their distance. We choose the $^{85}$Rb\,-$^{87}$Rb\,-$^{87}$Rb system as an example and use the mass of Rb atoms in our calculations. In the center-of-mass frame, six coordinates are needed to describe the three-particle system.
Three of these coordinates are taken to be the Euler angles---$\alpha$, $\beta$, and $\gamma$---which specify the orientation of the body-fixed frame relative to the space-fixed frame.
The remaining degrees of freedom can be represented by hyperradius $R$ and the two hyperangles $\theta$ and $\phi$.
In our method, we employ Delves' hyperspherical coordinates.
We introduce the mass-scaled Jacobi coordinates. $\vec{\rho}_1$ is the vector from atom 1 to atom 2, with the reduced mass denoted by $\mu_1$; the second Jacobi $\vec{\rho}_2$ is measured from the diatom center of mass to the third atom, with reduced mass $\mu_2$. $\theta$ is the angle between $\vec{\rho}_1$ and $\vec{\rho}_2$. The hyperradius $R$ and hyperangle $\phi$ are defined as
\begin{equation}
\label{1}
\mu R=\mu_1 \rho^2_1+\mu_2 \rho^2_2
\end{equation}
and
\begin{equation}
\label{2}
\tan \phi =\sqrt{\frac{\mu_2}{\mu_1}}\frac{\rho_2}{\rho_1},\;\; 0 \leq \phi\leq\frac{\pi}{2},
\end{equation}
respectively, where $\mu$ is an arbitrary scaling factor that is chosen as $\mu=\sqrt{\mu_1\mu_2}$ in our calculations.
$R$ is the only coordinate with the dimension of length, which represents the size of the three-body system.
$\theta$, $\phi$ and the three Euler angles $(\alpha,\beta,\gamma)$ can be collectively represented by $\Omega$ $[\Omega\equiv(\theta,\phi,\alpha,\beta,\gamma)]$, which describe the rotation of the plane that contains the three particles.
In hyperspherical coordinates, the Schr$\mathrm{\ddot{o}}$dinger equation can be written in terms of the rescaled wave function $\psi_{\upsilon'}(R;\Omega)=\Psi_{\upsilon'}(R;\Omega)R^{5/2}\sin\phi\cos\phi$:
\begin{equation}
\label{3}
\bigg[-\frac{1}{2\mu}\frac{d^2}{dR^2}+\bigg(\frac{\Lambda^2-\frac{1}{4}}{2\mu R^2}
+V(R;\theta,\phi)\bigg)\bigg]\psi_{\upsilon'}(R;\Omega) =E\psi_{\upsilon'}(R;\Omega)\,,
\end{equation}
where $\Lambda^2$ is the squared ``grand angular momentum operator``, whose expression is given in Ref\;\cite{CDLIN1995}.
The volume element relevant to integrals over $|\psi_{\upsilon'}(R;\Omega)|^2$ then becomes $dR \sin\theta d \theta d \phi d\alpha \sin \beta d\beta d\gamma$.
The index $\upsilon'$ labels the different independent solutions. The three-body interaction $V(R;\theta,\phi)$ in Eq.\;(\ref{3}) is taken to be a sum of the three pairwise two-body interactions $\upsilon(r_{ij})$:
\begin{equation}
\label{4}
V(R;\theta,\phi) = \upsilon(r_{12})+\upsilon(r_{13})+\upsilon(r_{23})\,.
\end{equation}
The interparticle distances $r_{ij}$ can be described in terms of the internal coordinates as follows:
\begin{align}
\label{5}
&r_{12}=R\sqrt{\frac{\mu}{\mu_1}}\cos\phi, \\
&r_{23}=R\Big(\frac{\mu}{\mu_2}\sin^2\phi
+\frac{1}{4}\frac{\mu}{\mu_1}\cos^2\phi
-\frac{1}{2}\sin 2\phi\cos\theta\Big)^{1/2},\\
&r_{13}=R\Big(\frac{\mu}{\mu_2}\sin^2\phi
+\frac{1}{4}\frac{\mu}{\mu_1}\cos^2\phi
+\frac{1}{2}\sin 2\phi\cos\theta\Big)^{1/2}.
\end{align}
The wave function $\psi_{\upsilon'}$ can then be expanded with the complete, orthonormal adiabatic channel functions $ \Phi_\upsilon$ as
\begin{align}
\label{8}
\psi_{\upsilon'}(R;\Omega)=\sum^\infty_{\nu=0}F_{\nu \upsilon'}(R)\Phi_\nu(R;\Omega)\,.
\end{align}
We determine the adiabatic potentials $U_{\nu}(R)$ and corresponding channel
functions $\Phi_{\nu}(R;\Omega)$ at a fixed $R$ by solving the following adiabatic eigenvalue equation:
\begin{equation}
\label{9}
\bigg(\frac{\Lambda^{2}-\frac{1}{4}}{2\mu R^{2}}+
V(R;\theta,\phi) \bigg)\Phi_{\nu}(R;\Omega)=U_{\nu}(R)\Phi_{\nu}(R;\Omega)\,.
\end{equation}
The channel function is further expanded on Wigner rotation matrices $D^J_{KM}$ as
\begin{align}
\label{10}
\Phi_\nu^{J\Pi M}(R;\Omega)=\sum^J_{K=0}u_{\nu K}(R;\theta,\phi)\overline{D}^{J\Pi}_{KM}(\alpha,\beta,\gamma),
\end{align}
\begin{align}
\label{11}
\overline{D}^{J\Pi}_{KM}=\frac{1}{4\pi}\sqrt{2J+1}
\left[D^J_{KM}+(-1)^{K+J}\Pi D^J_{-KM}\right],
\end{align}
where $J$ is the total nuclear orbital angular momentum, $M$ is its projection onto the laboratory-fixed axis, and $\Pi$ is the parity with respect to the inversion of the nuclear coordinates. The quantum number $K$ denotes the projection of $J$ onto the body-frame $z$ axis and takes the values $K=J,J-2,\ldots,-(J-2),-J$ for the ``parity-favored'' case, $\Pi=(-1)^J$, and $K=J-1,J-3,\ldots,-(J-3),-(J-1)$ for the ``parity-unfavored'' case, $\Pi=(-1)^{J+1}$.
For the $^{85}$Rb\,-$^{87}$Rb\,-$^{87}$Rb system, the wave function is symmetric with respect to the exchange of the two $^{87}$Rb atoms, and thus, this exchange symmetry can be built into the boundary conditions of the body-frame components as follows:
\begin{align}
\label{12}
&P_{12}\overline{D}^{J\Pi}_{KM}=\Pi(-1)^K\overline{D}^{J\Pi}_{KM},\\
&P_{12}\theta=\pi-\theta.
\end{align}
For even parity, $u_{\nu K}$ should be symmetric about $\pi/2$ with even $K$ and antisymmetric with odd $K$. For odd parity, $u_{\nu K}$ should be antisymmetric for even $K$ and symmetric for odd $K$. To satisfy the permutation requirements, $u_{\nu K}$ is expanded with symmetric B-spline basis sets
\begin{align}
\label{14}
u_{\nu K}(R;\theta,\phi)=\sum^{N_{\phi}}_i\sum^{N_{\theta}/2}_j
c_{i,j}B_i(\phi)\left(B_j(\theta)+B_{N_{\theta}+1-j}(\pi-\theta)\right),
\end{align}
or antisymmetric B-spline basis sets
\begin{align}
\label{15}
u_{\nu K}(R;\theta,\phi)=\sum^{N_{\phi}}_i\sum^{N_{\theta}/2}_j
c_{i,j}B_i(\phi)\left(B_j(\theta)-B_{N_{\theta}+1-j}(\pi-\theta)\right),
\end{align}
where $N_{\theta}$ and $N_{\phi}$ are the sizes of the basis sets in the $\theta$ direction and $\phi$ direction, respectively.
The constructed symmetric B-spline basis sets utilized in the $\theta$ direction reduce the number of basis functions to $N_{\theta}/2$.
The goal of our scattering study is to determine the scattering matrix $\underline{\mathcal{S}}$ from the solutions of Eq.\;(\ref{3}). We calculate the $\underline{\mathcal{R}}$ matrix, which is defined as
\begin{align}
\label{16}
\underline{\mathcal{R}}(R)=\underline{\textsl{F}}(R)[\widetilde{\underline{\textsl{F}}}(R)]^{-1}\,,
\end{align}
where matrices $\underline{\textsl{F}}$ and $\widetilde{\underline{\textsl{F}}}$ can be calculated from the solution of Eq.~(\ref{3}) and Eq.~(\ref{9}) by
\begin{align}
\label{17}
F_{\nu,\upsilon'}(R)=\int d\Omega \Phi_{\nu}(R;\Omega)^{*}\psi_{\upsilon'}(R;\Omega)\,,
\end{align}
\begin{align}
\label{18}
\widetilde{F}_{\nu,\upsilon'}(R)=\int d\Omega \Phi_{\nu}(R;\Omega)^{*}\frac{\partial}{\partial R}\psi_{\upsilon'}(R;\Omega)\,.
\end{align}
Following the method of Ref.\;\cite{PhysRevA.84.052721}, we divide the hyperradius into $(N-1)$ intervals with the set of grid points $R_1<R_2<\cdots R_N$. At a short distance, we use the SVD method to solve Eq.\;(\ref{3}) in the interval $[R_i,R_i+1]$. In the SVD method, the total wave function $\psi_{\upsilon'}(R;\Omega)$ is expanded in terms of the discrete variable representation (DVR) basis $\pi_i$ and the channel functions $\Phi_{\nu}(R;\Omega)$ as
\begin{align}
\label{19}
\psi_{\upsilon'}(R;\Omega)=\sum^{N_{DVR}}_i\sum^{N_{chan}}_{\nu}C^{ \upsilon'}_{i\nu}\pi_i(R)\Phi_\nu(R_i;\Omega)\,,
\end{align}
where $N_{DVR}$ is the number of DVR basis functions and $N_{chan}$ is the number of included channel functions.
Inserting $\psi_{\upsilon'}(R;\Omega)$ into the three-body Schrodinger equation yields the standard algebraic problem for the coefficients $C^{\upsilon'}_{i\nu}$:
\begin{align}
\label{20}
\sum^{N_{DVR}}_{j}\sum^{N_{chan}}_{\mu}\mathcal{T}_{ij}
\mathcal{O}_{i\nu,j\mu}C^{\upsilon'}_{i\nu}+U_\nu (R_i)C^{\upsilon'}_{i\nu}=E^{\upsilon'}C^{\upsilon'}_{i\nu},
\end{align}
where
\begin{align}
\label{21}
\mathcal{T}_{ij}=\frac{1}{2\mu}\int^{R_{i+1}}_{R_{i}}\frac{d}{dR}\pi_i(R)\frac{d}{dR}\pi_j(R)dR
\end{align}
are the kinetic energy matrix elements; ${R_{i}}$ and ${R_{i+1}}$ are the boundaries of the calculation box; and
\begin{align}
\label{22}
\mathcal{O}_{i\nu,j\mu}=\langle\Phi_\nu(R_i;\Omega)|\Phi_\mu(R_j;\Omega)\rangle
\end{align}
are the overlap matrix elements between the adiabatic channels defined at different quadrature points.
At large distances, the traditional adiabatic hyperspherical method is used to solve Eq.\;(\ref{3}).
When substituting the wave functions $\psi(R;\Omega)$ into Eq.\;(\ref{3}), a set of coupled ordinary differential equations is obtained:
\begin{align}
\label{28}
[-\frac{1}{2\mu}\frac{d^2}{dR^2}+U_\nu(R)- E]F_{\nu,\upsilon'}(R)
-\frac{1}{2\mu}\sum_{\mu}[2P_{\mu\nu}(R)\frac{d}{dR}+Q_{\mu\nu}(R)]F_{\mu \upsilon'}(R)=0\,,
\end{align}
where
\begin{align}
\label{29}
P_{\mu\nu}(R)=\int d\Omega \Phi_{\mu}(R;\Omega)^{*}\frac{\partial}{\partial R}\Phi_{\nu}(R;\Omega)
\end{align}
and
\begin{align}
\label{30}
Q_{\mu\nu}(R) = \int d\Omega \Phi_{\mu}(R;\Omega)^{*}\frac{\partial^{2}}{\partial R^{2}}\Phi_{\nu}(R;\Omega)
\end{align}
are the nonadiabatic couplings that control the inelastic transitions and the width of the resonance supported by adiabatic potential
$U_\nu(R)$ .
In our calculations, the relation between $\underline{P}$ and $\underline{Q}$ is $\frac{d}{d R} \underline{P}=-\underline{P^{2}}+\underline{Q}$, where
\begin{align}
\label{31}
P_{\nu \mu}^{2}(R)=-\int d \Omega \frac{\partial}{\partial R} \Phi_{\nu}(R ; \Omega)^{*} \frac{\partial}{\partial R} \Phi_{\mu}(R ; \Omega).
\end{align}
The coupling matrices have the following properties:
$P_{\nu \mu}=-P_{\mu \nu}$ and $P_{\nu \mu}^{2}=P_{\mu \nu}^{2}$, which leads to $P_{\nu \nu}=0,$ and $Q_{\nu \nu}=-P_{\nu \nu}^{2}$.
The effective hyperradial potentials, which include hyperradial kinetic energy contributions with the $P_{\nu\nu}^2$ term, are more physical than adiabatic hyperpotentials and are defined as
\begin{align}
\label{32}
W_{\nu \nu}(R)=U_{\nu}(R)-\frac{\hbar^{2}}{2 \mu} P_{\nu \nu}^{2}(R).
\end{align}
Next, the $R$-matrix propagation method is employed. Over the interval $[R_{1},R_{2}]$, for a given $\underline{\mathcal{R}}$ matrix Eq.\;(\ref{16}), the $R$-matrix propagation method can be used to calculate the corresponding $\underline{\mathcal{R}}$ matrix at another point $R = R_{2}$ as follows:
\begin{align}
\label{33}
\underline{\mathcal{R}}(R_2)=\underline{\mathcal{R}}_{22}
-\underline{\mathcal{R}}_{21}\left[\underline{\mathcal{R}}_{11}
+\underline{\mathcal{R}}(R_1)\right]^{-1}\underline{\mathcal{R}}_{12}.
\end{align}
The $\underline{\mathcal{K}}$ matrix can be expressed in the following matrix equation:
\begin{align}
\label{34}
\underline{\mathcal{K}}=
(\underline{f}-\underline{f}'\underline{\mathcal{R}})
(\underline{g}-\underline{g}'\mathcal{R})^{-1}\,,
\end{align}
where $f_{\nu \nu'}=\sqrt{\frac{2\mu }{\pi k_\nu}} k_\nu R j_{l_\nu}(k_\nu R)\delta_{\nu \nu'}$ and $g_{\nu \nu'}=\sqrt{\frac{2\mu }{\pi k_\nu}} k_\nu R n_{l_\nu}(k_\nu R)\delta_{\nu \nu'}$ are the diagonal matrices of energy-normalized spherical Bessel and Neumann functions.
For the recombination channel, $l_\nu$ is the angular momentum of the third atom relative to the dimer, and $k_\nu$ is given by $k_{\nu}=\sqrt{2 \mu\left(E-E_{2 b}\right)}$.
For the entrance channel, $l_{\nu}=\lambda_{\nu}+3 / 2$, and $k_{\nu}=\sqrt{2 \mu E}$.
The scattering matrix $\underline{\mathcal{S}}$ is related to $\underline{\mathcal{K}}$ as follows:
\begin{align}
\label{35}
\underline{\mathcal{S}}=(\underline{1}+i\underline{\mathcal{K}})(\underline{1}-i\underline{\mathcal{K}})^{-1}\,.
\end{align}
Using the convention of Mott and Massey\;\cite{PhysRevLett.103.153201}, the N-body cross-section in d dimensions is defined as
\begin{align}
\label{36}
\sigma_{fi}(J^\Pi)=N_p\bigg(\frac{2\pi}{k_i}\bigg)^{d-1}\frac{1}{\Omega(d)}\sum_i(2J+1)|S^{J\Pi}_{fi}-\delta_{fi}|^2\,,
\end{align}
where $\Omega(d)=2\pi^{d/2}/\Gamma(d/2)$ is the total solid angle in d dimensions and $N_p$ is the number of terms in the permutation symmetry projection operator.
In the $^{85}$Rb\,-$^{87}$Rb\,-$^{87}$Rb system, $N_p=2\,!$, $d=6$, and the total TBR rate is then
\begin{align}
\label{37}
K_3=\frac{k}{\mu}\sigma_3=\sum_{J,\Pi}K_3^{J,\Pi}=2!\sum_{J,\Pi}\sum_{f,i}\frac{32(2J+1)\pi^2}{\mu k^4}|S^{J,\Pi}_{i\rightarrow f}|^2,
\end{align}
where i and f label the three-body continuum (incident) channel and TBR (outgoing) channel, respectively.
$\sigma_3$ is the generalized TBR cross-section.
$K_3^{J,\Pi}$ is the partial recombination rate corresponding to $J^\Pi$ symmetry, and $k=(2\mu E)^{1/2}$ is the wave number in the incident channels.
Since experiments are performed at a fixed temperature instead of a fixed energy, the thermal average becomes crucial for proper comparison with the experiment.
Assuming a Boltzmann distribution, the thermally averaged recombination rates are given by
\begin{align}
\label{38}
\langle K_3\rangle(T)=\frac{\int K_3(E)E^2e^{-E/(k_BT)}dE}{\int E^2e^{-E/(k_BT)}dE}=\frac{1}{2(k_BT)^3}\int K_3(E)E^2e^{-E/(k_BT)}dE.
\end{align}
The results presented in \ref{subsectionC} are given for $T=120~nK$.
In our calculations, Eq.\;(\ref{9}) is solved with 134 SVD sectors and 10 SVD points in each sector for $R<2\,000 a_0$. In the interval $2\,000 a_0 < R < 22\,000 a_0$, we use the traditional adiabatic hyperspherical method with $P_{\mu\nu}$ and $Q_{\mu\nu}$ calculated by an improved method in Ref\;\cite{PhysRevA.86.062511}.
The matrix elements of coupling $P_{\mu\nu}$ and effective potential $W_{\nu\nu} $ can be fitted to an inverse polynomial series at a large distance, and the fitting results of $P_{\mu\nu}$ and $W_\nu$ are applied beyond $R =22\,000 a_0$.
\section{Results and discussion}
\subsection{Hyperspherical potential curves near the interspecies d-wave resonance}
\label{subsectionA}
We use the Lennard-Jones potential to model the interactions between two atoms. The advantages of the Lennard-Jones potential is that it has the van der Waals length and can also avoid numerical difficulties in a short range.\;\cite{jiawangPRL2012}. Thus, it has been widely used and
proven
to be an excellent model potential to explore van der Waals universality in Efimov physics \;\cite{yujunwangPRl2012,Ulmanis2016cslia+,PhysRevA.86.062511,jiawangPRL2012, PhysRevA.90.022106}. The potential is expressed in the following form:
\begin{equation}
\label{39}
\upsilon(r_{ij})= -\frac{C_{6,ij}}{r^6_{ij}}\bigg[1-\frac{1}{2}(\frac{\gamma_{ij}}{r_{ij}})^6\bigg]\,.
\end{equation}
In this study, $\gamma_{ij}$ is adjusted to give the desired scattering length and number of bound states.
The value of $C_6$ for two Rb atoms we adopted here is 4698 from Ref.\;\cite{PhysRevA.89.022703}.
The low-energy behavior of the $l$ th partial wave phase shift for scattering by a long-range central potential $1/r^s (s>2)$ satisfies
\begin{align}
\label{40}
\tan \delta_{l}(k, \infty) \sim-k^{2 l+1} \lambda_{l}-\frac{\pi}{2^{s}} \frac{\Gamma(s-1) \Gamma\left(l+\frac{3}{2}-\frac{1}{2} s\right)}{\Gamma^{2}\left(\frac{1}{2} s\right) \Gamma\left(l+\frac{1}{2}+\frac{1}{2} s\right)}2\mu_{2b} C_{s} k^{s-2},
\end{align}
with $2<s<2 l+3$\;\cite{doi:10.1063/1.1703889,PhysRevA.74.052715,PhysRevA.87.063629}. For the Lennard-Jones potential $s=6$, the scattering phase shift of the d-wave has the following expansion:
\begin{align}
\label{41}
\tan \delta_{2}(k, \infty) \sim-k^5 \lambda_{2}-\frac{\pi}{2^6} \frac{\Gamma(5) \Gamma\left(2\right)}{\Gamma^{2}\left(3\right) \Gamma\left(11/2\right)}2\mu_{2b} C_{6} k^4.
\end{align}
$a_d=\lambda_2^{1/5}$ is denoted as the ``d-wave scattering length'' to characterize the findings in terms of the d-wave interactions, which diverges when a d-wave dimer is almost bound.
Figure\;\ref{fig1} shows the s-wave scattering length $a_s$ (blue solid line) and d-wave scattering length $a_d$ (red dashed line) as a function of $\gamma_{ij}$. No obvious difference is observed between $^{85}$Rb\,-$^{87}$Rb and $^{87}$Rb\,-$^{87}$Rb, so we show the results for only one case. The vertical dotted lines enclose the $^{85}$Rb\,-$^{87}$Rb parameter range considered in our numerical calculations. In this range, one two-body s-wave bound state exists before the d-wave bound state emerges. For a homonuclear interaction, we focus on the case in which two $^{87}$Rb atoms are in d-wave resonance (point I in Fig.\;\ref{fig1}) and the more realistic case, in which $^{87}$Rb\,-$^{87}$Rb interact via the s-wave scattering length $a_s=100a_0$, which is indicated by arrows II and III. As shown in Fig.\;\ref{fig1}, the two $^{87}$Rb atoms have one s-wave bound state at point II and IV. At point III, the two $^{87}$Rb atoms have no bound states and are located away from the d-wave resonance.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{aa2}
\caption{(Color online) Two-body s-wave scattering length (blue solid line) and d-wave scattering length (red dashed line) as a function of the adjusting parameter $\gamma_{ij}$. The $^{85}$Rb\,-$^{87}$Rb interaction range is adjusted around the two-body d-wave resonance that is shown between the two black dotted lines. For the $^{87}$Rb\,-$^{87}$Rb interaction, points I, II, III and IV
are focused on
. }
\label{fig1}
\end{figure}
In the scattering process, the adiabatic potential curves $U_\nu(R)$ are important in understanding three-body physics. Figure\;\ref{fig2} shows the hyperspherical potential curves of the $^{85}$Rb\,-$^{87}$Rb--$^{87}$Rb system with the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length $a_{d} = -115 a_0$ (Fig.\;\ref{fig2a}) and $a_{d} = 90 a_0$ (Fig.\;\ref{fig2c}). $^{87}$Rb\,-$^{87}$Rb interact via the s-wave scattering length $a_s=100 a_0$ with the parameter $\gamma_{ij}$ adjusted to point II in Fig.\;\ref{fig1}. In each case, the solid potential curves correspond to the TBR channels and asymptotically approach the dimer binding energy. The effective potentials for these channels exhibit asymptotic behavior, given by
\begin{align}
\label{42}
W_f(R)=\frac{l_f(l_f+1)}{2 \mu R^2}+E^{(f)}_{2b},
\end{align}
where $E^{(f)}_{2b}$ is the dimer energy and $l_f$ is the relative orbital angular momentum between the atom and the dimer.
The subscript $f$ distinguishes the recombination channels. The dashed lines in Fig.\;\ref{fig2a} and Fig.\;\ref{fig2c} denote the three-body breakup channels (or entrance channels); i.e., all three atoms exist far from each other as $R\rightarrow \infty$, where the potentials behave as
\begin{align}
W_i(R)=\frac{\lambda_i(\lambda_i+4)+15/4}{2 \mu R^2}.
\label{43}
\end{align}
The values of $\lambda_i$ are nonnegative integers determined by $J^\Pi$ and the identical particle symmetry\;\cite{PhysRevA.65.010705}.
We use the dimensionless quantity of the nonadiabatic coupling strength defined by
\begin{align}
f_{v v^{\prime}}(R)=\frac{P_{v v^{\prime}}(R)^{2}}{2 \mu\left[U_{v}(R)-U_{v^{\prime}}(R)\right]}
\end{align}
to characterize the nonadiabatic coupling magnitude, which mainly controls the recombination process.
The coupling strength between the highest recombination channel and the lowest entrance channel is shown in Fig.\;\ref{fig2b} and Fig.\;\ref{fig2d} .
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=7.5cm]{2a.jpg}
\label{fig2a}
}
\subfigure{
\includegraphics[width=7.5cm]{2b.jpg}
\label{fig2b}
}
\subfigure{
\includegraphics[width=7.5cm]{2c.jpg}
\label{fig2c}
}
\subfigure{
\includegraphics[width=7.5cm]{2d.jpg}
\label{fig2d}
}
\caption{(Color online) Three-body adiabatic potential curves for the $^{85}$Rb\,-$^{87}$Rb-$^{87}$Rb system with $^{85}$Rb\,-$^{87}$Rb d-wave scattering length $a_d=-115a_0$ in (a) and $a_d=90a_0$ in (c). The $^{87}$Rb\,-$^{87}$Rb s-wave scattering length is $a_s=100a_0$ with the parameter $\gamma_{ij}$ adjusted to point II in Fig.\;\ref{fig1}. (b) and (d) show the avoided crossings between the highest recombination channel and the first entrance channel with their nonadiabatic coupling strength corresponding to (a) and (c), respectively.}
\label{fig2}
\end{figure}
When the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length $a_d$ is negative, as shown in Fig.\;\ref{fig2a} and \ref{fig2b}, the nonadiabatic coupling strength between the lowest entrance channel and the first recombination channel is localized at a short distance. Here, recombination occurs primarily by tunneling through the potential barrier in the lowest three-body entrance channel to reach the region of large coupling. Thus, the potential barrier in the lowest three-body entrance channel has an important role in the recombination process. Figure\;\ref{fig3a} shows the barrier for several negative $^{85}$Rb\,-$^{87}$Rb scattering lengths $a_d$. The height of the barrier decreases as the $^{85}$Rb\,-$^{87}$Rb d-wave interaction strengthens. Diminishing of the potential barrier in the entrance channel is responsible for the recombination enhancement on this scattering side\;\cite{Kartavtsev2009}. In addition, when the $^{85}$Rb\,-$^{87}$Rb interaction becomes sufficiently strong, a two-body d-wave shape resonance appears\;\cite{DIncao2009}. As a result, this resonance will produce a series of avoided crossings in the three-body potential curves at energies near the position of the resonance, as shown in Figs.\;\ref{fig3b}-\ref{fig3d}. Ref.\;\cite{Suno_2003} also found this phenomenon in a three-fermion system when the two-body scattering volume was negative. The positions of these avoided crossings might be expected to approach the three-body breakup threshold when the $^{85}$Rb\,-$^{87}$Rb d-wave interaction strengthens. This result can be demonstrated by Figs.\;\ref{fig3b}-\ref{fig3d}.
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=7cm]{3a.jpg}
\label{fig3a}
}
\subfigure{
\includegraphics[width=7cm]{3b.jpg}
\label{fig3b}
}
\subfigure{
\includegraphics[width=7cm]{3c.jpg}
\label{fig3c}
}
\subfigure{
\includegraphics[width=7cm]{3d.jpg}
\label{fig3d}
}
\caption{(Color online) (a) Lowest entrance channels with barriers in $^{85}$Rb\,-$^{87}$Rb-$^{87}$Rb for the different $^{85}$Rb\,-$^{87}$Rb d-wave scattering lengths; (b), (c) and (d) are potential curves that show a series of avoided crossings near the $^{85}$Rb\,-$^{87}$Rb two-body d-wave resonance. These avoided crossings approach the three-body breakup threshold when $a_d\rightarrow - \infty$. The $^{87}$Rb\,-$^{87}$Rb s-wave scattering length is $a_s=100a_0$ with the parameter $\gamma_{ij}$ adjusted to point II in Fig.\;\ref{fig1}.}
\label{fig3}
\end{figure}
For the positive $^{85}$Rb\,-$^{87}$Rb d-wave scattering length case, the important feature is the broad avoided crossing between the lowest entrance channel and the highest recombination channel, as shown in Fig.\;\ref{fig2c}. The shallowest recombination channel has an attractive well followed by a repulsive barrier at a larger distance. If the barrier is high enough, the atom and dimer could be trapped inside. Such almost bound states above the threshold of the potential are the shape resonance states. Figure\;\ref{fig4a} shows the lowest entrance channel and the shallowest recombination channel of the $^{85}$Rb\,-$^{87}$Rb\,-$^{87}$Rb system for different $^{85}$Rb\,-$^{87}$Rb d-wave scattering lengths with the parameter $\gamma_{ij}$ of the $^{87}$Rb\,-$^{87}$Rb interaction adjusted to point II in Fig.\;\ref{fig1}. The two-body threshold moves towards the three-body breakup threshold with an increase in the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length. Thus, when the top of the barrier is above the collision energy, recombination will be suppressed by this extra barrier, except possibly at energies that match atom-dimer three-body shape resonances behind this barrier. The existence of three-body shape resonance will lead to more or fewer sudden jumps of the phase shift by $\pi$. The phase shift is well described by the analytical expression:
\begin{equation}
\delta_{l}(E)=\delta_{\mathrm{bg}}-\arctan \left(\frac{\Gamma / 2}{E-E_{\mathrm{R}}}\right),
\end{equation}
where $E_R$ is the resonance position, $\Gamma$ is the resonance width, and $\delta_{bg}$ is a smoothly energy-dependent background phase shift\;\cite{PhysRevA.19.920}.
The phase shift $\delta_f$ for atom-dimer elastic scattering is related to the diagonal S-matrix element by the formula
\begin{equation}
S_{f\leftarrow f}^{0+}=\exp(2i\delta_f).
\end{equation}
The subscript f distinguishes the recombination channels.
In Figs.\;\ref{fig4b}-\ref{fig4d}, we plot the atom-dimer scattering phase shift $\delta$, $\tan \delta$ and and the analytical formula for $\tan\delta$ associated with the highest recombination channel as a function of energy. In Figs.\;\ref{fig4b} and \ref{fig4c}, the resonances appear as a jump of the atom-dimer scattering phase shift $0.993\pi$, which implies that shape resonances exist when the $^{85}$Rb\,-$^{87}$Rb scattering length is $a_d=114a_0$ and $a_d=90a_0$. By fitting the $\tan \delta$ points with the analytical expression, we obtain resonance positions of $-2.7\times10^{-12}$ (Fig.\;\ref{fig4b}) and $-3.5\times10^{-11}$ (Fig.\;\ref{fig4c}).
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=7cm]{4a.jpg}
\label{fig4a}
}
\subfigure{
\includegraphics[width=7cm]{4b.jpg}
\label{fig4b}
}
\subfigure{
\includegraphics[width=7cm]{4c.jpg}
\label{fig4c}
}
\subfigure{
\includegraphics[width=7cm]{4d.jpg}
\label{fig4d}
}
\caption{(Color online)(a) Lowest entrance and highest recombination channels for the three positive $^{85}$Rb\,-$^{87}$Rb d-wave scattering lengths $a_d$. The parameter of the $^{87}$Rb\,-$^{87}$Rb interaction is adjusted to point II in Fig.\;\ref{fig1}. (b), (c) and (d) are the corresponding atom-dimer scattering phase shifts $\delta$ and $\tan\delta$ and the analytical formula for $\tan\delta$ associated with the highest recombination channel as a function of energy. The red sold lines in (b) and (c) are the fitting results by the analytical formula $\tan\delta=\frac{\tan\delta_{bg}-\frac{\Gamma}{2(E-E_R)}}{1+\tan\delta_{bg}\frac{\Gamma}{2(E-E_R)}}$.}
\label{fig4}
\end{figure}
\subsection{Three-body recombination rates}
\label{subsectionB}
The partial rates $K_3^{J\Pi}$ and total $K_3$ for $J^\Pi=0^+, 1^-$ and $2^+$ symmetries as a function of the collision energy with the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length fixed at $a_d=-115a_0$ (Fig.\;\ref{fig5a}) and $a_d = 90 a_0$ (Fig.\;\ref{fig5b}) are shown in Fig.\;\ref{fig5}. The parameter of the $^{87}$Rb\,-$^{87}$Rb interaction is adjusted to point II in Fig.\;\ref{fig1}. In the zero-energy limit, the recombination rate obeys the threshold behavior $K^{J\Pi}_3 \sim E^{\lambda_{min}}$, where $\lambda_{min}$ is the minimum value of $\lambda$ in Eq.(\ref{43}). For $J^{\Pi}= 0^{+},1^{-},2^{+}$, we have $\lambda_{min} = 0, 1, 2$ in the $^{85}$Rb\,-$^{87}$Rb\,-$^{87}$Rb system\;\cite{PhysRevA.65.010705,PhysRevA.78.062701}, that is, the $J^\Pi=0^+$ partial rate increases, like $E^0$, from the threshold, while the rates $1^-$ and $2^+$ behave as $E^1$ and $E^2$, respectively. At high collision energies, $K_3$ decreases as $E^{-2}$, required by unitary. For the positive d-wave scattering length, Fig.\;\ref{fig5b} shows that the Wigner threshold law holds only at small energies $E< 10 ~\mu K$. Note that the Wigner threshold regime can be characterized as energies smaller than the smallest energy scale, which is typically a molecular binding energy\;\cite{PhysRevA.80.062702}. In this case, when $a_d=90a_0$, the newly formed d-wave dimer binding energy is approximately $12~\mu K$, which shows rough agreement with the attained threshold regime of $E< 10 ~\mu K$.
Although the $J^\Pi=0^+$ case is predicted to be the dominant symmetry in the zero collision energy limit, other symmetries may contribute substantially. Therefore, studying the energy-dependent partial recombination rates that correspond to various symmetries is interesting. Figure\;\ref{fig5} shows that the contributions from $J^\Pi=1^-$ and $J^\Pi=2^+$ partial waves are significant when the energy exceeds $40~\mu K$.
\begin{figure}[htbp]
\centering
\subfigure{
\includegraphics[width=7cm]{5a.jpg}
\label{fig5a}
}
\subfigure{
\includegraphics[width=7cm]{5b.jpg}
\label{fig5b}
}
\caption{(Color online) Partial rates $K_3^{J\Pi}$ and their total $K_3$ for $J^\Pi=0^+, 1^-$ and $2^+$ symmetries as a function of the collision energy when the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length is fixed at $a_d=-115a_0$ (Fig. (a)) and $a_d=90a_0$ (Fig. (b)) with the parameter of the $^{87}$Rb\,-$^{87}$Rb interaction adjusted to point II in Fig.\;\ref{fig1}. The dashed lines represent the threshold laws or unitary limit. }
\label{fig5}
\end{figure}
\subsection{Nonnegligible role of the intraspecies interaction in the heteronuclear system}
\label{subsectionC}
In a real ultracold atomic system, the inter- and intraspecies interactions are generally not controlled independently, and thus, a finite intraspecies scattering length exists. For the $^{85}$Rb\,-$^{87}$Rb\,-$^{87}$Rb system, near the interspecies d-wave Feshbach resonance, $^{87}$Rb\,-$^{87}$Rb interact with each other through a smaller s-wave scattering length $a_s=100a_0$. As shown in Fig\;\ref{fig1}, $^{87}$Rb\,-$^{87}$Rb has an s-wave bound state at point II but no s-wave bound state at point III with the same $^{87}$Rb\,-$^{87}$Rb s-wave scattering length.
To examine how the $^{87}$Rb\,-$^{87}$Rb interaction influences the TBR, we plot the hyperspherical potential curves for the first entrance channel and the highest-lying recombination channel for the same $^{85}$Rb\,-$^{87}$Rb\, d-wave scattering length $a_d = -115\,a_0$ but different $^{87}$Rb\,-$^{87}$Rb\. Interaction details are shown in Figs.\;\ref{fig6a1}, \ref{fig6b1}, \ref{fig6c1} and \ref{fig6d1}. The interaction details of $^{87}$Rb\,-$^{87}$Rb greatly affect the coupling between the lowest entrance channel and the highest recombination channel. The Landau-Zener parameter $T_{ij}$, which estimates the nonadiabatic transition probabilities\;\cite{CLARK1979295}, can quantitatively reflect this coupling strength and be calculated by
$T_{ij}=e^{-\delta_{ij}}=e^{-\frac{\pi \Delta^2_{ij}}{2\alpha_{ij}\nu}}$,
where $\Delta_{ij}=U_i-U_j$ is evaluated in the transition region and $\alpha_{ij}$ is obtained from P-matrix analysis.
When $^{87}$Rb\,-$^{87}$Rb are in d-wave resonance ($^{87}$Rb\,-$^{87}$Rb s-wave scattering length $a_s=84\,a_0$), point I in Fig.\;\ref{fig1}, the Landau-Zener parameter approaches $1$, as shown in Fig.\;\ref{fig6a1}, implying that a nonadiabatic transition occurs and the potential well of the recombination channel deepens. A d-wave-related trimer state (A peak in Fig.\;\ref{fig6a2}) can thus be supported, which will enable enhancement of the TBR rate. When we plot the total and partial $J^{\Pi} = 0^{+}$ TBR rates as a function of the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length at fixed $^{87}$Rb\,-$^{87}$Rb s-wave scattering length $a_s= 84\,a_0$, the total rate exhibits two clear enhancements, labeled ``A`` and ``B``, as shown in Fig.\;\ref{fig6a2}. This phenomenon was predicted by Wang \textit{et al.}\;\cite{PhysRevA.86.062511} in a three-identical-boson system. Peak A corresponds to the d-wave-related trimer state across the collision threshold. According to our analyses of the influence of the $^{87}$Rb\,-$^{87}$Rb interaction on the coupling between the first entrance channel and the highest-lying recombination channel, enhancement A will shift or disappear depending on the interaction details of the two identical atoms in the heteronuclear system. This result can be demonstrated in the following results of $K_3$, when $^{87}$Rb\,-$^{87}$Rb is not exactly in d-wave resonance.
Figures\;\ref{fig6b1} and \ref{fig6c1} show the hyperspherical potential curves with the same $^{85}$Rb\,-$^{87}$Rb two-body scattering length $a_d=-115 a_0$, as shown in Fig.\;\ref{fig6a1}, while $^{87}$Rb\,-$^{87}$Rb interact through s-wave scattering length $a_s= 100 a_0$ with different interaction details. We focus on the case in which the parameter of the $^{87}$Rb\,-$^{87}$Rb interaction is adjusted to point II in Fig.\;\ref{fig1}, where $^{87}$Rb\,-$^{87}$Rb has one s-wave bound state
near the d-wave resonance. In this case, we find that the position of enhancement A shifts from $1.19\,r_{\scriptscriptstyle\textsl{vdW}} $ to $1.08\,r_{\scriptscriptstyle\textsl{vdW}} $ ($r_{\scriptscriptstyle\textsl{vdW}} $ is the van der Waals length between $^{85}$Rb and $^{87}$Rb) compared to the case in which the homonuclear atoms are in d-wave resonance, as shown in Fig.\;\ref{fig6a2}. When the parameter $\gamma_{ij}$ is changed to point III in Fig.\;\ref{fig1}, where the interaction of $^{87}$Rb\,-$^{87}$Rb is away from the d-wave resonance, the coupling between the first entrance channel and the highest-lying recombination channel weakens. Figure\;\ref{fig6c1} shows that the Landau-Zener parameter is $0.186$ in this case, implying that an adiabatic transition occurs. As a result, the potential well in the recombination channel may not be deep enough to support the trimer state, leading to the absence of enhancement A in Fig.\;\ref{fig6c2}. Fig.\;\ref{fig6d1} show the hyperspherical potential curves, where the parameter $\gamma_{ij}$ is changed to point IV in Fig.\;\ref{fig1}, where $^{87}$Rb\,-$^{87}$Rb interact through s-wave scattering length $a_s= 1.3 a_0$ away from the d-wave resonance. We note that the Landau-Zener parameter is $0.345$ and enhancement A also disappears in this case.
Peak B formed after the d-wave dimer became bound. Our analysis in\;\ref{subsectionA} shows that this enhancement corresponds to the three-body shape resonance. Its position is also affected by the interaction details of the two homonuclear atoms, as shown in Figs.\;\ref{fig6a2}, \ref{fig6b2} , \ref{fig6c2} and \ref{fig6d2}.
Our results demonstrate that the intraspecies interaction has a significant role in determining the TBR in the heteronuclear system.
\begin{figure}[htbp]
\centering
\label{fig6}
\renewcommand{\thesubfigure}{(a\arabic{subfigure})}
\setcounter{subfigure}{0}
\subfigure{
\includegraphics[width=5.2cm]{6a1.jpg}
\label{fig6a1}
}
\subfigure{
\includegraphics[width=5.2cm]{6a2.jpg}
\label{fig6a2}
}
\renewcommand{\thesubfigure}{(b\arabic{subfigure})}
\setcounter{subfigure}{0}
\subfigure{
\includegraphics[width=5.2cm]{6b1.jpg}
\label{fig6b1}
}
\subfigure{
\includegraphics[width=5.2cm]{6b2.jpg}
\label{fig6b2}
}
\renewcommand{\thesubfigure}{(c\arabic{subfigure})}
\setcounter{subfigure}{0}
\subfigure{
\includegraphics[width=5.2cm]{6c1.jpg}
\label{fig6c1}
}
\subfigure{
\includegraphics[width=5.2cm]{6c2.jpg}
\label{fig6c2}
}
\renewcommand{\thesubfigure}{(d\arabic{subfigure})}
\setcounter{subfigure}{0}
\subfigure{
\includegraphics[width=5.2cm]{6d1.jpg}
\label{fig6d1}
}
\subfigure{
\includegraphics[width=5.2cm]{6d2.jpg}
\label{fig6d2}
}
\caption{(Color online) The first entrance channel and highest recombination channel hyperpotential curves and corresponding Landau-Zener parameter $T_{na}$ for the same $^{85}$Rb\,-$^{87}$Rb d-wave scattering length $a_d=-115a_0$ are shown in (a1), (b1), (c1) and (d1). The TBR rate $K_3$ varies with the $^{85}$Rb\,-$^{87}$Rb interaction when the $^{87}$Rb\,-$^{87}$Rb interaction is fixed at different points in Fig.\;\ref{fig1} are shown in (a2), (b2), (c2) and (d2).The vertical black dotted lines in (a2), (b2), (c2) and (d2) indicate the position of $^{85}$Rb\,-$^{87}$Rb d-wave resonances.}
\end{figure}
\section{Conclusions}
In summary, we have investigated the TBR rate for the heternuclear atomic system near the interspecies d-wave Feshbach resonance. The $^{85}$Rb\,-$^{87}$Rb-$^{87}$Rb system is chosen as an example and calculations are based on the Lennard-Jones model potential for the Rb-Rb interaction. The TBR rates are obtained using quantum calculations in the frame of the hyperspherical coordinates, which are based on a combination of the SVD method, traditional hyperspherical adiabatic method and
R-matrix propagation method. Our study reveals two different mechanisms of recombination rate enhancement for positive and negative $^{85}$Rb\,-$^{87}$Rb d-wave scattering lengths. When the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length is positive and large, a loosely bound dimer is produced, and enhancement occurs due to the existence of three-body shape resonance. We have identified two such shape resonances on the positive $^{85}$Rb\,-$^{87}$Rb d-wave scattering length side. For the case in which $^{87}$Rb\,-$^{85}$Rb interact via a negative d-wave scattering length, the coupling between the lowest entrance channel and the highest recombination channel is crucial to the formation of the three-body state. When the coupling strengthens, a nonadiabatic transition occurs, which deepens the potential well in the highest recombination channel and thus supports the three-body state. The enhancement on the negative interspecies scattering length side corresponds to the three-body state crossing the three-body threshold.
In addition, we investigated the influence of the finite $^{87}$Rb\,-$^{87}$Rb interaction on the recombination enhancement. With the same $^{87}$Rb\,-$^{87}$Rb s-wave scattering length $a_s=100 a_0$, when the interaction is near the d-wave resonance, the coupling between the lowest entrance channel and the highest recombination channel strengthens, enhancing the recombination rate. However, if the parameter of the $^{87}$Rb\,-$^{87}$Rb interaction is adjusted to a point away from the d-wave resonance, then the enhancement will disappear. Moreover, our study reveals that the intraspecies interaction affects the $^{85}$Rb\,-$^{87}$Rb d-wave scattering length values at which the enhancement appears. Our results have confirmed the main results of Ref.\;\cite{PhysRevA.86.062511} for the homonuclear case and provide numerical evidence that the TBR rates in heteronuclear systems are more complex than those in homonuclear systems.
\section{Acknowledgments}
We thank C. H. Greene, Jia Wang, Li You, Meng Khoon Tey and Cui Yue for helpful discussions.
Hui-Li Han was
supported by the National Natural Science Foundation of China under Grants No. 11874391 and No. 11634013 and the National Key Research and Development
Program of China under Grant No. 2016YFA0301503.
Ting-Yun Shi was supported by the Strategic Priority Research Program of the Chinese Academy of Sciences under Grant No. XDB21030300.
|
\section{Introduction}
Machine learning (ML) algorithms learn the information contained in the data and use it for the purpose of prediction, classification and clustering. As astronomical datasets are growing exponentially in size, ML techniques are becoming increasingly useful for creating models which will enable astronomers to expedite the process of astronomical discovery. For example, ML algorithms can be used to improve performance (in terms of time and processing capacity) - as we show in prediction of star formation properties in Section \ref{supervised}. They can be used to generate inferences by analysing large amounts of data to uncover their patterns. The outcome of the ML algorithm can be compared with the current model and be used to improve the understanding of the current models i.e. if the ML predictions for a small number of data points form outliers from the trend then they may point to the need of further investigation, as typically ML algorithms are designed to generalise on all data points.
Section~\ref{supervised} briefly describes the outcome of the supervised deep learning approach to predict three star formation properties of galaxies viz. stellar mass, star formation rate (SFR) and dust luminosity using broadband flux measurements. Section~\ref{unsupervised} presents our exploration of the unsupervised technique to discover information present in solar radio images.
\section{Discovering physical models based on supervised machine learning} \label{supervised}
In this work, our goal is to model and replicate the behavior of a specific stellar population synthesis code viz. \textsc{magphys}{} \citep{magphys}. The best fit parameters that characterise the star formation histories of galaxies are determined using these models. We implement a supervised machine learning technique - deep learning - to mimic the behavior of the \textsc{magphys}{} model to predict three important star formation properties - stellar mass, star formation rate and dust luminosity. The data used are from the GAMA (Galaxy And Mass Assembly) \citep{Driver2011} survey. We selected all galaxies with $0<z\leq0.5$ and stellar mass $\geq 10^9 M_{\odot}$ from the GAMA catalog. After applying filters to remove noise from the data, we had 76,455 galaxies which form our final sample set to train the deep learning model.
Estimating the three star-formation parameters using the \textsc{magphys}{} code for 10,000 galaxies would take $\sim100,000$ minutes (about 10 minutes per galaxy). As telescope technology improves, the size of galaxy samples with multi-wavelength data is ever increasing. A traditional stellar population synthesis code scales linearly with sample size. On the other hand, the deep learning model training, takes 3 to 30 mins depending on the free parameter being modelled, with most of the time taken up by the training phase, which is a one time effort. Once modelled, the time taken to predict the free parameters for new galaxies is negligible. This represents a huge savings in time, with potentially larger savings for samples from future large area imaging surveys.
This model can further be modified to also give the confidence level of each prediction. Galaxies for which the confidence level of prediction is very low, can be further investigated and run with the standard stellar population technique. This outcome can be investigated further and incorporated in the deep learning model, enriching the model as it encounters more and more data.
The error of the deep learning model is 0.0577, 0.1643 and 0.1143 (in their respective units) for the logarithm of stellar mass, star formation rate and dust luminosity respectively which is calculated using $$error = \sigma(y_{actual}-y_{predicted})$$
The comparison of predicted values against values from \textsc{magphys}{} is shown in Fig.~\ref{fig1}. More details on the prediction of star-formation properties is available in \citet{surana2020}.
\articlefigurethree{O4-4_stellar_mass_f1.eps}{O4-4_SFR_f2.eps}{O4-4_dust_luminosity_f3.eps}{fig1}{The scatter plots in the upper panels show the values predicted by the deep learning model compared to the \textsc{magphys}{} values, for stellar mass, star formation rate and dust luminosity. The dashed line shows the best linear fit through the scatter plot and the solid line represents the points where the predicted values equal the \textsc{magphys}{} model values. The lower panel shows the error in the prediction as a function of the respective parameter.}
\section{Discovering patterns and anomalies based on Unsupervised machine learning} \label{unsupervised}
Unsupervised ML algorithms are of particular importance in research,as they are used to uncover patterns, complex relationships and groupings that exist in the data. Quite often, the research explorations and approaches are biased by the knowledge, experience, expectations and intuitions of the researchers. Much of the analysis is also driven by visual inspection, which has severe limitations in terms of the number of dimensions that can be explored simultaneously, and the capacity of the human brain to assimilate information from large data volumes. We use unsupervised ML algorithms to understand data demography and observe groupings, patterns or anomalies in the data. We use solar radio images from the Murchison Widefield Array (MWA) to explore ML approaches. This opportunity has been enabled by the recent availability of an interferometric imaging pipeline \citep{Mondal-AIRCARS}, which provides solar radio images with unprecedented fidelity and dynamic range. The output of this pipeline is a hypercube - $I(\theta,\phi,\nu,t)$, where $I$ is the intensity of emission, $\theta$ and $\phi$ are the image coordinates, $\nu$ is the frequency, and $t$ is the time corresponding to the image. The MWA can produce up to $\sim 10^5$ images every minute. It quickly becomes infeasible to explore such large data sets using conventional human effort intensive approaches. Though, we are currently working with much slower image rate of $\sim 5.8\times 10^3$ images per minute, these are already challenging enough to require significant parallalization and processing power. This can be achieved both in the model implementation and the pre-processing of data required by the ML algorithm. In order to utilize this extensive amount of data and synthesise the information from them, we use an unsupervised ML technique called Self-Organising Maps (SOMs) \citep{kohonen-self-organized-formation-1982}. SOMs are a form of unsupervised neural networks that produce a lower dimensional (typically two) representation of the data, while preserving the topology. They can be used to explore similarities and/or dissimilarities in the solar images, pointing to the direction for further investigation. Visualization also forms a large part of the exploration, both before and after application of the unsupervised ML algorithms - before to understand the data and after to understand the patterns identified by the ML algorithm.
\section{Summary}
We have successfully applied Machine Learning to predict star formation properties of galaxies that has bought down the time to predict the specified parameters with very low error. Since ML models are designed to be generalised models, their results can be compared to existing models (simulations etc.) and the data points whose results do not match can be analysed further. This may help to better understand the underlying physical model, or help make the existing simulated models better.
On the other hand, ML can be applied to a vast set of unstructured data to discover phenomena which would otherwise take a lot of time and insight to uncover. Our work in using SOMs in solar radio images, is a promising example in using ML techniques to uncover patterns, clusters and anomalies in the data. As our ability to capture and store data increases, increased use of machine learning to understand the underlying information captured is inevitable.
\acknowledgements SS thanks National Centre for Radio Astrophysics for hosting her for a part of this work.
SS also thanks her colleagues from ThoughtWorks for the numerous interesting and useful discussions, and stimulating ideas.
|
\section{Conclusion}
In this article, we investigate a DL-enabled semantic communication system for speech transmission, named DeepSC-S, to improve the transmission efficiency by transmitting the semantic information only. Particularly, we jointly design the \emph{semantic encoder/decoder} and the \emph{channel encoder/decoder} to learn and extract the speech features, as well as to mitigate the channel distortion and attenuation for practical communication scenarios. Additionally, an attention mechanism based on a squeeze-and-excitation network is utilized to improve the recovery accuracy by minimizing the mean-square error of speech signals. Moreover, in order to enable DeepSC-S to work well over various physical channels, a DeepSC-S model with strong robustness to channel variations is developed. The adaptability of the proposed DeepSC-S is verified under the telephone systems and the multimedia transmission systems. Simulation results demonstrate that DeepSC-S outperforms the various benchmarks in the low SNR regime. Hence, the proposed DeepSC-S is a promising candidate for speech semantic communication systems.
\section{Experiment and Numerical Results}
In this section, we compare the performance of the proposed DeepSC-S, the traditional communication system, and the system with an extra feature coding for speech transmission under the additive white Gaussian noise (AWGN), Rayleigh, and Rician channels, where the accurate CSI is assumed. The details of the adopted benchmarks will be introduced in part A of this section. Moreover, in order to facilitate DeepSC-S with good adaptation to practical environments, it is tested over telephone systems and multimedia transmission systems.
In the whole experiment, we adopt the speech dataset from Edinburgh DataShare, which comprises more than 10,000 \emph{.wav} files trainset and 800 \emph{.wav} files testset with sampling rate 16KHz. In terms of the traditional telephone systems and multimedia transmission systems, the sampling rates for speech signals are 8KHz and 44.1KHz, respectively. Thus, for the experiment regarding telephone systems, the input samples are down-sampled to 8KHz and regarding multimedia transmission systems, the input samples are up-sampled to 44.1KHz. Note that the number of speech samples in different \emph{.wav} is inconsistent. In the simulation, we fix $W=16,384,$ and each sample sequence in $\boldsymbol S$ consists of frames $F=128$ with the frame length $L=128$.
\subsection{Neural Network Setting and Benchmarks}
In the proposed DeepSC-S, the number of SE-ResNet modules in the \emph{semantic encoder/decoder} is 6. For each SE-ResNet module, the number of blocks in the \emph{split} layer is 2, which is achieved by 2 CNN modules with 16 \emph{filters} in each module, and the \emph{transition} layer is implemented by a CNN module with 32 \emph{filters}. Moreover, a single CNN module is utilized in the \emph{channel encoder/decoder}, which contains 8 \emph{filters}. The learning rate is 0.001. The parameters settings of the proposed DeepSC-S in telephone systems are summarized in Table \ref{telephone DeepSC-S NN parameters}. For performance comparison, we provide the following three benchmarks.
\subsubsection{\textbf{Benchmark 1}}
The first benchmark includes the typical source and channel coding techniques. According to ITU-T G.711 standard, 64 Kbps pulse code modulation (PCM) is recommended for speech source coding in telephone systems with $2^8=256$ quantization levels\cite{cox1997three}. Moreover, 16-bits PCM is adopted in our work for speech transmission in multimedia transmission systems with $2^{16}=65,536$ quantization levels. The A-law PCM and uniform PCM are adopted in telephone systems and multimedia transmission systems, respectively. For the channel coding, turbo codes with soft output Viterbi algorithm (SOVA) is adopted\cite{wu1999influence}, in which the coding rate is 1/3, the block length is 512, and the number of decoding iterations is 5. 64-QAM is adopted to make the number of transmitted symbols in the traditional communication systems the same as that in DeepSC-S. The details are summarized in Table \ref{speech parameters}.
\renewcommand\arraystretch{1.15}
\begin{table}[tbp]
\footnotesize
\caption{Parameters settings of the proposed DeepSC-S for telephone systems.}
\label{telephone DeepSC-S NN parameters}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
& \textbf{Layer Name} & \textbf{Filters} & \textbf{Activation} \\
\hline
\multirow{2}{6.0em}{\textbf{\centering \ Transmitter}} & 6$\times$SE-ResNet & 6$\times$64 & ReLU \\
\cline{2-4}
& CNN layer & 8 & None \\
\cline{1-4}
\multirow{3}{6.0em}{\textbf{\centering \ \ \ Receiver}} & CNN layer & 8 & ReLU \\
\cline{2-4}
& 6$\times$SE-ResNet & 6$\times$64 & ReLU \\
\cline{2-4}
& Last layer (CNN) & 1 & None \\
\hline
\end{tabular}
\end{table}
\renewcommand\arraystretch{1.15}
\begin{table}[tbp]
\footnotesize
\caption{Parameters settings of the traditional communication systems.}
\label{speech parameters}
\centering
\begin{tabular}{|c|c|c|}
\hline
& \textbf{Telephone Systems} & \textbf{Multimedia Systems} \\
\cline{1-3}
\textbf{Sample rate} & 8KHz & 44.1KHz \\
\cline{1-3}
\textbf{Samples length} & 16384 & 16384 \\
\cline{1-3}
\textbf{Number of frames} & 128 & 128 \\
\cline{1-3}
\textbf{Frame length} & 128 & 128 \\
\cline{1-3}
\textbf{Source coding} & 8-bits PCM & 16-bits PCM \\
\cline{1-3}
\textbf{Channel coding} & Turbo codes & Turbo codes \\
\cline{1-3}
\textbf{Modulation} & 64-QAM & 64-QAM \\
\hline
\end{tabular}
\end{table}
\subsubsection{\textbf{Benchmark 2}}
The second benchmark combines feature learning with the traditional channel encoding, named semi-traditional system, as shown in Fig. \ref{feature learning}. At the training stage, the \emph{feature encoder} takes the speech samples, $\boldsymbol s$, as the inputs and its output is fed into the \emph{feature decoder} directly. The received signal is converted into the speech information, $\widehat{\boldsymbol s}$, by the \emph{feature decoder}. Based on signals $\boldsymbol s$ and $\widehat{\boldsymbol s}$, the MSE loss is computed at the end of the receiver, thus, the trainable parameters of the \emph{feature encoder} and the \emph{feature decoder} are updated via SGD at the same time.
For the E2E testing, the pre-trained feature leaning system is split into the \emph{feature encoder} and the \emph{feature decoder}, which are placed before the traditional transmitter and after the traditional receiver, respectively. Note that the signal processing blocks of the traditional transmitter and the traditional receiver are same as the settings as shown in Table \ref{speech parameters}. During the training stage, the \emph{feature encoder} and the \emph{feature decoder} are treated as the extraction and recovery operations without considering communication problems. During the testing stage, the system could yield efficient transmission and mitigate the channel effects. The 2D CNN is adopted in the \emph{feature encoder} and the \emph{feature decoder}. The learning rate is 0.001. The parameters settings are summarized in Table \ref{feature learning system NN parameters}.
\subsubsection{\textbf{Benchmark 3}}
In order to emphasize the gain from the attention mechanism, we consider a CNN-based semantic communication system without attention mechanism. The learning rate is 0.001. The parameters settings of the CNN-based system are similar to that of the proposed DeepSC-S as shown in Table \ref{telephone DeepSC-S NN parameters}. The only difference is to replace the SE-ResNet modules in the transmitter/receiver for implementing the \emph{semantic encoder/decoder} with the CNN modules. Each CNN module contains 32 \emph{filters}.
\subsection{Complexity Analysis}\label{complexity analysis}
The complexity of the semi-traditional system is higher than that of the traditional one due to the introduction of the \emph{feature encoder/decoder}. Moreover, due to the complexity of NN training, the CNN-based system and the developed DeepSC-S require higher computational cost than the traditional approaches, which can be measured by the floating point operations (FLOPs). For a single 2D CNN module, the required FLOPs can be expressed as\cite{molchanov2016pruning}:
\begin{equation}
{\mathrm{FLOPs}}_{\mathrm{CNN}}=2\times G\times H\times\left(C_{in}\times K^2+1\right)\times C_{out},
\label{cnn flops}
\end{equation}
where $G$ and $H$ are the width and height of the input feature map of the CNN, respectively. $K^2$ is the kernel size. $C_{in}$ and $C_{out}$ are the number of channels\footnote{Here, the channel refers to the parameter of CNN, not the wireless channels.} of the input and output feature maps, respectively.
\begin{figure}[tbp]
\includegraphics[width=0.45\textwidth]{Figures/benchmark/benchmark2.pdf}
\centering
\caption{The benchmark model by combing a feature encoder with the transmission systems.}
\label{feature learning}
\end{figure}
\renewcommand\arraystretch{1.15}
\begin{table}[tbp]
\footnotesize
\caption{Parameters settings of the benchmark 2.}
\label{feature learning system NN parameters}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
& \textbf{Layer Name} & \textbf{Filters} & \textbf{Activation}\\
\hline
\multirow{2}{7.3em}{\textbf{\centering Feature Encoder}} & 4$\times$CNN modules & 4$\times$32 & ReLU \\
\cline{2-4}
& CNN module & 8 & None \\
\cline{1-4}
\multirow{3}{7.3em}{\textbf{\centering Feature Decoder}} & CNN module & 8 & ReLU \\
\cline{2-4}
& 4$\times$CNN modules & 4$\times$32 & ReLU \\
\cline{2-4}
& CNN module & 1 & None \\
\hline
\end{tabular}
\end{table}
In the semi-traditional system, the number of \emph{filters} of the 5 cascaded CNN modules at the \emph{feature encoder} are 32, 32, 32, 32, 8, respectively. The input size of the \emph{feature encoder} is $128\times128\times1$, i.e., $G=128$ and $H=128$. $K=5$. $C_{in}$ of the 5 CNN modules are 1, 32, 32, 32, 32, respectively, and $C_{out}$ of the 5 CNN modules are 32, 32, 32, 32, 8, respectively. According to (\ref{cnn flops}), the FLOPs required by the \emph{feature encoder} can be calculated as
\begin{equation}
\begin{split}
\mathrm{FLOPs}_{\mathrm{Semi}}^{\mathcal T}&=2\times128\times128\times\left(1\times5\times5+1\right)\times32 \\
&+6\times128\times128\times\left(32\times5\times5+1\right)\times32 \\
&+2\times128\times128\times\left(32\times5\times5+1\right)\times8 \\
&=2.75\times10^9.
\end{split}
\label{flops feature encoder}
\end{equation}
At the receiver, the input size of the \emph{feature decoder} is $128\times128\times8$ and the number of \emph{filters} of the 6 CNN modules are 8, 32, 32, 32, 32, 1, respectively. Then the FLOPs required by the \emph{feature decoder} are $\mathrm{FLOPs}_{\mathrm{Semi}}^{\mathcal R}=2.81\times10^9$. Therefore, the total FLOPs required by the \emph{feature encoder} and the \emph{feature decoder} are $5.56\times10^9$, i.e., the semi-traditional system requires $5.56\times10^9$ FLOPs more than the traditional system.
In the CNN-based system, the number of \emph{filters} of the 7 CNN modules at the transmitter are 32, 32, 32, 32, 32, 32, 8, respectively. The input size of the CNN-based system is $128\times128\times1$, i.e., $G=128$ and $H=128$. $K=5$. $C_{in}$ of the 7 CNN modules are 1, 32, 32, 32, 32, 32, 32, respectively, and $C_{out}$ of the 7 CNN modules are 32, 32, 32, 32, 32, 32, 8, respectively. According to (\ref{cnn flops}), the FLOPs required by the transmitter are $\mathrm{FLOPs}_{\mathrm{CNN}}^{\mathcal T}=4.44\times10^9$. Similarly, the number of \emph{filters} of the 8 CNN modules at the receiver are 8, 32, 32, 32, 32, 32, 32, 1, respectively. Then the FLOPs required by the receiver are $\mathrm{FLOPs}_{\mathrm{CNN}}^{\mathcal R}=4.49\times10^9$. Therefore, the total FLOPs required by the CNN-based system are $8.93\times10^9$.
Furthermore, in the developed DeepSC-S, a single SE-ResNet module consists of 2 CNN modules with 16 \emph{filters} and kernel size is $5\times5$, as well as 1 CNN module with 32 \emph{filters} and kernel size is $1\times1$. Given the input size of a single SE-ResNet is $128\times128\times32$, so $G=128$ and $H=128$. $C_{in}$ of the 3 CNN modules are 32, 32, 32, respectively, and $C_{out}$ of the 3 CNN modules are 16, 16, 32, respectively. Then the FLOPs required by this SE-ResNet module are ${\mathrm{FLOPs}}_{\mathrm{SE}}=8.75\times10^8$. Accordingly, the total FLOPs required by DeepSC-S are $9.36\times10^9$, including $4.65\times10^9$ FLOPs at the transmitter and $4.71\times10^9$ FLOPs at the receiver, which achieves a $4.82\%$ increase over the CNN-based system.
\subsection{Experiments over Telephone Systems}\label{experiment telephpne}
In this experiment, we investigate a robust system to work under various channel conditions by training DeepSC-S under the fixed channel condition and then testing it under different fading channels. Note that during the training stage, the Gaussian noise contained in the three training channels is generated under a fixed SNR value, 8 dB, because we found that 8 dB is the most suitable value after comparing the models trained under different SNR values.
\begin{figure*}[tbp]
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/test_MSE_AWGN.pdf}
\subcaption{AWGN channels}
\label{MSE tested under AWGN channel}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/test_MSE_Rayleigh.pdf}
\subcaption{Rayleigh channels}
\label{MSE tested under Rayleigh channel}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/test_MSE_Rician.pdf}
\subcaption{Rician channels}
\label{MSE tested under Rician channel}
\end{minipage}
\caption{MSE loss tested under (a) AWGN channels, (b) Rayleigh channels, (c) Rician channels with the models trained under various channels.}
\label{MSE comparison for different testing channels}
\end{figure*}
According to Fig. \ref{MSE comparison for different testing channels}, when SNR is lower than 8 dB, DeepSC-S trained under the AWGN channels has higher MSE loss than the models trained under the Rayleigh channels and the Rician channels. As shown in Fig. \ref{MSE comparison for different testing channels} (a), in terms of the MSE loss tested under the AWGN channels, DeepSC-S trained under the AWGN channels outperforms the model trained under the Rayleigh channels and the Rician channels when SNR is higher than around 6 dB. Besides, according to Fig. \ref{MSE comparison for different testing channels} (b), DeepSC-S trained under the AWGN channels performs quite poor in terms of MSE loss when testing under the Rayleigh channels. Furthermore, Fig. \ref{MSE comparison for different testing channels} (c) shows the model trained under the three adopted channels can achieve MSE loss values under $9\times10^{-7}$ when testing under the Rician channels. Therefore, DeepSC-S trained under the Rician channels is adopted as a robust model that is capable of coping with various channel environments.
\subsubsection{SDR and PESQ Results}
Based on the roust model, i.e., DeepSC-S trained under the Rician channels with $\mathrm{SNR}=8\;\mathrm{dB}$, the relationship between the MSE loss values and the number of epochs is shown in Fig. \ref{loss vs epochs}. We can observe that the MSE loss reaches convergence after about 400 epochs.
\begin{figure}[htbp]
\includegraphics[width=0.40\textwidth]{Figures/telephone/MSE_loss.pdf}
\centering
\caption{The training MSE loss versus epoch under the Rician channels with $\mathrm{SNR}=8\;\mathrm{dB}$.}
\label{loss vs epochs}
\end{figure}
Fig. \ref{SDR result for telephone communcations} tests the SDR performance between the three benchmarks and the proposed DeepSC-S under the AWGN, Rayleigh, and Rician channels. From the figure, the semi-traditional system yields higher SDR score than the traditional one under all tested channel environments while its performance is unreliable when SNR is low. Besides, the CNN-based system and DeepSC-S obtain higher SDR score than the other two systems under the Rayleigh channels and the Rician channels, as well as the AWGN channels over most SNR regions. In addition, DeepSC-S performs steadily when coping with different fading channels and SNRs, however, for the semi-traditional system and the traditional system, the performances are quite poor under dynamic channel conditions, especially in the low SNR regime. Moreover, due to the attention mechanism, SE-ResNet, the proposed DeepSC-S achieves higher SDR score than the CNN-based system under all adopted SNRs and fading channels, which proves the effectiveness of DeepSC-S. Particularly, DeepSC-S achieves $10.92\%$ average ascent over the CNN-based system in terms of the SDR performance.
\begin{figure*}[htbp]
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/SDR_Rician_train_AWGN_test.pdf}
\subcaption{AWGN channels}
\label{SDR tested under AWGN channel telephone}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/SDR_Rician_train_Rayleigh_test.pdf}
\subcaption{Rayleigh channels}
\label{SDR tested under Rayleigh channel telephone}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/SDR_Rician_train_Rician_test.pdf}
\subcaption{Rician channels}
\label{SDR tested under Rician channel telephone}
\end{minipage}
\caption{SDR score versus SNR for speech-based telephone communications with the traditional system, the semi-traditional system, the CNN-based system, and DeepSC-S under (a) AWGN channels, (b) Rayleigh channels, (c) Rician channels.}
\label{SDR result for telephone communcations}
\end{figure*}
The PESQ score comparison is shown in Fig. \ref{PESQ result for telephone communcations}. From the figure, the CNN-based system and DeepSC-S outperform the other two systems under various fading channels and SNRs. Moreover, similar to the results of SDR, DeepSC-S obtains good PESQ when coping with channel variation while the traditional one provides poor scores in the low SNR regime. DeepSC-S also achieves higher score than the CNN-based system under all adopted channel conditions. Particularly, DeepSC-S achieves $7.34\%$ average increase over the CNN-basd system in terms of the PESQ performance. Based on the simulated results, the proposed DeepSC-S is able to yield better speech transmission for telephone systems under complicated communication scenarios than the traditional approaches, especially in the low SNR regime.
\begin{figure*}[htbp]
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/PESQ_Rician_train_AWGN_test.pdf}
\subcaption{AWGN channels}
\label{PESQ tested under AWGN channel telephone}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/PESQ_Rician_train_Rayleigh_test.pdf}
\subcaption{Rayleigh channels}
\label{PESQ tested under Rayleigh channel telephone}
\end{minipage}
\begin{minipage}[t]{0.33\linewidth}
\centering
\includegraphics[width=1\textwidth]{Figures/telephone/PESQ_Rician_train_Rician_test.pdf}
\subcaption{Rician channels}
\label{PESQ tested under Rician channel telephone}
\end{minipage}
\caption{PESQ score versus SNR for speech-based telephone communications with the traditional system, the semi-traditional system, the CNN-based system, and DeepSC-S under (a) AWGN channels, (b) Rayleigh channels, (c) Rician channels.}
\label{PESQ result for telephone communcations}
\end{figure*}
\subsection{Experiments over Multimedia Transmission Systems}
In this part, we present the SDR and PESQ performance comparison. The NN parameters settings of the CNN-based system and DeepSC-S are similar to that in the telephone communications experiment, but the number of \emph{filters} of the CNN layer in both systems is 16. Note that the SDR and PESQ results are also tested under the robust model, i.e., DeepSC-S trained under the Rician channels with $\mathrm{SNR}=8\;\mathrm{dB}$.
\begin{figure}[htbp]
\includegraphics[width=0.40\textwidth]{Figures/multimedia/SDR_Rician_train_AWGN_Rician_test.pdf}
\centering
\caption{SDR score versus SNR for speech-based multimedia communications with the traditional system, the semi-traditional system, the CNN-based system, and DeepSC-S under AWGN channels and Rician channels.}
\label{SDR result for multimedia communcations}
\end{figure}
\begin{figure}[htbp]
\includegraphics[width=0.40\textwidth]{Figures/multimedia/PESQ_Rician_train_AWGN_Rician_test.pdf}
\centering
\caption{PESQ score versus SNR for speech-based multimedia communications with the traditional system, the semi-traditional system, the CNN-based system, and DeepSC-S under AWGN channels and Rician channels.}
\label{PESQ result for multimedia communcations}
\end{figure}
Fig. \ref{SDR result for multimedia communcations} depicts the SDR performance comparison for multimedia communications between DeepSC-S and the three benchmarks under the AWGN channels and the Rician channels. Under the AWGN channels, the SDR score of the traditional communication system shows a sharp increase when SNR is higher than 8 dB. Moreover, it achieves SDR scores over 80 in high SNRs due to the high PCM quantization accuracy. However, DeepSC-S can reach universal strong SDR values for all tested SNRs and fading channels. Moreover, DeepSC-S outperforms the semi-traditional system and the traditional system under the Rician channels, as well as the AWGN channels in the low SNR regime. Furthermore, DeepSC-S has higher SDR score than the CNN-based system as the SE-ResNet module is utilized to learn and extract the essential information.
The simulation result of PESQ for multimedia communications is illustrated in Fig. \ref{PESQ result for multimedia communcations}. From the figure, DeepSC-S outperforms the three benchmarks under the Rician channels with any tested SNRs as well as the AWGN channels with low SNR values. Moreover, similar to the results of SDR, the proposed DeepSC-S achieves higher PESQ score than the CNN-based system under all adopted channel conditions. Thus, it is believed that the investigated DeepSC-S is with greater adaptability than the traditional system for speech-based multimedia communications when coping with channel variation.
\section{Introduction}
\IEEEPARstart{I}{nspired} by the success in various areas, deep learning (DL) has been considered as a promising candidate for communications to achieve higher system performance with more intelligence\cite{qin2019deep,qin2020federated}. Particularly, DL has shown its great potentials to solve the existing technical problems in both physical layer communications\cite{gruber2017deep,ye2017power,8227772} and wireless resource allocations\cite{sun2018learning,liang2019deep}.
Typically, a DL-based communication system is designed to reduce the complexity and/or improve the system performance, by merging one or multiple communication modules in the traditional block-wise architecture and using a neural network (NN) to represent the intelligent transceiver. However, even if the DL-enabled communication systems yield better performance and/or lower complexity for some scenarios and conditions, their state-of-the-art models mainly focus on performance improvement at the bit or symbol level, which usually takes bit-error rate (BER) or symbol-error rate (SER) as the performance metric. Particularly, the major task in the traditional communication systems and the developed DL-enabled systems is to recover the transmitted message accurately and effectively, represented by digital bit sequences. In the past decades, such type of wireless communication systems have experienced significant development from the first generation (1G) to the fifth generation (5G) with the system capacity approaching Shannon limit.
Shannon and Weaver\cite{weaver1953recent} categorized communications into three levels:
\begin{itemize}
\item \emph{Level A}: how accurately can the symbols of communication be transmitted? (The technical problem)
\item \emph{Level B}: how precisely do the transmitted symbols convey the desired meaning? (The semantic problem)
\item \emph{Level C}: how effectively does the received meaning affect conduct in the desired way? (The effectiveness problem)
\end{itemize}
This indicates the feasibility to transmit the semantic information, instead of the bits or symbols, to achieve higher system efficiency. Besides, due to the wide deployment of intelligent IoT applications, semantic-irrelative communications are no longer ideal as they transmit bit sequences, which contains information that could be relevant or not to the intelligent tasks at the receiver. Moreover, in the typical communication systems, the generated data is more than required, which limits the number of devices to be covered by the same network. Motivated by this, researchers are dedicating to develop a new system to process and exchange semantic information for more efficient communications.
Semantic theory takes into account the meaning and veracity of source information because they can be both informative and factual\cite{carnap1952outline}, which facilitates the semantic communication systems to transmit only the semantic information at the transmitter and to recover information at the receiver via minimizing the semantic error instead of BER/SER. Nevertheless, the exploration of semantic communications has gone through decades of stagnation since it was first identified because of the limitation of some fundamental problems, e.g., lack of mathematical model for semantic information. The semantic information refers to the information relevant to the transmission goal at the receiver, however, even the most cutting-edge work cannot define the semantic information or semantic features by a precise mathematical formula. Moreover, the semantic information varies for different transmission purposes, which could be in various formats, e.g., age of information\cite{maatouk2020age}, or more complicated semantic features.
Semantic data can be compressed to a proper size for transmission by using a lossless method\cite{basu2014preserving}, which utilizes the semantic relationship between different messages, while the traditional lossless source coding is to represent a signal with the minimum number of binary bits by exploring the dependencies or statistical properties of input signals. In addition, an end-to-end (E2E) communication system has been developed\cite{o2017introduction} in order to address the bottlenecks in traditional block-wise communication systems. Inspired by this, different types of sources have been considered in recent investigations on E2E semantic communication systems, which mainly focuses on the image and text transmission\cite{guler2018semantic,9398576,xie2020lite,bourtsoulatze2019deep,kurka2020deepjscc,jankowski2020wireless,lee2019deep,jankowski2020joint}. The investigation on semantic communication for speech signals transmission is still missed.
A semantic communication system with different inputs is shown in Fig. \ref{sound semantic communication}, which only transmits the semantic features highly relevant to the transmission task at the receiver. While the transmission tasks could be either the source message recovery or more intelligent tasks. For example, in speech signal processing, one intelligent task is to convert speech signals into text information, e.g., automatic speech recognition (ASR), but do not care about the characteristics of speech signals, e.g., the speaking speed and tone\cite{dahl2011context}. The core of ASR is to map each phoneme into a single alphabet, and then concatenate all alphabets into an understandable word sequence via a language model. In this case, the extracted semantic features only contain the text characteristics while the other features will not be transmitted by the transmitter. As a result, the network traffic is reduced significantly without performance degradation.
Without loss of generality, we consider a general model of semantic communication to restore the speech sources in this work, which could be extended to serve a specific intelligent task easily in the future. The intuition behind this work is to recover speech signals based on DL technique, which includes the recovery of speech characteristics. However, most DL algorithms pre-process speech signals to obtain magnitude, spectra, or Mel-Frequency Cepstrum by various operations, such as discrete cosine transform (DCT), before feeding into a learning system. Such extra operations capture the unique features of speech signals but runs counter to the motivation of intelligence. Therefore, a DL-enabled semantic communication system by learning semantic information directly from the raw speech signals is of great interest and importance.
In this paper, we propose a DL-enabled semantic communication system for speech signals, named DeepSC-S, by learning and extracting speech signals, and then recovering them at the receiver from the received features directly. The main contributions of this article can be summarized as fourfold:
\begin{itemize}
\item The DeepSC-S is first developed, which treats the transmitter and the receiver as two NNs. The joint semantic-channel coding is developed to deal with source distortions and channel effects.
\item Particularly, a squeeze-and-excitation (SE) network\cite{8701503} is employed in the developed DeepSC-S to learn and extract essential speech semantic information, as well assign high values to the weights corresponding to the essential information during the training phase. By exploiting the attention mechanism based on an SE network, DeepSC-S improves the accuracy of signal recovery.
\item Moreover, a trained model with high robustness to channel variations is developed by training DeepSC-S under a fixed channel condition, and then facilitating it with good performance when coping with different testing channel environments.
\item To verify the model adaptation to practical communication scenarios, the proposed DeepSC-S is applied to telephone systems and multimedia transmission systems, respectively. The performance is also compared with the traditional approaches to prove its superiority, especially in the low SNR regime.
\end{itemize}
\begin{figure}[tbp]
\includegraphics[width=0.45\textwidth]{Figures/model/semantic_communication.pdf}
\centering
\caption{A semantic communication system with different inputs.}
\label{sound semantic communication}
\end{figure}
The rest of this article is structured as follows. The related work is presented in Section \uppercase\expandafter{\romannumeral2}. Section \uppercase\expandafter{\romannumeral3} introduces the model of semantic communication system for speech transmission and the related performance metrics. In Section \uppercase\expandafter{\romannumeral4}, the details of the proposed DeepSC-S are presented. Simulation results are discussed in Section \uppercase\expandafter{\romannumeral5}. Section \uppercase\expandafter{\romannumeral6} draws conclusions.
\emph{Notation}: The single boldface letters are used to represent vectors or matrices and single plain capital letters denote integers. Given a vector $\boldsymbol x$, $x_i$ indicates its $i$th component, $\left\|\boldsymbol x\right\|$ denotes its Euclidean norm. Given a matrix $\boldsymbol Y$, $\boldsymbol Y\in\mathfrak R^{M\times N}$ indicates that $\boldsymbol Y$ is a matrix with real values and its size is $M\times N$. Superscript swash letters refer the blocks in the system, e.g., $\mathcal T$ in $\boldsymbol\theta^{\mathcal T}$ represents the parameter at the transmitter and $\mathcal R$ in $\boldsymbol\theta^{\mathcal R}$ represents the parameter at the receiver. $\mathcal{CN}(\boldsymbol m,\;\boldsymbol V)$ denotes multivariate circular complex Gaussian distribution with mean vector $\boldsymbol m$ and co-variance matrix $\boldsymbol V$. Moreover, $\boldsymbol a\ast\boldsymbol b$ represents the convolution operation on the vector $\boldsymbol a$ and the vector $\boldsymbol b$.
\section{Proposed Semantic Communication System for Speech Signals}
To address the aforementioned challenges, we design a DL-enabled semantic communication system for speech transmission, named DeepSC-S. Specifically, an attention-based two-dimension (2D) CNN is used for the \emph{semantic encoder/decoder} and a 2D CNN is adopted for the \emph{channel encoder/decoder}. The details of the developed DeepSC-S will be introduced in this section.
\subsection{Model Description}
As shown in Fig. \ref{proposed sys}, the input of the proposed DeepSC-S, denoted as $\boldsymbol S\in\mathfrak R^{B\times W}$, is a set of speech sample sequences, $\boldsymbol s$, which is drawn from the speech dataset, $\mathfrak S$, and $B$ is the batch size. $\mathfrak S$ consists of considerable speech sequences, which are collected by recording the speakings from different persons. The input sample sequences, $\boldsymbol S$, are framed into $\boldsymbol m\in\mathfrak R^{B\times F\times L}$ for training before passing through an attention-based encoder, i.e., the \emph{semantic encoder}, where $F$ indicates the number of frames and $L$ is the length of each frame. Note that the framing operation only reshapes $\boldsymbol S$ without any feature learning and extracting. The \emph{semantic encoder} directly learns the speech semantic information from $\boldsymbol m$ and outputs the learned features $\boldsymbol b\in\mathfrak R^{B\times F\times L\times D}$. The details of the \emph{semantic encoder} are presented in part B of this section. Afterwards, the \emph{channel encoder}, denoted as a CNN layer with 2D CNN modules, converts $\boldsymbol b$ into $\boldsymbol U\in\mathfrak R^{B\times F\times2N}$. In order to transmit $\boldsymbol U$ into a physical channel, it is reshaped into symbol sequences, $\boldsymbol X\in\mathfrak R^{B\times FN\times2}$, via a reshape layer.
The channel layer takes the reshaped symbol sequences, $\boldsymbol X$, as the input and produces $\boldsymbol Y$ at the receiver, which is given by
\begin{equation}
\boldsymbol Y=\boldsymbol H\boldsymbol X+\boldsymbol W,
\label{proposed channel}
\end{equation}
where $\boldsymbol H$ consists of $B$ number of channel coefficient vectors, $\boldsymbol h$, and $\boldsymbol W$ is Gaussian noise, which includes $B$ number of noise vectors, $\boldsymbol w$.
\begin{figure*}[htbp]
\includegraphics[width=1\textwidth]{Figures/model/proposed_sys.pdf}
\centering
\caption{The proposed system architecture for the speech semantic communication system.}
\label{proposed sys}
\end{figure*}
The received signal, $\boldsymbol Y$, is reshaped into $\boldsymbol V\in\mathfrak R^{B\times F\times2N}$ before feeding into the \emph{channel decoder}, represented by a CNN layer with 2D CNN modules. The output of the \emph{channel decoder} is $\widehat{\boldsymbol b}\in\mathfrak R^{B\times F\times L\times D}$. Afterwards, an attention-based decoder, i.e., the \emph{semantic decoder}, converts $\widehat{\boldsymbol b}$ into $\widehat{\boldsymbol m}\in\mathfrak R^{B\times F\times L}$ and $\widehat{\boldsymbol m}$ is recovered into $\widehat{\boldsymbol S}$ via the inverse operation of framing, named deframing. The size of $\widehat{\boldsymbol S}$ is same as that of $\boldsymbol S$ at the transmitter. The loss is calculated at the end of the receiver and backpropagated to the transmitter, thus, the trainable parameters in the whole system can be updated simultaneously.
\subsection{Semantic Encoder and Decoder}\label{section semantic encoder and decoder}
It is intuitive that speech signals at the silent time-slots carry no information, while speech signals at the speaking time-slots carry the essential information. Therefore, signals with zero magnitudes should be ignored, while signals with higher magnitudes should be processed with more attention. Moreover, at the speaking time-slots, people will speak loudly to emphasize the essential information. Correspondingly, the magnitudes of speech signals will increase abruptly. Similarly, people usually speak slowly to state the incomprehensible information, which results in a sudden drop on the frequency of corresponding speech signals. The essential information refers to the features that exist at the speaking time-slots, however, these features are hard or impossible to be captured by mathematical formula, which makes it difficult to measure the feature difference between the input speech signals and the recovered ones straightforwardly, or integrate them into loss function.
Inspired by this, we propose the DeepSC-S including the \emph{semantic encoder} and the \emph{semantic decoder} based on an attention mechanism, named SE-ResNet. In this work, SE-ResNet is utilized to learn and extract the essential information of speech signals. Particularly, SE-ResNet is capable to identify the essential information by assigning higher weights for them during the NN training process.
As shown in Fig. \ref{SE-ResNet}, for the SE-ResNet, a \emph{split} layer takes $\boldsymbol m$ as the input and produces multiples blocks, besides, all these blocks are concatenated. Then a \emph{transition} layer is utilized to reduce the dimension of these concatenated blocks and the output is denoted as $\boldsymbol p\in\mathfrak R^{M\times N\times C}$, which consists of $C$ features and each feature is in size of $M\times N$. For the SE layer, a \emph{squeeze} operation is employed to aggregate the 2D spatial dimension of each input feature, then an operation, named \emph{excitation}, intents to output the attention factor of each feature by learning the inter-dependencies of features in $\boldsymbol p$. The output of the SE layer, $\boldsymbol z\in\mathfrak R^{1\times1\times C}$, includes $C$ number of scale coefficients, which is utilized to scale the importance of the extracted features in $\boldsymbol p$, i.e., the features corresponding to the essential information in $\boldsymbol p$ are multiplied by the high scale coefficients in $\boldsymbol z$. By doing so, the weights of $\boldsymbol m$ are reassigned, i.e., the weights corresponding to the essential speech information are paid more attention. Note that the SE layer is an independent unit and one or multiple SE-ResNet modules can be sequentially connected. With more SE-ResNet modules, the performance of feature learning and extracting to the essential information will improve, however, it also increases the computational cost. Therefore, a trade-off between the learning performance and complexity should be considered during the training phase. Additionally, residual network is adopted to alleviate the problem of gradient vanishing due to the network depth by adding $\boldsymbol m$ into the output of SE-ResNet module, as shown in Fig. \ref{SE-ResNet}.
Particularly, the \emph{semantic encoder} is comprised by multiple SE-ResNet modules to convert input $\boldsymbol m$ into $\boldsymbol b$, corresponding to Fig. \ref{proposed sys}. For the \emph{semantic decoder}, in addition to several SE-ResNet modules, the \emph{last layer}, including a 2D CNN module with a single \emph{filter}, is utilized to reduce the size of output $\widehat{\boldsymbol m}$, because the size of $\boldsymbol m$ and $\widehat{\boldsymbol m}$ should be equal.
\begin{figure}[tbp]
\begin{minipage}[t]{1\linewidth}
\centering
\graphicspath{ {Figures/} }
\includegraphics[width=0.9\textwidth]{Figures/model/ResNet_encoder.pdf}
\subcaption{Attention-based semantic encoder.}
\label{Attention-based Encoder}
\end{minipage}
\begin{minipage}[t]{1\linewidth}
\centering
\graphicspath{ {Figures/} }
\includegraphics[width=0.9\textwidth]{Figures/model/ResNet_decoder.pdf}
\subcaption{Attention-based semantic decoder.}
\label{Attention-based Decoder}
\end{minipage}
\caption{The proposed \emph{semantic encoder} and \emph{semantic decoder} based on SE-ResNet.}
\label{SE-ResNet}
\end{figure}
\subsection{Model Training and Testing}
Based on the prior knowledge of CSI, the transmitter and the receiver parameters, $\boldsymbol\theta^{\mathcal T}$ and $\boldsymbol\theta^{\mathcal R}$, can be updated simultaneously. As aforementioned, the objective of the proposed DeepSC-S is to train a model to recover the speech signals and make it to work well under various fading channels and a wide SNR regime.
\subsubsection{Training Stage}
According to Fig. \ref{proposed sys}, the training algorithm of DeepSC-S is described in Algorithm \ref{training algorithm}. During the training stage, in order to facilitate the fast MSE loss convergence, the NN parameters, $\boldsymbol\theta=(\boldsymbol\theta^{\mathcal T},\boldsymbol\;\boldsymbol\theta^{\mathcal R})$, are initialized by a variance scaling initializer instead of 0. Besides, for achieving a valid training task, the MSE loss converges until the loss is no longer decreasing. The number of SE-ResNet modules is an important hyperparameter, which aims to facilitate the good performance of the \emph{semantic encoder/decoder} and the reasonable training time. Moreover, noise $\boldsymbol W$ in the channel layer is generated by a fixed SNR value.
After passing through the whole transceiver, the sample sequences set, $\boldsymbol S$, is recovered into $\widehat{\boldsymbol S}$, the size of $\boldsymbol S$ and $\widehat{\boldsymbol S}$ should be equal. Furthermore, the loss is computed at the end of the receiver according to (\ref{loss function}) and the parameters are updated by (\ref{SGD}).
\begin{algorithm}[htb]
\caption{Training algorithm of the proposed DeepSC-S}
\label{training algorithm}
\textbf{Initialization:} initialize parameters $\boldsymbol\theta^{\mathcal T{\boldsymbol(\mathbf0\boldsymbol)}}$ and $\boldsymbol\theta^{\mathcal R{\boldsymbol(\mathbf0\boldsymbol)}}$, $i=0$.
\begin{algorithmic}[1]
\State \textbf{Input:} Speech sample sequences $\boldsymbol S$ from speech dataset $\mathfrak S$, fading channel $\boldsymbol H$, noise $\boldsymbol W$ generated under a fixed SNR value.
\State Framing $\boldsymbol S$ into $\boldsymbol m$ with trainable size.
\While{Stop criterion is not meet}
\State $\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\boldsymbol m)\rightarrow\boldsymbol b$.
\State $\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\boldsymbol b)\rightarrow\boldsymbol X$.
\State Transmit $\boldsymbol X$ over physical channel and receive $\boldsymbol Y$ via \State (\ref{channel}).
\State $\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\boldsymbol Y)\rightarrow\widehat{\boldsymbol b}$.
\State $\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\widehat{\boldsymbol b})\rightarrow\widehat{\boldsymbol m}$.
\State Deframing $\widehat{\boldsymbol m}$ into $\widehat{\boldsymbol S}$.
\State Compute loss ${\mathcal L}_{MSE}(\boldsymbol\theta^{\mathcal T},\;\boldsymbol\theta^{\mathcal R})$ via (\ref{loss function}).
\State Update trainable parameters simultaneously via SGD:
\begin{equation}
\boldsymbol\theta^{\mathcal T\left(i+1\right)}\leftarrow\boldsymbol\theta^{\mathcal T\left(i\right)}-\eta\nabla_{\boldsymbol\theta^{\mathcal T\left(i\right)}}{\mathcal L}_{MSE}(\boldsymbol\theta^{\mathcal T},\;\boldsymbol\theta^{\mathcal R})
\label{transmitter paramter update}
\end{equation}
\begin{equation}
\boldsymbol\theta^{\mathcal R\left(i+1\right)}\leftarrow\boldsymbol\theta^{\mathcal R\left(i\right)}-\eta\nabla_{\boldsymbol\theta^{\mathcal R\left(i\right)}}{\mathcal L}_{MSE}(\boldsymbol\theta^{\mathcal T},\;\boldsymbol\theta^{\mathcal R})
\label{receiver paramter update}
\end{equation}
\State $i\leftarrow i+1$.
\EndWhile
\State \textbf{end while}
\State \textbf{Output:} Trained networks $\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\cdot)$, $\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\cdot)$, $\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\cdot)$, and $\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\cdot)$.
\end{algorithmic}
\end{algorithm}
\subsubsection{Testing Stage}
Based on the trained networks $\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\cdot)$, $\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\cdot)$, $\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\cdot)$, and $\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\cdot)$ from the outputs of Algorithm \ref{training algorithm}, the testing algorithm of DeepSC-S is illustrated in Algorithm \ref{testing algorithm}. Note that the speech sample sequences used for testing are different from that used for training.
The model is trained under a certain fading channel and a fixed SNR value, however, it is impractical to retrain the model for each possible channel condition and load all these models to the transmitter and the receiver. By comparing the testing results for models trained under various channel conditions, we adopt one of these models as the robust model, which could achieve good performance when coping with different channel environments. Accordingly, the robust model is employed to test the performance under various fading channels and different SNR values during the testing stage, as shown in Algorithm \ref{testing algorithm}.
\begin{algorithm}[htb]
\caption{Testing algorithm of the proposed DeepSC-S}
\label{testing algorithm}
\begin{algorithmic}[1]
\State \textbf{Input:} Speech sample sequences $\boldsymbol S$ from speech dataset $\mathfrak S$, trained networks $\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\cdot)$, $\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\cdot)$, $\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\cdot)$, and $\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\cdot)$, testing channel set $\mathcal H$, a wide range of SNR values.
\State Framing $\boldsymbol S$ into $\boldsymbol m$ with trainable size.
\For{each channel condition $\boldsymbol H$ drawn from $\mathcal H$}
\For{each SNR value}
\State Generate Gaussian noise $\boldsymbol W$ under the SNR value.
\State $\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\boldsymbol m)\rightarrow\boldsymbol b$.
\State $\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\boldsymbol b)\rightarrow\boldsymbol X$.
\State Transmit $\boldsymbol X$ over physical channel and receive $\boldsymbol Y$ \State via (\ref{channel}).
\State $\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\boldsymbol Y)\rightarrow\widehat{\boldsymbol b}$.
\State $\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\widehat{\boldsymbol b})\rightarrow\widehat{\boldsymbol m}$.
\State Deframing $\widehat{\boldsymbol m}$ into $\widehat{\boldsymbol S}$.
\EndFor
\State \textbf{end for}
\EndFor
\State \textbf{end for}
\State \textbf{Output:} Recovered speech sample sequences, $\widehat{\boldsymbol S}$, under different fading channels and various SNR values.
\end{algorithmic}
\end{algorithm}
\section{Related Work}
This section first introduce the related work on E2E communication systems to address the challenges in traditional systems. Then, we discuss the state-of-the-art semantic communications.
\subsection{End-to-End Communication Systems}
The DL-enabled E2E communication systems have achieved extremely competitive block-error rate (BLER) performance in comparison to the traditional block-wise communication systems in various scenarios\cite{o2017introduction}. In addition, it has shown great potentials in processing complicated communication tasks. For examples, the E2E learning has been employed in orthogonal frequency division multiplexing (OFDM) systems\cite{kim2017novel,felix2018ofdm}, as well as in multiple-input multiple-output (MIMO) systems\cite{o2017physical,he2020model}. Besides, channel estimation is a challenging problem in the DL-enabled E2E systems. In\cite{aoudia2019model}, reinforcement learning (RL) has been adopted to estimate the channel state information (CSI) through treating the channel layer and the receiver as the \emph{environment}. However, it requires a reliable channel to feedback the losses from the receiver to the transmitter during the training phase. Another novel channel agnostic solution has been proposed in\cite{ye2020deep}, which replaces the realistic channel with a deep neural network (DNN) by exploiting a conditional generative adversarial network (GAN).
Due to the complexity of NN training, a system with high training efficiency and low energy consumption is more than desired to make the E2E system applicable in practice. Transfer learning is a promising technology for adapting E2E communication systems to cope with the uncontrollable and unpredictable channel environments by training them over a statistical channel model\cite{pan2009survey}. Another appealing solution is to obtain a trained model yielding expected performance via the stochastic gradient descent (SGD) with a small number of iterations. Particularly, based on model-agnostic meta-learning (MAML), an E2E communication system has been developed in\cite{park2020meta}, which finds the initial NN parameters to achieve fast convergence for various channel conditions.
\subsection{Semantic Communications}
An initial research on semantic communication systems for text information was developed\cite{guler2018semantic}, which mitigates the semantic error by integrating the semantic inference and the physical layer communications to optimize the whole transceiver. However, such a text-based semantic communication system only measures the semantic error at the word level instead of the sentence level. Thus, a further investigation on semantic text transmission, named DeepSC, has been developed\cite{9398576} to deal with the semantic error at the sentence level with various length. Powered by the Transformer\cite{10.5555/3295222.3295349}, the semantic encoder and the channel encoder are co-designed to minimize the semantic error and to improve the system capacity. Moreover, the increasing deployment of smart IoT applications requires IoT devices to be capable of dealing with more complicated tasks, which runs counter to the limited computing capability of IoT devices. Inspired by this, a lite distributed semantic communication system for text transmission, named L-DeepSC\cite{xie2020lite}, has been further proposed to address the challenge of IoT devices by pruning and quantizing NN parameters. By doing so, the size of the trained model as well as the communication cost between the IoT devices and the server are reduced significantly, which makes it more suitable for IoT applications.
In semantic communications for image information, a DL-enabled semantic communication system has been developed\cite{bourtsoulatze2019deep}, which employs a convolutional neural network (CNN) at the transmitter to jointly design the source-channel coding. Besides, an image transmission system has been investigated to improve the accuracy of image reconstruction\cite{kurka2020deepjscc}, which backpropagates the channel output in order to generate a weight vector to the NN at the transmitter. In addition to perform the typical image reconstruction, more intelligent tasks have been considered. Particularly, an application based on a joint source-channel coding (JSCC) model to retrieve image has been proposed\cite{jankowski2020wireless}, which aims to reduce the transmission latency for IoT devices by achieving retrieval-oriented image compression. Besides, a joint image transmission-recognition system has been developed\cite{lee2019deep} for IoT applications, which has the superior recognition accuracy than the traditional approaches but requires the affordable computation resource. Moreover, by exploiting NN compression techniques, a deep JSCC\cite{jankowski2020joint} has been developed to perform image classification at the edge sever, which facilitates the IoT devices to process images with low computational complexity and to reduce the required transmission bandwidth. Note that the concept of semantic communications was not clearly stated by the authors in the aforementioned work on image transmission. However, we treat them as the pioneering works in the area as they share the spirit of semantic communications by extracting and transmitting the relevant information from the source for serving the transmission goals at the receiver.
\section{System Model}
In this section, we first introduce the considered system model, then the performance metrics are presented. The considered system transmits the original speech signals via a NN-based speech semantic communication system, which comprises two major tasks: \romannumeral1) semantic information learning and extracting from speech signals; \romannumeral2) mitigating the effects of wireless channels. For a practical communication scenario, the signal passing through the physical channel suffers from distortion and attenuation. Therefore, the considered DL-enabled system targets to recover the speech signals and to outperform the traditional approaches when coping with complicated channel conditions.
\subsection{Transmitter}\label{section system model-transmitter}
The proposed system model is shown in Fig. \ref{sys model}. From the figure, the input of the transmitter is a speech sample sequence, $\boldsymbol s=\left[s_1,\;s_2,\;...,\;s_W\right]$ with $W$ samples, where $s_w$ is $w$th item in $\boldsymbol s$ and it is a scalar value, i.e., a positive number, a negative number, or zero. At the transmitter, the input, $\boldsymbol s$, is mapped into symbols, $\boldsymbol x$, to be transmitted over physical channels. As shown in Fig. \ref{sys model}, the transmitter consists of two individual components: the \emph{semantic encoder} and the \emph{channel encoder}, each component is implemented by an independent NN. Denote the NN parameters of the \emph{semantic encoder} and the \emph{channel encoder} as $\boldsymbol\alpha$ and $\boldsymbol\beta$, respectively. Then the encoded symbol sequence, $\boldsymbol x$, can be expressed as
\begin{equation}
\boldsymbol x=\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\boldsymbol s)),
\label{auto-encoder}
\end{equation}
where $\mathbf T_{\boldsymbol\alpha}^{\mathcal S}(\cdot)$ and $\mathbf T_{\boldsymbol\beta}^{\mathcal C}(\cdot)$ indicate the \emph{semantic encoder} and the \emph{channel encoder} with respect to (w.r.t.) parameters $\boldsymbol\alpha$ and $\boldsymbol\beta$, respectively. Here we denote the NN parameters of the transmitter as $\boldsymbol\theta^{\mathcal T}=(\boldsymbol\alpha,\boldsymbol\;\boldsymbol\beta)$.
The mapped symbols, $\boldsymbol x$, are transmitted over a physical channel. Note that the normalization on transmitted symbols $\boldsymbol x$ is required to ensure the total transmit power constraint $\mathbb{E}\left\|\boldsymbol x\right\|^2=1$.
\begin{figure}[tbp]
\includegraphics[width=0.45\textwidth]{Figures/model/sys_model.pdf}
\centering
\caption{The model structure of DL-enabled speech semantic communication system.}
\label{sys model}
\end{figure}
The whole transceiver in Fig. \ref{sys model} is designed for a single communication link. The channel layer, represented by $p_h\left(\left.\boldsymbol y\right|\boldsymbol x\right)$, takes $\boldsymbol x$ as the input and produces the output as received signal $\boldsymbol y$. Denote the coefficients of a linear channel as $\boldsymbol h$, then the transmission process from the transmitter to the receiver can be modeled as
\begin{equation}
\boldsymbol y=\boldsymbol h\ast\boldsymbol x+\boldsymbol w,
\label{channel}
\end{equation}
where $\boldsymbol w\sim\mathcal{CN}(0,\;\sigma^2\mathbf I)$ indicates independent and identically distributed (i.i.d.) Gaussian noise, $\sigma^2$ is noise variance for each channel and $\mathbf I$ is the identity matrix.
\subsection{Receiver}
Similar to the transmitter, the receiver also consists of two cascaded parts, including the \emph{channel decoder} and the \emph{semantic decoder}. The \emph{channel decoder} is to mitigate the channel distortion and attenuation, and the \emph{semantic decoder} recovers speech signals based on the learned and extracted speech semantic features. Denote the NN parameters of the \emph{channel decoder} and the \emph{semantic decoder} as $\boldsymbol\chi$ and $\boldsymbol\delta$, respectively. As depicted in Fig. \ref{sys model}, the decoded signal, $\widehat{\boldsymbol s}$, can be obtained from the received signal, $\boldsymbol y$, by the following operation:
\begin{equation}
\widehat{\boldsymbol s}=\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\boldsymbol y)),
\label{auto-decoder}
\end{equation}
where $\mathbf R_{\boldsymbol\chi}^{\mathcal C}(\cdot)$ and $\mathbf R_{\boldsymbol\delta}^{\mathcal S}(\cdot)$ indicate the \emph{channel decoder} and the \emph{semantic decoder} w.r.t. parameters $\boldsymbol\chi$ and $\boldsymbol\delta$, respectively. Denote the NN parameters of the receiver as $\boldsymbol\theta^{\mathcal R}=(\boldsymbol\chi,\boldsymbol\;\boldsymbol\delta)$.
The objective of the whole transceiver system is to recover speech signals as close as to the original signals, which causes two challenges. The first one is the design of intelligent \emph{semantic encoder/decoder}, which utilizes the semantic information to recover speech signals, especially under the poor channel conditions, such as the low SNR regime. The second one is the design of the \emph{channel encoder/decoder} to alleviate symbol errors caused by the physical channels via adding redundancy information. For the traditional communications, the advanced channel coding techniques are achieved at the bit level to target a low BER. However, the bit-to-symbol transformation is not involved in our proposed system. The raw speech signals are directly mapped into a transmitted symbol stream by the \emph{semantic encoder} and the \emph{channel encoder}, and recovered at the receiver via inverse operations. Thus, we treat the speech recovery process as a signal reconstruction task to minimize the errors between the signal values in $\boldsymbol s$ and $\widehat{\boldsymbol s}$ by exploiting the characteristics of speech signals, then mean-squared error (MSE) is used as the loss function in our system to measure the difference between $\boldsymbol s$ and $\widehat{\boldsymbol s}$, denoted as
\begin{equation}
{\mathcal L}_{MSE}(\boldsymbol\theta^{\mathcal T},\;\boldsymbol\theta^{\mathcal R})=\frac1W\sum_{w=1}^W{(s_w-{\widehat s}_w)}^2,
\label{loss function}
\end{equation}
where $s_w$ and ${\widehat s}_w$ indicate the $w$th element of vectors $\boldsymbol s$ and $\widehat{\boldsymbol s}$, respectively. $W$ is the length of these two vectors.
Assume that the NN models of the whole transceiver are differentiable w.r.t. the corresponding parameters, which can be optimized via gradient descent based on (\ref{loss function}). It is worth to mention that the \emph{semantic encoder/decoder} and the \emph{channel encoder/decoder} are jointly designed. Besides, given prior CSI, both parameters sets $\boldsymbol\theta^{\mathcal T}$ and $\boldsymbol\theta^{\mathcal R}$ can be adjusted at the same time. Denote the NN parameters of the whole system model as $\boldsymbol\theta$, $\boldsymbol\theta=(\boldsymbol\theta^{\mathcal T},\boldsymbol\;\boldsymbol\theta^{\mathcal R})$, we adopt the SGD algorithm to train task in this paper, which iteratively updates the parameters $\boldsymbol\theta$ as follows:
\begin{equation}
\boldsymbol\theta^{(i+1)}\leftarrow\boldsymbol\theta^{(i)}-\eta\nabla_{\boldsymbol\theta^{(i)}}{\mathcal L}_{MSE}(\boldsymbol\theta^{\mathcal T},\;\boldsymbol\theta^{\mathcal R}),
\label{SGD}
\end{equation}
where $\eta>0$ is a learning rate and $\nabla$ indicates the differential operator.
\subsection{Performance Metrics}\label{section performance metric}
In our model, the system is committed to reconstruct the raw speech signals. Hence, the signal-to-distortion ration (SDR)\cite{vincent2006performance} is employed to measure the ${\mathcal L}_2$ error between $\boldsymbol s$ and $\widehat{\boldsymbol s}$, which is one of the commonly used metric for speech transmission and can be expressed as
\begin{equation}
SDR=10\log_{10}\left(\frac{\left\|\boldsymbol s\right\|^2}{\left\|\boldsymbol s-\widehat{\boldsymbol s}\right\|^2}\right).
\label{SDR}
\end{equation}
The higher SDR represents that the speech information is recovered with better quality, i.e., easier to understand for human beings. According to (\ref{loss function}), MSE loss could reflect the goodness of SDR. The lower the MSE, the higher the SDR.
Furthermore, the good recovery of speech signals is intuitively manifested by a satisfactory listening experience of the recovered speech signals, e.g., no latency and background noise. The perceptual evaluation of speech distortion (PESQ)\cite{rix2001perceptual} is adopted in the International Telecommunication Union (ITU-T) recommendation P.862\cite{ITU-T:p.862}, which is a good candidate for evaluating the quality of speech signals under various conditions, e.g., background noise, analog filtering, and variable delay, by scoring the quality from -0.5 to 4.5. The PESQ score is obtained by multiple operations, e.g., level align, time align and equalise, and disturbance processing\cite{rix2001perceptual}. In this work, we adopt an integrated open source PESQ assessment model developed by the ITU-T\cite{ITU-T:p.862}, which is able to evaluate the PESQ score in few milliseconds.
|
\section{Design}\label{S:design}
The core philosophy of MOOG is \textbf{"one task, one file."} Namely, a task should be implemented with a single configuration file.
This configuration file is a short ``recipe'' for the task, containing as little substantive code as possible, and should define a set of components to pass to the MOOG environment.
See Figure \ref{fig:env_schematic} for a schematic of these components.
\begin{figure}[t!]
\centering
\includegraphics[width=1.\linewidth]{env_schematic.png}
\caption{\textbf{Components of a MOOG environment.}
See main text for details.
}\label{fig:env_schematic}
\end{figure}
A MOOG environment receives the following components (or callables returning
them) from the configuration file:
\begin{itemize}
\item
\textbf{State}.
The state is a collection of sprites.
Sprites are polygonal shapes with color and physical attributes (position, velocity, angular velocity, and mass).
Sprites are 2-dimensional, and the state is 2.5-dimensional with z-ordering for occlusion.
The initial state can be procedurally generated from a custom distribution at the beginning of each episode.
The state is structured in terms of layers, which helps hierarchical organization.
See \href{https://jazlab.github.io/moog.github.io/moog/state_initialization/index.html}{state\_initialization} for procedural generation tools.
\item
\textbf{Physics}.
The physics component is a collection of forces that operate on the sprites.
There are a variety of forces built into MOOG (collisions, friction, gravity, rigid tethers, ...) and additional custom forces can also be used.
Forces perturb the velocity and angular velocity of sprites, and the sprite positions and angles are updated with Newton's method.
See \href{https://jazlab.github.io/moog.github.io/moog/physics/index.html}{physics} for more.
\item
\textbf{Task}.
The task defines the rewards and specifies when to terminate a trial.
See \href{https://jazlab.github.io/moog.github.io/moog/tasks/index.html}{tasks} for more.
\item
\textbf{Action Space}.
The action space allows the subject to control the environment.
Every environment step calls for an action from the subject.
Action spaces may impart a force on a sprite (like a joystick), move a sprite in a grid (like an arrow-key interface), set the position of a sprite (like a touch-screen), or be customized.
The action space may also be a composite of constituent action spaces, allowing for multi-agent tasks and multi-controller games.
See \href{https://jazlab.github.io/moog.github.io/moog/action_spaces/index.html}{action\_spaces} for more.
\item
\textbf{Observers}.
Observers transform the environment state into a observation for the subject/agent playing the task.
Typically, the observer includes a renderer producing an image.
However, it is possible to implement a custom observer that exposes any function of the environment state.
The environment can also have multiple observers.
See \href{https://jazlab.github.io/moog.github.io/moog/observers/index.html}{observers} for more.
\item
\textbf{Game Rules} (optional).
If provided, the game rules define dynamics or transitions not captured by the physics.
A variety of game rules are included, including rules to modify sprites when they come in contact, conditionally create new sprites, and control phase structure of trials (e.g. fixation phase to stimulus phase to response phase).
See \href{https://jazlab.github.io/moog.github.io/moog/game_rules/index.html}{game\_rules} for more.
\end{itemize}
Importantly, all of these components can be fully customized.
If a user would like a physics force, action space, or game rule not provided by MOOG, they can implement a custom one, inheriting from the abstract base class for that component.
This can typically be done with only a few lines of code.
The modularity of MOOG facilitates code re-use across experimental paradigms.
For example, if a user would like to both collect behavior data from humans using a continuous joystick and train RL agents with discrete action spaces on the same task, they can re-use all other components in the task configuration, only changing the action space.
For users interested in doing psychology or neurophysiology, we include an \href{https://github.com/jazlab/moog.github.io/tree/master/mworks}{example} of how to run MOOG through \href{https://mworks.github.io/}{MWorks}, a platform with precise timing control and interfaces for eye trackers, HID devices, electrophysiology software, and more.
\section{Example Tasks}\label{S:examples}
\vspace*{-5pt}
\begin{figure}[t!]
\centering
\includegraphics[width=1.\linewidth]{example_tasks.png}
\caption{\textbf{Example tasks.}
Time-lapse images of four example tasks.
Left-to-right:
(i) Pong - The subject aims to catch the yellow ball with the green paddle,
(ii) Red-Green - The subject tries to predict whether the blue ball with contact the red square or the green square,
(iii) Pac-Man - The subject moves the green agent to catch yellow pellets while avoiding the red ghosts,
(iv) Collisions - the green agent avoids touching the bouncing polygons.
}\label{fig:example_tasks}
\end{figure}
See the \href{https://github.com/jazlab/moog.github.io/tree/master/moog_demos/example_configs}{example\_configs} for a variety of task config files.
Four of those are shown in Figure \ref{fig:example_tasks}.
See the \href{https://github.com/jazlab/moog.github.io/tree/master/moog_demos}{demo documentation} for videos of them all and instructions for how to run them with a Python gui.
The MOOG codebase contains libraries of options for each of the components in Section \ref{S:design}, so implementing a task involves only combining the desired ingredients and feeding them to the environment.
For an example, the following code fully implements a navigate-to-goal task, where the subject must move an agent via a joystick action space to a goal location:
\vspace*{-5pt}
\begin{lstlisting}[language=Python]
"""Navigate-to-goal task."""
import collections
from moog import action_spaces, environment, observers, physics, sprite, tasks
# Initial state is a green agent in the center and red goal in the corner
def state_initializer():
# Goal is a red square with side-length 0.1 at position (0.1, 0.1)
# Color channels are [c0, c1, c2] arguments
goal = sprite.Sprite(x=0.1, y=0.1, shape='square', scale=0.1, c0=255)
# Agent is a green circle at position (0.5, 0.5)
agent = sprite.Sprite(x=0.5, y=0.5, shape='circle', scale=0.1, c1=255)
state = collections.OrderedDict([
('goal', [goal]),
('agent', [agent]),
])
return state
# Physics is a drag force on the agent to limit velocity
phys = physics.Physics((physics.Drag(coeff_friction=0.25), 'agent'))
# Task gives a reward 1 when the agent reaches the goal, and a new trial begins
# 5 timesteps later.
task = tasks.ContactReward(1., 'agent', 'goal', reset_steps_after_contact=5)
# Action space is a joystick with maximum force 0.01 arena widths per timestep^2
action_space = action_spaces.Joystick(0.01, 'agent')
# Observer is an image renderer
obs = observers.PILRenderer(image_size=(256, 256))
# Create the environment, ready to play!
env = environment.Environment(
state_initializer=state_initializer,
physics=phys,
task=task,
action_space=action_space,
observers={'image': obs},
}
\end{lstlisting}
\vspace*{-10pt}
This is an extremely simple task, but by complexifying the state initializer and adding additional forces and game rules, a \href{https://github.com/jazlab/moog.github.io/tree/master/moog_demos}{wide range of complex tasks} can be implemented with few lines of code.
\section{Intended Users}\label{S:users}
MOOG was designed for use by the following kinds of researchers:
\begin{itemize}
\item
Machine learning researchers studying reinforcement learning in 2.5-dimensional (2-D with occlusion) physical environments who want to quickly implement tasks in Python.
\item
Psychology researchers who want more flexibility than existing psychology platforms afford.
\item
Neurophysiology researchers who want to study interactive games yet still need to precisely control stimulus timing.
\item
Machine learning researchers studying unsupervised learning, particularly in the video domain.
MOOG can be used to procedurally generate video datasets with controlled statistics.
\end{itemize}
MOOG may be particularly useful for interdisciplinary researchers studying AI
agents, humans, and animals (or some subset thereof) all playing the same task.
\section{Introduction}\label{S:intro}
In recent years, trends towards studying object-based games have gained momentum in the fields of artificial intelligence, cognitive science, psychology, and neuroscience.
In artificial intelligence, interactive physical games are now a common testbed for reinforcement learning \citep{mnih2013playing, sutton2018reinforcement, leike2017ai, Francois_Lavet_2018} and object representations are of particular interest for sample efficient and generalizable AI \citep{battaglia2018relational, van2019perspective, greff2020binding}.
In cognitive science and psychology, object-based games are used to study a variety of cognitive capacities, such as planning, intuitive physics, and intuitive psychology \citep{ullman2017mind, Chabris2017}.
Developmental psychologists also use object-based visual stimuli to probe questions about object-oriented reasoning in infants and young animals \citep{spelke2007core, wood2020reverse}.
In neuroscience, object-based computer games have recently been used to study decision-making and physical reasoning in both human and non-human primates \citep{fischer2016functional, mcdonald2019bayesian, yoo2020neural, Rajalingham2021}.
Furthermore, a growing number of researchers are studying tasks using a combination of approaches from these fields.
Comparing artificial agents with humans or animals performing the same tasks can help constrain models of human/animal behavior, generate hypotheses for neural mechanisms, and may ultimately facilitate building more intelligent artificial agents \citep{lake2017building, willke2019comparison, hassabis2017neuroscience}.
However, building a task that can be played by AI agents, humans, and animals is a time-consuming undertaking because existing platforms are typically designed for only one of these purposes.
Professional game engines are designed for human play and are often complex libraries that are difficult to customize for training AI agents and animals.
Reinforcement learning platforms are designed for AI agents but are often too slow or inflexible for neuroscience work.
Existing psychology and neurophysiology platforms are too limited to easily support complex interactive games.
In this work we offer a game engine that is highly customizable and designed to support tasks that can be played by AI agents, humans, and animals.
\section{Limitations}\label{S:limitations}
Users should be aware of the following limitations of MOOG before choosing to use it for their research:
\begin{itemize}
\item
\textbf{Not 3D}.
MOOG environments are 2.5-dimensional, meaning that they render in 2-dimensions with z-ordering for occlusion.
MOOG does not support 3D sprites.
\item
\textbf{Very simple graphics}.
MOOG sprites are monochromatic polygons.
There are no textures, shadows, or other visual effects.
Composite sprites can be implemented by creating multiple overlapping sprites, but still the graphics complexity is very limited.
This has the benefit of a small and easily parameterizable set of factors of variation of the sprites, but does make MOOG environments visually unrealistic.
\item
\textbf{Imperfect collisions}.
MOOG's collision module implements Newtonian rotational mechanics, but it is not as robust as professional physics engines (e.g. can be unstable if object are moving very quickly and many collisions occur simultaneously).
\end{itemize}
\subsubsection*{Acknowledgments}
We thank Chris Stawarz and Erica Chiu for their contributions to the codebase.
We also thank Ruidong Chen, Setayesh Radkani, and Michael Yoo for their feedback as early users of OOG.
\clearpage
\bibliographystyle{abbrvnat}
\setlength{\bibsep}{5pt}
\setlength{\bibhang}{10pt}
\section{The COBRA model}\label{S:model}
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{figures/model_schematic_global.png}
\caption{\textbf{COBRA model schematic}\\
\textbf{A.} Entire model. The vision module (scene encoder and decoder), transition model, and exploration policy are all trained in a pure exploration phase with no reward. \textbf{B.} Transition model architecture. An action-conditioned slot-wise MLP learns one-step future-prediction. This is trained by applying the scene decoder to $\tilde \myvec{z}_{t+1}$, through which gradients from a pixel loss are passed. An auxiliary transition error prediction provides a more direct path to the pixel loss and makes adversarial training with the exploration policy more efficient. \textbf{C.} Adversarial training of transition model and exploration policy through which the behavior of moving objects emerges.
}
\label{fig:model_schematic}
\vspace*{-1em}
\end{figure*}
Our model (Figure~\ref{fig:model_schematic}) consists of four components: The \textbf{vision module}, \textbf{transition model}, and \textbf{exploration policy} are trained during an unsupervised \textit{exploration phase} without rewards, while the \textbf{reward predictor} is trained during a subsequent \textit{task phase} (with reward).
To present the architectures of these components, we must first briefly discuss the environment:
We are interested in environments containing objects that can be manipulated, in which agents can learn purposeful exploration without reward and a diverse distribution of tasks can be defined.
In all of our experiments we use a 2-dimensional virtual ``touch-screen'' environment that contains objects with configurable shape, position, and color.
The agent can move any visible object by clicking on the object and clicking in a direction for the object to move.
Hence the action space is continuous and 4-dimensional, namely a pair of clicks.
If the first click does not land in an object then nothing will move, regardless of the second click.
Despite its apparent simplicity, this environment supports highly diverse tasks that can challenge existing SOTA agents.
See Section \ref{S:env_tasks} for details about our environment and tasks.
We have open-sourced this environment \footnote{Available here: \url{https://github.com/deepmind/spriteworld}}, calling it ``Spriteworld'' \citep{spriteworld}.
\subsection{Unsupervised exploration phase}\label{SS:unsupervised_exploration_phase}
We want our agent to learn how its environment ``works'' in a task-free \textit{exploration phase}, where no extrinsic reward signals are available.
If this is done well, then the knowledge thus gained should be useful and generalizable for subsequent tasks.
Specifically, in the exploration phase COBRA learns representations of objects, dynamics, and its own action space, which it can subsequently use for any task.
For the exploration phase we let COBRA explore our Spriteworld environment in short episodes.
Each episode is initialized with 1-7 objects with randomly sampled shape, position, orientation and color.
\paragraph{Vision Module} We use MONet \citep{burgess2019} to learn a static scene representation.
MONet is an auto-regressive VAE that learns to decompose a scene into entities, such as objects, without supervision.
We can view it as having an encoder $V_{enc}$ and a decoder $V_{dec}$.
The encoder maps an image $\myvec{x}$ to a tensor $V_{enc}(\myvec{x}) = \myvec{z} \in \mathbb{R}^{K \times M}$ representing the full scene.
This representation consists of $K$ entities, each encoded by the mean of MONet's latent distribution of dimension $M$.
We hereafter call each row $k$ of the scene representation $\myvec{z}$ a ``slot.''
The decoder maps this tensor $\myvec{z}$ to a reconstructed image $V_{dec}(\myvec{z}) = \tilde{\myvec{x}}$.
Critically, MONet learns to put each object from a scene into a different slot.
Moreover, the MONet architecture ensures that each slot has the same representational semantics, so there is a common learned latent space to which each object is mapped.
Each slot represents meaningful properties of the objects, such as position, angle, color and shape.
See Figure~\ref{fig:supp_disentangling} in Appendix~\ref{S:supp_disentangling} for an example.
MONet can handle scene datasets with a variable number of objects by letting some slots' codes lie in a region of latent space that represents no object.
See Figure \ref{fig:scene_action_transition}-A for scene decomposition results, indicating that MONet successfully learns to accurately represent Spriteworld scenes in terms of objects.
While our Spriteworld environment is visually quite simple, in more complex environments MONet's slots learn to represent other elements of a scene (e.g. backgrounds or walls in 3D scenes) and its latent space can capture more complex visual properties (e.g. surface glossiness).
See \citep{burgess2019} for examples.
\begin{figure}[t!]
\centering
\includegraphics[width=0.99\linewidth]{figures/scene_action_transition.png}
\caption{\textbf{Scene decomposition, transition model, and exploration policy results}\\
\textbf{A.} Vision module (MONet) decomposing Spriteworld scenes into objects (one column per sample scene).
\textit{(First row)} Data samples.
\textit{(Second row)} Reconstruction of the full scene.
\textit{(Other rows)} Individual reconstructions of each slot in the learned scene representation.
Some slots are decoded as blank images by the decoder.
\textbf{B.} Rollouts of transition model, treated as an RNN, compared to ground truth on two scenes.
In each scene, one single item (indicated by dotted circle) is being moved along a line in multiple steps.
\textbf{C.} Exploration policy.
\textit{(Left)} Position click component of random samples from the trained exploration policy, which learns to click on (and hence move) objects.
\textit{(Middle)} Slice through the first two dimensions (position click) of the action sampler's quantile function, showing deformations applied on a grid of first clicks $\in [0, 1]^2$ with randomized second click.
\textit{(Right)} Slice through second two dimensions (motion click). There is virtually no deformation, indicating the exploration policy learns to sample motions randomly.
}\label{fig:scene_action_transition}
\end{figure}
\paragraph{Transition model} We introduce a method that maps an action $\myvec{a}_t$ and a scene representation $\myvec{z}_t$ at time step $t$ to a predicted next-step scene representation $\tilde \myvec{z}_{t+1}$.
This model applies a shared MLP $T_{net}$ slot-wise to the concatenations $[\myvec{z}_t^k; \myvec{a}_t]$ of each slot $\myvec{z}_t^k$ with the action $\myvec{a}_t$.
This is sufficient to predict $\myvec{z}_{t+1}$, because in Spriteworld there are no physical interactions between objects and any action will affect at most one object.
To train this transition model, we cannot easily use a loss in representation space, because the encoding $\myvec{z}_{t+1} = V_{enc}(\myvec{x}_{t + 1})$ of image $\myvec{x}_{t + 1}$ may have a different ordering of its slots than $\myvec{z}_t = V_{enc}(\myvec{x}_t)$ and solving the resulting matching problem in a general, robust way is non-trivial.
Instead, we circumvent this problem by applying the visual decoder (through which gradients can be passed) to $\tilde \myvec{z}_{t+1}$ and using a pixel loss in image space:
\begin{align}
\tilde \myvec{z}_{t+1} = T(\myvec{z}_t, \myvec{a}_t) &= [T_{net}(\myvec{z}_t^0, \myvec{a}_t), ..., T_{net}(\myvec{z}_t^K, \myvec{a}_t)]\nonumber\\
Loss_{T}(\myvec{z}_t, \myvec{a}_t) &= ||V_{dec}( T(\myvec{z}_{t}, \myvec{a}_t)) - \myvec{x}_{t + 1}||^2\nonumber
\end{align}
Additionally, we train an extra network $L_T(\myvec{z}, \myvec{a})$ to predict the output of the pixel loss, which we use as a measure of curiosity when training the exploration policy, as we found this to work better and be more stable than using the pixel loss directly.
See Appendix~\ref{S:supp_things_tried} for alternative transition models considered, including those that do not use pixel loss.
See Figure \ref{fig:scene_action_transition}-B for examples of transition model rollouts after training in tandem with the exploration policy (covered below).
\paragraph{Exploration Policy} In many environments a uniformly random policy is insufficient to produce action and observation sequences representative of those useful for tasks.
Consequently, in order for a agent to learn through pure exploration a transition model that transfers accurately to a task-driven setting, it should not take uniformly random actions.
In Spriteworld, this manifests itself in the sparseness of the action space.
If an agent were to take uniformly random actions, it would rarely move any object because the agent must click within an object to move it and the objects only take up a small portion of the image ($\approx 1.7 \%$ per object).
This is shown in Figure~\ref{fig:supp_exploration_behaviour} (top) and Appendix~\ref{S:supp_agent_videos}, and our transition model does not get enough examples of objects moving to be trained well in this condition.
Hence we need our agent to learn in the exploration phase a policy that clicks on and moves objects more frequently.
\begin{figure}[t!]
\centering
\includegraphics[width=0.55\linewidth]{figures/exploration_before_after_images.png}
\caption{
\textbf{Random policy and exploration policy}\\
Observations and actions taken by an agent during the unsupervised exploratory phase.
Actions are shown with small green arrows.
\textit{(Top)} Random agent, which rarely moves any object, provides a bad source of data for the transition model.
\textit{(Bottom)} trained exploration policy, which frequently moves objects, provides a good source of data for the transition model.
}\label{fig:supp_exploration_behaviour}
\end{figure}
Our approach is to train the transition model adversarially with an exploration policy that learns to take actions on which the transition model has a high error.
Such difficult-to-predict actions should be those that move objects (given that others would leave the scene unchanged).
In this way the exploration policy and the transition model learn together in a virtuous cycle.
This approach is a form of curiosity-driven exploration, as previously described in both the psychology \citep{gopnik1999} and the reinforcement learning literature \citep{schmidhuber1990line, schmidhuber1990making, pathak2017curiosity}.
We apply this adversarial training of the transition model and exploration policy in an exploration environment with randomly generated objects (see Appendix~\ref{S:supp_training}).
To put this idea into practice, we must learn a distribution over the 4-dimensional action space from which we can sample.
To do this, we take inspiration from distributional RL \citep{distributional_rl, dabney2017quantile, dabney2018quantile} and learn an approximate quantile function that maps uniform samples $\myvec{u}
\in [0, 1]^4$ to non-uniform samples in our action space $\myvec{a} \in [0, 1]^4$.
This can be thought of as a perturbation $D$ of each point in $[0, 1]^4$, parameterized by an MLP $D_{net}$.
We train $D_{net}$ to maximize the predicted error $L_T$ of the transition model's prediction of action $\myvec{a}$, subject to an $L_1$ regularization on the perturbation magnitude. Specifically, given $\myvec{u}\in [0, 1]^4$, the exploration policy is trained as:
\begin{align}
D(\myvec{u}, \myvec{z}) &= \myvec{u} + D_{net}(\myvec{u}, \myvec{z}) = \myvec{a} \nonumber\\
Loss(\myvec{u}, \myvec{z}) &= |D_{net}(\myvec{u}, \myvec{z})| - L_T(\myvec{z}, \myvec{a})\nonumber
\end{align}
Note that this method does not pressure $D$ to exactly be a quantile function.
Instead it is incentivized to learn a discontinuous approximation of the quantile function.
We tried various approaches to parameterizing $D$ (see Appendix~\ref{S:supp_things_tried}) and found this one to work best.
While the vision module, transition model, and exploration policy can in principle all be trained simultaneously, in practice for the sake of simplicity in this current work we pre-trained the vision module on frames generated by taking random actions in the exploration environment.
We then reloaded this vision module with frozen variables while training the transition model and exploration policy in tandem in the active adversarial exploration phase just described.
See Appendices~\ref{S:supp_model}-\ref{S:supp_training} for architecture and training details.
See Figure \ref{fig:scene_action_transition}-C for examples of samples from the trained exploration policy and a visualization of its deformation.
Images of the exploration policy in action can be seen in Figure~\ref{fig:supp_exploration_behaviour} (bottom), and see Appendix~\ref{S:supp_agent_videos} for videos.
\subsection{Task phase}\label{SS:task_phase}
Once the exploration phase is complete, COBRA enters the \textit{task phase}, in which it is trained to solve several tasks independently.
We freeze the trained vision module, transition model, and exploration policy, training only a separate reward predictor independently for each task.
See Section~\ref{S:env_tasks} for task details.
\paragraph{Reward Predictor}
For each task we train a model-based agent that uses the components learned in the exploration phase and only learns a reward predictor.
The reward predictor is trained from a replay buffer of (\texttt{observation}, \texttt{reward}) pairs, see Algorithm~\ref{alg:agent}.
Our agent acts with a simple 1-step search, sampling a batch of actions from the exploration policy, rolling them out for one step with the transition model, evaluating the predicted states with the reward predictor and selecting the best action (with epsilon-greedy).
This is effectively a myopic 1-step Model Predictive Control (MPC) policy.
It is sufficient to act optimally given the dense rewards in our tasks, but could readily be extended to multi-step MPC, other existing planning algorithms as in \cite{hafner2018learning} or Monte-Carlo Tree Search \citep{alphago}.
Also, see Appendix \ref{S:supp_sparse_rewards} for an extension that works with sparse rewards.
\begin{figure}[t!]
\centering
\begin{minipage}{.7\linewidth}
\begin{algorithm}[H]
\caption{Task phase agent training}
\label{alg:agent}
\begin{algorithmic}
\REQUIRE Branching factor $B$, Training factor $N$
\ENSURE Trained reward predictor $R_\theta$
\STATE ${\cal R} = \varnothing$
\COMMENT{initialize replay buffer}
\FOR{$t=1$ {\bfseries to} \texttt{num\_steps}}
\STATE $\myvec{x}_t, r_t = \texttt{environment.obs\_reward()}$
\STATE $\myvec{z}_t = V_{enc}(\myvec{x}_t)$
\COMMENT{scene representation}
\STATE ${\cal R} \leftarrow {\cal R} \cup (\myvec{z}_t, r_t)$
\STATE ${\cal A} = \varnothing$
\COMMENT{actions and predicted rewards}
\FOR{$b=1$ {\bfseries to} $B$}
\STATE $\myvec{u}_b \sim \text{Uniform}([0, 1]^4)$
\STATE $\myvec{a}_b = D(\myvec{z}_t, \myvec{u}_b)$
\COMMENT{exploration policy action}
\STATE $\tilde \myvec{z}_{t+1} = T(\myvec{z}_t, \myvec{a}_b)$
\COMMENT{transition model}
\STATE $\tilde r_b = R_\theta(\tilde \myvec{z}_{t+1})$
\COMMENT{reward predictions}
\STATE ${\cal A} \leftarrow {\cal A} \cup (\myvec{a}_b, r_b)$
\ENDFOR
\STATE $\myvec{a} = \text{epsilon-greedy}(\max_{\tilde r}({\cal A}))$
\COMMENT{best action}
\STATE $\texttt{environment.step}(\myvec{a})$
\FOR{$i=1$ {\bfseries to} $N$}
\STATE $\myvec{z}_i, \myvec{r}_i \sim {\cal R}$
\COMMENT{sample minibatch}
\STATE $\theta \leftarrow \theta - \partial_{\theta} ||R_\theta(\myvec{z}_i) - \myvec{r}_i||^2$
\COMMENT{train reward predictor}
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
\end{minipage}
\end{figure}
\section{Related Software}\label{S:related_software}
Professional game engines (e.g. \href{https://unity.com/}{Unity} and \href{https://www.unrealengine.com/}{Unreal}) and visual reinforcement learning platforms (e.g. \href{https://arxiv.org/abs/1612.03801}{DeepMind Lab} \citep{dm_lab}, \href{http://www.mujoco.org/}{Mujoco} \citep{mujoco}, and \href{http://vizdoom.cs.put.edu.pl/}{VizDoom}) are commonly used in the machine learning field for task implementation.
While MOOG has some limitations compared to these (see above), it does also offer some advantages:
\begin{itemize}
\item
\textbf{Python}.
MOOG tasks are written purely in Python, so users who are most comfortable with Python will find MOOG easy to use.
\item
\textbf{Procedural Generation}.
MOOG facilitates procedural generation, with a \href{https://github.com/jazlab/moog.github.io/tree/master/moog/state_initialization/distributions.py}{library of compositional distributions} to randomize conditions across trials.
\item
\textbf{Online Simulation}.
MOOG supports online model-based RL, with a \href{https://github.com/jazlab/moog.github.io/tree/master/moog/env_wrappers/simulation.py}{ground truth simulator} for tree search.
\item
\textbf{Psychophysics}.
MOOG can be run with MWorks, a psychophysics platform.
\item
\textbf{Speed}.
MOOG is fast on CPU.
While the speed depends on the task and rendering resolution, MOOG typically runs at ~200fps with 512x512 resolution on a CPU, much faster than DeepMind Lab and Mujoco and at least as fast as Unity and Unreal.
\end{itemize}
Python-based physics simulators, such \href{https://pybullet.org/wordpress/}{PyBullet} \citep{pybullet} and \href{http://www.pymunk.org/en/latest/}{Pymunk}, are sometimes used in the psychology literature.
While these offer highly accurate collision simulation, MOOG offers the following advantages:
\begin{itemize}
\item
\textbf{Customization}.
Custom forces and game rules can be easily implemented in MOOG.
\item
\textbf{Psychophysics, Procedural Generation, and Online Simulation},
as described above.
\item
\textbf{RL Interface}.
A task implemented in MOOG can be used out-of-the-box to train RL agents, since MOOG is Python-based and has DeepMind dm\_env and OpenAI Gym interfaces.
\end{itemize}
Psychology and neurophysiology researchers often use platforms such as
\href{https://www.psychopy.org/}{PsychoPy} \citep{psychopy},
\href{http://psychtoolbox.org/}{PsychToolbox} \citep{psychtoolbox}, and
\href{https://mworks.github.io/}{MWorks}.
These allow precise timing control and coordination with eye trackers and other controllers.
MOOG can interface with MWorks to leverage all of those features and offers the following additional benefits:
\begin{itemize}
\item
\textbf{Flexibility}.
MOOG offers a large scope of interactive tasks.
Existing psychophysics platforms are not easily customized for game-like tasks, action interfaces, and arbitrary object shapes.
\item
\textbf{Physics}.
Existing psychophysics platforms do not have built-in physics, such as forces, collisions, etc.
\item
\textbf{RL Interface},
as described above.
\end{itemize}
\section{Summary}\label{S:env_tasks}
The \href{https://jazlab.github.io/moog.github.io/index.html}{Modular Object-Oriented Games} library is a general-purpose Python-based platform for interactive games.
It aims to satisfy the following criteria:
\begin{itemize}
\item
Usable for reinforcement learning, psychology, and neurophysiology.
MOOG supports \href{https://github.com/deepmind/dm_env}{DeepMind dm\_env} and \href{https://gym.openai.com/}{OpenAI Gym} \citep{openai_gym} interfaces for RL agents and an \href{https://mworks.github.io/}{MWorks} interface for psychology and neurophysiology.
\item
Highly customizable.
Environment physics, reward structure, agent interface, and more can be customized.
\item
Easy to rapidly prototype tasks.
Tasks can be composed in a single short file.
\item
Light-weight and efficient.
Most tasks run quickly, almost always faster than 100 frames per second on CPU and often much faster than that.
\item
Facilitates procedural generation for randomizing task conditions every trial.
\end{itemize}
|
\section{Introduction}
Futures are standardized exchange-traded bilateral \mbox{contracts} of agreement to buy or sell an asset at a pre-determined price at a pre-specified time in the future. The underlying asset can be a physical commodity, like gold and silver, or oil and gas, but it can also be a market index like the S\&P 500 index or the CBOE volatility index.
Futures are an integral part of the derivatives market. The Chicago Mercantile Exchange (CME), which is the world's largest futures exchange, averages over 15 million futures contracts traded per day.\footnote[3]{Source: CME Group daily exchange volume and open interest report, available at \url{https://www.cmegroup.com/market-data/volume-open-interest/exchange-volume.html}} Within the universe of hedge funds and alternative investments, futures funds constitute a major part with hundreds of billions under management. This motivates us to investigate the problem of trading futures portfolio dynamically over time.
In this paper, we introduce a multiscale central tendency Ornstein-Uhlenbeck (MCTOU) model to describe the price dynamics of the underlying asset. This is a three-factor model that is driven by a fast mean-reverting OU process and a slow mean-reverting OU process. Similar multiscale framework has been widely used for modeling stochastic volatility of stock prices \cite{Fouque2000}. The flexibility of multifactor models permits good fit to empirical term structure as displayed in the market. Especially in deep and liquid futures markets, such as crude oil or gold, with over ten contracts of various maturities actively traded at any given time, multifactor models are particular useful. In the literature, we refer to \cite{nfactor} for a multifactor Gaussian model for pricing oil futures, and \cite{CORTAZAR2017182} for a multifactor stochastic volatility model for commodity prices to enhance calibration against observed option prices.
Under the MCTOU model, we first derive the no-arbitrage price formulae for the futures contracts. In turn, we solve a utility maximization problem to derive the optimal trading strategies over a finite trading horizon. This stochastic control approach leads to the analysis of the associated Hamilton-Jacobi-Bellman (HJB) partial differential equation satisfied by the investor's value function. We derive both the investor's value function and optimal strategy explicitly.
Our solution also yields the formula for the investor's certainty equivalent, which quantifies the value of the futures trading opportunity to the investor. Surprisingly the value function, optimal strategy and certainty equivalent depend not on the current spot and futures prices, but on the associated risk premia. In addition, we provide the numerical examples to illustrate the investor's optimal futures positions and optimal wealth over time.
In the literature, stochastic control approach has been widely applied to continuous-time dynamic optimization of stock portfolios dating back to \cite{Merton}, but much less has been done for portfolios of futures and other derivatives. For futures portfolios, one must account for the risk-neutral pricing before solving for the optimal trading strategies. To that end, our model falls within the multi-factor Gassian model for futures pricing, as used for oil futures in \cite{nfactor}. The utility maximization approach is used to derive dynamic futures trading strategies under two-factor models in \cite{firstpaper} and \cite{secondpaper}. A general regime-switching framework for dynamic futures trading can be found in \cite{Leung2019}. As an alternative approach for capturing futures and spot price dynamics, the stochastic basis model \cite{BahmanLeung, Angoshtari2019b} directly models the difference between the futures and underlying asset prices, and solve for the optimal trading strategies through utility maximization.
In comparison to these studies, we have extended the investigation of optimal trading in commodity futures market under two-factor models to a three-factor model. Closed-form expressions for the optimal controls and for the value function are obtained. Using these formulae, we illustrate the optimal strategies. Intuitively, it should be more beneficial to be able to access a larger set of securities, and this intuition is confirmed quantitatively. Therefore, we consider all available contracts of different maturities in that market. From our numerical example, the highest certainty equivalent is achieved from trading every contract that is available.
There are a few alternative approaches and applications of dynamic futures portfolios. In \cite{LeungLiLiZheng2015} and \cite{meanreversionbook2016}, an optimal stopping approach for futures trading is studied. In practice, dynamic futures portfolios are also commonly used to track a commodity index \cite{LeungWard}.
\section{The Multiscale Central Tendency Ornstein-Uhlenbeck Model}\label{sect-Multiscale-CTOU}
We now present the multiscale central tendency Ornstein-Uhlenbeck (MCTOU) model that describes the price dynamics of the underlying asset. This leads to the no-arbitrage pricing of the associated futures contracts. Hence, the dynamcis under both the physical measure $\mathbb{P}$ and risk-neutral pricing measure $\mathbb{Q}$ are discussed.
\subsection{Model Formulation}
We fix a probability space $(\Omega, \mathcal{F},\mathbb{P})$. The log-price of the underlying asset $S_t$ is denoted by $X_t^{(1)}$. Its evolution under the physical measure $\P$ is given by the system of stochastic differential equations:
\begin{align}
dX^{(1)}_t &= \kappa\left(X^{(2)}_t+X^{(3)}_t-X^{(1)}_t\right)dt+\sigma_1 dZ^{\P,1}_t,\label{model-multiscale-begin}\\
dX^{(2)}_t &= \frac{1}{\epsilon}\left(\alpha_2-X^{(2)}_t\right)dt\notag\\
&+\frac{1}{\sqrt \epsilon}\sigma_2 \left(\rho_{12} dZ^{\P,1}_t+\sqrt{1-\rho_{12}^2} dZ^{\P,2}_t\right),\label{model-multiscale-fast}\\
dX^{(3)}_t &= \delta\left(\alpha_3-X^{(3)}_t\right)dt+\sqrt{\delta}\sigma_3\bigg(\rho_{13} dZ^{\P,1}_t\notag\\
&+\rho_{23}dZ^{\P,2}_t+\sqrt{1-\rho_{13}^2-\rho_{23}^2}dZ^{\P,3}_t\bigg),\label{model-multiscale-end}
\end{align}
where $Z^{\P,1}$ , $Z^{\P,2}$ and $Z^{\P,3}$ are independent Brownian motions under the physical measure $\P$.
Under this model, the mean process of log-price $X^{(1)}$ is the sum of two stochastic factors, $X_t^{(2)}$ and $X_t^{(3)}$, modeled by two different OU processes. The first factor $X_t^{(2)}$ is
fast mean-reverting. The rate of mean reversion is represented by $1/\epsilon$, with $\epsilon > 0$ being a small parameter corresponding to the time scale of this process. $X_t^{(2)}$ is an ergodic process
and its invariant distribution is independent of $\epsilon$. This distribution is Gaussian with mean $\alpha_2$ and variance $\sigma_2^2/2$. In contrast, the second factor $X_t^{(3)}$ is a slowly mean-reverting OU process. The rate of mean reversion is represented by a small parameter $\delta>0$.
The three processes $X^{(1)}_t$, $X_t^{(2)}$, and $X_t^{(3)}$ can be correlated. The correlation coefficients $\rho_{12}$, $\rho_{13}$, and $\rho_{23}$ are constants, which satisfy $|\rho_{12}|<1$ and $\rho_{13}^2+\rho_{23}^2<1$.
We specify the market prices of risk as $\zeta_i$, for $i=1,2,3$, which satisfy
\begin{equation}
dZ^{\Q,i}_t=dZ_t^{\P,i}+\zeta_i\, dt,
\end{equation}
where $Z^{\Q,1}$, $Z^{\Q,2}$ and $Z^{\Q,3}$ are independent Brownian motions under risk-neutral pricing measure $\Q$. We introduce the combined market prices of risk $\lambda_i$, for $i=1,2,3$, defined by
\begin{align}
\lambda_1 &=\zeta_1 \sigma_1 ,\\
\lambda_2 &= \frac{1}{\sqrt \epsilon}\sigma_2 \left(\zeta_1\rho_{12} +\zeta_2\sqrt{1-\rho_{12}^2} \right),\\
\lambda_3 &= \sqrt{\delta}\sigma_3\left(\zeta_1\rho_{13} +\zeta_2\rho_{23}+\zeta_3\sqrt{1-\rho_{13}^2-\rho_{23}^2}\right).
\end{align}
Then, we write the evolution under the risk-neutral measure $\Q$ as:
\begin{align}
dX^{(1)}_t &= \kappa\left(X^{(2)}_t+X^{(3)}_t-X^{(1)}_t-\lambda_1/\kappa\right)dt\notag\\
&+\sigma_1 dZ^{\Q,1}_t,
\end{align}
\begin{align}
dX^{(2)}_t &= \frac{1}{\epsilon}\left(\alpha_2-X^{(2)}_t-\epsilon\lambda_2\right)dt\notag\\
&+\frac{1}{\sqrt \epsilon}\sigma_2 \left(\rho_{12} dZ^{\Q,1}_t+\sqrt{1-\rho_{12}^2} dZ^{\Q,2}_t\right),\\
dX^{(3)}_t &= \delta\left(\alpha_3-X^{(3)}_t-\lambda_3/\delta\right)dt+\sqrt{\delta}\sigma_3\bigg(\rho_{13} dZ^{\Q,1}_t\notag\\
&+\rho_{23}dZ^{\Q,2}_t+\sqrt{1-\rho_{13}^2-\rho_{23}^2}dZ^{\Q,3}_t\bigg).
\end{align}
For convenience, we define
\begin{align}
\bm{X}_t &= (X^{(1)}_t,X^{(2)}_t,X^{(3)}_t)',\\
\bm{Z}_t^\P &= (Z^{(\P,1)}_t,Z^{(\P,2)}_t,Z^{(\P,3)}_t)',\\
\bm{Z}_t^\Q &= (Z^{(\Q,1)}_t,Z^{(\Q,2)}_t,Z^{(\Q,3)}_t)',\\
\bm{\mu} &= (0,\alpha_2/\epsilon,\delta\alpha_3)',\\
\bm{\lambda} &= (\lambda_1,\lambda_2,\lambda_3)',\\
\bm{K} &= \bcm \kappa & -\kappa &-\kappa\\0 &1/\epsilon &0\\0&0&\delta
\ecm,\\
\bm{\Sigma}&=\bcm \sigma_1&0&0\\0&\sigma_2/\sqrt{\epsilon}&0\\0&0&\sqrt{\delta}\sigma_3\ecm,
\end{align}
and
\begin{equation}
\bm{C} = \bcm 1 &0&0\\\rho_{12}&\sqrt{1-\rho_{12}^2}&0\\\rho_{13}&\rho_{23}&\sqrt{1-\rho_{13}^2-\rho_{23}^2}\ecm.
\end{equation}
Then, the evolution for $\bm{X}_t$ under measures $\mathbb{P}$ and $\mathbb{Q}$ can be written concisely as
\begin{equation}
d\bm{X}_t = (\bm{\mu}- \bm{K}\bm{X}_t) dt + \bm{\Sigma C}d\bm{Z}_t^{\P},\label{N_Factor_SDE_P}
\end{equation}
and
\begin{equation}
d\bm{X}_t=\bigg(\bm{\mu}-\bm{\lambda}-\bm{K} \bm{X}_t\bigg) dt + \bm{\Sigma C}d\bm{Z}_t^{\Q}.\label{N_Factor_SDE_Q}
\end{equation}
\begin{remark}
If the stochastic mean of log price $X^{(1)}$ is only modeled by $X^{(2)}$ or $X^{(3)}$, instead of their sum, it will reduce to the CTOU model, which is used in \cite{MenciaSentana} for pricing VIX
futures. Under this model, the futures portfolio optimization problem has been studied in \cite{firstpaper}.
\end{remark}
\section{Futures Pricing and Futures Trading}\label{sect-formula-multiscale}
\subsection{Futures Pricing}
Let us consider three futures contracts $F^{(1)}$, $F^{(2)}$ and $F^{(3)}$, written on the same underlying asset $S$ with three arbitrarily chosen maturities $T_1$, $T_2$ and $T_3$ respectively. Recall that the asset price is given by \[S_t = \exp(X_t^{(1)}), \quad t\ge 0.\] Then, the futures price at time $t\in[0,T]$ is given by
\begin{equation}
F^{(k)}(t,\bm{x}):= \E^\Q\big[\exp(X^{(1)}_{T_k})\,|\,\bm{X}_t=\bm{x}\,\big],\label{Futures_Definition}
\end{equation}
for $k=1,2,3$.
Define the linear differential operator
\begin{equation}
\mathcal{L}^\Q\cdot = \bigg(\bm{\mu}-\bm{\lambda}-\bm{K} \bm{x}\bigg)'\nabla_{\bm{x}}\cdot+\frac{1}{2}\Tr \bigg(\bm{\Sigma\Omega \Sigma} \nabla_{\bm{xx}}\cdot \bigg), \label{L_Q}
\end{equation}
where $\nabla_{\bm{x}}\cdot = ( \partial_{x_1}\cdot,\cdots ,
\partial_{x_N}\cdot)'$ is the nabla operator and Hessian operator $\nabla_{\bm{xx}}\cdot$ satisfies
\begin{equation}
\nabla_{\bm{xx}} \cdot=
\begin{bmatrix}
\partial_{x_1}^2\cdot & \partial_{x_1x_2}\cdot & \dots & \partial_{x_1x_N}\cdot \\
\partial_{x_1x_2}\cdot & \partial_{x_2}^2\cdot & \dots & \partial_{x_2x_N}\cdot \\
\vdots & \vdots & \ddots & \vdots \\
\partial_{x_1x_N}\cdot & \partial_{x_2x_N}\cdot & \dots & \partial_{x_N}^2\cdot
\end{bmatrix}.
\end{equation}
Then, for $k=1,2,3$, the futures price function $F^{(k)}(t,\bm{x})$ solves the following PDE
\begin{equation}
(\partial_t + \mathcal{L}^\Q) F^{(k)}(t,\bm{x}) = 0,\label{Futures_PDE}
\end{equation}
for $(t,\bm{x})\in[0,T)\times \R^N$, with the terminal condition $F^{(k)}(T,\bm{x})=\exp(\bm{e_1}'\bm{x})$ for $\bm{x}\in \R^N$, where $\bm{e_1}=(1, 0 , 0)'$.
\begin{proposition}\label{prop-futures-price}
The futures price $F^{(k)}(t,\bm{x})$ is given by
\begin{equation}\label{Futures_Price_MultiScale}
F^{(k)}(t,\bm{x}) = \exp{\bigg(\bm{a}^{(k)}(t)'\bm{x}+\beta^{(k)}(t)\bigg)},
\end{equation}
where $\bm{a}^{(k)}(t)$ and $\beta^{(k)}(t)$ satisfy
\begin{align}\label{a(t)}
\bm{a}^{(k)}(t)&=\exp{\bigg(-(T-t)\bm{K}'\bigg)}\bm{e_1}\\
\beta^{(k)}(t) &= \int_t^T (\bm{\mu}-\bm{\lambda})'\bm{a}^{(k)}(s)\notag\\
&+\frac{1}{2}\Tr\bigg(\bm{\Sigma\Omega\Sigma}\bm{a}^{(k)}(s)\bm{a}^{(k)}(s)'\bigg)ds.\label{beta(t)}
\end{align}
\end{proposition}
\begin{proof}
We substitute the ansatz solution \eqref{Futures_Price_MultiScale} into PDE \eqref{Futures_PDE}. The $t$-derivative is given by
\begin{align}\label{F_t}
\partial_t F(t,\bm{x}) &= \left( \left(\frac{d\bm{a}^{(k)}(t)}{dt}\right)'\bm{x}+\frac{d\beta^{(k)}(t)}{dt}\right)\notag\\
&\times\exp{\bigg(\bm{a}^{(k)}(t)'\bm{x}+\beta^{(k)}(t)\bigg)}.
\end{align} Then, the first and second derivatives satisfy
\begin{align}\label{F_x}
\nabla_{\bm{x}} F(t,\bm{x}) &= \bm{a}^{(k)}(t)F(t,\bm{x}),\\
\label{F_xx}
\nabla_{\bm{xx}} F(t,\bm{x}) &= \bm{a}^{(k)}(t)\bm{a}^{(k)}(t)'F(t,\bm{x}).
\end{align}
By substituting \eqref{F_t}, \eqref{F_x}, and \eqref{F_xx} into PDE (\ref{Futures_PDE}), we obtain
\begin{equation}
\frac{d\bm{a}^{(k)}(t)}{dt} -\bm{K}'\bm{a}^{(k)}(t) = 0,\label{a_ODE}
\end{equation}
and
\begin{align}
\frac{d\beta^{(k)}(t)}{dt}+(\bm{\mu}-\bm{\lambda})'\bm{a}^{(k)}(t)\notag\\
+\frac{1}{2}\Tr\bigg(\bm{\Sigma\Omega\Sigma}\bm{a}^{(k)}(t)\bm{a}^{(k)}(t)'\bigg)=0.\label{beta_ODE}
\end{align}
The terminal conditions of $\bm{a}^{(k)}(t)$ and $\beta^{(k)}(t)$ are given by
\begin{equation}
\bm{a}^{(k)}(T)=\bm{e_1},\quad \beta^{(k)}(T) = 0.\label{a_beta_terminal_conditions}
\end{equation}
By direct substitution, the solutions to ODEs (\ref{a_ODE}) and (\ref{beta_ODE}) are given by \eqref{a(t)} and \eqref{beta(t)}.
\end{proof}
\subsection{Dynamic Futures Portfolio}
Now consider a collection of $M$ contracts of different maturities available to trade, where $M=1,$ 2, 3. We note that there are only three sources of randomness, so trading three contracts is sufficient. Any additional contract would be redundant in this model. By Proposition \ref{prop-futures-price}, we have
\begin{align}
\frac{dF_t^{(k)}}{F_t^{(k)}}&=\bm{a}^{(k)}(t)'\bm{\lambda}dt+\bm{a}^{(k)}(t)'\bm{\Sigma C} d\bm{Z}_t^{\P}\\
&\equiv \mu^{(k)}_F(t) dt + \bm{\sigma}^{(k)}_F(t)' d\bm{Z}_t^{\P},\label{mu_F^k}
\end{align}
where we have defined
\begin{equation}
\mu^{(k)}_F(t) \equiv \bm{a}^{(k)}(t)'\bm{\lambda},\quad\bm{\sigma}^{(k)}_F(t) \equiv \bm{C}'\bm{\Sigma}'\bm{a}^{(k)}(t).
\label{mu_F_Sigma_F_Definition_m}
\end{equation}
Define
\begin{equation}
d\bm{F}_t = \left( \frac{dF^{(1)}_t}{F^{(1)}_t}, \cdots, \frac{dF^{(M)}_t}{F^{(M)}_t} \right)'.
\end{equation} Then, in matrix form, the system of futures dynamics is given by the set of SDE:
\begin{equation}
d\bm{F}_t = \bm{\mu_F}(t) dt+\bm{\Sigma_F}(t)d\bm{Z}_t^\P ,
\end{equation}
where
\begin{align}\label{mu_F_Sigma_F_matrix_Definition_m}
\bm{\mu_F}(t) &= \left( \mu_F^{(1)}(t), \cdots , \mu_F^{(M)}(t) \right) ' \\
\bm{\Sigma_F}(t) &= \left(\bm{\sigma}^{(1)}_F(t),\cdots, \bm{\sigma}^{(M)}_F(t) \right)'.
\end{align}
Here, we assume there be no redundant futures contract, which means any futures contract could not be replicated by other $M-1$ futures contracts, indicating that $\rank\left(\bm{\Sigma_F}\right)=M$.
Next, we consider the trading problem for the investor. Let strategy $\bm{\pi}_t=\left(\pi^{(1)}_t,\cdots,\pi^{(M)}_t\right)'$,
where the element $\pi^{(k)}_t$
denotes the amount of money invested in $k$-th futures contract.
In addition, we assume the interest rate be zero for simplicity. Then, for any admissible strategy $\pi$, the wealth process is
\begin{align}
dW_t^\pi &= \sum_{k=1}^M \pi^{(k)}_t \frac{dF^{(k)}_t}{F^{(k)}_t}\notag\\
&= \bm{\pi}_t' \bm{\mu_F}(t) dt+\bm{\pi}_t'\bm{\Sigma_F}(t) d\bm{Z}_t^\P. \label{Wealth_SDE_m}
\end{align}
We note that the wealth process is only determined by the strategy $\bm{\pi}_t$ and it is not affected by factors variable $\bm{X}$ and futures prices $\bm{F}$.
The investor's risk preference is described by the exponential utility:
\begin{equation}
U(w)= -\exp(-\gamma w),\label{exponential_utility}
\end{equation}
where $\gamma>0$ denotes the coefficient of risk aversion. A strategy $\bm{\pi}$ is said to be admissible if $\bm{\pi}$ is real-valued
progressively measurable and satisfies the Novikov condition \cite{Novikov1972}:
\begin{equation}\label{integrability_condition}
\E^{\P}\bigg[\exp\left(\int_t^{\tilde T}\frac{\gamma^2}{2} \bm{\pi}_s'\bm{\Sigma_F}(s) \bm{\Sigma_F}'(s)\bm{\pi}_s ds \right)\bigg]< \infty.
\end{equation}
The investor fixes a finite optimization horizon $0< \tilde T \le T_1$, which means that $\tilde T$ has to be less than or equal to the maturity of the earliest expiring contract,
and seeks an admissible strategy $\bm{\pi}$ that maximizes the expected utility of wealth at $\tilde T$:
\begin{equation}
u(t,w)=\sup_{\bm{\pi}\in \mathcal{A}_t} \E[U(W_{\tilde T}^{\bm{\pi}})|W_t=w],
\label{Value_Function_m}
\end{equation}
where $\mathcal{A}_t$ denotes the set of admissible controls at the initial time $t$. Since the wealth SDE (\ref{Wealth_SDE_m}) does not depend on the factors variable $\bm{X}$ and futures prices $\bm{F}$, the value function does not depend on them either.
To facilitate presentation, we define
\begin{equation}
\mathcal{L}^{\bm{\pi}}\cdot = \bm{\pi}_t'\bm{\mu_F}(t) \partial_w\cdot+\frac{1}{2} \bm{\pi}_t'\bm{\Sigma_F}(t) \bm{\Sigma_F}'(t)\bm{\pi}_t\partial_{ww}\cdot.
\end{equation}
Then, following the standard verification approach to dynamic programming \cite{FlemingSoner93}, the candidate value function $u(t,w)$ and optimal trading strategy $\pi^*$ is found from the Hamilton-Jacobi-Bellman (HJB) equation
\begin{equation}
\partial_t u + \sup_{\bm{\pi}}
\mathcal{L}^{\bm{\pi}} u
= 0, \label{Value_Function_HJB_m}
\end{equation}
for $(t,w)\in[0,\tilde T)\times \mathbb{R}$, along with the terminal condition $u(T,w)=-e^{-\gamma w}$, for $w\in \mathbb{R}.$
\begin{theorem}\label{thm2} The unique solution to the HJB equation \eqref{Value_Function_HJB_m} is given by
\begin{equation}
u(t,w)=-\exp\left(
-\gamma w -\frac{1}{2}\int_t^{\tilde T}\Lambda^2(s)ds\right),
\label{u_Candidate_Solution_m}
\end{equation}
where
\begin{equation}\label{Lambda}
\Lambda^2(t)=\bm{\mu_F}(t)' \left(\bm{\Sigma_F}(t)\bm{\Sigma_F}(t)'\right)^{-1} \bm{\mu_F}(t)\,.
\end{equation}
The optimal futures trading strategy is explicitly given by
\begin{equation}
\bm{\pi}^*(t)=\frac{1}{\gamma}
\left( \bm{\Sigma_F}(t)\bm{\Sigma_F}'(t) \right)^{-1}\bm{\mu_F}(t)
.\label{Optimal_Strategy_m}
\end{equation}
\end{theorem}
\begin{proof}
We will first use the ansatz
\begin{equation}
u(t,w)= -e^{-\gamma w} h(t).
\end{equation}
Then, using the relations
\begin{equation}
\partial_t u = -e^{-\gamma w} \partial_t h(t),\quad \partial_w u=\gamma e^{-\gamma w} h(t),
\end{equation}
and
\begin{equation}
\partial_{ww} u=-\gamma^2 e^{-\gamma w} h(t),
\end{equation}
the PDE \eqref{Value_Function_HJB_m} becomes
\begin{align}
-\frac{d}{dt} h(t) + &\sup_{\bm{\pi}_t} \bigg[\gamma
\bm{\pi'}_t\bm{\mu_F} (t) h
\notag\\
&-\frac{1}{2} \gamma^2 \bm{\pi}_t'\bm{\Sigma_F}(t)\bm{\Sigma_F}'(t)\bm{\pi}_t h \bigg]
= 0, \label{h_HJB_m}
\end{align}
with terminal condition $h(\tilde T)=1$.
From the first-order condition, which is obtained from differentiating the terms inside the supremum with respect to $\bm{\pi}_t$ and setting the equation to zero, we have
\begin{equation}
\gamma \bm{\mu_F}(t)
-\gamma^2 \bm{\Sigma_F}(t)\bm{\Sigma_F}'(t) \bm{\pi}_t=0.
\end{equation}
Recall that $\rank(\bm{\Sigma_F}(t))=M$. Then, $\bm{\Sigma_F}(t)\bm{\Sigma_F}'(t)$ is an $M\times M$ invertible matrix. Accordingly, we have the optimal strategy \eqref{Optimal_Strategy_m}.
Given the fact that $A'A$ is the semi-positive definite matrix for any matrix $A$, the time-dependent component $\Lambda^2(t)=\bm{\mu_F}(t)' (\bm{\Sigma_F}(t)\bm{\Sigma_F}(t)')^{-1} \bm{\mu_F}(t)$ is always non-negative.
Substituting $\bm{\pi}^*$ back, the equation \eqref{h_HJB_m} becomes
\begin{align}
-\frac{d}{dt} h(t) &+
\frac{1}{2}\Lambda^2(t) h(t)=0.
\label{h_HJB_Reduced_m}
\end{align}
Accordingly, we have
\begin{equation}
h(t)=\exp{\bigg(-\frac{1}{2}\int_t^{\tilde T}\Lambda^2(s)ds\bigg)}.
\end{equation}
\end{proof}
\begin{example}
If there is only one futures contract $F^{(1)}$ available in the market. Then by \eqref{mu_F_Sigma_F_matrix_Definition_m}, we have
\begin{equation}\label{3factor-single-futures-mu}
\bm{\mu_F} = \mu_F^{(1)},\quad \bm{\Sigma_F} = \bm{\sigma}_F^{(1)}(t)'.
\end{equation}
Then, the optimal strategy \eqref{Optimal_Strategy_m} becomes
\begin{align}\label{Optimal_Strategy_MultiScale_Port1}
\pi^{(1)*}(t) &= \frac{1}{\gamma}\frac{\bm{\mu_F}(t)}{\bm{\Sigma_F}(t)\bm{\Sigma_F}'(t)} = \frac{1}{\gamma}\frac{\mu_F^{(1)}}{\bm{\sigma}_F^{(1)}(t)'\bm{\sigma}_F^{(1)}(t)},
\end{align}
where $\mu_F^{(1)}$ and $\bm{\sigma}_F^{(1)}(t)$ are shown as \eqref{mu_F^k}.
\end{example}
In order to quantify the value of trading futures to the investor, we define the investor's certainty equivalent associated with the utility maximization problem. The certainty equivalent is the guaranteed cash amount that would yield the same utility as that from dynamically trading futures according to \eqref{Value_Function_m}. This amounts to applying the inverse of the utility function to the value function in \eqref{u_Candidate_Solution_m}. Precisely, we define
\begin{align}
\mathcal{C}(t,w)&:=U^{-1}(u(t,w))\\
&=w+\frac{1}{2\gamma}\int_t^{\tilde T}\Lambda^2(s)ds.\label{nfactor:certequiv}
\end{align}
Therefore, the certainty equivalent is the sum of the investor's wealth $w$ and a non-negative time-dependent component $\frac{1}{2\gamma}\int_t^{\tilde T}\Lambda^2(s)ds$. The certainty equivalent is also inversely proportional to the risk aversion parameter $\gamma$, which means that a more risk averse investor has a lower certainty equivalent, valuing the futures trading opportunity less. From \eqref{a(t)}, \eqref{mu_F_Sigma_F_Definition_m}, \eqref{mu_F_Sigma_F_matrix_Definition_m} and \eqref{Lambda}, we see that the certainty equivalent depends on the constant matrix $\bm{K}$, volatility matrix $\bm{\Sigma}$, correlation matrix $\bm{C}$ and market prices of risk $\bm{\lambda}$. Nevertheless, the certainty equivalent does not depend on the current values of factors $\bm{X}_t$.
\begin{table}
\vspace*{.2cm}
\centering
\begin{scriptsize}
\begin{tabular}{c|c|c|c|c}
\hline
$X^{(1)}_0$ & $X^{(2)}_0$& $X^{(3)}_0$& $\alpha_2$& $\alpha_3$ \\
\hline
1& 0.5 &0.5 &0.5 & 0.5\\
\hline
\hline
$\epsilon$& $\delta$ & $\sigma_1$& $\sigma_2$&$\sigma_3$ \\
\hline
0.05& 0.01 & 0.8 & 0.02 & 0.3\\
\hline
\end{tabular}
\vspace*{.1 cm}
\begin{tabular}{c|c|c|c|c|c}
\hline
$\rho_{12}$&$\rho_{13}$ & $\rho_{23}$ & $\lambda_1$ & $\lambda_2$ & $\lambda_3$\\
\hline
0& 0& 0 & 0.02 & 0.02 & 0.02 \\
\hline
\hline
$T_1$&$T_2$&$T_3$ &$\tilde{T}$ & $\gamma$& $\kappa$\\
\hline
$1/12$&$2/12$&$3/12$& $1/12$ & $1$&5\\
\hline
\end{tabular}
\end{scriptsize}
\caption{Parameters for the MCTOU model.}
\label{Parameters_Table}
\end{table}
\begin{figure}
\centering
{\includegraphics[trim=0.4cm 1.5cm 1.5cm 2.5cm,clip, width=0.50\textwidth]{Prices1}}
\caption{
Top: simulation paths for asset's log-price $X^{(1)}$. Dashed curves represent $95\%$ confidence interval. Middle: simulation path for fast varying factor $X^{(2)}$ and slow-varying factor $X^{(3)}$. Dashed and dotted curves represent $95\%$ confidence interval for $X^{(2)}$ and $X^{(3)}$, respectively. Bottom: Sample price paths for the underlying asset and associated futures.}
\label{Prices}
\end{figure}
\section{Numerical Illustration}\label{sect-numeric}
In this section, we simulate the MCTOU process and illustrate the outputs from our trading model. With the closed-form expressions obtained in the Section \ref{sect-formula-multiscale}, we now generate the futures prices, optimal strategies and wealth processes numerically, using the parameters in Table \ref{Parameters_Table}. Primarily, we let $\epsilon$ and $\delta$ be small parameters and we consider trading three futures with maturities $T_1=1/12$ year, $T_2=2/12$ year and $T_3=3/12$ year. Then, our trading horizon will be $\tilde{T}=1/12$ year, no greater than the futures maturities. We assume 252 trading days in a year and 21 trading days in a month (or 1/12 year). In our figures, we show the corresponding trading days on the $x$ axis.
\begin{figure}
\centering
{\includegraphics[trim=0.4cm 1.8cm 1.5cm 2.2cm,clip, width=0.48\textwidth]{Optimal_Investment_MultiScale000}
}
\caption{
Optimal strategies $\pi^{(*)}$ for different futures combinations. Solid, dashed and dotted lines represent the optimal position (in \$) on $T_1$-futures, $T_2$-futures and $T_3$-futures, respectively.}
\label{Optimal_Investment_MultiScale}
\end{figure}
\begin{table*}
\vspace*{.2cm}
\centering\begin{scriptsize}
\begin{tabular}{cc|ccccccc}
\hline
\multicolumn{2}{c|}{\multirow{2}{*}{Parameters}} &\multicolumn{7}{c}{Futures Combinations (Maturity)}\\
\cline{3-9}
& & $T_1$ & $T_2$ & $T_3$ & $\{T_1, T_2\}$ & $\{T_1, T_3\}$ & $\{T_2, T_3\}$ & $\{T_1, T_2,T_3\}$ \\
\hline
\hline
\multirow{3}{*}{$\rho_{12}= 0 $}
& $\rho_{13}= -0.5 $ &
0.563 & 1.58 & 3.25 & 5.36 & 4.65 & 4.41 & 419 \\
& $\rho_{13}= 0 $ &
0.502 & 1.09 & 1.74 & 3.09 & 2.62 & 2.41 & 417 \\
& $\rho_{13}= 0.5 $ &
0.456 & 0.837 & 1.19 & 2.88 & 2.35 & 2.01 & 417 \\
\hline
\multirow{3}{*}{$\rho_{12}= 0.5 $}
& $\rho_{13}= -0.5 $ &
0.561 & 1.56 & 3.23 & 5.34 & 4.64 & 4.40 & 543 \\
& $\rho_{13}= 0 $ &
0.500 & 1.08 & 1.73 & 3.08 & 2.62 & 2.40 & 542 \\
& $\rho_{13}= 0.5 $ &
0.454 & 0.833 & 1.18 & 2.87 & 2.34 & 2.01 & 541 \\
\hline
\multirow{3}{*}{$\rho_{12}= -0.5 $}
& $\rho_{13}= -0.5 $ &
0.565 & 1.59 & 3.27 & 5.39 & 4.66 & 4.42 & 571 \\
& $\rho_{13}= 0 $ &
0.504 & 1.10 & 1.75 & 3.11 & 2.63 & 2.41 & 569 \\
& $\rho_{13}= 0.5 $ &
0.457 & 0.842 & 1.20 & 2.90 & 2.36 & 2.02 & 569 \\
\hline
\end{tabular}\end{scriptsize}
\caption{Certainty equivalents ($\times 10^{-4}$) for all possible futures combinations under different correlations.\label{CE_Table}}
\end{table*}
In Figure \ref{Prices}, we plot the simulation paths and $95\%$ confidence intervals for three factors in the top figure and middle figure. As shown in the middle panel, the $95\%$ confidence interval of the slow-varying factor $X^{(3)}$ is much narrower than the one for fast-varying factor $X^{(2)}$. At the bottom, we plot the spot price and futures prices. The three paths for the futures prices are highly correlated and $T_1$-futures price is equal to the asset's spot price at its maturity date $T_1$, which is the 21st trading day.
In Figure \ref{Optimal_Investment_MultiScale}, we plot the optimal strategies as functions of time for different portfolios and different correlation parameters. In each sub-figure, from top to bottom, we show the optimal strategies for one-contract portfolio, two-contract portfolio and three-contract portfolio respectively. The optimal investments on $T_1$-futures, dashed lines represent the optimal investment on $T_1$-futures, $T_2$-futures, and $T_3$-futures are represented by solid, dashed, and dotted lines respectively.
The optimal cash amount invested are deterministic functions for time, but the optimal units of futures held do vary continuously with the prevailing futures price.
Moreover, the investor takes large long/short positions in three-contract portfolio since all sources of risk can be hedged. We provide sample path for wealth process for three-contract portfolios in the Figure \ref{Wealth}.
In Figure \ref{nplotmn}, we see that the certainty equivalent increases as a function of trading horizon $\tilde T$, which means that the more time the investor has, the more valuable is the trading opportunity. As the trading horizon reduces to zero, the certainty equivalent converges to the initial wealth $w$, which is set to be 0 in this example, as expected from \eqref{nfactor:certequiv}. Also, with a lower risk aversion parameter $\gamma$, the investor has a higher certainty equivalent for any given trading horizon.
Table \ref{CE_Table} shows the certainty equivalents for all possible futures combinations under various correlation configurations. The certainty equivalent is much higher when more contracts are traded. In addition, if there is only one futures contract to trade, the certainty equivalent is increasing with respect to its maturity, see first three columns. The certainly equivalents tend to be higher when $\rho_{12}$ and $\rho_{13}$ are negative.
\begin{figure}[h]
\centering
{\includegraphics[trim=0.3cm 0cm 1.5cm 1.2cm,clip, width=0.45\textwidth]{Wealth2}}
\caption{
Sample path for wealth process for the three-futures portfolio. }
\label{Wealth}
\end{figure}
\begin{figure}[h]
\centering
{\includegraphics[trim=1cm 0.7cm 1.8cm 2cm,clip, width=0.45\textwidth]{nplotmn2}}
\caption{
Certainty equivalents for the three-futures portfolio as the trading horizon $\tilde T$ and risk aversion parameter $\gamma$ vary.}
\label{nplotmn}
\end{figure}
\section{Conclusion}\label{sect-conclusion}
We have studied the optimal trading of futures under a multiscale multifactor model. Closed-form expressions for the optimal controls and value function are derived through the analysis of the associated HJB equation. Using these, we have illustrated the path behaviors of the futures prices and optimal positions. We also quantify the values of the trading different combinations of futures under different model parameters.
\addtolength{\textheight}{-12cm}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
This paper extends recent work on boosting random forests [\cite{ghosal2020boosting}] to model non-Gaussian responses. Ensembles of trees have become one of the most successful general-purpose machine learning methods, with the advantage of being both computationally efficient and having few tuning parameters that require human intervention [\cite{breiman2001random}, \cite{friedman2001elements}]. Ensembles are made of multiple weak learners -- most commonly based on tree-structured models -- that are combined by one of two broad strategies: bagging or boosting. Bagging methods, in particular random forests [\cite{breiman2001random}], combine trees that are learned in parallel using the same processes: usually differing in the data samples given to the trees and any randomisation processes used in obtain them. In contrast, boosting methods, such gradient boosted forests [\cite{friedman2001greedy}], obtain trees in sequence: each tree used to improve the prediction of the current ensemble.
\cite{ghosal2020boosting} proposed combining bagging and boosting by boosting a random forest estimator: a second random forest was built to predict the out of bag residuals of the first random forest in a regression setting. This approach can be traced back to \cite{breiman2001random} and achieves a near-universal improvement in test-set error on real-world problems. A significant motivation in \cite{ghosal2020boosting} was to extend recent results on uncertainty quantification for random forests to boosting-type methods. \cite{mentch2016quantifying} and \cite{wager2017estimation} provided confidence intervals for the predictions of random forests by employing a particular bagging structure, when each tree is given by a strict subsample of the dataset. This result is obtained by representing the ensemble as an infinite order incomplete $U$-statistic for which a central limit theorem can be obtained with the limiting variance calculated either by a direct representation [\cite{mentch2016quantifying}] or the Infinitesimal Jackknife [\cite{wager2017estimation}], see also \cite{zhou2019asymptotic}. \cite{ghosal2020boosting} extended this framework to represent the combination of sequentially-built random forests as a sum of two jointly-normal $U$-statistics for which an extension of the Infinitesimal Jackknife could be applied.
This paper builds on the ideas in \cite{ghosal2020boosting} to the generalised regression setting where the responses follow an exponential family distribution. We use binary responses for classification and count data in which the response follows a Poisson distribution as motivating examples. In the general exponential family we assume that the expected value of the response is related to a function of the predictor but only through a suitable link function, i.e, $\mathds{E}[Y|X] = g^{-1}(f(X))$. When restricted to a linear model $f(x) = x^\top \beta$, this follows the generalized linear model framework, but is replaced in this paper by the sum of sequentially-constructed random forests. We estimate this model via Newton-boosting for each random forest step, which is essentially a Newton-Raphson update in gradient boosting [\cite{friedman2001greedy}]. This type of updating has been used in other iterative methods, for example word-embedding as in \cite{baer2018exponential}. We first obtain an optimal constant, and then fit two random forests in sequence, each time using a weighted fit to Newton residuals. This results in the same random forest process being used for each step; see \cref{sec:gbfdefine} for details.
We also extend the variance estimates derived in \cite{ghosal2020boosting} to this generalized regression model. This includes an extension of the Infinitesimal Jackknife estimate of covariance to include accounting for the initial constant; see \cref{sec:varest}. In \cref{sec:gbftheory} we discuss some theoretical properties of our algorithm, namely its asymptotic normality and consistency of the variance estimate under suitable regularity conditions. Finally in \cref{sec:gbfresults} we empirically demonstrate the utility of our algorithm - \cref{sec:simresults} focuses on results from a simulated dataset and \cref{sec:realresults} demonstrates the performance of our algorithm on some datasets from the UCI database (\cite{lichman2013uci}). In both cases we measure performance mainly in terms of test-set log likelihood, although other related metrics are also considered.
\section{Defining Generalised Boosted Forests}\label{sec:gbfdefine}
We consider the generalised model $\mathds{E}[Y|X] = g^{-1}(f(X))$ where $g$ is an appropriate link function; in this paper we consider the specific cases of the logit function if $Y$ is binary, or the log function if $Y$ is count data. Given data $(Z_i)_{i=1}^n = (Y_i, X_i)_{i=1}^n$ our goal is to construct an estimate $\hat{f}$ for $f$. Note that in the generalized linear model literature [\cite{mccullagh1989generalized}] the estimate $\hat{f}$ is from the family of linear functions but all of our calculations below can apply to any family of functions. Throughout this paper we will refer to the estimate $\hat{f}$ and its properties to be in the "link" space, and we will refer to $g^{-1}(\hat{f})$ to be in the "response" space.
We define the log-likelihood for the dataset $(Y_i, X_i)_{i=1}^n$ to be $\sum_{i=1}^n \ell(\eta_i; Y_i, X_i)$, where $\eta_i = f(X_i)$ is in the link space. For notational simplicity we will refer to the likelihood function for one point to be simply $\ell_i(\eta_i)$ without explicit reference to the dependence on the data $(Y_i, X_i)$. To estimate $\hat{f}$, first we will obtain an MLE-type initial value $\hat{\eta}_{MLE}^{(0)}$ given by
$$
\hat{\eta}_{MLE}^{(0)} = \argmax_{t} \sum_{i=1}^n \ell_i(t)
$$
To improve upon this estimate we will define "residuals" and then fit a function $\hat{f}^{(1)}$ with these residuals as responses. In \cite{ghosal2020boosting} we used the Gaussian regression setting with the model being $Y = f(X) + \epsilon$, where $\epsilon \sim N(0, \sigma^2)$. Then the residuals are given by $r_i = Y_i - \hat{f}(X_i)$, where $\hat{f}$ is some estimator of $f$. To generalise the definition of residuals to the non-Gaussian setting we will first note the relationship between residuals and the log-likelihood in the Gaussian setting. The log-likelihood in this case is given by $\sum_{i=1}^n \ell_i(\hat{f}(X_i)) = C - \sum_{i=1}^n \frac{(Y_i - \hat{f}(X_i))^2}{2\sigma^2}$, where $C$ is a constant. Then we see that $r_i \propto \ell_i'(\hat{f}(X_i))$.
Our goal now is to find a function $\hat{f}^{(1)}$ such that the log-likelihood has improved from its previous value while also utilising some notion of "residuals" as related to the derivative of the log-likelihood. So we must try to maximise $\sum_{i=1}^n \ell_i(\eta_i)$, i.e., minimise $\sum_{i=1}^n -\ell_i(\eta_i)$. Using a second order Taylor expansion on $-\ell_i$ we note that:
\begin{align*}
\sum_{i=1}^n -\ell_i(\eta_i) &\approx \sum_{i=1}^n \left[ - \ell_i(\hat{\eta}_i^{(0)}) - \ell_i'(\hat{\eta}_i^{(0)}) \cdot (\eta_i - \hat{\eta}_i^{(0)}) - \ell_i''(\hat{\eta}_i^{(0)}) \cdot \frac{(\eta_i - \hat{\eta}_i^{(0)})^2}{2} \right] \\
&= \sum_{i=1}^n -\ell_i''(\hat{\eta}_i^{(0)}) \left( \eta_i - \hat{\eta}_i^{(0)} + \frac{\ell_i'(\hat{\eta}_i^{(0)})}{\ell_i''(\hat{\eta}_i^{(0)})} \right)^2 + C,
\end{align*}
where $C$ is some constant. We can then approximate $\xi_i = \eta_i - \hat{\eta}_i^{(0)}$ using a random forest trained with predictors $X_i$, training responses $\ell_i'(\hat{\eta}_i^{(0)})/\big(-\ell_i''(\hat{\eta}_i^{(0)})\big)$ (the weighted residuals) and training weights $-\ell_i''(\hat{\eta}_i^{(0)})$. This sort of "weighted fit" is akin a Newton-Raphson update. Denote this base random forest by $\hat{f}^{(1)}$. Here we note that:
\begin{itemize}
\item This calculation holds for any initial estimate $\hat{f}^{(0)}(X_i) = \hat{\eta}_i^{(0)}$ but for the purposes of this paper the initial estimate $\hat{\eta}_i^{(0)} = \hat{\eta}_{MLE}^{(0)}$ is constant over $i$
\item There may be different ways of implementing the training weights while fitting a random forest. This paper employs the \textsf{ranger} package in \textsf{R} where weights are used as the sampling weights of the datapoints given to each tree. Alternative approaches might include utilising the weights in a weighted squared error loss function to be used for the splitting criteria in construction of each tree. We will assume that these two procedures (and any procedure using reasonable re-weighting) are equivalent.
\item In the Gaussian regression case the second derivatives $\ell_i''(\hat{\eta}_i^{(0)})$ are constant. In the non-Gaussian case we could also make the same assumption and simply fit random forests with the residuals $\ell_i'(\hat{\eta}_i^{(0)})$ as the response (with weights assumed to be 1 without loss of generality). Initial experiments (not shown) suggest that this procedure does not perform better than Newton weighting.
\end{itemize}
Once we have $\hat{\xi}_i^{(1)} = \hat{f}^{(1)}(X_i)$ we can also fit another random forest as a boosted step. Following similar calculations as above we will use predictors $X_i$, weighted residuals $\ell_i'(\hat{\eta}_i^{(0)}+\hat{\xi}_i^{(1)})/\big(-\ell_i''(\hat{\eta}_i^{(0)}+\hat{\xi}_i^{(1)})\big)$ as the training response and training weights $-\ell_i''(\hat{\eta}_i^{(0)}+\hat{\xi}_i^{(1)})$. Denote this boosted-stage random forest by $\hat{f}^{(2)}$.
The final predictor for a test point $x$ will be given by
$$
\hat{f}(x) = \hat{\eta}_{MLE}^{(0)} + \hat{f}^{(1)}(x) + \hat{f}^{(2)}(x)
$$
in the link space and by $g^{-1}(\hat{f}(x))$ in the response space. Similar to the Gaussian case discussed in \cite{ghosal2020boosting} we could also continue this boosting for more than one step; however we expect diminishing returns (and increasing computational burden) with further boosting steps as the size of the remaining signal decreases relative to the intrinsic variance in the data.
\section{Uncertainty Quantification}
In this section we derive a variance estimate that can be calculated at no additional cost. We also show that the variance estimate is consistent under some regularity conditions and motivate a central limit theorem. Finally we end the section with an algorithmic description of the generalised boosted forest.
\subsection{Variance Estimation}\label{sec:varest}
\paragraph{The Infinitesimal Jackknife}
We use the Infinitesimal Jackknife (IJ) method to calculate the variance estimates of each stage and also for the final predictor (in the link space). In general \cite{efron1982jackknife} defines the Infinitesimal Jackknife for an estimate of the form $\hat{\theta}(P^0)$ where $P^0$ is the uniform probability distribution over the empirical dataset. First we consider the slightly more general estimate $\hat{\theta}(P^*)$ where $P^*$ is some probability distribution over the empirical dataset, i.e., it is a vector of the same length as the size of the dataset and has positive elements that add to 1. This is then approximated by the hyperplane tangent to the surface $\hat{\theta}(P^*)$ at the point $P^* = P^0$, i.e., $\hat{\theta}(P^*) \approx \hat{\theta}_{\text{TAN}}(P^*) = \hat{\theta}(P^0) + (P^*-P^0)^\top U$, where $U$ is a vector of the directional derivatives given by
$$
U_i = \lim_{\epsilon \to 0} \frac{\hat{\theta}(P^0 + \epsilon(\delta_i - P^0)) - \hat{\theta}(P^0)}{\epsilon}, i = 1,\ldots,n
$$
Under a suitable asymptotic normal distribution for $P^*-P^0$ we can obtain the variance for $\hat{\theta}_{\text{TAN}}(P^*)$ to be $\frac1{n^2} \sum_{i=1}^n U_i^2$. This is the IJ variance estimate for the estimator $\hat{\theta}(P^0)$.
\paragraph{In the link space}
Borrowing the same notation as above suppose $\hat{\eta}_{MLE}^{(0)} = \hat{\theta}(P^0)$, where $P^0$ is the uniform probability vector and its directional derivatives are given by $U_i$. Then the IJ variance estimate of $\hat{\eta}_{MLE}^{(0)}$ will be given by $\widehat{var}_{IJ}(\hat{\eta}_{MLE}^{(0)}) = \frac1{n^2} \sum_{i=1}^n U_i^2$. \cite{wager2014confidence} demonstrate that the directional derivatives for random forests are given by
$$
U'_i = n \cdot cov_b (N_{i,b}, T_b(x))
$$
where $N_{i,b}$ is the number of times the $i$\textsuperscript{th} datapoint is used in training the $b$\textsuperscript{th} tree, $T_b(x)$ is the prediction for a testpoint $x$ from the $b$\textsuperscript{th} tree, and the covariance is calculated over all trees in the forest. Thus for our predictor $\hat{f}$ the IJ variance estimate is given by
\begin{align}
\hat{V}(x) &= \frac1{n^2} \sum_{i=1}^n \left( U_i + n \cdot cov_b (N_{i,b}^{(1)}, T_b^{(1)}(x)) + n \cdot cov_b (N_{i,b}^{(2)}, T_b^{(2)}(x)) \right)^2 \nonumber \\
&\qquad\qquad + \frac1B \big(var_b(T_b^{(1)}(x)) + var_b(T_b^{(2)}(x)) \big) \label{eq:varest}
\end{align}
where the second term accounts for the extra variability from the random forest being an incomplete U-statistic rather than a complete U-statistic [\cite{ghosal2020boosting}]. In \cref{sec:varcons} we show that this variance estimate is consistent under some regularity assumptions detailed in \cref{sec:gbftheory}. We would especially like to note one of the cross terms in the first term of the above expression given by $\frac1n \sum_{i=1}^n U_i \cdot cov_b (N_{i,b}^{(1)}, T_b^{(1)}(x))$ which is the IJ estimate of the covariance between the constant ($\hat{\eta}_{MLE}^{(0)}$) and the base stage random forest ($\hat{f}^{(1)}$). Similar cross-product terms will be estimates of co-variance between $\hat{\eta}_{MLE}^{(0)}$ (and $\hat{f}^{(1)}$) with the boosted stage random forest $\hat{f}^{(2)}$.
\paragraph{Bias correction}
The IJ variance estimate is known to have an upward bias, and so we will employ a correction term before further exploration. Our bias correction is inspired by \cite{zhou2019asymptotic}, namely the final formula for the variance estimate of a U-statistic in appendix E of that paper. Although note that we don't need to use the multiplicative factor because we already account for the variance between the trees in the second term of our variance estimate above \eqref{eq:varest}. Our final variance estimate is given by
\begin{align}
\hat{V}(x) &= \frac1{n^2} \sum_{i=1}^n \left( U_i + n \cdot cov_b (N_{i,b}^{(1)}, T_b^{(1)}(x)) + n \cdot cov_b (N_{i,b}^{(2)}, T_b^{(2)}(x)) \right)^2 \nonumber \\
&\qquad\qquad + \frac1B \cdot \left(1 - \frac{n}{k}\right) \cdot \big(var_b(T_b^{(1)}(x)) + var_b(T_b^{(2)}(x)) \big) \label{eq:varestcorr}
\end{align}
where $k$ is the size of the subsamples (without replacement) for each tree in the random forests $\hat{f}^{(1)}$ and $\hat{f}^{(2)}$.
\paragraph{Examples}
The closed form of $U_i$ will depend on the type of response being used. For example we can calculate $U_i$ for two cases below:
\begin{itemize}
\item For the binomial response family ($g^{-1}$ is the logit function) our model is $Y_i \sim binomial(n_i, p_i)$ where $p_i = g^{-1}(f(X_i))$ and $n_i$ are given positive integers. So we estimate $\hat{\eta}_i^{(0)}$ as follows ($C$ is a constant)
\begin{align*}
\ell_i(t) &= y_i \log\left(\frac{e^t}{1+e^t}\right) + (n_i-y_i)\log\left(\frac1{1+e^t}\right) + C \\
&= y_i(t - \log(1+e^t)) - (n_i-y_i)\log(1+e^t) + C \\
&= y_i t - n_i\log(1+e^t) + C\\
\implies \ell_i'(t) &= y_i - n_i \cdot \frac{e^t}{1+e^t}
\end{align*}
Then $\displaystyle\sum_{i=1}^n \ell_i'(t) = 0 \implies \displaystyle\sum_{i=1}^n y_i - \frac{e^t}{1+e^t} \cdot \displaystyle\sum_{i=1}^n n_i = 0 \implies \hat{\eta}_i^{(0)} = t = \log\left(\frac{\sum_{i=1}^n y_i}{\sum_{i=1}^n (n_i-y_i)}\right)$. Thus $\hat{\eta}_i^{(0)} = \hat{\theta}(P^0)$ where $P^0$ is the uniform probability vector and $\hat{\theta}(P) = \log\left(\frac{Y^\top P}{(N-Y)^\top P}\right)$ for a general probability vector $P$. So
\begin{align*}
U_i &= \lim_{\epsilon \to 0} \frac{\hat{\theta}(P^0 + \epsilon(\delta_i - P^0)) - \hat{\theta}(P^0)}{\epsilon} \\
&= \lim_{\epsilon \to 0} \frac{\log(Y^\top (P^0 + \epsilon(\delta_i - P^0))) + \log((N-Y)^\top (P^0 + \epsilon(\delta_i - P^0)))}{\epsilon} \\
&\qquad\qquad - \frac{\log(Y^\top P^0) + \log((N-Y)^\top P^0)}{\epsilon} \\
&= \lim_{\epsilon \to 0} \frac{\log(Y^\top P^0 + \epsilon\cdot Y^\top(\delta_i - P^0))) - \log(Y^\top P^0)}{\epsilon} \\
&\qquad\qquad - \frac{\log((N-Y)^\top P^0 + \epsilon\cdot (N-Y)^\top(\delta_i - P^0))) - \log((N-Y)^\top P^0)}{\epsilon} \\
&= \frac{Y^\top (\delta_i - P^0)}{Y^\top P^0} - \frac{(N-Y)^\top (\delta_i - P^0)}{(N-Y)^\top P^0} \\
&= \frac{y_i - \bar{y}}{\bar{y}} - \frac{(n_i - y_i) - (\bar{n} - \bar{y})}{\bar{n} - \bar{y}} \\
&= \frac{\bar{n}y_i - n_i \bar{y}}{\bar{y}(\bar{n} - \bar{y})}
\end{align*}
\item In case where $Y$ is count data and we use a Poisson response family with our model being $Y_i \sim Poisson(\lambda_i)$, where $\lambda_i = \exp(f(X_i))$ and we get the following calculations for estimating $\hat{\eta}_i^{(0)}$ ($C$ is a constant)
\begin{align*}
\ell_i(t) &= -e^t + y_i \log(e^t) + C = y_i t - e^t + C \implies \ell_i'(t) = y_i - e^t \\
\text{Then } \sum_{i=1}^n \ell_i'(t) &= 0 \implies \sum_{i=1}^n (y_i - e^t) = 0 \implies \hat{\eta}_i^{(0)} = t = \log(\bar{y})
\end{align*}
So we have $\hat{\eta}_i^{(0)} = \hat{\theta}(P^0)$ with $\hat{\theta}(P) = \log(Y^\top P)$ and thus
\begin{align*}
U_i &= \lim_{\epsilon \to 0} \frac{\hat{\theta}(P^0 + \epsilon(\delta_i - P^0)) - \hat{\theta}(P^0)}{\epsilon} \\
&= \lim_{\epsilon \to 0} \frac{\log(Y^\top (P^0 + \epsilon(\delta_i - P^0))) - \log(Y^\top P^0)}{\epsilon} \\
&= \lim_{\epsilon \to 0} \frac{\log(Y^\top P^0 + \epsilon\cdot Y^\top(\delta_i - P^0))) - \log(Y^\top P^0)}{\epsilon} \\
&= \frac{Y^\top (\delta_i - P^0)}{Y^\top P^0} = \frac{y_i - \bar{y}}{\bar{y}}
\end{align*}
\end{itemize}
\paragraph{In the response space}
Note that in the response space (by transformation with $g^{-1}$) we can use the delta-method to get the corresponding variance estimates. Using the delta method is valid under decreasing $\mbox{var}(\hat{f})$. Thus for a testpoint $x$ if $\hat{f}(x)$ is the prediction and $\hat{V}(x)$ is the IJ variance estimate in the link space [\eqref{eq:varestcorr}] then in the response space the variance estimate will be $\hat{V}(x) \cdot \big((g^{-1})'(\hat{f}(x))\big)^2$.
\subsection{Some theoretical exploration} \label{sec:gbftheory}
We will now briefly explore some theoretical properties of our estimator and the variance estimate.
\paragraph{A Regularity Condition}
First we need a regularity condition very similar to Condition 1 in \cite{ghosal2020boosting}. Note that the residuals $\ell_i'(\hat{\eta}_i^{(0)})/\big(-\ell_i''(\hat{\eta}_i^{(0)})\big)$ is a function of $\hat{\eta}_i^{(0)} = \hat{\eta}_{MLE}^{(0)}$ and thus depends on the whole dataset. Thus each of the trees in the random forest $\hat{f}^{(1)}$ depends on the whole dataset as well and so it's not a valid U-statistic kernel. This makes $\hat{f}^{(1)}$ not a valid U-statistic. A similar argument holds for $\hat{f}^{(2)}$. However, if we instead defined the random forests based on ``noise free'' residuals such as $\ell_i'(\mathds{E}[\hat{\eta}_i^{(0)}])/\big(-\ell_i''(\mathds{E}[\hat{\eta}_i^{(0)}])\big)$ then the base random forest would be a U-statistic (to make the second (boost) random forest a U-statistic as well we need a slightly more modified version of $\eta_i^{(1)}$). We should first make the following general assumption showing asymptotic equivalence of these two types of random forests.
\begin{cond} \label{cond:regularity}
For each of $j$ = $1,2$, define $\hat{f}^{(j)}$ to be the predictor obtained when using the residuals $\ell'_i(\hat{\eta}_i^{(j-1)})/\big(-\ell_i''(\hat{\eta}_i^{(j-1)})\big)$ and weights $-\ell_i''(\hat{\eta}_i^{(j-1)})$. We assume there are fixed functions $g^{(j-1)}$ such that for $\check{\eta}_i^{(j-1)} = g^{(j-1)}(x_i)$, the forest $\check{f}^{(j)}(x)$ obtained when using residuals $\ell'_i(\check{\eta}^{(j-1)})/\big(-\ell_i''(\check{\eta}_i^{(j-1)})\big)$ and weights $-\ell_i''(\check{\eta}_i^{(j-1)})$ follows
$$
\frac{\hat{f}^{(j)}(x) - \check{f}^{(j)}(x)}{\sqrt{var(\hat{f}^{(j)}(x))}} \xrightarrow{p} 0.
$$
\end{cond}
This condition is slightly more general than we need. A natural choice for the functions $g^{(j-1)}$ could be $g^{(0)}(x) = \mathds{E} \left[\hat{\eta}_{MLE}^{(0)}\right]$ and $g^{(1)}(x) = \mathds{E} \left[\hat{\eta}_{MLE}^{(0)} + \hat{f}^{(1)}(x)\right]$. However, $g^{(1)}(x) = E\left[\hat{\eta}_{MLE}^{(0)} + \check{f}^{(1)}(x)\right]$ could also be used so that $\check{\eta}_{MLE}^{(0)} + \check{f}^{(1)}(x) + \check{f}^{(2)}(x)$ represents a boosting analogue that is exactly a sum of $U$-statistics. Our main requirement is that the generalised boosted forest be approximated to within its standard error by functions that represent a sum of $U$ statistics.
In the rest of this section we will state and prove our theoretical results for the estimator $\check{f}(x) = \hat{\eta}_{MLE}^{(0)} + \check{f}^{(1)}(x) + \check{f}^{(2)}(x)$. A discussion regarding this condition can be found in $\S 3$ of \cite{ghosal2020boosting}.
\paragraph{Asymptotic Normality and Variance Consistency}
Suppose the random forests $\hat{f}^{(1)}$ and $\hat{f}^{(2)}$ both have $B_n$ trees and each tree $T(x; Z_I)$ is constructed with a subsample $I$ of size $k_n$ (without replacement). Define the variance of this tree kernel to be $\zeta_{k_n,k_n}$ and we also know that the variance of the complete U-statistic with $T$ as the kernel is given by $\frac{k_n^2}{n} \zeta_{1,k_n}$ [\cite{hoeffding1948class, lee1990u}]. Without loss of generality we can also assume $\mathds{E} T(x; Z_1,\dots,Z_{k_n}) = 0$.
Finally we assume the following Lindeberg-Feller type condition [due to \cite{mentch2016quantifying}]
\begin{cond}\label{cond:mainLF}
Suppose that the dataset $Z_1,Z_2,\dots \overset{iid}{\sim} D_Z$ and let $T(Z_1,\dots,Z_{k_n})$ be the tree kernel of a random forest based on a subsample of size $k_n$. Define $T_{1,k_n}(z) = \mathds{E} T(z,Z_2,\dots,Z_{k_n})$. We assume that for all $\delta > 0$
$$
\lim_{n \to \infty} \frac1{\zeta_{1,k_n}} \mathds{E}\left[ T^2_{1,k_n}(Z_1) \mathds{1}\{|T_{1,k_n}(Z_1)| > \delta\zeta_{1,k_n}\} \right] = 0
$$
\end{cond}
Then under the conditions \ref{cond:regularity} and \ref{cond:mainLF} we can prove the following result
\begin{thm}\label{thm:main}
Assume that the dataset $Z_1,Z_2,\dots \overset{iid}{\sim} D_Z$ and that $\mathds{E} T^4(x; Z_1,\dots,Z_{k_n}) \leq C < \infty$ for all $x, n$ and some constant $C$. If $k_n, B_n \to \infty$ such that $\frac{k_n}{n} \to 0$ and $\frac{n}{B_n} \to 0$ as $n \to \infty$ as well as $\lim\limits_{n \to \infty} \frac{k_n \zeta_{1,k_n}}{\zeta_{k_n,k_n}} \neq 0$ then we have
$$
\frac{\hat{f}(x)}{\sigma_n(x)} \overset{\mathscr{D}}{\to} \mathscr{N}(0,1)
$$
for some sequence $\sigma_n(x)$. Further $\sigma_n^2(x)$ is consistently estimated by $\hat{V}(x)$ as given by \eqref{eq:varest}.
\end{thm}
Our assumptions and conditions for this theorem is almost exactly the same as Theorem 1 in \cite{ghosal2020boosting} and so we will refer to that paper for the proof of asymptotic normality. The proof for the variance consistency part can be found in \cref{sec:varcons}.
Also note that the above result is for the link space estimate and variance estimate. But we can easily observe that $\sigma_n(x) \to 0$ and so we can use an appropriate second-order Taylor expansion and a delta-method type argument to prove asymptotic normality of the response space estimate (mentioned at the end of \cref{sec:gbfdefine}) and consistency of the response space variance estimate (mentioned at the end of \cref{sec:varest}).
\newpage
\subsection{Algorithm}
\begin{algorithm}[ht]
\caption{Generalised Boosted Forest (\textit{and its variance estimate})}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\SetKwInOut{Obtain}{Obtain}
\SetKwInOut{Calculate}{Calculate}
\Input{The data $\big(Z_i = (Y_i, X_i)\big)_{i=1}^n$, the link function $g^{-1}$, the log-likelihood function $\ell$, the tree function $T$, the number of trees in both forests $B$, the subsample size for each tree $k$, and the test point $x$.}
\Obtain{The MLE-type constant $\hat{\eta}_{MLE}^{(0)} = \argmax_{t} \sum_{i=1}^n \ell(t; Z_i)$. This may have a closed-form or may be done numerically.}
\Calculate{The directional derivatives of the MLE-type constant $U_i^{(0)}$.}
Calculate weighted residuals $r_i^{(1)} = \frac{\ell_i'(\hat{\eta}_{MLE}^{(0)})}{-\ell_i''(\hat{\eta}_{MLE}^{(0)})}$ and training weights $w_i^{(1)} = -\ell_i''(\hat{\eta}_{MLE}^{(0)})$
\For{$b = 1$ \KwTo $B$}{
Choose a subset of size $k$ from $[n]$ such that the probability of the $i$\textsuperscript{th} datapoint being included is proportional to $w_i^{(1)}$.
\Calculate{$T^{(1)}_b$ as the tree with the training data $r_I^{(1)}$ and $N_{i,b}^{(1)} = \mathds{1}\{i \in I\}$.}
}
\Obtain{The first (base) random forest $\hat{f}^{(1)}(x) = \frac1B \sum_{b=1}^B T^{(1)}_b(x)$.}
\Calculate{The directional derivatives of $\hat{f}^{(1)}(x)$ as $U_i^{(1)}(x) = n\cdot cov_b(N_{i,b}^{(1)}, T^{(1)}_b(x))$.}
Calculate weighted residuals $r_i^{(2)} = \frac{\ell_i'(\hat{\eta}_{MLE}^{(0)} + \hat{\eta}_i^{(1)})}{-\ell_i''(\hat{\eta}_{MLE}^{(0)} + \hat{\eta}_i^{(1)})}$ and training weights $w_i^{(1)} = -\ell_i''(\hat{\eta}_{MLE}^{(0)} + \hat{\eta}_i^{(1)})$, where $\hat{\eta}_i^{(1)} = \hat{f}^{(1)}(X_i)$.
\For{$b = 1$ \KwTo $B$}{
Choose a subset of size $k$ from $[n]$ such that the probability of the $i$\textsuperscript{th} datapoint being included is proportional to $w_i^{(2)}$.
\Calculate{$T^{(2)}_b$ as the tree with the training data $r_I^{(2)}$ and $N_{i,b}^{(2)} = \mathds{1}\{i \in I\}$.}
}
\Obtain{The second (boosted) random forest $\hat{f}^{(2)}(x) = \frac1B \sum_{b=1}^B T^{(2)}_b(x)$.}
\Calculate{The directional derivatives of $\hat{f}^{(2)}(x)$ as $U_i^{(2)}(x) = n\cdot cov_b(N_{i,b}^{(2)}, T^{(2)}_b(x))$.}
\Output{The \textit{generalised boosted forest} estimate in the \textbf{link} space at the test point $x$ given by $\hat{f}(x) = \hat{\eta}_{MLE}^{(0)} + \hat{f}^{(1)}(x) + \hat{f}^{(2)}(x)$ and its (bias-corrected) variance estimate given by $\hat{V}(x) = \frac1{n^2} \sum_{i=1}^n \left( U_i^{(0)} + U_i^{(1)} + U_i^{(2)} \right)^2 + \frac1B \cdot \left(1 - \frac{n}{k}\right) \cdot \big(var_b(T_b^{(1)}(x)) + var_b(T_b^{(2)}(x)) \big)$.}
\Output{The \textit{generalised boosted forest} estimate in the \textbf{response} space at the test point $x$ given by $g^{-1}(\hat{f}(x))$ and its variance estimate given by $\hat{V}(x) \cdot \big( g^{-1}(\hat{f}(x)) \big)^2$.}
\end{algorithm}
\section{Empirical Studies for Generalised Boosted Forest} \label{sec:gbfresults}
We use the random forest implementation in the \verb|ranger| package in \verb|R| [\cite{rangerRpackage}]. In this package the weights (the second derivatives of the likelihood function) are used as probabilities for the corresponding datapoints when selected in the training set of a particular tree.
\subsection{Performance on Simulated Datasets}\label{sec:simresults}
We use a signal given by $f(X) = \sum_{i=1}^5 X_i$, where $X \sim U([-1,1]^{15})$.
Our simulation runs for 200 replicates - in each of them we generate a dataset of size $n = 1000$ and train a generalised boosted forest with varying levels of subsample sizes (20\%, 40\%, 60\% or 80\% of the size of the training data) and $B = 1000$ trees for each of the two random forests. For each replication we also use a separate fixed set of 100 test points from $U([-1,1]^{15})$ and then append 5 more fixed points to them, which are given by
$$
p_1 = \mathbf{0}_{15},\; p_2 = \left(\frac13, \mathbf{0}_{14}^\top \right)^\top,\; p_3 = \frac1{3\sqrt{15}}*\mathbf{1}_{15},\; p_4 = 2p_3,\; p_5 = 3 p_3
$$
Experiments in \cite{ghosal2020boosting} indicate that the signal to noise ratio affects the performance of boosted forests (referred to as the Gaussian case above) since if the noise in the dataset increase so does that in the estimate and the benefit from boosting decreases. The same is true for the generalised boosted forest. We have therefore experimented with different combinations of signal-to-noise ratios as given below.
\begin{itemize}
\item \textbf{Binomial family:} One way to improve signal was to have the success probabilities far away from each other (since the noise is given by $p(1-p)$ and is minimised for $p=0,1$). For this purpose we multiplied the signal $f(X)$ by a scaling factor $s \in \{1,2,4,8,16\}$ before applying the logistic function to transform into probabilities. Thus the probabilities are pushed towards 0 and 1 as the scale increases.
We can also achieve higher signal by having a higher number of trials ($n_i$) - for this purpose we generated $n_i$ uniformly from the set $[M] = \{1,2,\dots,M\}$ where $M$ was varied to be in the set $\{1,2,4,8,16\}$. Once we have the values of $n_i$, we generate $Y_i \sim Binomial(n_i, p_i)$ where $p_i = g^{-1}(s \cdot f(X_i))$ for an appropriate scale $s$, and $g^{-1}$ being the logistic function.
\item \textbf{Poisson family:} If $Y_i \sim Poisson(\lambda_i)$ then $var(\mathds{E}(Y_i))/mean(var(Y_i)) = var(\lambda_i)/mean(\lambda_i)$ is the signal-to-noise ratio. So if we multiply all $\lambda_i$ values with a scaling factor then the signal-to-noise ratio increases. Thus we set $Y_i \sim Poisson(\lambda_i)$, where $\lambda_i = g^{-1}(f(X_i) + \log(s)) = s \cdot g^{-1}(f(X_i))$, where $s \in \{1,2,4,8,16\}$ and $g^{-1}$ is the exponential function.
\end{itemize}
For each of these settings we obtain predictions and a variance estimates in the link space, which we can also transform into their corresponding values in the response space. We use the following metrics to measure our performance
\begin{enumerate}
\item The improvement in test-set log-likelihood from the MLE-type stage to fitting the base forest, and also from fitting the base forest to the boosted stage. In the leftmost panel in Figure \ref{fig:testsetplots} we show the binomial case only for the value of $M=4$ since all the other number of trials have a very similar shape to the plot (when suitably scaled). Note that the y-axis of the plots are in the pseudo log scale (inverse hyperbolic sin)
\item We can define MSE in two ways - by $\frac1n \sum (\hat{f}(X_i) - f(X_i))^2$ in the link space and by $\frac1n \sum (g^{-1}(\hat{f}(X_i)) - g^{-1}(f(X_i)))^2$ in the response space. Consider the three different estimators $\hat{f}$ corresponding to the MLE-type, base random forest and final boosted estimator and the improvement in MSE that can be calculated between these stages in the same way as for log-likelihood above. In the second and third diagrams in Figure \ref{fig:testsetplots} we've also restricted the binomial family to the case $M=4$ (the other values of $M$ show plots with similar shape and scale). Here too the y-axis of the plots are in the pseudo log scale (inverse hyperbolic sin).
Note that each point in these plots are a summary of 200 numbers and for the Poisson family with high subsample fractions and high scales an unusual situation may arise. In this case the response space MSE may \textit{increase} between the base and first boosting stage - although this doesn't always happen, the amount of increment has a very heavy right tail, and the base to boost stage almost always corrects for this increase in MSE. For this reason, we show the median MSE improvement in this plot rather than the mean. Note that this is in spite of the link space showing \textit{decreases} in MSE between the different stages for all cases - this is be because the response space estimate is an exponentiation of the link space estimate and has too much variability, i.e, outliers play a larger role in skewing MSE. We have included a more detailed explanation for this phenomenon in \cref{sec:poissonMSE}.
\begin{figure}[ht]
\centering
\caption{Improvements in loglikelihood and MSE (link and response spaces) in the pseudo-log scale}
\label{fig:testsetplots}
\includegraphics[width=\columnwidth]{LL_and_MSE.pdf}
\end{figure}
\item Confidence interval coverage as the average number of times (out of the 200 replicates) the real signal is contain within the 95\% confidence interval. We plot the coverage as a function of the true value of the signal (shown in the pseudo-log scale) in Figure \ref{fig:testsetcoverage}. We restrict ourselves to the link space, a few representative families and scales, and a sample fraction of 0.4 (since we've observed that coverage doesn't really depend on the sample fraction).
We see that the coverage always improves with the second (boosted) random forest. For the binomial family there seems to be a threshold beyond which the confidence interval doesn't cover the true signal at all and this threshold is independent of both ways of trying to improve the signal-to-noise ratio. A similar phenomenon is apparent for the Poisson family although here the threshold depends on the scale. Possible reasons for this behaviour are discussed in \cref{sec:gbflimits}.
\begin{figure}[ht]
\centering
\caption{Coverage of 95\% confidence intervals vs the true signal}
\label{fig:testsetcoverage}
\includegraphics[width=\columnwidth]{signal_vs_coverage.pdf}
\end{figure}
\end{enumerate}
More detailed plots and discussion using further metrics can be found in \cref{sec:detailedlinearplots}.
For the 5 fixed points $p_1,\dots,p_5$ we can also find improvements in expected values of log-likelihood and MSE but the effects won't be very pronounced for individual points. Instead we will look at the following metrics
\begin{enumerate}
\item Absolute bias averaged over the 200 replicates. In Figure \ref{fig:5ptsbias} we again focus only on $M=4$ for the binomial family (the other values of $M$ show plots with similar shape and scale). Note that for similar reasons as MSE above, the average absolute bias can be calculated in the link (top two) and response (bottom two) spaces.
We can see that in the link space and for the binomial family (top left of Figure \ref{fig:5ptsbias}), the absolute bias doesn't depend on the subsample fraction but increases with scale and the distance of the test point from the origin. For the other three cases relationship with scale and distance from origin is similar although increasing subsample fractions generally decreases the absolute bias. Most importantly in all cases we can see that boosting reduces absolute bias.
\begin{figure}[p]
\centering
\caption{Absolute bias for the 5 fixed points - link space on top, response space on bottom}
\label{fig:5ptsbias}
\includegraphics[width=\columnwidth]{bias_curves.pdf}
\end{figure}
\item Consistency of the variance estimate (bias corrected) calculated by taking the ratio of \textit{average of the variance estimates} over the \textit{variance of the estimates} (over the 200 replicates). In Figure \ref{fig:5ptsvarcons} we restrict the binomial family to $M=4$ since the other values of $M$ have plots with similar shape and scale. This plot is only for the link space although the plot for the response space is very similar.
Here we see that the variance is more consistent for higher subsample fraction. For the binomial family variance consistency is slightly better for points nearer to the origin although the opposite effect is observed for the Poisson family. The effect of scale on variance consistency is difficult to ascertain. Lastly the effect of the boosting step on this metric seems to depend on the test-point and the family - for test-points farther from the origin this metric becomes better with boosting for the binomial, but worse for the Poisson family.
\begin{figure}[ht]
\centering
\caption{Variance consistency for the 5 fixed points (link space)}
\label{fig:5ptsvarcons}
\includegraphics[width=\columnwidth]{var_cons_curves.pdf}
\end{figure}
\item Kolmogorov-Smirnov statistic for normality. This can also be calculated for the link and response spaces (i.e., before and after the delta transformation). The values for the K-S statistic in the link space fall in the range of 0.1-0.3 - it doesn't change too much for the different points, but it increases slightly with subsample fraction and scales $s$, it is also slightly lower for the boosted stage than the base forest stage. A detailed plot for this metric can be found in \cref{sec:detailedlinearplots}.
\end{enumerate}
See \cref{sec:detailedlinearplots} for more detailed plots and discussions with other performance metrics.
\subsection{Performance on Real Datasets}\label{sec:realresults}
We present below results for some real datasets from the UCI database. Given the data $(Y_i, X_i)_{i=1}^n$ we employed a 10-fold cross-validation to get predictions $\hat{f}(X_i)$ and variance estimates $\hat{V}_i$ in the link space for all three stages. For performance metrics we use the log-likelihood (defined similar to above) MSE, average variance and prediction intervals (in the response space). The last three are given by
$$
\text{MSE} = \frac1n \sum_{i=1}^n (Y_i - g^{-1}(\hat{f}(X_i)))^2, \; \text{Avg Var} = \frac1n \sum_{i=1}^n \hat{V}_i^{(R)} = \frac1n \sum_{i=1}^n \hat{V}_i \cdot \big(g^{-1}(\hat{f}(X_i))\big)^2,
$$
$$
95\% \text{ prediction interval for } Y_i, \text{i.e.}, P_i = \left( g^{-1}(\hat{f}(X_i)) \pm \Phi^{-1}(0.975) \sqrt{\hat{V}_i^{(R)} + \hat{V}_e^{(R)}} \right),
$$
where $g^{-1}$ is the appropriate link function and $V_e^{(R)}$ is the error variance (in the response space) estimated by $\hat{V}_e^{(R)} =$ MSE. Once we have the prediction intervals we calculate the prediction coverage (PC) to be $\sum_{i=1}^n \mathds{1}_{\{Y_i \in P_i\}}$.
We've used the \textsf{Abalone} and \textsf{Solar flare} datasets to fit the Poisson family, with results reported in Table \ref{tab:poisson}.
\begin{table}[ht]
\caption{Performance statistics for the \textsf{Abalone} and \textsf{Solar Flare} datasets.}
\label{tab:poisson}
\begin{center}
\begin{tabular}{lrrrr}
\toprule
\textbf{\textsf{Abalone}} & MSE & Avg Var & PC & LL\\
\midrule
stage0 & 10.413 & 0.003 & 0.9494 & 12.8729\\
stage1 & 4.688 & 3.583 & 0.9825 & 13.1637\\
stage2 & 4.549 & 5.848 & 0.9882 & 13.1745\\
\bottomrule
\end{tabular}
\quad
\begin{tabular}{lrrrr}
\toprule
\textbf{\textsf{Solar flare}} & MSE & Avg Var & PC & LL\\
\midrule
stage0 & 0.703 & 0.001 & 0.9340 & -0.6659\\
stage1 & 5.195 & 20.312 & 0.9972 & -0.8382\\
stage2 & 1.768 & 6.661 & 0.9849 & -0.6526\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
We can see that the boosting stages improve all the measures for the \textbf{Abalone} data (except for the average variance). But in the \textbf{Solar flare} data the response takes the value 0 a lot of the time and only occasionally has values between 1 and 8 - that is why the first stage random forest fails to improve log-likelihood but further boosting helps in this case.
We've also used the \textsf{Spam} dataset for the binomial (Bernoulli) family where our results are given in Table \ref{tab:binom}.
\begin{table}[ht]
\caption{Performance statistics for the \textsf{Spam} dataset.}
\label{tab:binom}
\begin{center}
\begin{tabular}{lrrrr}
\toprule
\textbf{\textsf{Spam}} & MSE & Avg Var & PC & LL\\
\midrule
stage0 & 0.239 & 0.00006 & 1.0000 & -0.6709\\
stage1 & 0.059 & 0.02725 & 0.9813 & -0.2484\\
stage2 & 0.040 & 0.03162 & 0.9837 & -0.1648\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
Here boosting improves both MSE and the log-likelihood. It also makes the coverage for a 95\% prediction interval less conservative. Further note that the log-likelihood using the probability predictions from the \verb|ranger| package in \verb|R| is found to be -0.1957 which is less that the final log-likelihood given by our method, i.e, our method fits the dataset better than traditional classification forests in the log-likelihood metric.
\section{Conclusion}
We demonstrated a new method, \textit{Generalised Boosted Forest}, to fit responses which can be modelled by an exponential family. With the goal of maximising the training set log-likelihood we start with an MLE-type estimator in the link space (as opposed to the actual MLE which is in the response space). Then with a second order Taylor series approximation to the log-likelihood we showed that a random forest, fitted with pseudo-residuals and appropriate weights, can improve this log-likelihood. Then, similar to \cite{ghosal2020boosting}, we add another random forest as a boosting step which further improves the log-likelihood. Our method also uses the Infinitesimal Jackknife to get variance estimates, at no additional computational cost.
Through simulations and real data examples we've shown the effectiveness of our algorithm. The random forests always improve the test-set log-likelihood in all of the cases we've tested. We also see noticeable improvement in test-set MSE. Our estimates for uncertainty quantification shows conservative coverage of the resulting confidence/prediction intervals and the near-Gaussianity of the distribution of predictions.
Although outside the scope of this paper, the method could be improved in many ways. We could of course try more boosting steps; the boosting steps could also have a multiplier determined either by hyper-parameter tuning or with Newton-Raphson. Also note that we have not used over-dispersed models for variance, but obtaining an overdispersion parameter may be important in providing appropriate confidence intervals in some cases. We may also wish to consider alternative base-stage models such as a generalised linear model rather than a straightforward constant. In this case, performance improvement from boosting this model with random forests can also be thought of as a goodness of fit test for the base model.
\section*{Acknowledgements}
We are indebted to \href{mailto:<EMAIL>}{Benjamin Baer} for the idea of Newton boosting updates.
\printbibliography
\afterpage{\clearpage}
\newpage
|
\section{Appendix}
In general, we want to maximize complete log-likelihood $\log p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}|\boldsymbol{\theta}, \boldsymbol{\epsilon})$, where $\boldsymbol{\theta}$ and $\boldsymbol{\epsilon}$ are parameters of data / mask generative models.
We have ELBO (Eq.~(\ref{eq:elbo_final})) that lower bounds the complete data log-likelihood and let's treat unobserved $\mathbf{x}_\mathbf{u}$ as latent variables used in EM.
Next we introduce a distribution $q(\mathbf{x}_\mathbf{u})$ defined over the unobserved attributes. For any choice of $q(\mathbf{x}_\mathbf{u})$, the following decomposition holds,
\begin{align}
\log p&(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}|\boldsymbol{\theta}, \boldsymbol{\epsilon})
= \\
&\mathcal{L}(q| \boldsymbol{\theta}, \boldsymbol{\epsilon}) + D_\text{KL}(q || p(\mathbf{x}_\mathbf{u}|\mathbf{x}_\mathbf{o}, \mathbf{m}, \boldsymbol{\theta}, \boldsymbol{\epsilon}))
\end{align}
where
\begin{align*}
\mathcal{L}(q| \boldsymbol{\theta}, \boldsymbol{\epsilon})
&=\int q(\mathbf{x}_\mathbf{u}) \log p(\mathbf{x}_\mathbf{u},\mathbf{x}_\mathbf{o},\mathbf{m}|\boldsymbol{\theta}, \boldsymbol{\epsilon}) d \mathbf{x}_\mathbf{u} \\
&+ \int q(\mathbf{x}_\mathbf{u}) \log q(\mathbf{x}_\mathbf{u}) d \mathbf{x}_\mathbf{u}
\end{align*}
At $E$ step, fix $\boldsymbol{\theta}, \boldsymbol{\epsilon}$ as $\boldsymbol{\theta}^*, \boldsymbol{\epsilon}^*$, and lower bound
\begin{align*}
\mathcal{L}(q| \boldsymbol{\theta}^*, \boldsymbol{\epsilon}^*) =
\log p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}|\boldsymbol{\theta}^*, \boldsymbol{\epsilon}^*) - D_{KL}
\end{align*}
Maximum is obtained if the proposal distribution $q^*=p(\mathbf{x}_\mathbf{u}|\mathbf{x}_\mathbf{o}, \mathbf{m}, \boldsymbol{\theta}^*)$, yielding
\begin{align*}
\mathcal{L}(q^*| \boldsymbol{\theta}^*, \boldsymbol{\epsilon}^*) =
\log p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}|\boldsymbol{\theta}^*, \boldsymbol{\epsilon}^*)
\end{align*}
At $M$ step, fix $q=p(\mathbf{x}_\mathbf{u}|\mathbf{x}_\mathbf{o}, \mathbf{m}, \boldsymbol{\theta}^*)$ and to maximize lower bound w.r.t parameters:
\begin{align*}
\mathcal{L}(q^*| \boldsymbol{\theta}, \boldsymbol{\epsilon}) &=
\int q^*(\mathbf{x}_\mathbf{u}) \log p(\mathbf{x}_\mathbf{u},\mathbf{x}_\mathbf{o},\mathbf{m}|\boldsymbol{\theta}, \boldsymbol{\epsilon}) d \mathbf{x}_\mathbf{u}
\\& + \int q^*(\mathbf{x}_\mathbf{u}) \log q^*(\mathbf{x}_\mathbf{u}) d \mathbf{x}_\mathbf{u}
\end{align*}
The second term of RHS does not depend on $\boldsymbol{\theta}, \boldsymbol{\epsilon}$ therefore equivalent to maximize the first term of RHS. As we derived in varitional inference that $\log p(\mathbf{x}_\mathbf{u},\mathbf{x}_\mathbf{o},\mathbf{m}|\boldsymbol{\theta}, \boldsymbol{\epsilon})\geq \mathcal{L}_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\boldsymbol{\theta}}, \boldsymbol{\boldsymbol{\epsilon}}}(\mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u},\mathbf{m}) $, where $\boldsymbol{\phi}, \boldsymbol{\psi}$ are parameters of proposal distribution of variational inference, we have:
\begin{align*}
&\int q^*(\mathbf{x}_\mathbf{u}) \log p(\mathbf{x}_\mathbf{u},\mathbf{x}_\mathbf{o},\mathbf{m}|\boldsymbol{\theta}, \boldsymbol{\epsilon}) d \mathbf{x}_\mathbf{u}
\\&\geq \int q^*(\mathbf{x}_\mathbf{u}) \mathcal{L}_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\boldsymbol{\theta}}, \boldsymbol{\boldsymbol{\epsilon}}}(\mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u},\mathbf{m}) d \mathbf{x}_\mathbf{u}
\end{align*}
If we generate $\mathbf{x}_\mathbf{u}$ from $q^*(\mathbf{x}_\mathbf{u})=p(\mathbf{x}_\mathbf{u}|\mathbf{x}_\mathbf{o}, \mathbf{m}, \boldsymbol{\theta}^*)\approx
\int p_{\boldsymbol{\theta}^*}(\mathbf{x}_{\mathbf{u}}|\mathbf{m},\mathbf{z})
q_{\boldsymbol{\phi}^*,\boldsymbol{\psi}^*}(\mathbf{z}|\mathbf{x}_{\mathbf{o}},\mathbf{m}) d\mathbf{z}$, this holds for standard EM methods. However,
if we generate $\mathbf{x}_\mathbf{u}$ from prior that $
\int p_{\boldsymbol{\theta}^*}(\mathbf{x}_{\mathbf{u}}|\mathbf{m},\mathbf{z})
p(\mathbf{z}) d\mathbf{z}$, $\mathbf{x}_\mathbf{u}$ is not conditional on $\mathbf{x}_\mathbf{o}$ but only $\mathbf{m}$, it might violate the derivation from EM.
\section{Conclusion}
\label{sec:conclusions}
In this work, we propose VSAE, a novel latent variable model to learn from partially-observed heterogeneous data. The proposed VSAE handles missingness effectively by introducing a selective proposal distribution which is factorized w.r.t the data attributes. Further, VSAE is a general framework and capable of performing multiple tasks including data imputation, data generation and mask generation.
Extensive experiments with comparison to state-of-the-art deep latent variable models demonstrated the effectiveness of VSAE on a variety of tasks.
We summarize our contributions in the partially-observed heterogeneous setting as follows:
\textbf{{{Heterogeneity.}}} The {factorization w.r.t. attributes in the latent space} reduces the negative impact from the heterogeneity in the raw data space.
\textbf{{{Partial observations.}}} VSAE approximates the true posterior given partially-observed data with a novel {selective proposal distribution}. The automatic encoder selection between observed and unobserved attributes enables VSAE to ignore noisy information
and learn from partial observations.
\textbf{{{No MCAR assumption.}}} The independence assumption between data and mask can be restrictive. VSAE relaxes this assumption and models the joint distribution of data and mask together. To handle the intractablility of $\mathbf{x}_\mathbf{u}$, an expected lower bound is maximized during training.
\section{Experiments}
\label{experiment}
\subsection{Tabular Data}\label{sec::tabular_data}
\begin{table*}[t]
\centering
\caption{\textbf{MCAR Data Imputation on UCI datasets.} We consider categorical, numerical and mixed tabular datasets in the missing ratio of 0.5. Categorical and numerical attributes are evaluated by \textbf{PFC} and \textbf{NRMSE} respectively, lower is better for both. We show mean and standard deviation over 3 independent trials. $\Delta<0.0005$.}
\label{table:uci_mcar}
\scalebox{0.89}{\begin{tabular}{lcccccccc}
\toprule
& \textbf{Phishing}&\textbf{Mushroom} & &\textbf{Yeast}&\textbf{Whitewine} && \multicolumn{2}{c}{\textbf{Heart (mixed)}} \\
\cmidrule{1-3}\cmidrule{5-6}\cmidrule{8-9}
{Attribute type} &categorical&categorical &&numerical&numerical &&categorical&numerical \\
\midrule
AE
& $0.348\pm0.004$ & $0.556\pm0.009$ &
& $0.737\pm0.035$ & $0.3772\pm\Delta$&
& $0.514\pm0.026$ & $0.650\pm0.006$ \\
VAE
& $0.274\pm\Delta$ & $0.470\pm0.016$ &
& $0.461\pm\Delta$ & $0.3724\pm\Delta$ &
& $\bf{0.485\pm0.016}$ & $0.630\pm 0.025$\\
CVAE w/ mask
& $0.241\pm\Delta$ & $0.445\pm\Delta$ &
& $0.445\pm\Delta$ & $0.3726\pm\Delta$&
& $0.516\pm 0.015$& $0.605\pm0.005$\\
MVAE
& $0.308\pm0.016$ & $0.586\pm0.017$ &
& $0.442\pm0.016$ & $0.3732\pm\Delta$&
& $0.517\pm0.010$ & $0.636\pm0.009$\\
HI-VAE
& $0.238\pm\Delta$ & $0.470\pm0.008$ &
& $0.429\pm0.005$ & $\bf{0.3720\pm\Delta}$&
& $\bf{0.481\pm0.012}$ & $0.603\pm0.008$\\
VSAE (ours)
& $\bf{0.237\pm\Delta}$ & $\bf{0.416\pm0.009}$ &
& $\bf{0.419\pm0.008}$ & $\bf{0.3719\pm\Delta}$ &
& $\bf{0.482\pm0.014}$& $\bf{0.579\pm 0.015}$\\
\bottomrule
\end{tabular}}
\end{table*}
Tabular data are ordered arrangements of rows and columns. Each row is a sample with multiple attributes (typically low-dimensional) and each column is a single attribute collected heterogeneously. Due to communication or privacy issues, those attributes of data samples often are partially-observed.
For this scenario, we choose \textbf{{UCI repository}} which contains various tabular datasets of numerical or categorical attributes.
In all experiments, min-max normalization is applied to pre-process the numerical data and the unobserved dimensions are replaced by standard normal noise.
We split the training and test set with size ratio 4:1 and use 20$\%$ of training data as a validation set to choose the best model. Mean-squared error, cross-entropy and binary cross-entropy are used as reconstruction loss for numerical, categorical and mask variables, respectively.
\paragraph{Data Imputation.}
We first consider a data imputation experiment---imputing unobserved attributes given observed attributes and mask. VSAE can be used in this case as in Sec.~\ref{subsec:model_app}. We report the standard measures: \textbf{{NRMSE}} (RMSE normalized by the standard deviation of ground truth features, averaged over all features) and \textbf{{PFC}} (proportion of falsely classified attributes of each feature, averaged over all features) for numerical and categorical attributes. The evaluation is under various missing mechanisms by synthesizing masks following different rules.
\textbf{\textit{MCAR masking.}} We randomly sample from independent Bernoulli distributions with predefined missing ratios to mimic MCAR missing mechanism.
VSAE is compared with the deterministic autoencoder (AE), VAE, conditional VAE (CVAE)~\citep{NIPS2015_5775} conditioned on the mask, multi-modal MVAE~\citep{Wu2018MGM} and HI-VAE~\citep{nazabal2018handling}.
We use publicly released codes of MVAE/HI-VAE, and implement other baselines with the same backbone and at least as many parameters as VSAE.
\begin{table}[t
\caption{\textbf{Non-MCAR Data Imputation.} We show mean and standard deviation of \textbf{NRMSE} over 3 independent trials, lower is better. $\Delta<0.0005$. } \vspace{-5pt}\label{table:uci_nonmcar}
\scalebox{0.83}{\begin{tabular}{c c c c}
\toprule
&\textbf{Method} & MAR & NMAR \\
\midrule
\multirow{2}{*}{\textbf{Yeast}}
& MIWAE & $\bf{0.475\pm0.005}$ & $0.456 \pm 0.036$\\
& VSAE (ours) & $\bf{0.472\pm0.006}$ & $\bf{0.425\pm0.007}$\\
\midrule
\multirow{2}{*}{\textbf{Whitewine}}
& MIWAE & $0.3834\pm\Delta$ & $0.3723 \pm \Delta$\\
& VSAE (ours) &$\bf{0.3825\pm\Delta}$ & $\bf{0.3717\pm\Delta}$\\
\bottomrule
\end{tabular}}
\end{table}
Table~\ref{table:uci_mcar} shows that VSAE outperforms other methods on all datasets under MCAR missing mechanism in the missing ratio of 0.5.
Fig.~\ref{fig:uci_ratios} illustrates VSAE generally achieves lower error along with relatively lower variance in all missing ratios.
When the missing ratio increases (i.e.\ more data attributes become unobserved), VSAE consistantly maintains stable performance on most of the datasets. Conversely, we can observe a performance drop along with higher variance in the case of baselines.
As the missing ratio increases, the attributive proposal network of VSAE maintains the same input, while the encoders of other methods have to learn to focus on the valuable information.
We believe the selection mechanism of proposal distribution in VSAE mitigates this negative effect.
\begin{table*}[t]
\centering
\caption{\textbf{Data Imputation on multi-modal datasets.} Missing ratio is $0.5$. We evaluate each dataset w.r.t.\ each attribute---label attribute is evaluated by \textbf{PFC}, image attributes of MNIST and FashionMNIST are evaluated by \textbf{MSE} averaged over pixels, other attributes are evaluated by \textbf{MSE}. Lower is better for all. We show mean and standard deviation over 3 independent trials. $\Delta < 0.001$. }
\scalebox{0.8}{\begin{tabular}{lcccccccccc}
\toprule
& \multicolumn{2}{c}{\textbf{FashionMNIST + label (PFC)}} && \multicolumn{2}{c}{\textbf{MNIST + MNIST}}&& \multicolumn{3}{c}{\textbf{CMU-MOSI}} \\
\cmidrule{1-3}\cmidrule{5-6}\cmidrule{8-10}
{Attribute type}&Image & Label && Digit--1 & Digit--2 && Text & Audio & Image \\
\midrule
AE
& $0.1105\pm0.001$ & $0.366\pm0.01$&
& $0.1077\pm\Delta$ & $0.1070\pm\Delta$ &
&$0.035\pm0.003$ & $0.224\pm0.025$ & $0.019\pm0.003$\\
VAE
& $0.0885\pm\Delta$&$0.411\pm0.01$ &
& $0.0734\pm\Delta$ & $0.0682\pm\Delta$ &
&$0.034\pm\Delta$ & $0.202\pm0.003$ & $\bf{0.017\pm\Delta}$\\
CVAE w/ mask
& $0.0887\pm\Delta$ & $0.412\pm0.01$&
& $0.0733\pm\Delta$ & $0.0679\pm\Delta$ &
&$0.043\pm\Delta$ & $0.257\pm0.002$ & $0.020\pm\Delta$ \\
MVAE
& $0.1402\pm0.026$ & $0.374\pm0.07$&
& $0.0760\pm\Delta$ & $0.0802\pm\Delta$ &
& $0.44\pm\Delta$ & $0.213\pm0.001$ & $0.025 \pm \Delta$ \\
HI-VAE
& $0.1575\pm0.006$ & $0.405\pm0.01$&
& $0.0772\pm\Delta$ & $0.0725\pm\Delta$ &
& $0.047\pm\Delta$ & $0.211\pm0.005$ & $ 0.0267\pm\Delta$ \\
VSAE (ours)
& $\bf{0.0874\pm\Delta}$ & $\bf{0.356\pm0.01}$&
& $\bf{0.0712\pm\Delta}$ & $\bf{0.0663\pm\Delta}$ &
& \color{black}$\bf{0.033\pm\Delta}$ & \color{black}$\bf{0.200\pm\Delta}$ & \color{black}$\bf{0.017\pm\Delta}$ \\
\bottomrule
\end{tabular}}
\label{mnist_mnist_svhn_0.5}
\end{table*}
\begin{figure*}[t]
\begin{floatrow}
\ffigbox[\FBwidth]{%
\includegraphics[width=0.985\linewidth]{figures/data_impt.pdf}
}{%
\caption{\textbf{Data Imputation on MNIST+MNIST.} Top is the labels of unobserved digit via pre-defined rules; middle is the observed attribute; bottom shows the imputation of unobserved attribute from VSAE.}
\label{fig:mm_imp}
}
\ffigbox[\FBwidth]{%
\includegraphics[width=0.94\linewidth]{figures/data_gen.pdf}
}{%
\caption{\textbf{Data Generation on MNIST+MNIST.} Images are generated without conditional information. As shown, the correspondence between attributes are preserved in stochastic data generation. }
\label{fig:mm_gen}
}
\end{floatrow}
\end{figure*}
In Fig.~\ref{fig:uci_ratios}, a seemingly strange behaviour is observed --- some models (e.g. HI-VAE for \textit{Phishing}) perform better with more missing data. We hypothesize that:
{(i)} The missingness provides a natural way of dropout on the raw data space.
Under certain missing ratios, it possibly reduces overfitting. Hence, it can help the model to generalize;
{(ii)} To synthesize the missing mechanism, we sample and fix a mask on each dataset for all models, but the model itself may also be sensitive to the mask randomness.
Under a certain random sampling of the mask, some corrupted data or outliers, where one model is sensitive, might be observed and degrade the performance of this model.
For example, on \textit{Phishing} dataset, HI-VAE performs slightly better at 0.5 than 0.3). The error bar is wider at 0.3, but relatively narrow at 0.5 and 0.7.
Some outliers observed at 0.3 may contribute to this variance.
It performs worse again at 0.7, as the positive effect by masking out the outliers is not dominant when the missing ratio is too high.
\textbf{\textit{Non-MCAR masking.}}
MIWAE~\citep{mattei2019miwae} assumes data are MAR and approximates the conditional expectation of the unobserved attributes using importance sampling.
We follow it to mimic non-MCAR missing mechanisms on UCI numerical datasets --- {MAR}: 25\% attributes are default observed, then sample the remaining mask from $\operatorname{sigmoid}(\frac{1}{M}\sum_{k=1}^K \mathbf{x}_k)$,
where $K$ is the number of observed attributes; {NMAR}: sample the mask for $\mathbf{x}_i$ from $\operatorname{sigmoid}(\mathbf{x}_i)$. We use the public code of MIWAE and the same imputation estimates (single imputation of 1000 importance samplings) for fair comparison.
Table~\ref{table:uci_nonmcar} indicates VSAE can outperform the state-of-the-art non-MCAR model MIWAE in the non-MCAR setting, particularly for the NMAR, as VSAE models the joint distribution of attributes and mask without introducing the presumably false independence among them.
\paragraph{Mask Generation.}
VSAE enables us to generate data and mask from the learned generative model $p_{\boldsymbol{\theta},\boldsymbol{\epsilon}}(\mathbf{x},\mathbf{m},\mathbf{z})$.
We show mask generation results on UCI and data generation on multi-modal datasets (Sec.~\ref{sec:gen_multimodal}), since the sole data generation is not qualitatively or quantitatively measurable in the case of UCI datasets.
After training, we can sample from the prior to generate the mask on the corresponding dataset.
We evaluate mask generation by calculating the average proportion of missing attributes ($m_i=0$) on 100 sampled masks. In the \textbf{MCAR} setting, averaged over datasets, we get 0.312$\pm$0.016, 0.496$\pm$0.009, 0.692$\pm$0.005 for the ground truth missing ratios of 0.3, 0.5, 0.7. In the \textbf{MAR} setting where the mask depends on the dataset, we get 0.631$\pm$0.026 and 0.665$\pm$0.019 for the ground truth missing ratios of 0.626 (Whitewine) and 0.676 (Yeast). Similarly, in the \textbf{NMAR} setting we get 0.513$\pm$0.021 and 0.592$\pm$0.013 for the ground truth missing ratios of 0.502 (Whitewine) and 0.583 (Yeast). It indicates VSAE is capable of accurately learning the mask distribution.
Interestingly, we find improvement in the performance by conditioning the reconstructed mask variable on the data decoders. We speculate that this may be because the learned distributional mask variable can inform the data decoder of the missingness distributed in the data space, which in turn allows the potential missing mechanism to guide the data generative process.
\begin{table*}[t]
\centering
\caption{\textbf{Ablation study on modules of VSAE.} We consider experiments on \textit{Yeast} and \textit{Whitewine}. The masks of MCAR, MAR and NMAR are synthesized by the rules mentioned in Sec.~\ref{sec::tabular_data}. We evaluate each dataset by \textbf{NRMSE}. Lower is better. We show mean and standard deviation over 3 independent trials. $\Delta<0.0005$.}
\label{table:ablation}
\scalebox{0.83}{\begin{tabular}{lcccccccc}
\toprule
& \multicolumn{2}{c}{\textbf{MCAR}} & &\multicolumn{2}{c}{\textbf{MAR}} && \multicolumn{2}{c}{\textbf{NMAR}} \\
\cmidrule{2-3}\cmidrule{5-6}\cmidrule{8-9}
& \textbf{Yeast}&\textbf{Whitewine} & &\textbf{Yeast}&\textbf{Whitewine} && \textbf{Yeast}&\textbf{Whitewine} \\
\midrule
{Missing ratio}
& 0.5 & 0.5 &
& 0.676 & 0.626 &
& 0.583 & 0.502 \\
\midrule
\textcolor{black}{VSAE (collective only)}
& ${0.426\pm0.001}$ & ${0.3730\pm\Delta}$ &
& ${0.480\pm\Delta}$ & ${0.3837\pm\Delta}$&
& $\bf{0.427\pm0.005}$ & ${0.3737\pm\Delta}$ \\
\textcolor{black}{VSAE (attributive only)}
& ${0.457\pm0.017}$ & ${0.3728\pm\Delta}$ &
& ${0.493\pm0.005}$ & ${0.3827\pm\Delta}$ &
& ${0.434\pm0.006}$ & ${0.3729\pm\Delta}$ \\
\midrule
\textcolor{black}{VSAE (w/o mask modeling)}
& $\bf{0.418\pm0.004}$ & ${0.3720\pm\Delta}$ &
& ${0.485\pm0.006}$ & ${0.3834\pm\Delta}$ &
& ${0.437\pm\Delta}$ & ${0.3724\pm\Delta}$ \\
\midrule
No EM
& $\bf{0.426\pm0.012}$ & ${0.3723\pm\Delta}$ &
& ${0.476\pm0.001}$ & $\bf{0.3821\pm\Delta}$ &
& ${0.432\pm0.003}$ & ${0.3727\pm\Delta}$ \\
EM with 1 sampling
& ${0.433\pm0.002}$ & $\bf{0.3719\pm\Delta}$ &
& $\bf{0.473\pm\Delta}$ & ${0.3825\pm\Delta}$ &
& $\bf{0.426\pm0.003}$ & ${0.3721\pm\Delta}$ \\
EM with 1000 samplings
& $\bf{0.418\pm0.008}$ & $\bf{0.3718\pm\Delta}$ &
& $\bf{0.471\pm0.007}$ & ${0.3826\pm\Delta}$ &
& $\bf{0.424\pm0.005}$ & $\bf{0.3716\pm\Delta}$ \\
\midrule
VSAE (ours)
& $\bf{0.419\pm0.008}$ & $\bf{0.3719\pm\Delta}$ &
& $\bf{0.472\pm0.006}$ & ${0.3825\pm\Delta}$ &
& $\bf{0.425\pm0.007}$ & $\bf{0.3717\pm\Delta}$\\
\bottomrule
\end{tabular}}
\end{table*}
\subsection{Multi-modal Data}
\cite{BaltrusaitisAM17} defined \textit{multi-modal data} as data of multiple \textit{modalities}, where each modality is a way to sense the world---seeing, hearing, feeling, etc.
However, our definition of \textit{multi-modal} covers a wider spectrum where the data could be of the same type (e.g.\ image) but from different distributions (e.g. different shapes).
In the manner multi-modal data are collected or represented, we can safely treat multi-modal data (typically high-dimensional) as a type of heterogeneous data. In the following, we use \textit{attribute} and \textit{modality} interchangeably as a notion of heterogeneity.
We design experiments on three types of multi-modal data:
(i) image/label pair--- \textbf{{FashionMNIST}} images and labels;
(ii) image/image pair---synthesized bimodal MNIST+MNIST datasets by pairing two different digits from \textbf{{MNIST}} as \{(0,9),(1,8),(2,7),(3,6),(4,5)\};
(iii) standard multi-modal dataset \textbf{{CMU-MOSI}}~\citep{zadeh2018multi} including visual, textual and acoustic signals. See Appendix E for more results.
For all we use the standard training/validation/test split and all masking follows MCAR. We evaluate the performance on labels with \textbf{{PFC}} (proportion of falsely classified attributes), images with \textbf{{MSE}} (mean-squared error) averaged over pixels and other attributes with \textbf{MSE}.
\paragraph{Data Imputation.}
Table~\ref{mnist_mnist_svhn_0.5} demonstrates VSAE can achieve superior performance for multi-modal data imputation on all modalities with lower variance. Fig.~\ref{fig:mm_imp} presents the qualitative results of imputations on MNIST+MNIST image pairs.
To demonstrate robustness to missing ratio, we conducted experiments with missing ratio of 0.3, 0.5, 0.7 on the MNIST+MNIST dataset and the sum errors by VSAE are 0.1371$\pm$0.0001, 0.1376$\pm$0.0002 and 0.1379$\pm$0.0001 respectively. This indicates that VSAE also stays robust under different missing ratios for multi-modal datasets.
\paragraph{Data Generation.}\label{sec:gen_multimodal}
Fig.~\ref{fig:mm_gen} shows that VSAE is capable of generating image-based attributes following the underlying correlation.
The learning process does not require any supervision and can be effectively carried out with only access to partially-observed data.
We believe the underlying mechanism of selective proposal distribution benefits the performance. Though unobserved attributes in one data sample are not available even during training, they could be the observed attributes in others.
Thus, the collective proposal networks are able to construct a stochastic mapping from observable to unobservable information among the whole training set.
The separate structure of {attributive} and {collective proposal network} enforces VSAE to attend to the observed attributes, by ignoring unobserved attributes and performing efficient heterogeneity aggregation in the latent space. Thus it shows consistent robustness to various missing ratios.
In contrast, baselines primarily approximate the posterior by a single proposal distribution inferred straight from the whole input.
As a result baselines rely heavily on neural networks to extract expressive information from the data, which can be dominated by missing information.
\subsection{Ablation Study}
In Table~\ref{table:ablation}, we perform extensive ablation studies to demonstrate the effectiveness of three critical modules: selective proposal network, mask distribution modeling, and expectation over generated $\mathbf{x}_\mathbf{u}$ during training.
\textbf{Selective proposal distribution.} We compare VSAE to its variants with only collective or only attributive proposal network, which shows both are essential to the improvement.
Empirically we observe the collective structure tends to outperform its attributive counterpart on \textit{Yeast} but reversely worse on \textit{Whitewine}. In fact, if we analyze the datasets, \textit{Yeast} has highly correlated attributes (e.g. scores from different methods for the same task), whereas \textit{Whitewine} has more independent attributes (e.g. density, acidity or sugar).
\textbf{Modeling the mask.} We exclude mask conditional log-likelihood and make data decoder conditioned on the input mask. We observe the performance drops for non-MCAR, and do not observe significant improvement for MCAR. It indicates the mask reconstruction does not sacrifice model capacity in the case of MCAR, where the mask distribution can be easily learned.
\textbf{Expectation over generated $\mathbf{x}_\mathbf{u}$.} This learning procedure mitigates the intractability of $\mathbf{x}_\mathbf{u}$, therefore allowing to model the dependence between $\mathbf{x}_\mathbf{u}$ and $\mathbf{m}$. We first consider removing this expectation procedure and ignoring $\mathbf{x}_\mathbf{u}$ during training. The result (No EM) indicates this procedure plays an important role in NMAR ($\mathbf{x}_\mathbf{u}$ depends on $\mathbf{m}$) and does not significantly sacrifice the performance of MCAR and MAR. Compared to larger sampling size (of 1000), the result shows our proposed method (which generates 100 samples of $\mathbf{x}_\mathbf{u}$ to take expectation during training) is sufficient to obtain a good estimation of the expectation.
\section{Introduction}
\label{sec:intro}
Learning from data is an integral part of artificial intelligence.
A typical assumption of learning algorithms that the data is fully-observed is clearly unrealistic in many settings, therefore handling missing data has been a long-standing problem~\citep{NIPS1993_767, Schafer:1997}. The data generation process combined with the observation mechanism by which data are hidden makes learning from those data much more complex.
Moreover, conventional algorithms rely heavily on clean \textit{homogeneous} data, yet varied, \textit{heterogeneous} data are a common setting for learning. In fact, heterogeneity is ubiquitous in a variety of platforms from healthcare to finance to social networks to manufacturing systems~\citep{he2017learning}.
By the nature of the observation mechanism of the data, partial observability is often associated with heterogeneity. For example, a bank client is more likely to conceal the income amount than the attributes like gender or age; a doctor can never perform all medical tests for a patient, but choose to collect test results based on particular symptoms and expertise.
In this work we present a deep latent variable model for representation learning from {partially-observed heterogeneous} data.
Deep generative models have been shown to be effective in a variety of homogeneous data learning~\citep{bando2018statistical, gulrajani2017improved, mescheder2017adversarial, yang2017improved}. However, learning these models from heterogeneous data introduces new challenges. In particular, our definition of heterogeneity spans a wide range of forms from data type to data source/modality. For example, we may have categorical or numerical data from different distributions, or mixed modalities representing images, text, and audio. The main challenge is how to align and integrate heterogeneous data to model the joint distribution. Our proposed latent variable model handles this effectively by selecting appropriate proposal distributions, and performing the integration in a latent space instead of the input space.
Learning from partially-observed data is another challenge in deep generative models.
Naive solutions such as ignoring or zero-imputing missing data will likely degrade performance by introducing sparsity bias~\citep{abs-1906-00150}.
Having a model designed to learn from incomplete data not only increases the application spectrum of deep learning algorithms but also benefits down-stream tasks such as data imputation, which remains an open and challenging area of research.
Some prior work requires fully-observed data for training~\citep{suzuki2016joint,ivanov2018variational}, or makes the assumption that data are missing completely at random (MCAR)~\citep{yoon2018gain,li2019misgan}, which assumes \textit{missingness} (the manner in which data are missing) occurs independently from the data. Our method relaxes these assumptions by learning the joint distribution of data and missingness patterns (or \textit{mask}).
In this work, we propose the variational selective autoencoder (VSAE), a general and flexible model for representation learning from partially-observed heterogeneous data. The proposed deep latent variable model is capable of capturing hidden dependencies within partially-observed heterogeneous data by performing selection and integration in the latent representation.
It learns the joint distribution of data/mask without strong assumptions about missingness mechanism, resulting in applications for data generation and imputation. In particular, it can be trained effectively with a single objective to impute missing data from any combination of observed data.
Extensive evaluation on challenging low-dimensional and high-dimensional heterogeneous data shows improvement over state-of-the-art models.
The contributions are summarized as follows:
\vspace{-0.06in}
\begin{itemize}[leftmargin=*]
\item A novel selective proposal distribution efficiently learns representations from partially-observed heterogeneous data.
\item The proposed method models the joint distribution of the data and the imputation mask, resulting in a unified model for various tasks including generation and imputation.
\item VSAE does not make restrictive assumptions on the missingness mechanism, expanding the scope of scenarios in which data imputation can be effectively learned.
\end{itemize}
\section{Background}
\label{sec:method}
\textbf{Problem Statement.}
We represent any \textit{heterogeneous} data point as a set of random variables $\mathbf{x}=[\mathbf{x}_{1},\mathbf{x}_{2}...,\mathbf{x}_{M}]$ representing different \textit{attributes} collected from multiple sources. The type and size of each attribute $\mathbf{x}_i$ can vary. It can be either high-dimensional (e.g.\ multimedia data) or low-dimensional (e.g.\ tabular data).
We define an $M$-dimensional binary \textit{mask} variable $\mathbf{m} \in \mathbf\{0,1\}^M $ to represent the missingness: for the $i$-th attribute, $m_i=1$ if it is observed and $0$ otherwise.
Thus we can induce \textit{observed attributes} by the set $\mathbb{O} = \{i|m_i=1\}$ and \textit{unobserved attributes} by the complementary set $\mathbb{U}=\{i|m_i=0\}$.
Accordingly, we denote the collective representation of the observed attribute with $\mathbf{x}_\mathbf{o} = [\mathbf{x}_i|m_i=1]$ and unobserved attributes with $\mathbf{x}_\mathbf{u} = [\mathbf{x}_i|m_i=0]$.
In general, every instance has a different set of $\mathbf{x}_\mathbf{o}$ as well as $\mathbf{x}_\mathbf{u}$, determined by the mask variable $\mathbf{m}$.
\textbf{Missingness Mechanism. }
\label{sec:pre_imp-process}
The generative process of incomplete data can be modeled by the joint distribution
$ p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m})$.
\cite{little2019statistical} categorize the missingness mechanism into three types based on the dependence between the data and mask as follows,
\textbf{\textit{Missing Completely At Random (MCAR).}} Missingness is completely independent of data,
\vspace{-1mm}
\begin{align}\label{eq:mcar}
p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}) = p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}) p(\mathbf{m})
\end{align}
\textbf{\textit{Missing At Random (MAR).}} Missingness depends only on observed attributes,
\vspace{-1mm}
\begin{align}\label{eq:mar}
p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}) = p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}) p(\mathbf{m}|\mathbf{x}_\mathbf{o})
\end{align}
\textbf{\textit{Not Missing At Random (NMAR).}} Missingness depends on both observed and unobserved attributes,
\vspace{-1mm}
\begin{align}\label{eq:nmar}
p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}, \mathbf{m}) = p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u}) p(\mathbf{m}|\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u})
\end{align}
Most prior work on learning from partially-observed data follows the MCAR or MAR assumption since the factorization in Eq.~\eqref{eq:mcar} and Eq.~\eqref{eq:mar} decouples the mask $\mathbf{m}$ from $\mathbf{x}_\mathbf{u}$ in the integral of the likelihood function. A common approach is to define the log-likelihood of an incomplete dataset by marginalizing over the unobserved attributes $\log [\int p( \mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u}, \mathbf{m}) d\mathbf{x}_\mathbf{u}] = \log p(\mathbf{x}_\mathbf{o}, \mathbf{m})$ and ignoring the underlying missing mechanism. Our goal is to go beyond this simple but restrictive solution and model the joint distribution of the data and the mask.
\section{Proposed Method}
\label{sec:math-vsae}
\textcolor{black}
Variational Selective Autoencoder (VSAE) aims to relax the strong missingness assumption and model the joint distribution $p(\mathbf{x}, \mathbf{m})=\int p(\mathbf{x}, \mathbf{m}| \mathbf{z})p(\mathbf{z})d\mathbf{z}$. The latent variables $\mathbf{z}$ are inferred from data instances with the mask to capture the dependence among the attributes and the mask. The data and missingness information embedded in the latent variables can together guide data imputation and generation.
At a high level, as illustrated in Fig.~\ref{fig_arc}, VSAE is formulated with individual encoders for observed attributes and a collective encoder for unobserved ones to construct the selective proposal distribution. The selected latent variables of each attribute are aggregated and decoded to reconstruct the mask and all attributes independently. To tackle the intractability of unobserved attributes $\mathbf{x}_\mathbf{u}$ and latent variables $\mathbf{z}$, we formulate the learning procedure as an external expectation maximization (EM) nested with variational inference (VI). In the following, we provide details on the VSAE model.
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{figures/architetcure_new_v2.pdf}
\caption{\textbf{Model overview.} The input includes data \textit{attributes} (denoted by different colors --- $\mathbf{x}_1,\mathbf{x}_2$ are unobserved; $\mathbf{x}_3$ is observed) and \textit{mask}. The \textbf{\textit{attributive proposal network}} takes the attributes as input, while the \textbf{\textit{collective proposal network}} takes the observed attributes and the mask as input. We use $\operatorname{if}$-condition to denote the selective proposal distribution (Eq.~\ref{eq:selective_proposal}, the arrowed one is selected). The aggregated latent codes are fed to both \textbf{\textit{mask generative network}} and \textbf{\textit{data generative network}}. The output of the mask generative network will be fed to each decoder of the data generative network as extra condition. Standard normal prior is not plotted for simplicity.
All components are trained simultaneously in an end-to-end manner.}
\label{fig_arc}
\end{figure*}
\subsection{Model description}
Following the VAE\footnote{Refer to Appendix A for a detailed description of VAE. }~\citep{KingmaW13,pmlr-v32-rezende14}, we construct a proposal distribution $q(\mathbf{z}|\mathbf{x}, \mathbf{m})$ to approximate the intractable true posterior.
With the inclusion of the novel selective proposal distribution, we expand the parameters of inference networks to $\{\boldsymbol{\phi},\boldsymbol{\psi}\}$, where $\boldsymbol{\phi}$ and $\boldsymbol{\psi}$ represent encoder parameter for observed and unobserved attributes. Following the same fashion, the parameters of generative networks are expanded to $\{\boldsymbol{\theta},\boldsymbol{\epsilon}\}$, with $\boldsymbol{\theta}$ denoting the decoder parameter for the data, and $\boldsymbol{\epsilon}$ for the mask.
The variational evidence lower bound of $\log p(\mathbf{x}, \mathbf{m})$ can thus be derived as
\begin{align}\label{eq:elbo_init}
\begin{aligned}
&\mathcal{L}_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\theta}, \boldsymbol{\epsilon}}(\mathbf{x},\mathbf{m})
\underbrace{\mathbb{E}_{\mathbf{z}\sim q_{\boldsymbol{\phi}, \boldsymbol{\psi}}(\mathbf{z}|\mathbf{x}, \mathbf{m})}[{\log p_{\boldsymbol{\theta}, \boldsymbol{\epsilon}}(\mathbf{x},\mathbf{m}|\mathbf{z})}]}_\text{cond. LL}\\&-
\underbrace{{D}_\text{KL}
{(q_{\boldsymbol{\phi}, \boldsymbol{\psi}}(\mathbf{z} | \mathbf{x}, \mathbf{m}) || p(\mathbf{z})),}}_\text{KL Regularizer}
\end{aligned}
\end{align}
\textbf{Factorized Latent Space.} We assume the latent space can be factorized w.r.t.\ $M$ attributes
\begin{align}\label{eq:factorization}
p(\mathbf{z})&=\prod_{i=1}^M p(\mathbf{z}_i), \quad
q(\mathbf{z}|\mathbf{x}, \mathbf{m})=\prod_{i=1}^M q(\mathbf{z}_i|\mathbf{x}, \mathbf{m})
\end{align}
Priors $p(\mathbf{z}_i)$ are standard Gaussians and proposal distributions $q(\mathbf{z}_i|\mathbf{x}, \mathbf{m})$ are Gaussians with inferred means and diagonal covariances. This factorization separates the encoding of each attribute and efficiently yields a distribution for latent variables by assuming the latent variables are conditionally independent given the data and mask. Hence, it provides a mechanism to decouple the heterogeneity in the raw data space while integrating them efficiently in the latent space.
\textbf{Selective Proposal Distribution.} The standard proposal distribution of VAEs, inferred from fully-observed data, is not applicable for partially-observed input. To circumvent this, we introduce our \textit{selective proposal distribution} for each latent variable:
\begin{equation}\label{eq:selective_proposal}
q_{\boldsymbol{\phi},\boldsymbol{\psi}}(\mathbf{z}_i|\mathbf{x}, \mathbf{m}) =
\begin{cases}
q_{\boldsymbol{\phi}}(\mathbf{z}_i|\mathbf{x}_i) & \text{if $m_i=1$}\\
q_{\boldsymbol{\psi}}(\mathbf{z}_i|\mathbf{x}_\mathbf{o},\mathbf{m}) & \text{if $m_i=0$}\\
\end{cases}
\end{equation}
This conditional selection of proposal distribution is determined by the mask variable. Accordingly, we subdivide the inference network into two types:
\textbf{\textit{Attributive Proposal Network.}}
$q_{\boldsymbol{\phi}}(\mathbf{z}_i | \mathbf{x}_i)$, inferred merely from the individual observed attribute and selected for an observed attribute;
\textbf{\textit{Collective Proposal Network.}} $q_{\boldsymbol{\psi}}(\mathbf{z}_i | \mathbf{x}_\mathbf{o}, \mathbf{m})$, collecting all observed values and the mask to produce the proposal distribution and selected for an unobserved attribute.
This formulation aids VAE encoders by explicitly focusing on the relevant inputs and ignoring the less informative ones.
\textbf{Latent Variable Aggregation.}
We sample the latent variables for all attributes using Eq.~\eqref{eq:selective_proposal}.
Next, to capture the dependencies between observed attributes, unobserved attributes and mask, an aggregation function $\mathcal{F}(\cdot)$ is performed before the decoders.
We use \textit{concatenation} as $\mathcal{F}(\cdot)$, though it can be any aggregation function in general.
The conventional VAEs, however, often aggregate the attributes naively in the raw data space. Consequently, the heterogeneity and partially-observed nature will restrain those models from learning informative representations.
\textbf{Data \& Mask Generative Networks.}
By applying the chain rule, the conditional log-likelihood $\log p_{\boldsymbol{\theta}, \boldsymbol{\epsilon}}(\mathbf{x},\mathbf{m}|\mathbf{z})$ in Eq.~\eqref{eq:elbo_init} is decomposed as {mask conditional log-likelihood} $\log p_{\boldsymbol{\epsilon}}(\mathbf{m}|\mathbf{z})$ and {data conditional log-likelihood} $\log p_{\boldsymbol{\theta}}(\mathbf{x}|\mathbf{m}, \mathbf{z})$. The mask and data are reconstructed from shared $\mathbf{z}$ through the \textbf{\textit{mask generative network}} and \textbf{\textit{data generative network}} shown in Fig.~\ref{fig_arc}.
Further, the data conditional log-likelihood factorizes over the attributes assuming the reconstructions are conditionally independent given $\mathbf{m}$ and $\mathbf{z}$:
\begin{align}\label{eq:generative_independence}
\medmath{\log p_{\boldsymbol{\theta}}(\mathbf{x}|\mathbf{m}, \mathbf{z})
=\underbrace{\sum_{i \in \mathbb{O}} \log p_{\boldsymbol{\theta}}(\mathbf{x}_i|\mathbf{m}, \mathbf{z})}_\text{Observed} + \underbrace{\sum_{j \in \mathbb{U}} \log p_{\boldsymbol{\theta}}(\mathbf{x}_j|\mathbf{m}, \mathbf{z})}_\text{Unobserved}
}
\end{align}
\textcolor{black}{\textbf{Expectation Maximization.}} The ELBO (Eq.~\eqref{eq:elbo_init}) is hard to maximize since $\mathbf{x}_\mathbf{u}$ is unobserved. \textcolor{black}{We can use expectation maximization (EM) algorithm~\citep{Dempster77maximumlikelihood} to handle its intractability (refer to Appendix B for full derivation). EM alternates between inferring the unobserved data given the parameters ($E$ step) and optimizing the parameters given the “filled in” data ($M$ step).}
\textbullet {\textbf{ $E$ step:}} By taking an expectation over $\mathbf{x}_\mathbf{u}$,
\begin{align}\label{eq:expected_obj}
\mathcal{L}'_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\boldsymbol{\theta}}, \boldsymbol{\boldsymbol{\epsilon}}}(\mathbf{x}_\mathbf{o},\mathbf{m})= \mathbb{E}_{\mathbf{x}_\mathbf{u}}[\mathcal{L}_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\boldsymbol{\theta}}, \boldsymbol{\boldsymbol{\epsilon}}}(\mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u},\mathbf{m})]
\end{align}
We now expand Eq.~\eqref{eq:expected_obj} into conditional log-likelihood and KL divergence terms.
Since only the unobserved attributes conditional log-likelihood depends on $\mathbf{x}_\mathbf{u}$, with $ \mathbf{z}_i\sim q_{\boldsymbol{\phi}, \boldsymbol{\psi}}(\mathbf{z}_i|\mathbf{x}, \mathbf{m})$ given by Eq.~(\ref{eq:selective_proposal}), we obtain
\begin{align}\label{eq:elbo_final}
&\mathcal{L}'_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\boldsymbol{\theta}}, \boldsymbol{\boldsymbol{\epsilon}}}(\mathbf{x}_\mathbf{o},\mathbf{m})=
\underbrace{\mathbb{E}_{\mathbf{z}}\big[
\sum_{j\in \mathbb{U}} \mathbb{E}_{\mathbf{x}_j}[\log p_{\boldsymbol{\theta}}(\mathbf{x}_j|\mathbf{m},\mathbf{z})]}_\text{Unobserved attributes cond. LL} \big]\nonumber
+\\&
\underbrace{ \mathbb{E}_{\mathbf{z}}
\big[\sum_{i\in \mathbb{O}} \log p_{\boldsymbol{\theta}}(\mathbf{x}_i|\mathbf{m},\mathbf{z})\big]}_\text{Observed attributes cond. LL}
+
\underbrace{\mathbb{E}_{\mathbf{z}}[
\log p_{\boldsymbol{\epsilon}}(\mathbf{m}|\mathbf{z})]}_\text{Mask cond. LL}\nonumber -\\
&
\underbrace{\sum_{i=1}^{M} \mathbb{E}_{\mathbf{z}_i}[\log { q_{\boldsymbol{\phi}, \boldsymbol{\psi}}(\mathbf{z}_i|\mathbf{x}, \mathbf{m})}-\log { p(\mathbf{z}_i)}]}_\text{KL Regularizer}
\end{align}
In Eq.~(\ref{eq:elbo_final}), direct calculation of unobserved attributes conditional log-likelihood is intractable. Instead, during training given the learned generative model $p(\mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u},\mathbf{m},\mathbf{z})$ with VI, we can generate conditional distribution of unobserved attribute $\mathbf{x}_j$ from:
\begin{align}\label{eq:syn_unob}
\medmath{
\hat{p}({\mathbf{x}}_j|\mathbf{x}_\mathbf{o}, \mathbf{m})= \int
\Phi(\mathbf{z})
p_{\boldsymbol{\theta}^*}(\mathbf{x}_j|\mathbf{m}, \mathbf{z})
d\mathbf{z}
}
\end{align}
where $\Phi(\mathbf{z}):=
\prod_{i \in \mathbb{O}} q_{\boldsymbol{\phi}^*}(\mathbf{z}_i|\mathbf{x}_i)
\prod_{j \in \mathbb{U}} p(\mathbf{z}_i)$ and $*$ denotes the parameters learned up to the current iteration. It enables us to estimate each unobserved terms with
$\mathbb{E}_{\mathbf{x}_j\sim\hat{p}_(\mathbf{x}_
j|\mathbf{x}_\mathbf{o}, \mathbf{m})}[\log p_{\boldsymbol{\theta}}(\mathbf{x}_j|\mathbf{m},\mathbf{z})]$.
\textbullet\textbf{ $M$ step:} We can therefore maximize the final objective function (Eq.~(\ref{eq:elbo_final})) with $\mathbf{x}_\mathbf{u}$ sampled from Eq.~(\ref{eq:syn_unob}).
Conventional methods of learning from partially-observed data simply marginalizes and maximizes $\log \int p(\mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u},\mathbf{m})d\mathbf{x}_\mathbf{u} = \log p(\mathbf{x}_\mathbf{o}, \mathbf{m})$ by ignoring $\mathbf{x}_\mathbf{u}$. Our procedure can be viewed as maximizing the expected lower bound $\mathbb{E}_{\mathbf{x}_\mathbf{u}}[\mathcal{L}_{\boldsymbol{\phi}, \boldsymbol{\psi}, \boldsymbol{\boldsymbol{\theta}}, \boldsymbol{\boldsymbol{\epsilon}}}(\mathbf{x}_\mathbf{o},\mathbf{x}_\mathbf{u},\mathbf{m})]$
to approximately maximize
$\mathbb{E}_{\mathbf{x}_\mathbf{u}} [\log p(\mathbf{x}_\mathbf{o}, \mathbf{x}_\mathbf{u},\mathbf{m})]$.
Empirically, given a partially-observed batch, our two-stage scheme is:
(1) decode the latent codes drawn from the prior $p(\mathbf{z}_i)$ of unobserved attributes and the attributive proposal distribution $q_{\boldsymbol{\phi}^*}(\mathbf{z}_i|\mathbf{x}_i)$ of observed attributes to generate $\mathbf{x}_\mathbf{u}$ by the decoders learned so far;
(2) re-input the same incomplete batch to calculate all observed terms and estimate the unobserved term with Eq.~(\ref{eq:syn_unob}), then optimize over all parameters.
We generate 100 samples of $\mathbf{x}_\mathbf{u}$ to take expectation and experiments show that it gives an effective estimation to the full expectation.
Another possible alternative in $E$ step is to infer the latent variables of unobserved attributes from the collective proposal distribution $q_{\boldsymbol{\psi^*}}(\mathbf{z}|\mathbf{x}_\mathbf{o}, \mathbf{m})$ but empirically we find the prior performs better. The possible explanation lies in two-fold: (i) Approximating the true posterior $p(\mathbf{z}|\mathbf{x}_\mathbf{u}, \mathbf{x}_\mathbf{o}, \mathbf{m})$ is challenging without $\mathbf{x}_\mathbf{u}$. At early stages of training, the parametric proposal distribution is not learned well to match the posterior. Hence, the parameter-free prior is a more stable choice for the noisy unobserved attributes;
{(ii)} If sampling from the prior for unobserved attributes, the aggregated latent variables are still conditioned on $\mathbf{x}_\mathbf{o}$ from the attributive proposal distribution. Compared to collective proposal distribution $q_{\boldsymbol{\psi^*}}(\mathbf{z}|\mathbf{x}_\mathbf{o}, \mathbf{m})$ for $\mathbf{x}_\mathbf{u}$, sampling from the prior does not depend on $\mathbf{m}$. Since both data and mask generative processes have not yet been accurately captured, it can encourage focus on generating $\mathbf{x}_\mathbf{u}$.
\subsection{Model applications}\label{subsec:model_app}
Unlike conventional data imputation models, the generative model $p(\mathbf{x},\mathbf{m},\mathbf{z})$ is learned. Therefore, VSAE constructs a unified framework for data imputation, data generation and mask generation.
\textbf{{Data Imputation. }}The aim is to impute the missing data given the observed data, thus it can be viewed as conditional generation of the unobserved attributes.
This can be performed by sampling the latent codes for all attributes using $q_{\boldsymbol{\phi}, \boldsymbol{\psi}}(\mathbf{z}_i|\mathbf{x}, \mathbf{m})$ in Eq.~(\ref{eq:selective_proposal}). Next, the aggregated latent codes and mask are given to the decoders of the unobserved attributes for generation.
This process can be described as
$ p(\mathbf{x}_{\mathbf{u}} | \mathbf{x}_{\mathbf{o}}, \mathbf{m}) \approx
\int p_{\boldsymbol{\theta}}(\mathbf{x}_{\mathbf{u}}|\mathbf{m},\mathbf{z})
q_{\boldsymbol{\phi},\boldsymbol{\psi}}(\mathbf{z}|\mathbf{x}_{\mathbf{o}},\mathbf{m}) d\mathbf{z}$.
\textbf{{Data \& Mask Generation.}}
Given random samples from a standard Gaussian prior $p(\mathbf{z})$, we can generate masks using the mask generative network. Next, the sampled latent codes and the generated mask are given to the data generative network to generate data.
Modeling the mask distribution in our framework not only helps to inform the data generation of the relevant missingness mechanism but also enables to sample the mask itself. MisGAN~\citep{li2019misgan} also employs a mask generation process. However, MisGAN generates the mask assuming the mask is MCAR, whereas VSAE generates the mask assuming the mask is NMAR. Mask generation may have applications like synthesizing incomplete data or obtaining the potential mask if the real mask is not available (e.g.\ if data is corrupted rather than only missing).
\subsection{Scalability}
The complexity grows linearly w.r.t the number of attributes $M$ and their dimensions. In practice, heterogeneous data are often either \textit{tabular} (large $M$ but low-dimensional) or \textit{multi-modal} (high-dimensional but small $M$), thus practically tractable. For large $M$ with high-dimensional attributes (e.g. URLs), one may use embedding layers to densify these categorical attributes to low-dimensional vectors, or manually group subset of the attributes to constrain the $M$ as a feasible constant.
\section{Related Work}
\label{sec:related_work}
\paragraph{Learning from Heterogeneous Data.}
Most existing approaches modeling statistical dependencies in unstructured heterogeneous data focus on obtaining alignments and subsequently modeling relationships between different domains \citep{kim2017learning,zhu2017unpaired,castrejon2016learning}. However, there has been little progress in learning the joint distribution of the full data comprising different domains.
Other methods~\citep{misra2016cross,zhang2018learning} handle heterogeneity in labels or datasets in weakly-supervised learning settings.
MVAE~\citep{Wu2018MGM} uses
a product-of-experts inference network to solve the inference problem in multi-model setting.
Our work focuses on modeling all types of heterogeneity and demonstrates effectiveness in data generation and imputation applications.
\paragraph{Learning from Partially-Observed Data.}
Classical methods dealing with missing data such as MICE~\citep{buuren2010mice} and MissForest~\citep{10.1093/bioinformatics/btr597} typically learn discriminative models to impute missing features. Advanced by deep neural networks, several models have also been developed to address data imputation based on autoencoders ~\citep{gondara2017multiple,vincent2008extracting}, generative adversarial networks (GANs)~\citep{ li2019misgan,yoon2018gain}, and autoregressive models \citep{bachman2015data}.
In this work, we focus on improving deep latent variable models to efficiently learn from partially-observed heterogeneous data.
Deep latent variable models (DLVMs) are generative models that can map complex raw input to a flexible latent representation and have recently gained attention on handling partially-observed data due to the flexibility of generative modeling and representation learning.
To tackle the intractable posterior of DLVMs, variational autoencoder (VAE) first uses deep neural networks to approximate the posterior and maximizes a variational evidence lower bound (ELBO).
Based on VAE, prior work~\citep{ivanov2018variational, MaTPHNZ19, mattei2019miwae, nazabal2018handling, collier2020vaes} attempted to improve DLVMs under strong missingness mechanism assumptions.
VAEAC~\citep{ivanov2018variational} imputed attributes conditional on observed ones by learning from fully-observed data;
Partial VAE~\citep{MaTPHNZ19} encoded observed data with a permutation invariant set function;
MIWAE~\citep{mattei2019miwae} introduced a tighter bound using importance sampling under a relaxed Missing At Random (MAR, refer to Sec.~\ref{sec:pre_imp-process}) assumption;
HI-VAE~\citep{nazabal2018handling} factored decoders on low-dimensional heterogeneous data under MCAR.
These methods jointly map the sources (a.k.a.\ \textit{attribute}) of heterogeneous data into a holistic latent space, which adds unavoidable noise to the latent space as the distribution of one attribute can be far from the others.
\subsubsection*{\bibname}}
\begin{document}
\twocolumn[
\aistatstitle{Variational Selective Autoencoder: Learning from Partially-Observed Heterogeneous Data}
\aistatsauthor{ Yu Gong\textsuperscript{1,2} \And
Hossein Hajimirsadeghi\textsuperscript{1} \And
Jiawei He\textsuperscript{1} \And
Thibaut Durand\textsuperscript{1} \And
Greg Mori\textsuperscript{1,2}}
\aistatsaddress{ \textsuperscript{1}Borealis AI \And
\textsuperscript{2}Simon Fraser University
} ]
\begin{abstract}
Learning from heterogeneous data poses challenges such as combining data from various sources and of different types.
Meanwhile, heterogeneous data are often associated with missingness in real-world applications due to heterogeneity and noise of input sources.
In this work, we propose the variational selective autoencoder (VSAE), a general framework to learn representations from partially-observed heterogeneous data.
VSAE learns the latent dependencies in heterogeneous data by modeling the joint distribution of observed data, unobserved data, and the imputation mask which represents how the data are missing.
It results in a unified model for various downstream tasks including data generation and imputation. Evaluation on both low-dimensional and high-dimensional heterogeneous datasets for these two tasks shows improvement over state-of-the-art models.
\end{abstract}
\input{chapters/introduction.tex}
\input{chapters/related_work.tex}
\input{chapters/methods.tex}
\input{chapters/experiment.tex}
\input{chapters/conclusion.tex}
\bibliographystyle{plainnat}
|
\section{Introduction}
The globular
clusters AL~3, NGC~6558, and HP~1 share the characteristics of
having a metallicity of [Fe/H]$\sim$-1.0 and of being located in the Galactic bulge. They are old and could represent the earliest stellar populations in the Galaxy (Ortolani et al. 2006; Barbuy et al. 2018a; Kerber et al. 2019).
The star cluster AL~3 was discovered by Andrews \& Lindsay (1967) and was also cataloged as
BH~261 by van den Bergh \& Hagen (1975), reported as a faint open cluster.
It is reported in the ESO/Uppsala catalogue (Lauberts 1982) as ESO 456-SC78.
Ortolani et al. (2006) showed that the star cluster shows B, V, I
colour-magnitude diagrams (CMD) typical of a globular cluster.
It is centred at
J2000 $\alpha$ = 18$^{\rm h}$14$^{\rm m}$06.6$^{\rm s}$,
$\delta$ = $-$28$^{\rm o}$$38'06''$, with Galactic coordinates
l = 3.36$^{\rm o}$, b = $-$5.27$^{\rm o}$,
and located at 6\fdg25 and 2 kpc from
the Galactic centre, hence in the inner bulge volume. The cluster
has a depleted red giant branch (RGB), similarly to low-mass Palomar clusters,
indicating it to have been stripped along its lifetime.
This cluster has not been further observed so far.
The NGC~6558 cluster is located in a window, identified by Blanco (1988),
with equatorial coordinates (J2000)
$\alpha$ = 18$^{\rm h}$ 10$^{\rm m}$18.4$^{\rm s}$,
$\delta$ = -31$^{\circ}$ 45' 49" and
Galactic coordinates l = 0.201$^{\circ}$, b = -6.025$^{\circ}$.
It was analysed in terms of CMD by
Rich et al. (1998). Rossi et al. (2015) obtained a proper-motion-cleaned
CMD and presented a proper motion analysis, from which a study
of its orbits was given in P\'erez-Villegas et al. (2018, 2020).
The globular cluster Cl Haute-Provence 1 or HP~1, also
designated BH 229 and ESO 455-SC11, was discovered by
Dufay et al. (1954). It is located at
J2000 $\alpha$ = 17$^{\rm h}$31$^{\rm m}$05.2$^{\rm s}$,
$\delta$ = $-$29$^{\rm o}$$58'54''$, with Galactic coordinates
l = 357.42$^{\rm o}$, b = 2.12$^{\rm o}$.
In the present work, we studied individual stars of these clusters in a limited region of the spectrum in the H-band
corresponding to the wavelength region of the Phoenix spectrograph
at the Gemini South telescope,
centred at 15555 {\rm \AA}, and covering 15520-15590 {\rm \AA}, with
a high spectral resolution of R$\sim$75,000.
This region was chosen for containing prominent lines of CN, CO, and OH.
In Sect. 2, we examine the list of atomic and molecular lines in the region.
In Sect. 3, the observations are described. In Sect. 4,
the CMD is analysed. In Sect. 5,
the stellar parameters are given. In Sect. 6, the C, N, O abundances
are derived. Conclusions are drawn in Sect. 7.
\section{Spectroscopy in the H-band: Atomic and molecular lines}
The H-band will be intensely observed in the near future, given the new
instruments placing emphasis on the near-infrared region, such as the
James Webb Space Telescope (JWST), and new spectrographs on ground-based telescopes such as MOONS@VLT (presently CRIRES@VLT is available) and MOSAIC@ELT. The project APOGEE (Apache Point Observatory Galactic Evolution Experiment), with observations at a resolution of R$\sim$22,000 carried out at the 2.5-metre Sloan Foundation Telescope at the Apache Point Observatory in New Mexico (APOGEE-2N), and the 2.5-metre du Pont Telescope at Las Campanas Observatory in Chile (APOGEE-2S), as described in Majewski et al. (2017) has shown the power of the H-band.
Given the short wavelength range covering only 70 {\rm \AA} of the Phoenix spectrograph, we faced the challenge of identifying the lines in moderately metal-poor
stars of the Galactic bulge, for which only a few lines
are available. Because the available lamps did not include lines in this region, and experience proved that sky lines yielded a better wavelength calibration, and given the short wavelength range, it is not straightforward to identify the lines.
For this reason, we proceeded to a line identification, in the spectra of
the reference stars Arcturus and $\mu$ Leo, and created a
shortened version of a line list, containing only detectable lines.
Mel\'endez \& Barbuy (1999, hereafter MB99) worked on a list of atomic lines in the
J and H bands. The list of lines mostly corresponded to the detectable
lines.
That previous line list needed to be largely completed. Upon checking the lines detectable in the wavelength range 15520-16000
\rm \AA,
this was done by verifying the line lists from APOGEE (Shetrone et al. 2015) and VALD (Piskunov et al. 1995, Ryabchikova et al. 2015).
We note that astrophysical oscillator log~gf strengths
were applied to the APOGEE line list, and these should be suitable for abundance derivation.
Through a line-by-line check of its detectability in the Arcturus spectrum,
we identified lines of \ion{Mg}{I}, \ion{Si}{I},
\ion{Ca}{I}, \ion{Ti}{I}, \ion{Mn}{I}, \ion{and Ni}{I},
and we were not able to find detectable lines from the species \ion{C}{I},
\ion{O}{I}, \ion{Sc}{I}, \ion{V}{I}, \ion{Cr}{I}, \ion{Co}{I}, \ion{Cu}{I},
\ion{Y}{I}, \ion{and Y}{II}. The spectra computed including all lines
of all these elements are entirely equivalent to the one computed with the shortened line list, therefore to ensure practicality when identifying which lines really contribute
to a feature, we created
a table containing the detectable lines only. In this table, available upon request,
we report the oscillator strengths from MB99, APOGEE, VALD, and adopted values,
where in order of preference we adopted NIST and MB99.
Molecular electronic transition lines of CN A$^{2}\Pi$-X$^{2}\Sigma$,
and vibration-rotation
CO X$^{1}\Sigma^{+}$, OH X$^{2}\Pi$
lines were included in the synthetic spectra
calculations.
The line lists for CN were made available by
S. P. Davis, the CO line lists were adopted from
Goorvitch (1994), and the OH line list was made available by
S. P. Davis and A. Goldman (Goldman et al. 1998).
For more details on CN, CO, and OH molecular lines, we refer the reader to
Mel\'endez \& Barbuy (1999), Mel\'endez et al. (2001), and
Mel\'endez et al. (2002).
TiO $\phi$-system b$^{1}$$\Pi$-d$^{1}$$\Sigma$ lines are also present in the region. The line list by Jorgensen (1994) is included in the
calculations as described in
Schiavon \& Barbuy (1999) and Barbuy et al. (2018a).
The adopted dissociation potential of OH is 4.392 eV, D$_0$ = 11.092 eV for CO and D$_0$ = 7.65
eV for CN (Huber \& Herzberg 1979).
The PFANT code employed here for calculations of synthetic spectra is described in Barbuy et al.
(2018b), and it is available together with the atomic and molecular line lists.\footnote{http://trevisanj.github.io/PFANT}
We identified the lines in the reference stars Arcturus and $\mu$ Leo.
For the reference star Arcturus,
the spectrum atlas from Hinkle et al. (1995) is used,
and for the metal-rich reference giant star $\mu$ Leo,
APOGEE spectra are used, and their studies are be presented elsewhere.
The adopted stellar parameters for Arcturus and $\mu$ Leo
are from Mel\'endez et al. (2003) and Lecureur et al. (2007),
and they are reported in Sect. 5.
\begin{table}
\small
\caption[2]{\label{logbook}
Log book of observations. Proposals GS-2006A-C9 on 15-16/07/2006,
and GS-2008A-Q-23 on 10/04/2008 and 22/05/2008}
\begin{flushleft}
\begin{tabular}{lrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{Object} & Date & LST & Exp. & Seeing & \\
& & h:m & (m) & (") & \\
\noalign{\vskip 0.2cm}
\noalign{\hrule\vskip 0.2cm}
\noalign{\vskip 0.2cm}
NGC~6558-11 & 15-07-2006 & 22:00 & 2x25 & 0.6 & \\
NGC~6558-64 & 16-07-2006 & 14:48 & 3x30 & 0.7 & \\
NGC~6558-11 & 16-07-2006 & 16:33 & 2x20 & 0.5 & \\
NGC~6558-42 & 16-07-2006 & 19:13 & 2x30 & 0.7 & \\
NGC~6558-97 & 16-07-2006 & 20:25 & 3x20 & 0.6 & \\
NGC~6558-73 & 16-07-2006 & 21:38 & 2x30 & 0.9 & \\
HP~1-4 & 15-07-2006 & 13:42 & 3x30 & 0.6 & \\
HP~1-6 & 15-07-2006 & 20:12 & 3x30 & 0.6 & \\
HP~1-5 & 16-07-2006 & 13:03 & 3x30 & 0.8 & \\
HP~1-3 & 16-07-2006 & 18:02 & 2x30 & 0.6 & \\
AL~3-3 & 10-04-2008 & 15:58 & 4x15 & --- & \\
AL~3-6 & 22-05-2008 & 15:09 & 4x15 & --- & \\
AL~3-7 & 22-05-2008 & 16:21 & 4x15 & --- & \\
\noalign{\smallskip} \hline \end{tabular}
\end{flushleft}
\end{table}
\section {Observations}
The spectra of red giant stars of the bulge globular clusters
NGC~6558, AL~3, and HP~1 were observed with the Phoenix spectrograph
installed on the 8m Gemini South telescope.
The program was tri-national from
Brazil (PI: B. Barbuy), Chile (PI: M. Zoccali),
and Australia (PI: J. Mel\'endez).
The final suitable spectra include three stars of AL~3, two stars of NGC~6558,
and one star of HP~1, centred at 1.555 $\mu$m in the H band.
Another three stars in NGC~6558, and three in HP~1
were also observed; however, these showed a low Signal-to-Noise (S/N), due to clouds or high airmass.
The log of observations is provided in Table \ref{logbook}.
This is the first spectroscopic work on AL~3, except for Gaia
data (Gaia Collaboration 2020).
In Fig. \ref{chart}, a 3 min $B$ exposure of AL~3
is shown for a field extraction of 3.3'$\times$3.3'
(510$\times$510 pixels).
The sample stars of AL~3 are identified in the chart.
Charts and identifications of the observed stars in NGC~6558 and
HP~1 are given in Barbuy et al. (2009, 2018) and Barbuy et al. (2006,
2016), respectively.
\begin{figure}
\includegraphics[angle=0,scale=.50]{39761f1.pdf}
\caption{AL~3: 3 min B image with the three sample stars identified.
Extraction of 3'$\times$3'.
North is up and east is left.}
\label{chart}
\end{figure}
\subsection{Gaia cross-check}
In order to verify the corresponding membership probability of observed stars in AL~3, we performed the cross-match with Gaia Early Data Release 3 (EDR3; Gaia Collaboration 2020). We selected stars within 20' from the cluster centre and used the renormalised unit weight error (RUWE) $\le 2.4$ to ensure the kinematics precision and the minimum match separation.
Having the high-precision EDR3 proper motions ($\mu_\alpha^*$ = $\mu_\alpha \cos \delta$ and $\mu_\delta$), we obtain the mean proper motions for the cluster of $\mu_\alpha^* = 3.59\pm0.03$ mas yr$^{-1}$ and $\mu_\delta = -3.54\pm0.04$ mas yr$^{-1}$. These values are compatible with those given in Baumgardt et al. (2019). We also computed the Gaussian membership probability distribution of AL~3. We found that the stars AL3-6 and AL3-7 have membership probabilities of 100\%. Finally, the star AL3-3 has a relatively low membership probability of $\sim$ 60\%, but still, it could be considered a member. Therefore, all three observed stars are probable members of AL~3. Table \ref{gaia} provides the Gaia EDR3 cross-match and the membership probabilities.
\begin{table*}
\small
\caption[2]{\label{gaia}
Gaia magnitudes, proper motions and membership probability.}
\begin{center}
\begin{tabular}{lccccccc
\noalign{\smallskip}
\hline
\noalign{\smallskip}
Star & $\alpha$(J2000) & $\delta$(J2000) & $G$ & $G_{RP}$ & $\mu_{\alpha}^*$ & $\mu_{\delta}$ & Memb \\
& deg & deg & mag & mag & mas yr$^{-1}$ & mas yr$^{-1}$ & \% \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
AL3-3 & 273.5288067 & -28.6357960 & $14.186\pm0.003$ & $13.323\pm0.004$ & $3.78\pm0.09$ & $-4.04\pm0.07$ & 59 \\
AL3-6 & 273.5247404 & -28.6346067 & $15.228\pm0.003$ & $14.303\pm0.011$ & $3.56\pm0.06$ & $-3.64\pm0.04$ & 100 \\
AL3-7 & 273.5220767 & -28.6380356 & $13.924\pm0.003$ & $13.062\pm0.005$ & $3.58\pm0.03$ & $-3.48\pm0.03$ & 100 \\
\noalign{\smallskip} \hline \end{tabular}
\end{center}
\end{table*}
\subsection{Radial velocity of AL~3}
We were able to derive radial velocities
for the sample stars.
We used the low S/N individual observations of each star (S/N$\sim$10.0) combined to increase the signal-to-noise to S/N$\sim$18.0, and S/N$\sim$22.0 for AL3-3 and AL3-7, respectively.
Due to the stacking process, the most prominent features identified are FeI 15,534.26, OH 15,542.10, TiI 15,543.758, TiO/NiI 15,55.25 blend, CN 15,555.25, and FeI 15,591.49. We also used the OH sky lines, as listed
in Table 2 by Mel\'endez et al. (2003).
These features were used for AL3-3, giving a radial velocity of -67.65 $\pm$ 3.65 kms$^{-1}$. In the combined spectrum of AL3-7, the same features result in a radial velocity of -68.93 $\pm$ 4.83 kms$^{-1}$.
The corresponding heliocentric velocities of -57.29 km.s$^{-1}$ and -58.57 km.s$^{-1}$ lead to a final mean heliocentric velocity of -57.93 km.s$^{-1}$ $\pm$ 4.28 for AL~3.
Figures \ref{rval33} and \ref{rval37} show the line identification and radial velocity derivation.
The star AL3-6 shows a very noisy spectrum, and we verified
that it was observed under a high airmass of over 1.3, which
also explains that it is plagued with telluric features.
For AL3-6, we obtained a
different heliocentric radial velocity, as shown in Fig. \ref{rval36}, of -29.57 $\pm$ 5.85 km.s$^{-1}$,
compatible with the value given by Baumgardt et al. (2019) of -29.38 $\pm$ 0.60 km.s$^{-1}$.
The derived radial velocity is of crucial importance for
the computation of the cluster's orbits - see Sect. 5.
However, we
obtained two different figures: v$_{\rm r}^{hel}$ = -57.93 and -29.57
km.s$^{-1}$. In Fig. \ref{rval36}, we show the spectrum of AL3-6 compared with that of AL3-7.
Therefore, we obtained two different radial velocities for
AL~3, and consequently we considered both of them for the
calculation of orbits.
Since we had already computed the orbits for the lower value
(from Baumgardt et al. 2019), we show the orbits for the higher value here.
Clearly, new observations of these stars in a more extended wavelength range and with a higher S/N would be of great interest
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,scale=.50]{39761f2.pdf}
\caption{AL~3-3: Radial velocity derivation. The solid black line is the observed spectrum, the solid grey line is the noise spectrum, the solid red line is the synthetic spectrum, the dashed red lines are those used to derive the radial velocity, and the dashed black lines are the OH sky lines.}
\label{rval33}
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,scale=.50]{39761f3.pdf}
\caption{AL~3-7: Radial velocity derivation. The solid black line is the observed spectrum, the solid grey line is the noise spectrum, the solid red line is the synthetic spectrum, the dashed red lines are those used to derive the radial velocity, and the dashed black lines are the OH sky lines.}
\label{rval37}
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,scale=.50]{39761f4.pdf}
\caption{AL~3-6: Radial velocity derivation. The solid black line is the observed spectrum, the
solid grey line is the spectrum of star AL3-7, the dashed red lines are those used to derive the radial velocity, and the dashed black lines are the OH sky lines.}
\label{rval36}
\end{figure*}
\section{Colour-magnitude diagram of AL~3}
Ortolani et al. (2006) presented $B$, $V$, and Cousins $I$ images of AL~3, observed on 2000 March 6 using the 1.54m Danish telescope at the European Southern Observatory (ESO) at La Silla. They derived a reddening of E(B-V) = $0.36\pm0.03$ and a distance of $6.0\pm0.5$ kpc, based on the magnitude of the horizontal branch (HB) of AL~3, and suggested a metallicity of [Fe/H]$=-1.3\pm0.3$ from a comparison of the CMD with the mean locus of the cluster M5.
Rossi et al. (2015) described the observations with EFOSC at the NTT (ESO), La Silla in May 2012. They reported a distance of $\sim 6.5$ kpc, in agreement with Harris (1996, 2010 edition). P\'erez-Villegas et al. (2020) adopted both distances of 6.0 and 6.5 kpc and the radial velocity from Baumgardt et al. (2019) for the calculation of orbits.
We carried out the isochrone fitting using the SIRIUS code (Souza et al. 2020) on the same AL~3 data presented in Ortolani et al. (2006). The metallicity was limited by using a Gaussian prior with the value of Ortolani et al. (2006). We adopted
the Dartmouth Stellar Evolution Database (DSED - Dotter et al. 2008), with [$\alpha$/Fe]=0.4 and primordial helium. The asymptotic giant branch (AGB) is also shown, based on BaSTI isochrones (Pietrinferni et al. 2006), since they are not available in DSED.
The result is in very good agreement with results from Ortolani et al. (2006) within $1-\sigma$. We obtain a reddening of $E(B-V)=0.38\pm0.04$, a distance of d$_{\odot}=6.0\pm0.6$ kpc, and a metallicity of [Fe/H]$= -1.34\pm0.18$. Our age determination indicates an old age of $13.4^{+1.0}_{-1.2}$ Gyr, indicating that AL~3 is another relic fossil. It is important to stress that our distance of 6.0 kpc is also in agreement, within $1-\sigma$, with the value of 6.5 kpc from Harris (1996, 2010), Rossi et al. (2015), and Baumgardt et al. (2019).
Finally, we note that the dispersion of the data could be due to differential reddening, together with contamination and blends.
The reddening of AL~3 is relatively high, and the differential reddening is certainly an issue, as it is in other bulge globular clusters with similar reddening.
We expect an amount of about 20\% of differential reddening.
In principle it can be corrected, but the standard procedures for differential reddening correction in this cluster cannot be applied due to the few bona fide reference stars in the CMDs.
Figure \ref{fig:al3-cmd} shows the solution of isochrone fitting. The solid blue line represents the median solution, while the shaded regions indicate the solutions within $1-\sigma$. The red stars are the three sample stars. Finally, Figure \ref{fig:al3-corner} exhibits the corner plots showing the (anti)correlations between the parameters.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{39761f5.pdf}
\caption{AL~3 $V$ vs. $V-I$ CMD. The black dots are the stars within 120 pixel from the cluster centre (see Ortolani et al. 2006). The red stars are the observed stars of the present work. The solid blue line represents the median solution of the isochrone fitting, while the blue region reveals the solutions within $1-\sigma$.}
\label{fig:al3-cmd}
\end{figure}
\section{Stellar parameters}
\subsection{NGC 6558 and HP~1}
Individual stars of NGC~6558 were analysed
with high-resolution spectroscopy by
Barbuy et al. (2007, 2018b) and with moderate-resolution spectroscopy
by Dias et al. (2015). The stars NGC6558-42 and NGC6558-64
are studied here.
Similar studies of HP~1 were carried out in Barbuy et al. (2006, 2016)
and Dias et al. (2016).
In the 2006 article, the bright red giants were labelled with numbers 1 to 6,
for the purpose of identifying them in the cluster chart.
In 2016, we adopted the identification numbers corresponding to the
photometric reductions
relative to observations obtained at the New Technology Telescope (NTT)
at ESO, in 1994, as described in Ortolani et al. (1997).
HP1-4 and HP1-5 are stars 2115 and 2939 in Barbuy et al. (2016).
HP1-2 is the same as in Barbuy et al. (2006).
In our study, we only analysed HP1-5.
\subsection{AL~3}
The magnitudes and colours as follows are indicated in Table 3:
B, V from Ortolani et al. (2006), V, I from Rossi et al. (2015),
JHK from the 2MASS catalogue (Skrutskie et al. 2006),\footnote{
$\mathtt{http://ipac.caltech.edu/2mass/releases/allsky/}$;
$\mathtt{https://irsa.iapc.caltech.edu}$}
and JHK from the VVV survey
(Saito et al. 2012).\footnote{$\mathtt{horus.roe.ac.uk/vsa}$}
Effective temperatures were initially derived from $B-V$, $V-I$, $V-K$, and $J-K$ using
the colour-temperature calibrations of Alonso et al. (1999).
V,I Cousins were transformed to V,I Johnson using
($V-I$)$_{C}$=0.778($V-I$)$_{J}$ (Bessell 1979).
The $J,H,K_S$ magnitudes and colours were transformed from the 2MASS
system to California Institute of Technology (CIT), and from this to
Telescopio Carlos S\'anchez (TCS), using the relations established by
Carpenter (2001) and Alonso et al. (1998).
The conversion of JHK VVV colours to the JHK 2MASS system was done using
relations by Soto et al. (2013).
The temperatures resulting from photometry are of the order
of 5000 K for the three stars.
These temperatures, however, are not compatible with
another indicator, which is the Hydrogen Brackett 16
line, centred at 15556.457 {\rm \AA}. A fit of
this line for both AL~3 stars was carried out iteratively, after deriving their CNO abundances.
The resulting temperatures, adopted in the following analysis, are 4250 K and 4500 K for AL3-3 and AL3-7, respectively. The fits to the
hydrogen line are shown in Figure \ref{hline1}.
For AL3-6, the low quality of the spectrum does not allow the fit of
the hydrogen line, in particular due to strong telluric
absorptions in the region. It appears to be cooler and compatible with 4150 K. This incompatibly between photometric and hydrogen-wing-derived temperatures is
a main source of uncertainty in the present study.
To derive the gravity, we used the PARSEC isochrones
(Bressan et al. 2012).\footnote{http://stev.oapd.inaf.it/cgi-bin/cmd} To inspect the isochrones,
we adopted a metallicity of [Fe/H]$=-1.0$, or
overall metallicity Z$ = 0.00152$ (10 times below solar), and an age of 12 Gyr.
Assuming a reddening of $E(B-V) = 0.36$ (Ortolani et al. 2006, and present results), leading to $E(V-I) = 0.478$
and A$_V = 1.12$, we transformed the apparent magnitudes to absolute magnitudes,
as well as the colours ($V-I_{corr}$ = $V-I$-$E(V-I)$),
and we identified the correspondence of the observed stars to the theoretical isochrone.
The metallicity resulting from the CMD fitting is [Fe/H]=-1.34, which was imposed as a prior. We inspected individual lines of Fe in the AL3-3 spectrum and the fits are more compatible with
[Fe/H]=-1.0. There is also the evidence from other similar bulge globular clusters such as NGC 6558, NGC 6522, HP~1, and Terzan 9, which are found to have
[Fe/H]$\sim$-1.0 from high-resolution spectroscopy. Bica (2016) showed that there is a peak in metallicity at
[Fe/H]$\sim$-1.0 in the bulge, which we also adopted for AL~3. An isochrone fitting with this higher metallicity was tried, but appeared difficult to converge.
This is a second source of uncertainty of the present study.
Final adopted stellar parameters for program stars,
and of the reference stars Arcturus (Mel\'endez et al. 2003) and
$\mu$ Leo (Lecureur et al. 2007), are reported in Table \ref{param}.
\begin{table*}
\small
\caption[3]{\label{logmag}
AL~3: coordinates, magnitudes, and colours of sample stars.}
\begin{flushleft}
\begin{tabular}{lrrrrrrrrrrrr}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{Star}& \hbox{2MASS}& \hbox{$\alpha$(J2000)} & \hbox{$\delta$(J2000)}&
\hbox{V}& \hbox{I} & \hbox{J} & \hbox{H} & \hbox{K} & \hbox{J} & \hbox{H} & \hbox{K} &\\
& & & & & & & \multispan2 {2MASS} & & \multispan2 {VVV} \\
\noalign{\vskip 0.2cm}
\noalign{\hrule\vskip 0.2cm}
\noalign{\vskip 0.2cm}
B11 & 18101902-3144506 & 18 10 19.01 & -31 44 50.64 & 15.902 & 14.275 & 13.136& 12.440 & 12.280 & 13.017 & 12.397 & 12.248 & \\
B64 & 18101803-3145435 & 18 10 18.03 & -31 45 43.55 & 15.703 & 14.180 & 13.064 & 12.456 & 12.277 & 13.055 & 12.529& 12.384 & \\
B73 & 18102150-3145268 & 18 10 21.50 & -31 45 26.77 & 15.709 & 14.187 & 13.128 & 12.449 & 12.316 & 13.047 & 13.047& 12.313 & \\
F42 & --- & 18 10 17.65 & -31 45 38.93 & 16.054 & 14.442 & ---& --- & --- & --- & --- & --- & \\
F97 & 18101520-3146014 & 18 10 15.21 & -31 46 00.67 & 16.037 & 14.467 & 13.183 & 12.481 & 12.338 & --- & 12.503& 12.378 & \\
\hline
HP1-2 & 17310585-2958354 & 17 31 05.60 & -29 58 34.00 & 16.982 & 14.332 & 12.210 & 11.268 & 10.969 & 14.588 & 13.675& 13.368& \\
HP1-4 & 17310538-2959199 & 17 31 05.30 & -29 59 20.00 & 17.070 & 14.281 & --- & 11.67 & --- & 11.258 & 11.392& 10.688 & \\
HP1-5 & 17310729-2959021 & 17 31 07.20 & -29 59 02.00 & 17.131 & 14.395 & 11.901 & 10.869 & 10.595 & 12.021 & 11.285& 10.898 & \\
\hline
AL3-3 & 18140691-2839087& 18 14 06.90 & -28 38 09.0 & 14.524 &13.204 & 12.214 & 11.631 &11.469 & 12.211 & 11.714& 11.544 & \\
AL3-6 & 18140592-2838049& 18 14 05.80 & -28 38 06.0 & 15.563 &14.203 & 12.763 & 12.272 & 12.256 & --- & --- & --- & \\
AL3-7 & 18140529-2838168& 18 14 05.30 & -28 38 19.0 & 14.313 &12.963&
11.878 & 11.296 & 11.170 & 11.920 &11.472 & 11.195 \\
\noalign{\smallskip} \hline \end{tabular}
\end{flushleft}
\end{table*}
\begin{table}
\caption[4]{\label{param}
Adopted stellar parameters for individual stars in NGC~6558,
AL~3, and HP~1, and resulting C, N, O abundances. For NGC~6558, the stellar parameters are from Barbuy et al. (2007), and for HP~1 they are from Barbuy et al. (2016). Stellar parameters for the Sun, Arcturus, and $\mu$ Leo
are also included. }
\begin{tabular}{l@{}r@{}r@{}rrr@{}r@{}r@{}r@{}}
\noalign{\smallskip}
\hline
\noalign{\smallskip}
\hbox{Name}&
\hbox{T$_{\rm eff}$} &\hbox{\phantom{-}\phantom{-}log~g}& \hbox{\phantom{-}\phantom{-}[Fe/H]} & \hbox{v$_t$} & [C/Fe] & [N/Fe] & [O/Fe] & \\
& \hbox{(K)} & & & \hbox{km/s} & & & & \\
\noalign{\smallskip}
\noalign{\hrule}
\noalign{\smallskip}
& \multicolumn{7}{c}{ Program stars } \\
\hline
F42 & \phantom{-}3800 & \phantom{-}0.5 & \phantom{-}$-$1.05 & 1.65 & $-$0.5 & +0.8 & +0.5&\\
HP1-5 & \phantom{-}4525 & \phantom{-}2.0 & \phantom{-}$-$1.07 & 1.55 & $-$0.1 & 0.8 &0.5 & \\
AL3-3 & \phantom{-}4250 & \phantom{-}1.7 &\phantom{-}$-$1.00 &1.2 & $-$0.1 & 0.0 &0.0 & \\
AL3-6 & 4150 & 1.3 &-1.00 &1.2 & +0.8 & +1.0 &+0.8 & \\
AL3-7 & \phantom{-}4500 & \phantom{-}1.5 &\phantom{-}$-$1.00 &1.2 & $-$0.2 & -0.1 &0.0 & \\
\hline
\noalign{\smallskip}
& \multicolumn{7}{c}{ Reference stars } \\
\hline
Sun & \phantom{-}5770 & \phantom{-}4.44 & \phantom{-}+0.00 & 1.0 & +0.00 & +0.00 & +0.00 & \\
Arcturus & \phantom{-}4275 & \phantom{-}1.55 & \phantom{-}$-$0.54 & 1.65 & +0.11 & +0.40 & +0.43 & \\
$\mu$ Leo & \phantom{-}4540 & \phantom{-}2.3 & \phantom{-}+0.30 & 1.3 & $-$0.1 & +0.65 & +0.00 & \\
\noalign{\smallskip} \hline
\end{tabular}
\end{table}
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,scale=.95]{39761f6.pdf}
\caption{AL~3-3 and AL~3-7: Hydrogen Brackett 16 line computed
for T$_{eff}$ = 4250, 4500, 4750, 5000 K (upper panels) and adopted values of 4250 K and 4750 K,
respectively (lower panels).
The dashed line is the observed spectrum}
\label{hline1}
\end{figure*}
\begin{figure}[h!]
\centering
\includegraphics[angle=0,scale=.45]{39761f7.pdf}
\caption{AL~3-6: Tentative fit to this low S/N spectrum,
showing that it is compatible with a high CNO abundance.
Blue and black lines: Same observed spectrum normalized in
two different ways.
Red and cyan lines: Synthetic spectra computed
with [C/Fe]=0.7 (red), 0.8 (cyan), [N/Fe]=+1.0, [O/Fe]=+0.8.}
\label{al36cd}
\end{figure}
\section{CNO abundances}
The atmospheric models were interpolated in the grid of models by Gustafsson et al. (2008).
The synthetic spectra were computed employing the PFANT code
described in Barbuy et al. (2018b).
In order to derive the C, N, O abundances, we
fitted the CN, OH, and CO lines iteratively.
\subsection{The cool red giant N6558-42}
The cool red giant, NGC6558-42, shows strong lines
and is a typical red giant. For this reason, we show the fits to the spectrum of this star in detail in Figures
\ref{n655842ab} and \ref{n655842cd}.
The star NGC6558-64 instead, which would have an effective
temperature of 4850 K according to the analysis from optical
spectra by Barbuy et al. (2018b), could be as hot as 5500 K.
This is seen from the profile of its Hydrogen Brackett 16 line;
however, this should be taken with caution due to defects in the observed spectrum.
For this reason, we could not converge on CNO abundances for this
star.
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,scale=.90]{39761f8.pdf}
\caption{NGC 6558-42: Line identification
in the range 15527-15555 {\rm \AA}. Dashed line: Observed spectrum. Solid red line: Synthetic spectrum.
Synthetic spectrum computed with [C/Fe]=-0.5, [N/Fe]=0.8, [O/Fe]=+0.5.}
\label{n655842ab}
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[angle=0,scale=.90]{39761f9.pdf}
\caption{NGC 6558-42: Same as Fig. \ref{n655842ab}, in the range 15555-15587 {\rm \AA}.}
\label{n655842cd}
\end{figure*}
\subsection{AL-3 and HP~1}
There is a clear contrast between the spectra of
HP1-5, AL3-3, and AL3-7, that have shallow lines, and AL3-6 and
NGC6558-42, that show strong lines.
Whereas NGC6558-42 is a typical red
giant, the stars AL3-3, AL3-7, and HP1-5 show weak molecular lines.
From the location of stars AL3-3 and AL3-7 in the CMD of Fig. \ref{fig:al3-cmd}, they should be AGB stars.
In Fig. \ref{hp15eal}, we show the observed spectrum in the selected
wavelength regions containing CN, OH lines, and the CO bandhead at 15578 {\rm \AA} for stars AL3-3, AL3-7, and HP1-5.
The molecular lines are very shallow, due to a combination of warm temperatures and
low metallicities. Clearly, the CNO abundances derived for these stars are less
reliable than for the cool star NGC6558-F42.
Their CNO abundances are compatible with being close to solar,
but given the shallowness of the lines, it is clear that
the molecular lines are not reliable for abundance measurements.
AL3-6 instead shows very strong CNO lines. Figure \ref{al36cd} indicates
that [C/Fe]= +0.7, 0.8, [N/Fe]=+1.0, [O/Fe]=+0.8 for this star.
We show two different renormalisations to illustrate the difficulty
in analysing this spectrum. Additionally, the computations with two different
carbon abundances illustrate the extreme sensitivity of the lines.
Clearly, however, there is an urgent need to observe this star in the optical
and/or in a more extended wavelength region in the H-band to obtain firm conclusions on the CNO abundances of AL~3.
\begin{figure*}
\includegraphics[angle=0,scale=.95]{39761f10.pdf}
\caption{HP~1-5, AL~3-3, and AL~3-7: Spectrum in selected wavelength regions containing CN, OH lines and the CO bandhead.
Synthetic spectra are computed for the CNO abundances
given in Table \ref{param}.}
\label{hp15eal}
\end{figure*}
\subsection{Errors}
The main uncertainty in the derivation of CNO abundances in AL~3 stems from
the effective temperatures. Adopting a colour excess E(B-V)=0.36
the photometric magnitudes results in temperatures of $\sim$5000 K.
In order to be compatible with the wings of hydrogen lines,
we would have to adopt E(B-V)=0.2, but
the present fit of the CMD confirms the high reddening value.
Hydrogen wings that we adopted are a very good temperature indicator, and the temperatures can be roughly derived for the two
stars showing the line, whereas for the cooler star AL3-6, the H line is not strong, indicating that it is cooler than the other two stars. For this star, we adopted T$_{\rm eff}$ = 4150 K, compatible with its almost absent hydrogen-profile and a temperature compatible with its location in the RGB. With this temperature, we obtained high CNO values and could not converge with lower values.
We note that the CN, CO, and OH dissociation equilibrium interplay
gives a strong constraint on the result, but at the same time, these are very sensitive to abundance variations.
We adopted effective temperature errors of $\pm$250 K for AL3-6, and $\pm$150 K for
AL3-3 and AL3-7, errors of $\pm$0.8 for the gravity,
and $\pm$0.3 for metallicity.
\subsection{Orbits of AL~3}
In order to investigate whether AL~3 belongs to the bulge, thick disc, or halo, given its solar CNO abundances, we carried out the calculations of orbits for the cluster. We employed a Galactic model that includes an exponential disc made by the superposition of three Miyamoto-Nagai potentials (Smith et al. 2015), a dark matter halo with a Navarro-Frenk-White (NFW) density profile (Navarro, Frenk \& White 1997), and a triaxial Ferrers bar. The total mass of the bar is $1.2 \times 10^{10}$ M$_{\odot}$, an angle of $25^{\circ}$ with the Sun-major axis of the bar, and a major axis extension of 3.5 kpc. We assumed three pattern speeds of the bar $\Omega_b= 40$, 45, and 50 km s$^{-1}$ kpc$^{-1}$. We kept the same bar extension, even though we changed the bar pattern speed. Our Galactic model has a circular velocity $V_0=241$ km s$^{-1}$ at $R_0=8.2$ kpc (Bland-Hawthorn \& Gerhard 2016; which is also compatible with the Gravity Collaboration et al. (2019)), and we assumed a peculiar motion of the Sun with respect to the local circular orbit of $(U,V,W)_{\odot}= (11.1, 12.24, 7.25)$ km s$^{-1}$ (Sch{\"o}nrich, Binney \& Dehnen 2010).
To take into account the effects of the observational uncertainties associated with the cluster's parameters, we generated a set of of 1000 initial conditions using a Monte Carlo method considering the errors of distance, heliocentric radial velocity, and absolute proper motion components. With such initial conditions, we integrated the orbits forward for 10 Gyr using the NIGO tool (Rossi 2015).
In Fig. \ref{orbits}, we show the probability density map of the orbits of AL 3 in the $x-y$ and $R-z$ projections co-rotating with the bar, using the two adopted distances, 6.0 kpc (Ortolani et al. 2006) and 6.5 kpc (Baumgardt et al. 2019). The yellow colour displays the space region that the orbits of AL 3 cross more frequently. The black curves are the orbits using the central values of the observational parameters.
In Figure \ref{histograms}, we show histograms relating to the number
of probable orbits as a function of pericentric distance
(r$_{min}$), apocentric distance (r$_{max}$), maximum height above the plane (|Z|$_{max}$), and eccentricity ($e$). In Table \ref{Orb_para}, we give the median values of the orbital parameters, and the errors provided in each column are derived considering the 16th and 84th percentiles of the distribution. The orbital parameters are similar for the three pattern speeds.
From the apocentric distance, we can see that AL~3 is mostly confined within $\sim 4$ kpc, and it has a high probability to belong to the bulge component ($\sim 88 \%$) when the adopted distance is 6.5 kpc. However, a significant fraction of orbits can reach apocentric distances up to $\sim6$ kpc, indicating that AL 3 could belong to the disc component, with a non-negligible probability ($\sim 12 \%$).
In Fig. \ref{histograms}, we show histograms relating
to the number of probable orbits as a function of
pericentric distance (r$_{min}$), apocentric distance
(r$_{max}$), maximum height above the plane
(|Z|$_{max}$), and eccentricity.
If a distance of 6.5 kpc is adopted, then
it is clear that AL~3 is a very central cluster,
with a maximum height reaching at most
$Z\sim1.5$ kpc, and with a high eccentricity orbit.
This is typical of the very old, moderately metal-poor
globular clusters of the Galactic bulge, similar
to NGC~6522, NGC~6558, and HP~1.
On the other hand, if a distance of 6.0 kpc is adopted,
according to Ortolani et al. (2006), the cluster reaches distances farther from the Galactic centre with lower eccentricities (see left and bottom and top panels of Figures \ref{orbits} and \ref{histograms}, respectively).
The probabilities also change to an $\sim$60 \%
probability of belonging to the bulge and $\sim$40\%
to the disc. With this distance, the probability of being part of the disc increases significantly, and maybe this result could be more consistent with the solar CNO abundances.
\begin{figure*}
\includegraphics[angle=0,scale=.65]{39761f11a.pdf}
\includegraphics[angle=0,scale=.65]{39761f11b.pdf}
\caption{Probability density map for $x-y$ and $R-z$ projections of the set of orbits for AL 3 for distances
of 6.0 kpc (left panels) and 6.5 kpc (right panels), using three different values of $\Omega_b= 40, 45,$ and $50$ km s$^{-1}$ kpc$^{-1}$. The orbits are co-rotating with the bar frame. Yellow corresponds to the larger probabilities. The black lines show the orbits using the central values.}
\label{orbits}
\end{figure*}
\begin{figure*}
\includegraphics[angle=0,scale=.8]{39761f12a.pdf}
\includegraphics[angle=0,scale=.8]{39761f12b.pdf}
\caption{Distribution of orbital parameters for AL 3, for distances of 6.0 kpc (top panels) and 6.5 kpc (bottom panels), with pericentric distance $r_{\rm min}$, apocentric distance $r_{\rm max}$, maximum vertical excursion from the
Galactic plane $|z|_{\rm max}$, and eccentricity $e$. The colours show the different pattern speed of the bar, $\Omega_b=40$ (blue), 45 (orange), and 50 (green) km s$^{-1}$ kpc$^{-1}$.}
\label{histograms}
\end{figure*}
\begin{table*}
\begin{center}
\caption{Median orbital parameters and membership probability of AL 3.}
\begin{tabular}{@{}ccccccc@{}}
\hline
$\Omega_b$ & $\langle r_{\rm min}\rangle$ & $\langle r_{\rm max} \rangle$ & $\langle |z|_{\rm max}\rangle $ &$\langle e\rangle$ & $\mathcal{P}_{\rm bulge}$& $\mathcal{P}_{\rm disc}$\\
(km s$^{-1}$ kpc$^{-1}$) & (kpc) & (kpc) & (kpc) & & (\%) & (\%) \\
\hline
\multicolumn{7} {c} {$d_\odot= 6.0$ kpc} \\
\hline
40 & $1.00 ^{+0.43}_{-0.61} $ & $3.94 ^{+0.52}_{-0.66} $ & $ 1.36 ^{+0.19} _{-0.15 }$ & $0.60^{+0.20}_{-0.10} $ & 61.7 & 38.3\\
45 & $ 0.84^{+ 0.47}_{-0.45} $ & $ 3.97^{+0.55 } _{-0.65 }$ & $ 1.36 ^{+0.16}_{-0.13 } $ & $ 0.66 ^{+0.14 } _{-0.15} $ & 60.1 &39.9 \\
50 & $ 0.75^{+0.89 } _{- 0.34}$ & $ 3.98 ^{+0.72 } _{-0.65 }$ & $ 1.31 ^{+0.15 } _{-0.10 } $ & $ 0.68 ^{+ 0.12}_{- 0.20} $ & 60.6 & 39.4\\
\hline
\multicolumn{7} {c} {$d_\odot= 6.5$ kpc} \\
\hline
40 & $ 0.46^{+0.69 }_{-0.36 } $ & $ 3.34^{+0.69 }_{-0.61 } $ & $1.41 ^{+0.21 } _{-0.19 }$ & $ 0.75^{+0.18 }_{-0.19 } $ & 89.2 & 10.8\\
45 & $ 0.47^{+0.50 }_{-0.36 } $ & $ 3.39 ^{+0.76 } _{-0.67 }$ & $ 1.40^{+0.17 }_{-0.17 } $ & $ 0.76^{+0.17 } _{-0.15 } $ & 87.9 & 12.1\\
50 & $ 0.53^{+0.35 } _{-0.40 }$ & $ 3.42 ^{+0.75 } _{-0.73 }$ & $ 1.33^{+0.15 } _{-0.17 } $ & $ 0.74 ^{+0.17 }_{-0.12 } $ & 88.5 & 11.5 \\
\hline
\label{Orb_para}
\end{tabular}
\end{center}
\end{table*}
\section{Conclusions}
We analysed spectra of individual stars of the globular clusters
AL~3, NGC~6558, and HP~1, obtained with the PHOENIX spectrograph at the
Gemini South telescope. With a high spectral resolution of R$\sim$75,000,
in the H band centred at 15555 {\rm \AA}, the wavelength coverage is short (15520 - 15590 {\rm \AA}).
In AL~3,
this limited wavelength range means that it is difficult to use atomic lines
to deduce the stellar parameters effective temperature, gravity, and metallicity.
For this reason,
we obtained the effective temperature from the Hydrogen Brackett 16 line,
gravity from photometric data, and isochrones. The metallicity
[Fe/H]$\sim$-1.3$\pm$0.3 was deduced from the observed CMD given in Ortolani et al. (2006) and is confirmed in the present work
through analysis of the same CMD.
We note that we adopted [Fe/H]=-1.0 for the analysed stars,
due to spectroscopic evidence.
For NGC~6558 and HP~1, the stellar parameters were adopted
from previous analyses from optical spectra (Barbuy et al. 2007, 2018),
and Barbuy et al. (2006, 2016) respectively.
Adopting these stellar parameters, we computed the synthetic spectra in order
to derive the abundances of C, N, and O. Since they vary interdependently,
the fit was done iteratively, where particular attention was given to the
CO bandhead.
The stars analysed in NGC~6558 and HP~1 show typical CNO abundances of
red giants, and confirm previous oxygen abundance derivation.
AL~3 is a more complex case:
two stars analysed in AL~3 show solar CNO abundance ratios, but based on
very shallow lines, and the location of these two stars
in the CMD point to them being AGB stars. The star AL3-6 shows instead very strong
CNO abundances of the order of [C/Fe]=+0.8, [N/Fe]=+1.0, [O/Fe]=+0.8.
A strong CNO abundance indicated by this cooler star shows that AL~3
appears to be an extremely interesting old cluster.
In conclusion, further investigation of this cluster are clearly needed.
We also derived
the cluster's radial velocity, which in turn allowed us to compute the cluster orbits. For
the two AGB stars, we found a higher velocity, whereas
for the third cooler star, the radial velocity is compatible with the value from
Baumgardt et al. (2019).
The orbital behaviour of AL~3 indicates that it is a
typical inner bulge, moderately metal-poor globular cluster if its
distance to the Sun is 6.5 kpc (Baumgardt 2019), but if the
distance is 6.0 kpc (Ortolani et al. 2006 and present work),
there is an increased probability of AL~3 belonging to a diss component.
We derive a very old age for AL~3 of 13.4 Gyr, which would point towards
an old bulge cluster.
Therefore, we conclude that the cluster AL~3 appears to
be an extremely interesting cluster that should be
further investigated through more wavelength-extended spectra,
and including larger samples of member stars.
\begin{acknowledgements}
BB, HE, RR, JM, and EB acknowledge grants from FAPESP, CNPq and CAPES - Financial code 001.
TM acknowledges FAPESP postdoctoral fellowship no.
2018/03480-7.
APV acknowledges FAPESP postdoctoral fellowship no.
2017/15893-1.
Support for MZ is provided by Fondecyt Regular 1191505, the BASAL CATA Center
for Astrophysics and Associated Technologies through grant PFB-06, and ANID -
Millennium Science Initiative Project ICN12\_009, awarded to the Millenium
Institute of Astrophysics (MAS).
S.O. acknowledges the partial support of the research
program DOR1901029, 2019, and the project BIRD191235, 2019 of the University of Padova.
\end{acknowledgements}
|
\section{Introduction}
\label{sec:intro}
One of the central aims in the field of chemical reaction dynamics
is the accurate determination of reaction rates.
This is not just an abstract problem of academic (or basic) concern%
~\cite{truh83, marc92, truh96, KomatsuzakiBerry01a},
but also a practical problem
with many potential applications in complex
reactions~\cite{dougherty06, green11, henkelman14, Ezra2015, hoffmann17}.
The possibility of optimizing reaction rates by external driving
could perhaps take these applications further
in offering improvements to throughput and efficiency.
Multi-barrier reactions were considered early~\cite{mill68, mill79b}
in the context of quantum mechanical tunneling through barriers
at constant energy.
In this \emph{M-problem}, the periodic orbit between the barriers
gives rise to the possibility of an infinite number of
returns to the turning point from which tunneling can proceed.
The return times are usually not commensurate with the period,
either because of coupling to
other degrees of freedom---such as from the bath---or
because of variations in the potential.
In such cases, the
coherence in the returns is altered, changing the nature of the dynamics
in ways that we address in this work.
Problems involving fluctuating~\cite{DG92, Mad92}
or oscillating~\cite{Lehmann00a, Lehmann00b, Lehmann03} barriers
have also received significant attention leading to, for example,
the identification of the phenomenon of resonant activation~\cite{VDB93, Hang94}.
While the approaches originally focused
on the overdamped regime~\cite{DG92, Lehmann00a},
underdamped systems were later examined~\cite{hern01d, Lehmann03}.
For example, mean first passage times have been employed
to calculate (diffusion) rates in spatially periodic multi-barrier potentials.
Therein, various static~\cite{hern02b}
as well as stochastically driven~\cite{hern02c, hern04c} cases
have been characterized primarily through numerical methods.
In the current context, the main challenge in a multi-saddle system
comes from the unpredictability of states in the intermediate basin.
A reactant entering this region may leave
either as a reactant or product depending on the exact
initial conditions~\cite{DeVogelaere1955, mill76b, pollak80, Carpenter2005a}.
Historically, this challenge has been approached by
categorizing reactions into two classes~\cite{mill76b, pech76, pech81}:
\emph{Direct} reactions exhibit a single \ac{TS}.
\emph{Complex} reactions, on the other hand,
have two clearly separated \acp{TS}.
The potential well between those barriers
is assumed to be sufficiently deep
that it gives rise to a long-lived collision complex.
Trajectories passing through one \ac{TS} enter this collision complex and
hence cannot be correlated to trajectories passing through the other \ac{TS}.
In reality, however, a reaction cannot always be uniquely classified.
These concerns were addressed in a unified theory by Miller~\cite{mill76b},
and later refined by Pollak and Pechukas~\cite{pollak79}
so as to address shallow potential wells.
While an important advancement,
this theory still treats the saddle's interactions statistically,
thereby neglecting dynamical effects like resonances.
Moreover, they considered
multi-step reactions in which the positions and heights of the barriers
are time-independent.
Last, there are numerous publications on valley-ridge inflection points,
which are typically described by a normal \ac{TS} followed by
a \emph{shared} one~\cite{Carpenter2005a, Rehbein2011, Collins2013}.
Craven and Hernandez~\cite{hern16d} recently examined
a four-saddle model of ketene isomerization
influenced by a time-dependent external field.
They encountered complicated phase space structures similar to
those in systems with closed reactant or product basins~\cite{hern17e}.
As a result, their analysis was limited to local \acp{DS}
and no reaction rates were calculated.
Moreover, successfully calculating instantaneous rates
based on a globally recrossing-free \ac{DS}
attached to the \ac{NHIM} of a time-dependent multi-saddle system
has---to our knowledge---not yet been reported.
In this paper, we address
the challenge of determining the instantaneous \ac{TS} decay rate
for systems that not only feature multiple barriers along the reaction path,
but that are also time-dependent.
Specifically,
we investigate an open, time-dependent two-saddle model with one \ac{DoF}
as introduced in Sec.~\ref{sec:materials/model}.
The theoretical framework
along with the numerical methods used throughout the paper
are described in Secs.~\ref{sec:materials/ds_nhim}
through~\ref{sec:materials/rates}.
We then discuss the system's phase space structure
under the influence of periodic driving at different frequencies
in Sec.~\ref{sec:ps}.
By leveraging unstable trajectories bound between the saddles,
we can construct an almost globally recrossing-free \ac{DS}
associated with the so-called geometric cross
in Sec.~\ref{sec:cross}.
The \ac{DS} is used
in Sec.~\ref{sec:rates}
to calculate instantaneous decay rates and averaged rate constants
for the underlying activated complex using different methods.
Thus we report a
detailed analysis of the nontrivial phase space structure
of the chemical M-problem
and the calculation of its associated decay rates.
\section{Methods and materials}
\label{sec:materials}
\subsection{Time-dependent two-saddle model system}
\label{sec:materials/model}
In this paper, we investigate
the properties of multi-barrier systems by considering
a 1-\ac{DoF} model potential featuring two Gaussian barriers whose
saddle points are centered at $x = \num{+-1}$.
Initially, both barriers are placed at the same level.
As we are interested in considering the time-dependent case, however,
we drive the barrier's heights $B_\varphi(t)$ sinusoidally in opposite phases.
That is, we use the same amplitude and frequency $\omega$ for both saddles,
but opposite initial phases $\varphi \in \qty{0, \pi}$.
This leads to the potential
\begin{subequations}
\label{eq:materials/model}
\begin{align}
V(x, t) &= B_0(t) \ee^{-(x + 1)^2} + B_\pi(t) \ee^{-(x - 1)^2} \\
\qq*{with}
B_\varphi(t) &= \frac{7}{4} + \frac{1}{4} \sin(\omega t + \varphi)
\MathPeriod
\end{align}
\end{subequations}
The oscillation frequency $\omega$ is a free parameter
that can be varied relative to the other natural time scales of the system
at a given fixed total energy.
At an arbitrary time, one of the barriers will be larger than the other.
For example, when the second barrier is larger,
the potential takes a shape such as that shown at the top
of Fig.~\ref{fig:static_manifolds}(b).
Throughout this paper,
dimensionless units are used to explore the range of phenomena that can arise
from varying the relative timescales of the system and the driving.
\subsection{Dividing surfaces and the NHIM}
\label{sec:materials/ds_nhim}
\begin{figure}
\includegraphics[width=\figurewide]{figure1}
\caption{%
Typical structures of static potentials $V(x)$
and their corresponding phase spaces $x$--$v_x$
with one and two barriers in panels (a) and (b), respectively.
The potential barriers separate reactant (R) and product (P) states.
The two-barrier case features an additional intermediate (I) state
in between.
The maxima are associated with a hyperbolic fixed point (diamonds)
and a dividing surface (indicated by dashed vertical lines) each.
The corresponding manifolds divide the phase space into
four distinct, numbered regions in panel (a)
and six regions in panel (b).
See Secs.~\ref{sec:materials} and~\ref{sec:ps/limits}
for details.}
\label{fig:static_manifolds}
\end{figure}
Reactants and products on the canonical \ac{PES}
are usually separated by a \ac{DS} often associated with a rank-1 saddle,
such as that shown in Fig.~\ref{fig:static_manifolds}(a),
whose unstable direction identifies the reaction coordinate.
The maximum energy configuration along
the corresponding one-dimensional minimum energy
path~\cite{Glasstone1941, Fukui1970, Fukui1976}
$x$ is called the \ac{TS}~\cite{eyring35, wigner37, pech81, truh96}.
In \ac{TST}, the local decay rate is obtained from the flux through the \ac{DS}.
A \ac{DS} is locally recrossing-free if no particle pierces it more than once
before leaving some pre-determined interaction region around the saddle.
In this case, the decay rates (to exit the interaction region)
as determined by the \ac{DS} are locally exact.
We refer to a \ac{DS} as being
globally exact or recrossing-free
if the above is true independent of the choice of the interaction region
as long as said regions do not overlap
with the stable reactant or product regions~\cite{hern19a}.
Using this definition avoids inherent recrossings
caused by reflections in closed reactant or product basins~\cite{hern17e}.
In this paper,
we only address transitions over barriers in series,
and we do not address the parallel case in which a reaction could
access more than one distant barrier.
The scope of the definitions of globally exact and recrossing-free
is therefore limited accordingly.
Every saddle point of a $d$-\ac{DoF} potential
is associated with $(2 d - 1)$-dimensional
stable and unstable manifolds \Ws\ and \Wu\ in phase
space---\cf\ Fig.~\ref{fig:static_manifolds}(a).
Their $(2 d - 2)$-dimensional intersection is called the
\ac{NHIM}~\cite{Lichtenberg82, hern93b, hern93c, Ott2002a, wiggins2013normally}
and describes the unstable subspace of particles trapped on the saddle.
It is noted with a diamond in Fig.~\ref{fig:static_manifolds}(a).
Depending on the convention used,
the time-evolution of the \ac{NHIM} or of a single point on the \ac{NHIM}
forms the \ac{TS} trajectory~\cite{hern19a}.
A ($2 d - 1$)-dimensional \ac{DS} can be constructed
by attaching it to the \ac{NHIM},
which works even for time-dependent driving.
This particular \ac{DS} is locally recrossing-free
as long as particles do not cross far away from the
\ac{NHIM}~\cite{KomatsuzakiBerry99, komatsuzaki06b, komatsuzaki06a, Li09}.
However, it does not have to be globally recrossing-free if, \eg,
the system features multiple barriers or valleys with sharp turns.
In \ac{TST}, one typically uses the saddle
point as the \ac{TS} itself, but the variational principle
clearly suggests that the \ac{DS} can be moved away from it.
In practice, this has led to applications in which the \ac{DS}
is associated with the saddle point.
However, not only is this strong association not necessary,
it is even possible that the \ac{NHIM} (which anchors the \ac{DS})
can be disconnected.
Indeed, we find in the current problem that the optimal \ac{NHIM}
consists of multiple, disjoint sets.
As seen in Sec.~\ref{sec:ps},
such structure can emerge
from the interaction of multiple saddles.
Those parts of the \ac{NHIM} not associated with a saddle
can nevertheless be used to construct a \ac{DS} as before,
and may feature fewer global recrossings compared to
a \ac{DS} associated uniquely with a single saddle point.
The question of whether such a \ac{DS} with fewer or even no recrossings
can be found in a driven multi-saddle system
has led to the central results of this work.
\subsection{Revealing geometric structures}
\label{sec:materials/geometry}
The geometric structure of the phase space can be revealed
using the \ac{LD}~\cite{Mancho2010, Mancho2013, hern15e, hern16a, hern16d}
defined by
\begin{equation}
\label{eq:materials/ld}
\LD(x_0, v_0, t_0) = \int_{t_0 - \tau}^{t_0 + \tau} \dd{t} \norm{v(t)}
\end{equation}
for a given initial position $x_0$, velocity $v_0$, and time $t_0$.
It measures the arc length of a trajectory $x(t)$
in the time interval $t_0 - \tau \le t \le t_0 + \tau$.
A local minimum in the \ac{LD} arises
when the particle covers the minimum distance
in the interval $t_0 - \tau \le t \le t_0 + \tau$.
It consequently remains longer in the interaction region
when integrating forwards or backwards in time,
and thus provides a signature for the presence of a stable or unstable manifold,
respectively.
The \ac{LD} has the advantage that it is conceptually very simple
and that it can be applied to practically any system.
This makes it suitable for a first visual inspection.
As discussed in Ref.~\cite{hern18g}, however,
it features a nontrivial internal structure.
Numerically determining the exact position of stable and unstable manifolds
is therefore difficult.
A numerically simpler scheme is based on
the concept of reactive (and nonreactive) regions
as described in Refs.~\cite{hern18g, hern19a}.
It discriminates initial conditions
by first defining an interaction region in position space
that encompasses the relevant dynamics.
Particles are then propagated forwards and backwards in time
until they leave said interaction region.
In both directions of time,
a particle can end up as either reactant (R) or product (P).
This leads to four possible classifications for a
given initial condition
as shown in Fig.~\ref{fig:static_manifolds}(a)
for each of the four regions:
\begin{enumerate*}[label=\arabic*.]
\item nonreactive reactants \RR{R}{R},
\item nonreactive products \RR{P}{P},
\item reactive reactants \RR{R}{P}, and
\item reactive products \RR{P}{R}.
\end{enumerate*}
Similar concepts have been introduced in, \eg, Ref.~\cite{hern16d}.
Such a classification for the initial conditions has to be extended
to include the consequences of a local minimum between the saddles of the
reacting system of Eq.~\eqref{eq:materials/model}.
A low-energy particle trapped near this local minimum
[\cf\ Fig.~\ref{fig:static_manifolds}(b)]
would lead to diverging computation times
because it may never leave the interaction region.
To solve this problem, an additional termination condition is introduced,
whereby any particle that crosses the potential minimum
a specified number of times \nmc\
is classified as an intermediate (I) particle.
Consequently, up to nine different regions in phase space can be distinguished
for any given value of \nmc.
Using the concept of reactive regions,
stable and unstable manifolds can be revealed
as borders between adjacent regions.
Their closure's intersections form the \ac{NHIM}.
The algorithm used to calculate points of the \ac{NHIM}
is based on the \ac{BCM} introduced in Ref.~\cite{hern18g}.
It starts by defining a quadrangle
with one corner in each of the phase space regions
surrounding the manifold intersection.
The quadrangle is then contracted
by successively determining the region corresponding to an edge's midpoint
and moving the appropriate adjacent corner there.
The \ac{BCM} is therefore effectively composed of
four intertwined classical bisection algorithms.
To reliably identify the initial corners,
we modify the \ac{BCM} slightly as detailed in Appendix~\ref{sec:bcm}.
\subsection{Calculating decay rates}
\label{sec:materials/rates}
The existence of a \ac{NHIM} of codimension 2 and its role in determining
the chemical reaction rate brings an additional concern.
Namely, what is the degree of instability of the \ac{TS} as determined
by the decay of trajectories that start in the proximity of the \ac{NHIM}?
In a time-dependent---\eg\ driven---environment,
this instantaneous decay will be time-dependent as well.
Nevertheless, it can be assigned a single characteristic decay rate constant
when the time-dependence is periodic
by taking the average over the period~\cite{Lehmann00a, hern19e}.
In this paper, we implement three different
methods, summarized in Appendix~\ref{sec:ratemethods},
for calculating decay rates:
\begin{enumerate*}[label=\emph{(\roman*)}]
\item
The ensemble method~\cite{Lehmann00a, hern19e}
yields instantaneous (time-resolved) rates
by propagating a large number of particles.
It is computationally expensive but conceptually simple.
\item
The \ac{LMA}~\cite{hern19e, hern20m}
accelerates the computation of instantaneous rates
by leveraging the linearized dynamics near the \ac{NHIM}.
\item
If only average rate constants are desired,
the Floquet rate method~\cite{hern14f, hern19e} can be used instead
while requiring even less computational resources.
\end{enumerate*}
In the cases resolving the dynamics of
the two-barrier problem of Eq.~\eqref{eq:materials/model},
all three generally converge within reasonable time.
However, they each involve different assumptions which might
have led to different results, and which can provide
complementary interpretations about the underlying dynamics.
As shown in the results sections, all three lead to
decay rates in excellent agreement.
The repetition thus also serves to provide assurance in the
reported values.
\section{Geometric structure of the two-saddle system}
\label{sec:ps}
The phase space structure of
the model system introduced in Sec.~\ref{sec:materials/model}
is highly dependent on the driving frequency $\omega$.
In the following we will give
a qualitative overview of the behavior the system can exhibit.
\subsection{Limiting cases}
\label{sec:ps/limits}
A static two-saddle system
akin to Eq.~\eqref{eq:materials/model} with $\omega = 0$
exhibits the phase space structure shown
in Fig.~\ref{fig:static_manifolds}(b).
The saddle tops are associated with hyperbolic fixed points
whose stable and unstable manifolds each form a cross.
If the first saddle is smaller than the second one,
two of its manifolds constitute a homoclinic orbit.
In this case, the phase space is composed of six regions, namely
\begin{enumerate*}[label=\arabic*.]
\item nonreactive reactants \RR{R}{R},
\item nonreactive products \RR{P}{P},
\item reactive reactants \RR{R}{P},
\item reactive products \RR{P}{R},
\item particles that react over the first saddle
but get reflected at the second \RR{R}[I]{R}, and
\item intermediate particles that are trapped between the saddles \RR{I}{I}.
\end{enumerate*}
Likewise, if the driving frequency is sufficiently large ($\omega \to \infty$),
the particle will effectively see an average static potential
in which it must cross two similar static barriers of equal height.
As the energy is conserved,
once the particle crosses the first barrier,
it necessarily crosses the second barrier.
This results in a phase space structure similar to that for the static case of
Fig.~\ref{fig:static_manifolds}(b)---although with
heteroclinic orbits connecting the hyperbolic fixed points.
In such (effectively) static cases,
it is straightforward to define a global recrossing-free \ac{DS}.
In a 1-\ac{DoF} constant energy system,
if (and only if) a particle crosses the highest saddle,
it has demonstrated to have enough energy to react over all saddles.
Since the largest barrier therefore unambiguously determines
whether a particle reacts or not,
its associated local \ac{DS} becomes the global (recrossing-free) \ac{DS}.
While this holds true for static systems,
dynamically driven systems may exhibit much richer dynamics.
For example, in the case of an alternating pair of dominant barriers,
as in the model of Eq.~\eqref{eq:materials/model},
the naive \ac{DS} jumps discontinuously
from one side to the other twice per period.
As a result there exist reactive trajectories that never cross the \ac{DS}.
Instead, the latter jumps over the former
leading to an inconsistent description of the reaction.
Addressing this issue by defining particles between the local \acp{DS}
as reacting the moment the dominant saddle changes would lead to more problems,
\eg, unphysical Dirac delta peaks in the reaction rate.
To solve this issue, the system has to be treated as a whole.
\subsection{Intermediate driving frequency}
\label{sec:ps/medium}
\begin{figure}
\includegraphics[width=\figurewide]{figure2}
\caption{%
Phase space structure for the time-dependent potential
in Eq.~\eqref{eq:materials/model}
with $\omega = \pi$ at $t_0 = 3 / 2$
as revealed by the \acl{LD} \LD\ given in
Eq.~\eqref{eq:materials/ld} with $\tau = 16$.
Although the two geometric crosses
seen in Fig.~\ref{fig:static_manifolds}(b)
are still present in panel (a),
they now exhibit a complicated substructure
involving a vast number of homoclinic and heteroclinic points
as well as homoclinic and heteroclinic orbits.
The progressively zoomed cutouts
shown in subpanels (b) through (d) exhibit self-recurring structures.
Labels in the bottom right corners
indicate the corresponding enlargement from the previous zoom level.
The potential at time $t_0$
is indicated in the top left inset of panel (a).}
\label{fig:fractal_ld}
\end{figure}
The limiting cases discussed so far result in effectively static systems.
Since we are interested in novel and nontrivial behavior, however,
we will now turn to intermediate driving frequencies.
These can exhibit varying degrees of complexity
as a function of the driving frequency $\omega$.
An example of such non-trivial behavior with a highly complex phase space
is shown in Fig.~\ref{fig:fractal_ld}.
We use the \ac{LD} defined in Sec.~\ref{sec:materials/geometry}
for visualization
since it is very general and requires little knowledge about the system.
The geometric structure
was obtained for the driven potential $V(x, t)$
of Eq.~\eqref{eq:materials/model}
at an intermediate driving frequency $\omega = \pi$.
The general shape of the boundaries
separating reactive and nonreactive regions
(\cf\ Fig.~\ref{fig:static_manifolds}(b))
is still vaguely visible.
However, the precise position
of the crossing points between the stable and unstable manifolds
can no longer be determined.
This family of crossing points
together with the associated stable and unstable manifolds
within their vicinity
appears as a cross that has arisen from all of these geometric considerations.
For simplicity, we define it as a \emph{geometric cross} throughout this work.
Note that this term is not meant to be a precise mathematical
structure but rather an illustrative concept for describing the
complex phase space of the system under study.
The fractal-like geometric crosses
seen in the series of Figs.~\ref{fig:fractal_ld}(a)--(d)
for a finite $\tau$
suggests a fractal structure at all scales
for $\tau \to \infty$.
This structure emerges from particles trapped between the two saddles.
For example,
a reactant can enter the intermediate region over the left saddle,
be reflected multiple times at both saddles,
and finally leave the interaction region over the right saddle as a product.
The number of reflections
is in this case highly dependent on the particular initial conditions
as a result of the system being chaotic.
In turn, this leads to a discontinuity in the \ac{LD},
and eventually to the self-recurring patterns of a fractal structure.
Figure~\ref{fig:fractal_ld} also supports
the observation of geometric structures at a given time
that are thinner near the dominant saddle compared to the lower energy saddle.
In Fig.~\ref{fig:fractal_ld}(a), when the right barrier is dominant,
particles initialized near the higher saddle
start with higher potential energy
and therefore have a lower chance of being reflected.
As a result, fewer of these particles linger in the interaction region and
the phase space structures are thinner.
While this may be an interesting observation,
the geometric cross near the dominant barrier is still highly fractal.
There are no isolated, weakly fractal geometric crosses
that could reasonably be tracked numerically.
Consequently, we cannot make meaningful statements
about a globally recrossing-free \ac{DS}.
\subsection{Slow driving frequency}
\label{sec:ps/slow}
\begin{figure*}
\includegraphics[width=\textwidth]{figure3}
\caption{%
Reactive regions for potential~\eqref{eq:materials/model}
with $\omega = \pi / 10$ and $\nmc = 8$
as a function of time $t_0$ (\cf\ inset labels).
The color (or shaded) legend at the top
indicates whether a particle starts or ends
in the reactant (R), intermediate (I), or product (P) state.
The position of the point on the \acs{NHIM}
associated with the primary geometric cross (black dot)
is tracked across a full period
(\acs{TS} trajectory, black dashed line).
The insets in the upper right corner of the top row panels and
the lower right corner of the bottom row panels illustrate
the potential at the corresponding initial time $t_0$.}
\label{fig:reactive_regions}
\end{figure*}
The previous sections have shown
the range of complexity the model system~\eqref{eq:materials/model} can exhibit.
We now need to move
from the aesthetically pleasing structures of Fig.~\ref{fig:fractal_ld}
to a more rigorous identification of the globally recrossing-free \ac{DS}.
To do so, we switch to a lower driving frequency $\omega = \pi / 10$
which is simpler to analyze but still exhibits nontrivial behavior.
Additionally, we employ the concept of reactive regions
as described in Sec.~\ref{sec:materials/geometry}
instead of the \ac{LD}.
The partitioning of the phase space into nine distinct regions
allows to make quantitative assessments more easily.
Application of this analysis to $V(x, t)$ with $\omega = \pi / 10$
leads to the time-dependent regions
shown in Fig.~\ref{fig:reactive_regions}.
Although fractal-like structures still remain,
they are less pronounced
and mostly concentrated around whichever saddle happens to be the
lower saddle at a given instance.
The higher saddle, on the other hand, is accompanied by a
clearly visible geometric cross,
where the four regions known from the one-saddle case
(no intermediate states) meet.
The regions are also arranged in the same way:
\RR{R}{P} on top, \RR{R}{R} to the left,
\RR{P}{P} to the right, and \RR{P}{R} below.
In the following,
we refer to this geometric cross as the \emph{primary geometric cross}.
\section{The geometric cross}
\label{sec:cross}
We will now analyze the primary geometric cross
and its associated \ac{TS} trajectory in more detail.
Its time-dependent position can be tracked precisely over a full period
using the algorithm described
in Sec.~\ref{sec:materials/geometry} and Appendix~\ref{sec:bcm}.
The result is marked as a series of black dots
in Fig.~\ref{fig:reactive_regions}.
The corresponding trajectory connecting those dots
is indicated as a black dashed line.
\subsection{Global transition state trajectory}
\label{sec:cross/ds}
The primary geometric cross associated with the
instantaneously higher saddle
in Fig.~\ref{fig:reactive_regions}
remains on the barrier nearly
as long as the barrier remains dominant.
However, when the barriers' heights approach each other,
the geometric cross quickly moves from one barrier to the other
in the following way.
The geometric cross begins to rapidly accelerate towards the middle ($x = 0$).
It crosses the local potential minimum
exactly when both saddle points are level (\eg\ $t_0 = 10$)
and continues in the same direction
until it is located near the now higher saddle (\eg\ $t_0 = 15$).
The reverse happens in the following half period,
thereby forming a closed trajectory
with the same period $T_1 = 20$ as the potential $V(x, t)$
(dashed line in Fig.~\ref{fig:reactive_regions}).
\begin{figure}
\includegraphics[width=\figurewide]{figure4}
\caption{%
Examples of typical periodic trajectories $x(t)$
with periods $T_2 = 40$ (darker gray or blue)
and $T_3 = 60$ (lighter gray or orange)
for potential~\eqref{eq:materials/model} with $\omega = \pi / 10$.
The primary \acs{TS} trajectory
from Fig.~\ref{fig:reactive_regions}
with period $T_1 = 20$ (black dashed line)
is shown for comparison.}
\label{fig:periodic}
\end{figure}
This primary geometric cross
marks the position of a particle on an unstable periodic trajectory
trapped in the interaction region.
The particle on this trajectory
oscillates between the saddles with a period of $T_1 = 20$
so that it is always located near the higher saddle.
Many other unstable periodic trajectories associated
with geometric crosses---\ie\ hyperbolic fixed points---in phase space
have also been found for particles in the interaction region.
Two such trajectories are shown in Fig.~\ref{fig:periodic}.
They are typical of an increased degree of structure
relative to the primary \ac{TS} trajectory.
All such trajectories together
form the system's disjoint, fractal-like time-dependent \ac{NHIM}.
In contrast to the primary \ac{TS} trajectory, however,
all other trajectories have periods larger than $T_1$.
The only exceptions to this observation
are the local \ac{TS} trajectories in the vicinities of the saddle maxima,
which are also part of the global \ac{NHIM}.
These trajectories have the same period $T_1$ as the potential by construction.
\begin{figure}
\includegraphics[width=\figurewide]{figure5}
\caption{%
(a)~Position \xDS\ and
(b)~percentage of trajectories with recrossings or classification errors
as a function of time $t_0$ for various choices of their assignment
as reactive or nonreactive.
The assignment is performed according to the crossing of a specified
\acs{DS} associated with either
the local (solid) or global (thick dashed) \acs{TS} trajectories,
the instantaneous potential minimum between the saddles (dotted),
or the discontinuous \acs{TS} trajectory jumping
between the local ones (dash-dotted).
For every $t_0$, an ensemble of \num{e6} particles with
uniformly distributed velocities $\num{1.7} \le v_x(t_0) \le \num{2.1}$
is initialized at $x(t_0) = \num{-3}$
and propagated for $\Delta t = 80$ time units.
The attached \acs{DS} is parallel to $v_x$ for all times.
The right local \acs{DS} is recrossing-free by construction
since particles are started solely to the left of the saddles
and is therefore not shown in (b).
This specific choice for the ensemble
is also the reason for the graph's asymmetry.}
\label{fig:ds_comparison}
\end{figure}
The complex dynamics that arises from the moving barriers is also affected
by the degree to which a given trajectory is decoupled from the barriers
as it traverses the well between them.
Although the energy of the minimum stays roughly constant, as indicated
in Fig.~\ref{fig:reactive_regions}, its position moves back and forth
between the barriers as shown in Fig.~\ref{fig:ds_comparison}(a)
so that it is always closer to the lower barrier.
This is a manifestation of the Hammond postulate~\cite{hammond55}
but now applied to the negative potential.
When the first (left) barrier dominates the dynamics,
a locally nonrecrossing \ac{DS} associated with it identifies
the trajectories with sufficient energy to cross it, and those continue
unabated across the well and the second barrier.
As a consequence, a \ac{DS} located at the well identifies
the reactive trajectories equally well (or badly) in this regime.
This is indicated in Fig.~\ref{fig:ds_comparison}(b)
by good identification when the activated particle continue past the second
barrier
or disagreement when misidentification
of trajectories begin to be reflected across both.
When the second (right) barrier dominates the dynamics,
identifying reactive trajectories at a \ac{DS} at the distant well
allows the evolving trajectories to be reflected by the barrier leading to
recrossings.
Thus we find that the reaction dynamics in between the
barriers---\eg\ at the well---does not go through a single
identifiable doorway.
In turn, this points to the need
for describing the dynamics---even in a local sense---through
a geometric picture that spans the two barriers,
\eg\ the global NHIM.
Meanwhile, since the \ac{NHIM} now consists of more than one trajectory,
it is not necessarily obvious which of these is most suited
for attaching a \emph{global} \ac{DS}.
We can, however, set conditions
the global \ac{TS} trajectory should fulfill.
First, for symmetry reasons,
the trajectory should have the same period $T_1 = 20$ as the potential.
Second, the global \ac{TS} trajectory should approach
the higher saddle's local \ac{TS} trajectory
in the (quasi-)static limit $\omega \to 0$ (\cf\ Sec.~\ref{sec:ps}).
The only trajectory found that matches both criteria
is the primary \ac{TS} trajectory introduced previously,
see Fig.~\ref{fig:ds_comparison}(a).
The large featureless regions surrounding the trajectory in phase space
additionally suggest that it affects
a significant fraction of the system's dynamics.
We will therefore refer to this trajectory as the global \ac{TS} trajectory.
\subsection{Comparison of dividing surfaces}
\label{sec:cross/quality}
The next task is to determine the
degree to which the global \ac{TS} trajectory
gives rise to a recrossing-free \ac{DS}.
Numerically, this can be tested by attaching
a \ac{DS} to it as specified in the caption of Fig.~\ref{fig:ds_comparison},
propagating an ensemble initialized near it,
and recording the number and direction of \ac{DS} crossings (or not)
that transpire thereafter in the propagated trajectories.
For simplicity, we consider only
the most challenging cases in which
the ensemble's initial energies are chosen to be
between the saddles' minimum and maximum heights.
The usual error in the \ac{DS} is signaled by the
existence of more than one crossing for the trajectories,
and the fraction of such recrossings is used below
as a measure for the \ac{DS}'s quality.
For simplicity, we limit ourselves to \acp{DS} defined by
$x = \xDS(t)$, \ie, parallel to $v_x$.
The results of this analysis are shown in Fig.~\ref{fig:ds_comparison}(b).
As can be seen, the global \ac{DS}
associated with the time-dependent geometric cross
features error rates that are significantly reduced
compared to the local \ac{DS} fixed at the left barrier.
One possible \ac{DS} could be constructed by placing it at
the instantaneous potential minimum---shown as the dotted curve
in Fig.~\ref{fig:ds_comparison}(a).
It would be expected to be ineffective given that rates are usually
determined by rate-limiting barriers, not valleys, in between
reactants and products.
Indeed, the recrossing errors found for this
\ac{DS}, shown in Fig.~\ref{fig:ds_comparison}(b),
were high and even worse than those from the use of the \ac{DS}
fixed at the left barrier.
But the highest error rate comes from the naive attempt to
treat the \ac{DS} associated with the instantaneously higher saddle point
as the global one.
In this case, errors can arise from events beyond the recrossing of the \ac{DS}.
That is, there now exists the possibility that
the discontinuous instantaneous \ac{DS} can jump over the trajectory.
It is the combination of recrossing and classification errors that leads
to the jagged and large deviations in the \si{\percent} error
seen for the naive discontinuous \ac{DS}.
As discussed in Appendix~\ref{sec:discont_classerr},
this can lead to
\emph{misclassification} of the reactivity of the trajectory.
Finally, we consider a local \ac{DS} fixed at the right barrier.
This choice would lead to no recrossing or classification errors
for trajectories moving in the forward direction (from reactants to products).
However, it would fail badly for trajectories moving in the backward direction
by symmetry with our finding for the forward trajectories crossing the
\ac{DS} at the left barrier.
Thus, the use of the global \ac{DS} associated with the \ac{TS} trajectory
best captures the time-dependent geometry of the reaction.
\begin{figure}
\includegraphics[width=\figurewide]{figure6}
\caption{%
Reactive regions analogous to Fig.~\ref{fig:reactive_regions}
at $t_0 = \num{0.25}$
with the addition of the \acs{DS} (solid black line).
(a)~The simplest choice of a \acs{DS} parallel to $v_x$
leads inevitably to recrossings as indicated.
(b)~The immediate region of the \acs{TS} trajectory
[shown as a dot in panel (a)],
is enlarged, as indicated, to reveal
the geometry of the ensemble (densely dotted)
used in the rate calculations of Sec.~\ref{sec:rates}.
The ensemble is sampled equidistantly
on a line parallel to the unstable manifold \Wu\
at distance $x - \xDS = \num{-1e-3}$
from the \acs{TS} trajectory at $\xDS(t_0)$.
The manifolds \Wu\ and \Ws\ are given
by the boundaries between reactive and nonreactive regions.
The differential of \Wu\ is indicated by
the sides $\Delta x$ and $\Delta v_x^\mathrm{u}$
of the slope.}
\label{fig:recross_regions}
\end{figure}
Although this \ac{TS}-trajectory \ac{DS}
is much better than any alternatives considered so far,
it still exhibits an amount of recrossings
that cannot be explained by numerical imprecision alone.
Instead,
recrossings are caused by the fractal-like phase space structure of the system:
Figure~\ref{fig:recross_regions}(a), for example,
shows the phase space structure at $t_0 = \num{0.25}$.
We can see a relatively large patch of nonreactive reactants
(labeled \RR{R}{R}, light blue)
to the right of the \ac{DS}.
Particles in this patch leave the interaction region to the reactant (left) side
forwards and backwards in time.
Consequently, they need to cross the \ac{DS} at least twice,
which counts as a recrossing.
An analogous argument can be applied
to \RR{P}{P} regions to the left of the \ac{DS}.
The fractal-like phase space structure thus leads to
numerous problematic patches of vastly different sizes,
hence the two distinct peaks in Fig.~\ref{fig:ds_comparison}(b).
A totally recrossing-free \ac{DS}, by contrast,
would necessarily have to divide the phase space
such that \RR{R}{R} regions are always on the reactant side,
and \RR{P}{P} regions always on the product side.
This is not possible with a planar \ac{DS} of any orientation
due to the system's fractal-like nature.
A globally recrossing-free \ac{DS}---if it exists---would have to
curve as time passes
because the entirety of the phase space between the saddle points
has a clockwise rotating structure
(including periodic trajectories on the \ac{NHIM},
\cf\ Figs.~\ref{fig:reactive_regions} and~\ref{fig:periodic}).
The periodicity of this system
would then result in a fractal, spiral-like \ac{DS}.
Thus the next step in generalizing this theory would require
the identification of a non-planar \ac{DS} anchored at the
\ac{TS} trajectory which we leave as a challenge to future work.
\section{Reaction rate constants}
\label{sec:rates}
\begin{figure}
\includegraphics[width=\figurewide]{figure7}
\caption{%
Various instantaneous reaction rates $k$ parameterized by time $t$
and associated with the global \acs{TS} trajectory
of the potential defined in Eq.~\eqref{eq:materials/model}
with $\omega = \pi / 10$, and driving period $T_1 = \num{20}$.
The ensemble rate $\kEnse(t)$ (blue solid)
is obtained by propagation of 20 ensembles of \num{e5} particles each
as described in Appendix~\ref{sec:ratemethods/ensemble}.
The manifold rate $\kMani(t)$ (orange dashed)
is obtained using Eq.~\eqref{eq:ratemethods/k_m}.
The Floquet rate constant \kFloq\ (green dotted)
is obtained from Eq.~\eqref{eq:ratemethods/k_f}.
The mean rates, \kEnseAvrg\ and \kManiAvrg\
(thick horizontal lines)
are averaged over the period
as defined in Eq.~\eqref{eq:materials/k_e_avrg}.
For comparison, the Floquet rate constant \kFloqLoc\ (red dash-dotted)
of a single barrier's local \acs{TS} trajectory is also shown.}
\label{fig:rate_constant}
\end{figure}
We can now calculate decay rate constants $k$
for the activated complex
using the global \ac{TS}-trajectory \ac{DS} defined in Sec.~\ref{sec:cross}.
The patches leading to recrossing
(\cf\ Sec.~\ref{sec:cross/quality})
are unproblematic in this case
because the ensemble was selected close to the \ac{NHIM} and
thereby necessarily far from them,
as can be seen in Fig.~\ref{fig:recross_regions}.
The few recrossings that do still occur
are artifacts from the numerical error in the propagation,
and are sufficiently small in number that their effect is smaller than the
numerical precision of the calculation.
In the following,
we consider the three different possible approaches
defined in Sec.~\ref{sec:materials/rates}
in order to demonstrate their equivalence in multi-saddle systems.
An example for the initial reactant ensemble and manifold geometry
can be found in Fig.~\ref{fig:recross_regions}(b).
While the application of the \ac{LMA} and the Floquet method
to our model system are straightforward,
applying the ensemble method poses a challenge:
A finite ensemble of reactants---\eg\ of size \num{e5} as implemented here---%
will mostly react within a short time%
---\eg\ \num{2} to \num{5} units of time in the case shown here---%
compared to the period of driving $T_1 = 20$.
Resolving the whole period with a single ensemble
would therefore require an exponentially growing ensemble size.
This would not be numerically feasible.
Instead, multiple ensembles have been started at
times $t_j$ incremented at equal intervals
$\Delta t$, and set to $1$ in the current case.
An instantaneous rate $\kEnse(t; t_j)$ can be obtained for each ensemble $j$.
The instantaneous rate for the whole period
$\kEnse(t)$---independent of $t_j$---can then be recovered
by concatenating the segments of each $\kEnse(t; t_j)$ for $t$
from $t_j$ to $t_{j + 1}$.
Compared to the first option,
this approach scales linearly with the system's period instead of exponentially
leading to vastly decreased computing times and increased numerical stability.
The results are shown in Fig.~\ref{fig:rate_constant}.
Since $\kEnse(t)$ and $\kMani(t)$ are not constant in time,
we additionally show the average
\begin{equation}
\label{eq:materials/k_e_avrg}
\kAvrg = \frac{1}{T_1} \int_0^{T_1} \dd{t} k(t)
\end{equation}
over one period $T_1$ of the \ac{TS} trajectory.
Figure~\ref{fig:rate_constant} shows a large variation
in the instantaneous reaction rate
in the interval $\num{0.13} \lesssim k(t) \lesssim \num{3.85}$.
Two features are distinctive.
First, the rate $k(t)$ shows mostly flat plateaus
during the time the \ac{TS} trajectory is located near a saddle.
This can be explained using Fig.~\ref{fig:reactive_regions}:
Although the phase space structure as a whole is undergoing
significant changes during, \eg, $\num{2.5} \le t \le \num{7.5}$,
the local vicinity around the \ac{TS} trajectory stays almost unaffected.
Second, there are deep dips in $k(t)$
while the \ac{TS} trajectory moves between the saddles
(as seen at around $t \in \qty{10 j \mid j \in \mathbb{Z}}$).
As can be seen in Figs.~\ref{fig:reactive_regions} and~\ref{fig:recross_regions},
these times are characterized by much more shallow geometric crosses
with a low difference in the slopes of the stable and unstable manifolds.
This effect is particularly apparent in Fig.~\ref{fig:recross_regions}(b).
The same observations can also be interpreted another way.
By comparing Fig.~\ref{fig:rate_constant} to Fig.~\ref{fig:ds_comparison}(a),
we can see a clear correlation
between the velocity of the \ac{TS} trajectory and the instantaneous rate $k(t)$:
the faster the \ac{TS} trajectory moves, the lower the rate drops.
\begin{table}
\caption{%
Values of the reaction rate constants
discussed in Fig.~\ref{fig:rate_constant}.
The averaged ensemble and \acs{LMA} rate constants
\kEnseAvrg\ and \kManiAvrg\
match the Floquet rate constants
to within less than \SI{0.1}{\percent}.
The local (single barrier) Floquet rate constant, however,
differs by \SI[retain-explicit-plus]{+27}{\percent}.}
\label{tab:rate_constant}
\begin{tabular}
{l l S[round-mode=figures, round-precision=5, table-format=2.4]}
\toprule
Method & Symbol & {Value} \\
\midrule
ensemble propagation & \kEnseAvrg & 2.7055 \\
manifold geometry (\acs{LMA}) & \kManiAvrg & 2.7062045871555145 \\
Floquet stability analysis & \kFloq & 2.7036088391591266 \\
\midrule
local saddle (Floquet analysis) & \kFloqLoc & 3.43837170284471 \\
\bottomrule
\end{tabular}
\end{table}
As can be seen
in Fig.~\ref{fig:rate_constant} and Table~\ref{tab:rate_constant},
\kEnseAvrg, \kManiAvrg, and \kFloq\ are in excellent agreement,
which illustrates the equivalence of all three methods.
The local Floquet rate constant \kFloqLoc\ of a single saddle,
on the other hand, differs significantly from \kFloq,
even though both saddles are identical.
While the local rate constant can thus be used
as an upper limit for the overall rate constant,
there is no straightforward way to derive a global rate constant from it.
Thus global methods employing the full \ac{TS} trajectory
are necessary if accurate rates for multi-saddle systems are desired.
All three methods in this section satisfy this requirement,
and are consequently in agreement.
\section{Conclusion and outlook}
\label{sec:conclusion}
In this paper, we have
fully characterized the reaction geometry and
determined the associated decay rates
in oscillatory (or time-dependent) two-saddle systems
in the framework of \ac{TST}.
The first set of central results of this work
lies in revealing the phase space structure of
the two-saddle model system.
While the structure of stable and unstable manifolds
is straightforward for (quasi-)static ($\omega \to 0$)
or very fast oscillating ($\omega \to \infty$) systems,
intermediate frequencies lead to a fractal-like phase space.
In this case,
the existence of a completely recrossing-free \ac{DS} is questionable.
For lower oscillation frequencies, however,
an isolated geometric cross
with negligible substructure---%
referred to as the \emph{primary} geometric cross---emerges.
This structure is part of the \ac{NHIM}
and can now be referred to as a \emph{global} \ac{TS} trajectory
in contrast to the \emph{local} \ac{TS} trajectories associated
with the respective single barriers.
The global \ac{TS} trajectory
oscillates between the two local \ac{TS} trajectories
with the same frequency as the potential
and can be used to attach a mostly recrossing-free \ac{DS}.
The second set of central results of this work
involves the determination of the decay rate constants of
the oscillatory (or time-dependent) two-saddle model system.
Using the \ac{DS} acquired in the first part,
we can propagate ensembles of particles,
record a time-dependent reactant population,
and finally derive an instantaneous reaction rate parameterized by time
according to Ref.~\cite{hern19e}.
Alternatively, the same result can be achieved
purely by analyzing the time-dependent phase space geometry.
For comparison,
a rate constant can be obtained from the global \ac{TS} trajectory
by means of Floquet stability analysis.
This method is in excellent agreement
with the average of each instantaneous rate.
While these results mark an important step
in the treatment of time-dependent multi-saddle systems,
many questions still remain unanswered:
First, we restricted ourselves to two-saddle systems with one \ac{DoF}.
To be applicable to real-world systems, however,
the methods presented here
will have to be generalized to at least more \acp{DoF}
because few chemical reactions can be treated exactly, or nearly so,
when reduced to just one coordinate.
Second, it will be important
to investigate the influence of minor manifold crossings
on the rate constant.
This is particularly necessary for
cases of time-dependent barriers found here in which
there is no longer an equivalent to the primary geometric cross.
This is even more challenging when the alternation between barriers
is driven at high frequencies (\cf\ Fig.~\ref{fig:fractal_ld}).
The applicability of our results to real-world systems also remains to
be demonstrated.
It remains unclear whether it is possible
to treat systems without a primary geometric cross.
That chemical reactions can be represented with potentials
exhibiting the challenges discussed here, is illustrated by
the isomerization of ketene via formylmethylene and oxirene
which has been modeled via a four-saddle
potential~\cite{gezelter1995, hern13c, wiggins14b, hern14e, hern16d}.
The isomerization reaction of triangular KCN
via a metastable linear K-CN configuration
can similarly be described by a two-saddle system~\cite{borondo13a, borondo18a}.
Thus the analysis of time-dependent driven potentials resolved here,
when applied to these and other chemical reactions,
should provide new predictive rates for driven chemical reactions of interest.
\begin{acknowledgments}
Useful discussions with Robin Bardakcioglu are gratefully acknowledged.
The German portion of this collaborative work was partially supported
by the Deutsche Forschungsgemeinschaft (DFG) through Grant No.~MA1639/14-1.
The US portion was partially supported
by the National Science Foundation (NSF) through Grant No.~CHE 1700749.
MF is grateful for support from
the Landesgraduiertenförderung of the Land Baden-Württemberg.
This collaboration has also benefited from support
by the European Union's Horizon 2020 Research and Innovation Program
under the Marie Skłodowska-Curie Grant Agreement No.~734557.
\end{acknowledgments}
|
\section{\textbf {Introduction}}
The value distribution of $L$-functions plays an important role in analytic number theory. A celebrated theorem of Selberg \cite{SCR} states that $\log \zeta(\frac{1}{2}+it)$ is approximately Gaussian distributed, namely,
\begin{align}\label{SCLT}
\lim_{T\rightarrow \infty}\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\frac{\log |\zeta(\tfrac{1}{2}+it)|}{\sqrt{\tfrac{1}{2}\log\log T}} > V}
=\int_{V}^\infty e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}
\end{align}
for any fixed $V$. Throughout this paper, $\operatorname{meas}(\cdot)$ stands for the one-dimensional Lebesgue measure.
Selberg also stated that primitive $L$-functions in the Selberg class are ``statistically independent"
without precise description for the independence (see \cite[p.7]{S1992}). Later,
Bombieri and Hejhal \cite{BH1995} indeed showed that normalized values of $L$-functions satisfying certain assumptions
behave like independent Gaussian variables, precisely,
\begin{align}\label{BHJVD}
& \lim_{T\rightarrow \infty}\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{
\frac{\log |L_j(\frac{1}{2}+it)|}{\sqrt{\frac{n_{j}}{2}\log\log T}} > V_{j} \text{ for } j = 1, \dots, r}
=\prod_{j=1}^r\int_{V_{j}}^\infty e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}
\end{align}
for any fixed $V_{1}, \dots, V_{r} \in \mathbb{R}$ under certain conditions of $\{L_j\}_{j=1}^r$,
where $n_{j}$ are certain positive integers determined by $L_{j}$.
It is natural to ask if \eqref{SCLT} and \eqref{BHJVD} still hold for a larger range of $V$.
In fact, if estimates similar to \eqref{SCLT} holds for a larger range of $V$,
then one could expect the growth of moments of $\zeta(\frac{1}{2} +it)$ as
\begin{align*}
\int_T^{2T}|\zeta(\tfrac{1}{2}+it)|^{2k} dt
\asymp_{k} T (\log T)^{k^2}.
\end{align*}
This idea has been used by Soundararajan \cite{SM2009}, where he proved the upper bound for all $k \geq 0$ conditionally on the Riemann Hypothesis, up to a factor of $(\log T)^\epsilon$. However, the asymptotic \eqref{SCLT} should not hold for arbitrarily large $V$, as observed by Radziwi\l\l\mbox{} \cite{Ra2011}, where he proved that \eqref{SCLT} holds for $V=o((\log\log T)^{1/10-\epsilon})$ and conjectured that \eqref{SCLT} holds for $V=o(\sqrt{\log\log T})$.
In this paper, we derivate large deviations results for joint value distribution of $L$-functions and give some further evidence of the ``statistical independence" of $L$-functions in a general class, including the Riemann zeta function, Dirichlet $L$-functions, and $L$-functions attached to holomorphic and Maa\ss \ cusp forms. We obtain the expected upper bound \eqref{BHJVD} for
$V = o\l((\log\log T)^{1/6}\r)$ (see Therorem \ref{Main_Thm_LD_JVD}).
We also obtain weaker bounds when $V\ll \sqrt{\log\log T}$ (see Theorem \ref{GJu} and Theorem \ref{RH_LD_JVD}). As an application, we have the following result.
\begin{theorem}\label{ThmDirichlet}
Let $\{\chi_j\}_{j=1}^r$ be distinct primitive Dirichlet characters (where the case $\chi_{i} = 1$ for some $1 \leq i \leq r$ is allowed).
Then there exists some positive constant $B$ depending on $\{\chi_j\}_{j=1}^r$ such that for any fixed sufficiently small positive real number $k$,
\begin{align}
\label{EMVTRD}
\int_{T}^{2T}\l( \min_{1\leq j \leq r} |L(\tfrac{1}{2}+it, \chi_j)| \r)^{2k}dt
\ll T(\log{T})^{k^2/r + Bk^{3}},
\end{align}
\begin{align}
\label{EMVTRD2}
\int_{T}^{2T}\l( \max_{1\leq j \leq r} |L(\tfrac{1}{2}+it, \chi_j)| \r)^{-2k}dt
\gg T(\log{T})^{k^2/r - Bk^{3}},
\end{align}
\begin{align}
T(\log{T})^{k^2/r + Bk^{3}}
\ll \int_{T}^{2T}\exp\l( 2k\min_{1\leq j \leq r} \Im\log L(\tfrac{1}{2}+it, \chi_j)\r)dt
\ll T(\log{T})^{k^2/r + Bk^{3}},
\end{align}
and
\begin{align}
T(\log{T})^{k^2/r - B k^{3}}
\ll \int_{T}^{2T}\exp\l( -2k\max_{1\leq j \leq r} \Im\log L(\tfrac{1}{2}+it, \chi_j)\r)dt
\ll T(\log{T})^{k^2/r + B k^{3}}.
\end{align}
The above implicit constants depend on $\{\chi_j\}_{j=1}^r$.
If we assume the Riemann Hypothesis for $L(s, \chi_j)$ for all $j$, then we have for any $k > 0$ and $\varepsilon > 0$,
\begin{align}
\int_{T}^{2T}\l( \min_{1\leq j \leq r}|L(\tfrac{1}{2}+it, \chi_j)| \r)^{2k}dt
\ll T(\log{T})^{k^2/r + \varepsilon},
\end{align}
\begin{align}
\int_{T}^{2T}\l( \max_{1\leq j \leq r} |L(\tfrac{1}{2}+it, \chi_j)| \r)^{-2k}dt
\gg T(\log{T})^{k^2/r - \varepsilon},
\end{align}
\begin{align} \label{EMVTRDRH3}
T(\log{T})^{k^2/r - \varepsilon}
\ll \int_{T}^{2T}\exp\l( 2k\min_{1\leq j \leq r} \Im\log L(\tfrac{1}{2}+it, \chi_j)\r)dt
\ll T(\log{T})^{k^2/r + \varepsilon},
\end{align}
and
\begin{align} \label{EMVTRDRH4}
T(\log{T})^{k^2/r - \varepsilon}
\ll \int_{T}^{2T}\exp\l( -2k\max_{1\leq j \leq r} \Im\log L(\tfrac{1}{2}+it, \chi_j)\r)dt
\ll T(\log{T})^{k^2/r + \varepsilon}.
\end{align}
Here, the above implicit constants depend on $\{\chi_j\}_{j=1}^r$, $k$, and $\varepsilon$.
\end{theorem}
\begin{remark}
Theorem \ref{ThmDirichlet} is a special case for Theorem \ref{New_MVT} and Theorem \ref{New_MVT_RH}, which are applicable to more general $L$-functions such as $L$-functions associated with $GL(2)$ cusp forms.
\end{remark}
\begin{remark}
It is now known that for $0\leq k \leq 2$
\begin{align*}
\int_{T}^{2T}|\zeta(\tfrac{1}{2}+it)|^{2k}dt
\asymp_{k} T(\log{T})^{k^2},
\end{align*}
by the works of Heap-Radziwi\l\l-Soundararajan \cite{HRS2019} and Heap-Soundararajan \cite{HeapSound},
so the unconditional result \eqref{EMVTRD} is nontrivial when $k$ is sufficiently small and $0 < k < B^{-1}/2$,
if one of the $L$-functions is chosen to be the Riemann zeta-function and $r\geq 2$.
In particular, when $k$ is sufficiently small and $0 < k < B^{-1}/2$, we obtain the relation
\begin{align*}
\int_{T}^{2T}\l( \min\l\{ |\zeta(\tfrac{1}{2} + it)|, |L(\tfrac{1}{2} + it, \chi)| \r\} \r)^{2k}dt
= o\l( \int_{T}^{2T}|\zeta(\tfrac{1}{2} + it)|^{2k}dt \r) \quad \text{as $T \rightarrow + \infty$}.
\end{align*}
This relation also holds for any $k > 0$ if the Riemann Hypothesis for $\zeta(s)$ and $L(s, \chi)$ is true.
\end{remark}
\begin{remark}
The conditional upper bound for the $2k$-th moment for $\Im \log \zeta(\frac{1}{2}+it)$ was recently proved by Najnudel \cite{Naj}
using a different argument, but the lower bound remained unproved.
Estimates \eqref{EMVTRDRH3}, \eqref{EMVTRDRH4} give the lower bound and also recover the original Najnudel's result.
\end{remark}
\begin{remark}
Gonek \cite{G1989} showed a lower bound of the negative moment of the Riemann zeta-function under the Riemann Hypothesis.
On the other hand, there were no unconditional results for the lower bound.
Estimate \eqref{EMVTRD2} gives an unconditional result for the negative moment for sufficiently small $k$.
\end{remark}
\subsection{Proof sketch and outline of the paper}
Our proof starts with approximate formulae for $L$-functions that combines the formula of Selberg \cite{SCR} and the hybrid formula of Gonek-Hughes-Keating \cite{GHK2007} developed for the study of $\zeta(s)$. These formulae essentially contain a Dirichlet polynomial involving primes and explicit expressions involving zeros of these $L$-functions. After examining the contribution from the zeros, the joint value distribution of $L$-functions can be reduce to the joint value distribution of Dirichlet polynomials. To make clear of the scope of our methods, we work within a general class of $L$-functions satisfying suitable assumptions. These assumptions can be verified for Dirichlet $L$-functions and $L$-functions associated with $GL(2)$ cusp forms. Note that we do not assume the Ramanujan conjecture, but Hypothesis H of Rudnick-Sarnak \cite{RS1996}. The weaker assumption makes our proof different in several aspects (see Section \ref{appL} and \ref{dpoly}).
The rest of the paper is organized as follows: In Section \ref{General} we give the definition of the class of $L$-functions and state our results for these $L$-functions and their corresponding Dirichlet polynomials.
In Section \ref{appL} we give the approximate formulas for $L$-functions in our class.
In Section \ref{dpoly} we give the proofs of results for Dirichlet polynomials.
In Section \ref{uncond} and Section \ref{cond} we give the proof for results for $L$-functions
without and with the Riemann Hypothesis respectively. In Section \ref{finalremarks}, we give some remarks on other applications of our method.
\section*{acknowledgement}
The authors would like to thank Winston Heap for his comments and remarks on earlier version of the paper.
The authors also thank Kenta Endo and Masahiro Mine for their useful comments.
The first author is supported by Grant-in-Aid for JSPS Research Fellow (Grant Number: 19J11223). The second author thanks the Max Planck Institute for the financial support when attending the conference on zeta functions in
Centre International de Rencontres Math\'ematiques, where the project was started.
\section{\textbf{Definitions and Results} }\label{General}
We consider the class $\mathcal{S}^\dagger$ consisting of $L$-functions $F(s)$
that satisfy the following conditions (S1)--(S5).
\begin{enumerate}
\setlength{\parskip}{2mm}
\item [(S1)] (Dirichlet series) $F(s)$ has a Dirichlet series ${F(s) = \sum_{n = 1}^{\infty}\frac{a_{F}(n)}{n^{s}}}$,
which converges absolutely for $\sigma > 1$.
\item[(S2)] (Analytic continuation) \quad
There exists $m_F \in \mathbb{Z}_{\geq 0}$ such that
$(s - 1)^{m_F}F(s)$ is entire of finite order.
\item[(S3)](Functional equation) \quad
$F(s)$ satisfies the following functional equation
\begin{align*}
\P_{F}(s) = \omega_F \overline{\P_{F}}(1 - s),
\end{align*}
where $\P_{F}(s) = \gamma(s)F(s)$ and
\gamma(s)= Q^{s}\prod_{\ell = 1}^{k}\Gamma(\lambda_{\ell}s + \mu_{\ell}),
$ with
$\lambda_{\ell} > 0$, $Q > 0$, $\Re \mu_{\ell} \geq -\frac{1}{2}$,
and $|\omega_F |= 1$. Here we use the notation $ \overline{\P_{F}}(s)=\overline{\P_F(\overline s)}$.
The quantity $d_{F}=2\sum_{\ell = 1}^{k}\lambda_{\ell}$ is often called the degree of $F$.
\item[(S4)] (Euler product)
$F(s)$ can be written as
\begin{align*}
F(s) = \prod_{p}F_{p}(s),
\quad
F_{p}(s) = \exp\l( \sum_{\ell = 1}^{\infty}\frac{b_{F}\l( p^{\ell} \r)}{p^{\ell s}} \r),
\end{align*}
where
$b_{F}(n) = 0$ unless $n = p^{\ell}$ with $\ell \in \mathbb{Z}_{\geq 1}$, and
$b_{F}(n) \ll n^{\vartheta_{F}}$ for some $\vartheta_{F} < \frac{1}{2}$.
\item[(S5)](Hypothesis H)
For any $\ell \geq 2$,
\begin{align}
\label{HHRS}
\sum_{p}\frac{|b_{F}(p^{\ell}) \log{p^{\ell}}|^2}{p^{\ell}} < +\infty.
\end{align}
\end{enumerate}
This class of $L$-functions is different from the Selberg class in that we do not assume the Ramanujan Conjecture,
which says
\begin{itemize}
\item [(S3')] $F(s)$ has the same functional equation as in (S3)
with the condition $\Re \mu_{\ell} \geq 0$ for $1\leq \ell \leq k$.
\item[(S5')]
For every $\varepsilon > 0$, the inequality
$a_{F}(n) \ll_{F} n^{\varepsilon}$ holds.
\end{itemize}
The set of $L$-functions satisfying (S1), (S2), (S3'), (S4), and (S5') is called the Selberg class $\mathcal{S}$. It is believed that automorphic $L$-functions of $GL(n)$ are in the Selberg class $\mathcal{S}$, but this hasn't been proved due to the lack of the Ramanujan Conjecture. The weaker conditions in $\mathcal S^{\dagger}$ allow us to include more automorphic $L$-functions of $GL(n)$ unconditionally.
The properties (S1)-(S4) for automorphic $L$-functions of $GL(n)$ as well as
their Rankin-Selberg convolutions can be found in \cite[Section 2]{RS1996}.
Hypothesis H have been proved for $1\leq n\leq 4$
( The case for $n = 1$ is trivial and the case for $n=2$ follows from the work of Kim-Sarnak \cite{KimSarnak}.
See Rudnick-Sarnak \cite[Proposition 2.4]{RS1996} for $n= 3$ and Kim \cite{Kim06} for $n=4$).
To study the joint value distribution of $L$-functions in $\mathcal{S}^\dagger$, we need the following assumption $\mathscr A$.
\begin{assumption*}[$\mathscr{A}$]
An $r$-tuple of Dirichlet series $\bm{F} = (F_{1}, \dots, F_{r})$
and an $r$-tuple of the numbers $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in \mathbb{R}^{r}$
satisfy $\mathscr{A}$ if and only if
$\bm{F}$, $\bm{\theta}$ satisfy the following properties.
\begin{itemize}
\item[(A1)] (Selberg Orthonormality Conjecture) \quad
For any $F_j$, we have
\begin{align} \label{SNC}
\sum_{p \leq x}\frac{|a_{F_{j}}(p)|^2}{p}
= n_{F_{j}} \log{\log{x}} + O_{F_{j}}(1), \quad x\rightarrow \infty.
\end{align}
for some constant $n_{F_{j}}>0$.
For any pair $F_{i} \not= F_{j}$,
\begin{align}
\label{SOC2}
\sum_{p \leq x}\frac{a_{F_{i}}(p) \overline{a_{F_{j}}(p)}}{p}
= O_{F_{i}, F_{j}}(1), \quad x\rightarrow \infty.
\end{align}
\item[(A2)] For every $i$, there is at most one $j\not=i$ such that $F_{i} = F_{j}$ and in this case $|\theta_{i} - \theta_{j}| \equiv \pi/2\pmod{ \pi }$.
\item[(A3)](Zero Density Estimate)
For every $F_{j}$, there exists a positive constant $\kappa_{F_{j}}$ such that, uniformly for any $T \geq 3$ and $1/2 \leq \sigma \leq 1$,
\begin{align} \label{ZDC1}
N_{F_{j}}(\sigma, T) \ll_{F_{j}} T^{1 - \kappa_{F_{j}}(\sigma - 1/2)}\log{T},
\end{align}
where $N_{F}(\sigma, T)$ is the number of nontrivial zeros $\rho_{F} = \b_{F} + i\gamma_{F}$ of $F \in \mathcal{S}^{\dagger}$
with $\b_{F} \geq \sigma$ and $0 \leq \gamma_{F} \leq T$.
\end{itemize}
\end{assumption*}
\begin{remark}
The Selberg Orthonormality Conjecture has been proved
for $L$-functions associated with cuspidal automorphic representations of $GL(n)$ unconditionally for $n\leq 4$ by Liu-Ye in \cite{LiuYe},
(See \cite{AS, Liu05}) and in full generality in \cite{LiuYe06}.
\end{remark}
\begin{remark}
It is natural to assume (A2). This allows us to consider the joint distribution of $\log |F(s)|$ and $\Im\log F(s)$.
It can be seen that $\Re e^{-i\theta_1}\log F(s)$ and $\Re e^{-i\theta_2}\log F(s)$ can not be independent
when $|\theta_1-\theta_2| \not\equiv \frac{\pi}{2} \pmod{\pi}$
\end{remark}
\begin{remark} We require a strong zero density estimate (the exponent of $\log T$ is $1$) close to the line $\sigma=1/2$.
Results of the shape \eqref{ZDC1} have been obtained for the Riemann zeta-function by Selberg \cite{SCR} and for Dirichlet $L$-functions by Fujii \cite{F1974}.
For $GL(2)$ $L$-functions, (A3) has been established by Luo \cite{Luo95} for holomorphic cusp forms
of the full modular group (see \cite[Section 7]{FZ} for other congruence subgroups).
The proof should also be applicable for Maa\ss\ forms (see the remark in \cite[p 141]{Luo95}, see also a weaker estimate in \cite{SSmaass}).
If we assume the Riemann Hypothesis for $F_j$, then \eqref{ZDC1} holds for any $\kappa_{F_j}>0$.
\end{remark}
Before we state our theorems, we need some notation.
Let $r$ be a fixed positive integer. For $\bm V=(V_1, \dots, V_r)\in \mathbb{R}^r$,
$\bm \theta=(\theta_1, \dots, \theta_r) \in \mathbb{R}^{r}$, and $\bm F=(F_1, \dots, F_r)\in (\mathcal{S}^\dagger)^r$
satisfying assumption $\mathscr A$, we define
\begin{align}
\S(T, \bm{V}; \bm{F}, \bm{\theta})
:= \set{t \in [T, 2T]}
{\frac{\Re e^{-i\theta_{j}}\log{F_{j}(\frac{1}{2} + it)}}{
\sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}} > V_j \text{ for } j = 1, \dots, r},
\end{align}
where the constants $n_{F_j}$ are defined in \eqref{SNC}. Let $\a_{\bm{F}} := \min\{ 2r, \frac{1 - 2\vartheta_{\bm{F}}}{2\vartheta_{\bm{F}}} \}$, where $\vartheta_{\bm{F}}=\max_{1\leq j\leq r}\vartheta_{F_{j}}$ as defined in (S4).
Here $\a_{\bm{F}} = 2r$ if $\vartheta_{\bm{F}} = 0$.
We denote $\norm[]{\bm{z}}=\max_{1 \leq j \leq r}|z_{j}|$.
Throughout this paper, we write $\log_{3}x$ for $\log{\log{\log}}x$.
\subsection{Results for large deviations}
The following theorem extends the result of Bombieri and Hejhal \cite{BH1995},
where we show \eqref{BHJVD} holds for a larger range of $V$.
\begin{theorem} \label{Main_Thm_LD_JVD}
Let $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in \mathbb{R}^{r}$ and
$\bm{F} = (F_1, \dots, F_{r}) \in (\mathcal{S}^{\dagger} )^{r}$ satisfy assumption $\mathscr{A}$.
Let $A \geq 1$ be a fixed constant.
For any large $T$ and any $\bm{V} = (V_{1}, \dots, V_{r}) \in \mathbb{R}^{r}$
with $\norm[]{\bm{V}} \leq A(\log{\log{T}})^{1/10}$, we have
\begin{align} \label{Main_Thm_LD_JVD1}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& =\l( 1 + O_{\bm{F}, A}\l(\frac{(\norm[]{\bm{V}}^{4} + (\log_{3}{T})^2)(\norm[]{\bm{V}} + 1)}{\sqrt{\log{\log{T}}}}
+ \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}\r)\r)
\prod_{j = 1}^{r}\int_{V_j}^{\infty} e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}.
\end{align}
Moreover, if $\bm \theta \in [-\frac{\pi}{2}, \frac{\pi}{2}]^r$ and $\|\bm V\| \leq A (\log{\log{T}})^{1/6}$
we have
\begin{align} \label{Main_Thm_LD_JVD2}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \leq\l ( 1 + O_{\bm{F}, A}\l(\frac{(\norm[]{\bm{V}}^{2} + (\log_{3}{T})^2)(\norm[]{\bm{V}} + 1)}{\sqrt{\log{\log{T}}}}
+ \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}\r)\r)
\prod_{j = 1}^{r}\int_{V_j}^{\infty} e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}},
\end{align}
and if $\bm \theta \in [\frac{\pi}{2}, \frac{3\pi}{2}]^r$, $\|\bm V\| \leq a_1(\log{\log{T}})^{1/6}$, and
$\prod_{j = 1}^{r} (1 + |V_{j}|) \leq a_1(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}$ with $a_1= a_1(\bm{F}) > 0$ small enough, we have
\begin{align} \label{Main_Thm_LD_JVD3}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \geq \l ( 1 - O_{\bm{F}}\l(\frac{(\norm[]{\bm{V}}^{2} + (\log_{3}{T})^2)(\norm[]{\bm{V}} + 1)}{\sqrt{\log{\log{T}}}}
+ \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}\r)\r)
\prod_{j = 1}^{r}\int_{V_j}^{\infty} e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}.
\end{align}
\end{theorem}
\begin{remark}
For $r=1$, $F_1=\zeta$, and $\theta_{1} = 0$, the asymptotic for $\|\bm V\|\ll (\log\log T)^{10}$ in \eqref{Main_Thm_LD_JVD1}
was obtained by Radziwi\l\l\mbox{} \cite{Ra2011}, and the bound for $\|\bm V\|\ll (\log\log T)^{1/6}$ in \eqref{Main_Thm_LD_JVD2}
was obtained by Inoue \cite{II2019}.
\end{remark}
It is reasonable to conjecture that the asymptotic in \eqref{Main_Thm_LD_JVD1} holds
for $\|\bm V\|=o(\sqrt{\log\log T})$ as speculated in \cite{Ra2011} for $\zeta(s)$. If we are only concerned with upper and lower bounds, we could extend the range of $\|\bm V\|$ further, which yields applications to moments of $L$-functions.
We have the following unconditional results which extends the range of $\|\bm V\|$ in Theorem \ref{Main_Thm_LD_JVD}.
\begin{theorem} \label{GJu}
Let
$\bm{F} = (F_1, \dots, F_{r}) \in (\mathcal{S}^{\dagger})^{r}$
and $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in [-\tfrac{\pi}{2}, \tfrac{3\pi}{2}]^r$ satisfy assumption $\mathscr{A}$.
Let $T$ be large.
There exists some positive constant $a_{2} = a_{2}(\bm{F})$ such that
if $\bm{\theta} \in [-\tfrac{\pi}{2}, \tfrac{\pi}{2}]^r$, we have
\begin{align*}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \ll_{\bm{F}} \l\{\l(\prod_{j = 1}^{r}\frac{1}{1 + V_{j}}\r) + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r\}
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
+ O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}}} \r) \r)
\end{align*}
for any $\bm{V} = (V_{1}, \dots, V_{r}) \in (\mathbb{R}_{\geq 0})^{r}$
satisfying $\norm[]{\bm{V}} \leq a_{2}(1 + V_{m}^{1/2})(\log{\log{T}})^{1/4}$ with $V_{m} = \min_{1 \leq j \leq r}V_{j}$,
and if $\bm{\theta} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]^{r}$, we have
\begin{align*}
\frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta}))
\gg_{\bm{F}} \l(\prod_{j = 1}^{r}\frac{1}{1 + V_{j}}\r)
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
+ O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}}} \r) \r)
\end{align*}
for $\norm[]{\bm{V}} \leq a_{2}(1 + V_{m}^{1/2})(\log{\log{T}})^{1/4}$ with
$\prod_{j = 1}^{r}(1 + V_{j}) \leq a_{2}(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}$.
\end{theorem}
Substituting
$
\bm{V} = \l(\frac{V}{\sqrt{\frac{n_{F_{1}}}{2}\log{\log{T}}}}, \dots, \frac{V}{\sqrt{\frac{n_{F_{r}}}{2}\log{\log{T}}}}\r)
$ to Theorem \ref{GJu}, we obtain the following corollary.
\begin{corollary} \label{NNGJu}
Let
$\bm{F} = (F_1, \dots, F_{r}) \in (\mathcal{S}^{\dagger})^{r}$
and $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in [-\tfrac{\pi}{2}, \tfrac{3\pi}{2}]^r$ satisfy assumption $\mathscr{A}$.
Set $h_{\bm{F}} = n_{F_{1}}^{-1} + \cdots + n_{F_{r}}^{-1}$.
There exists a small constant $a_{3} = a_{3}(\bm{F})$ such that
if $\bm{\theta} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]^{r}$, we have
\begin{align} \label{NNGJu1}
& \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)} > V} \\
& \ll_{\bm{F}}
\l(\frac{1}{(1 + V / \sqrt{\log{\log{T}}})^{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r)
\exp\l( -h_{\bm{F}}\frac{V^2}{\log{\log{T}}}
\l( 1 + O_{\bm{F}}\l( \frac{V}{\log{\log{T}}} \r) \r) \r)
\end{align}
for any $0 \leq V \leq a_{3}\log{\log{T}}$, and if $\bm{\theta} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]^{r}$, we have
\begin{align} \label{NNGJu2}
& \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)} > V} \\
& \gg_{\bm{F}} \frac{1}{(1 + V / \sqrt{\log{\log{T}}})^{r}}
\exp\l( -h_{\bm{F}}\frac{V^2}{\log{\log{T}}}
\l( 1 + O_{\bm{F}}\l( \frac{V}{\log{\log{T}}} \r) \r) \r)
\end{align}
for any $0 \leq V \leq a_{3}\min\{\log{\log{T}}, (\log \log T)^{\frac{\a_{\bm{F}}}{r} + \frac{1}{2} + \frac{1}{2r}}\}$.
\end{corollary}
\begin{remark}
When $r=1$, $F_1=\zeta$, and $\theta_{1} = 0$, Jutila \cite{Ju1983}, using bounds on moments of $\zeta(s)$, has proved
\begin{align}
& \operatorname{meas}\set{t \in [T, 2T]}{ \log |\zeta(\tfrac{1}{2} + it) |> V}\ll T
\exp\l( -\frac{V^2}{\log{\log{T}}}\l(1 + O\l(\frac{V}{\log{\log{T}}}\r) \r) \r)
\end{align}
uniformly for $0\leq V\leq \log \log T$. More recently, Heap and Soundararajan\cite[Corollary 2]{HeapSound} obtained
\begin{align}
& \operatorname{meas}\set{t \in [T, 2T]}{ \log |\zeta(\tfrac{1}{2} + it) |> V}\ll T
\exp\l( -\frac{V^2}{\log{\log{T}}}+ O\l(\frac{V\log_3T}{\sqrt{\log \log T}}\r) \r)
\end{align}
for $\sqrt{\log \log T}\log_3T\leq V\leq (2-o(1))\log\log T$
using sharp bounds on moments of $\zeta(s)$.
Our Theorem \ref{GJu} slightly improves Jutila's bound by a factor of $\frac{\sqrt{\log \log T}}{V}$ when $\sqrt{\log\log T}\leq V\leq a_3\log\log T$ for some small constant $a_3$, but is weaker than \cite{HeapSound} when $\sqrt{\log \log T}\log_3T\leq V\leq (2-o(1))\log\log T$.
\end{remark}
If we assume the Riemann Hypothesis for the corresponding $L$-functions, we can extend the range of $\|\bm V\|$ in Theorem \ref{GJu} even further.
\begin{theorem}\label{RH_LD_JVD}
Let $\bm{F} = (F_{1}, \dots, F_{r}) \in (\mathcal{S}^{\dagger} )^{r}$ and $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in \mathbb{R}^{r}$
satisfy assumption $\mathscr{A}$ and assume that the Riemann Hypothesis is true for $F_{1}, \dots, F_{r}$. Let $\bm{V} = (V_{1}, \dots, V_{r}) \in (\mathbb{R}_{\geq 3})^{r}$ and denote $V_{m} = \min_{1 \leq j \leq r}V_{j}$. Then there exists $a_4=a_4(\bm F)$ such that the following holds.
If $\bm{\theta} \in [-\tfrac{\pi}{2}, \tfrac{\pi}{2}]^{r}$ and $\|\bm{V}\| \leq a_{4}V_{m}^{1/2}(\log{\log{T}})^{1/4}(\log_{3}{T})^{1/2}$, we have
\begin{align}
\label{RH_LD_JVD1u}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \ll_{\bm{F}}\l(\frac{1}{V_{1} \cdots V_{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
+ O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}}} \r) \r).
\end{align}
If $\bm{\theta} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]^{r}$, $\|\bm{V}\| \leq a_{4}V_{m}^{1/2}(\log{\log{T}})^{1/4}(\log_{3}{T})^{1/2}$
and
$\prod_{j = 1}^{r}V_{j} \leq a_{4}(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}$, we have
\begin{align}
\label{RH_LD_JVD1l}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \gg_{\bm{F}} \frac{1}{V_{1} \cdots V_{r}}
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
- O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}}} \r) \r).
\end{align}
Moreover, there exist some positive constants $a_{5} = a_{5}(\bm{F})$
such that for any $\bm{V} \in (\mathbb{R}_{\geq 3})^{r}$ with $\norm[]{\bm{V}} \geq \sqrt{\log\log{{T}}}$ and $\bm{\theta} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]^r$,
\begin{align}
\label{RH_LD_JVD2}
\frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta}))
\ll_{\bm{F}}
\exp\l( -a_{5}\norm[]{\bm{V}}^2 \r) + \exp\l( -a_{5} \norm[]{\bm{V}}\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}} \r).
\end{align}
\end{theorem}
With $r=1$, Theorem \ref{RH_LD_JVD} slightly improves the bound in \cite[Proposition 4.1]{MT} in the range of the following corollary.
\begin{corollary} \label{CBSP}
Let $F \in \mathcal{S}^{\dagger} $, and assume the Riemann Hypothesis for $F$.
Let $A \geq 1$, $\theta \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]$.
Then, for any real number $V$ with
$\sqrt{\log{\log{T}}} \leq V \leq A (\log{\log{T}})^{{2/3}} (\log_{3}{T})^{1/3}$, we have
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\Re e^{-i\theta} \log{F(\tfrac{1}{2}+it)|} > V}
\ll_{A, F} \frac{\sqrt{\log{\log{T}}}}{V}\exp\l( -\frac{V^2}{n_{F}\log{\log{T}}} \r).
\end{align*}
as $T\rightarrow \infty$.
\end{corollary}
\subsection{Moments of $L$-functions}
We apply the large deviation results to obtain bounds for moments of $L$-functions without and with the Riemann Hypothesis of the corresponding $L$-functions.
The following theorem is a consequence of Theorem \ref{GJu} without the Riemann Hypothesis.
\begin{theorem} \label{New_MVT}
Let $\bm{F} = (F_{1}, \dots, F_{r}) \in (\mathcal{S}^{\dagger})^{r}$ and $\bm{\theta}=(\theta_1, \dots, \theta_r) \in \mathbb{R}^{r}$ satisfy assumption $\mathscr{A}$.
Set $h_{\bm{F}} = n_{F_{1}}^{-1} + \cdots + n_{F_{r}}^{-1}$. Then there exist some positive constants $a_{6} = a_{6}(\bm{F})$
and $B=B(\bm{F})$ such that
for any fixed $0 < k \leq a_{6}$,
\begin{align} \label{New_MVT1}
\int_{T}^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Re e^{-\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}\r)dt
\ll_{\bm F} T (\log{T})^{k^2 / h_{\bm{F}} + Bk^{3}}
\end{align}
when $\bm{\theta} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]^r$,
and if $\vartheta_{\bm{F}} \leq \frac{1}{r + 1}$, we have
\begin{align} \label{New_MVT2}
\int_{T}^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}\r)dt
\gg _{\bm F} T (\log{T})^{k^2 / h_{\bm{F}} - Bk^{3}}
\end{align}
when $\bm{\theta} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]^r$.
Here, the above implicit constants depend only on $\bm{F}$.
In particular, if $\vartheta_{\bm{F}} \leq \frac{1}{r + 1}$, it holds that, for any $0 < k \leq a_{6}$,
\begin{align*}
\int_{T}^{2T}\l(\min_{1 \leq j \leq r}|F_{j}(\tfrac{1}{2}+it)|\r)^{2k}dt
\ll_{\bm F} T (\log{T})^{k^2 / h_{\bm{F}} + Bk^{3}}
\end{align*}
\begin{align*}
\int_{T}^{2T}\l(\max_{1 \leq j \leq r}|F_{j}(\tfrac{1}{2}+it)|\r)^{-2k}dt
\gg_{\bm F} T (\log{T})^{k^2 / h_{\bm{F}} - Bk^{3}}
\end{align*}
\begin{align*}
T (\log{T})^{k^2 / h_{\bm{F}} - Bk^{3}}
\ll_{\bm F} \int_{T}^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Im \log F_{j}(\tfrac{1}{2}+it)\r)dt
\ll_{\bm F} T (\log{T})^{k^2 / h_{\bm{F}} + Bk^{3}}
\end{align*}
and
\begin{align*}
T (\log{T})^{k^2 / h_{\bm{F}} - Bk^{3}}
\ll_{ \bm{F}} \int_{T}^{2T}\exp\l(-2k\max_{1 \leq j \leq r}\Im \log F_{j}(\tfrac{1}{2}+it)\r)dt
\ll_{ \bm{F}} T (\log{T})^{k^2 / h_{\bm{F}} + Bk^{3}}.
\end{align*}
\end{theorem}
\begin{remark}
The constant $a_6$ could be computed explicitly by keeping track of the constants in the proof. It depends on the size of $a_{F_i}(p)$,
the constants ($n_{F{j}}, O_{F_i}(1), O_{F_i, F_j}(1)$) in condition (A1), the degree $d_{F_{j}} = 2\sum_{\ell = 1}^{k}\lambda_{\ell}$,
the constant $\kappa_{F_j}$, and the implicit constant in \eqref{ZDC1}. We also need $a_6 \ll_{\bm F}\frac{1}{r}$.
\end{remark}
\begin{remark}
We could also allow $k\rightarrow 0$ as $T\rightarrow \infty$. In fact, as shown in the proof, the dependency of the implicit constant in $k$ is of size $\frac{1 + k\sqrt{\log\log T}}{1+(k\sqrt{\log\log T})^r}$.
\end{remark}
The restriction on $k$ in Theorem \ref{New_MVT} is due to the fact that the range of $\| \bm V\|$ in Theorem \ref{GJu} is only a small multiple of $\log\log T$. If we assume the Riemann Hypothesis for the corresponding $L$-functions, we can apply Theorem \ref{RH_LD_JVD} to establish moment results for all $k>0$.
\begin{theorem} \label{New_MVT_RH}
Let
$\bm{F} = (F_{1}, \dots, F_{r}) \in (\mathcal{S}^{\dagger} )^{r}$
and $\bm{\theta} = (\theta_{1}, \dots, \theta) \in [-\tfrac{\pi}{2}, \tfrac{3\pi}{2}]^r$ satisfy assumption $\mathscr{A}$,
and assume that the Riemann Hypothesis is true for $F_{1}, \dots, F_{r}$.
Let $T$ be large, and put $\varepsilon(T) = (\log_{3}{T})^{-1}$.
Then there exists some positive constant $B = B(\bm{F})$ such that for any $k > 0$,
if $\bm{\theta} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]^r$, we have
\begin{align} \label{New_MVT_RH1}
\int_{T}^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Re e^{-\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}\r)dt
\ll_{k, \bm{F}} T(\log{T})^{k^{2}/h_{\bm{F}} + Bk^{3}\varepsilon(T)}
\end{align}
and if $\bm{\theta} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]^r$ and $\vartheta_{\bm{F}} < \frac{1}{r + 1}$, we have
\begin{align} \label{New_MVT_RH2}
\int_{T}^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Re e^{-\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}\r)dt
\gg _{k, \bm{F}} T + T(\log{T})^{k^{2}/h_{\bm{F}} - Bk^{3}\varepsilon(T)}
\end{align}
In particular, if $\vartheta_{\bm{F}} < \frac{1}{r + 1}$, it holds that, for any $k > 0$, $\varepsilon > 0$,
\begin{align*}
\int_{T}^{2T}\l(\min_{1 \leq j \leq r}|F_{j}(\tfrac{1}{2}+it)|\r)^{2k}dt
\ll_{\varepsilon, k, \bm{F}} T(\log{T})^{k^{2}/h_{\bm{F}} + \varepsilon},
\end{align*}
\begin{align*}
\int_{T}^{2T}\l(\max_{1 \leq j \leq r}|F_{j}(\tfrac{1}{2}+it)|\r)^{-2k}dt
\gg_{\varepsilon, k, \bm{F}} T(\log{T})^{k^{2}/h_{\bm{F}} - \varepsilon},
\end{align*}
\begin{align*}
T(\log{T})^{k^{2}/h_{\bm{F}} - \varepsilon}
\ll_{\varepsilon, k, \bm{F}}\int_{T}^{2T} \exp\l(2k\min_{1 \leq j \leq r}\Im \log F_{j}(\tfrac{1}{2}+it)\r)dt
\ll_{\varepsilon, k, \bm{F}} T(\log{T})^{k^{2}/h_{\bm{F}} + \varepsilon},
\end{align*}
and
\begin{align*}
T(\log{T})^{k^{2}/h_{\bm{F}} - \varepsilon}
\ll_{\varepsilon, k, \bm{F}}\int_{T}^{2T} \exp\l(-2k\max_{1 \leq j \leq r}\Im \log F_{j}(\tfrac{1}{2}+it)\r)dt
\ll_{\varepsilon, k, \bm{F}} T(\log{T})^{k^{2}/h_{\bm{F}} + \varepsilon}.
\end{align*}
\end{theorem}
\begin{remark}
One would expect that for any $k>0$, it holds that
\begin{align
\int_T^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log F_{j}(\tfrac{1}{2}+it)\r)dt
\asymp_{k} T \frac{(\log T)^{k^2 / h_{\bm{F}}}}{(\log{\log{T}})^{(r-1)/2}}.
\end{align}
It remains an interesting question to see if the bounds in Theorem \ref{New_MVT_RH} can be made sharp using techniques from Harper \cite{H2013} under the corresponding Riemann Hypothesis.
\end{remark}
\begin{remark}
An $L$-function is called primitive if it cannot be factored into $L$-functions of smaller degree. It is conjectured that
\begin{align}
\int_0^T | F(\tfrac{1}{2}+it)|^{2k} dt \sim C(F, k)T(\log T)^{k^2}
\end{align}
for some constant $C(F, k)$ as $T\rightarrow \infty$, see \cite{Conrey05,Heap13}.
It is expected that the values of distinct primitive $L$-functions are uncorrelated, which leads to the conjecture
\begin{align}
\int_0^T |F_1(\tfrac{1}{2}+it)|^{2k_1} \cdots |F_r(\tfrac{1}{2}+it)|^{2k_r} dt
\sim C(\bm F, \bm k) T(\log T)^{k_1^2 + \dots + k_r^2}
\end{align}
for some constant $C(\bm F, \bm k)$ as $T\rightarrow \infty$ if $F_i\not=F_j$ for $i\not=j$.
This has be established for the product of two Dirichlet $L$-functions for $k_1=k_2=1$ (see \cite{Heap13, M1970, Top} and
for some degree two $L$-functions when $k=1$ and $r=1$ (see \cite{Good, Zhang1, Zhang2}).
For higher degree $L$-functions and higher values of $k$, obtaining the asymptotic formula seems to be beyond the scope of
current techniques.
An upper bound of this kind has been established by Milinovich and Turnage-Butterbaugh \cite{MT} for automorphic $L$-functions of $GL(n)$
under the Riemann Hypothesis for these $L$-functions. Modifying our approach, we can recover their results (see Section \ref{finalremarks}) and obtain some weaker unconditional results for sufficiently small $k_i$'s when $n\leq 2$.
Our results give some further evidence that distinct primitive $L$-functions are `` statistically independent".
\end{remark}
\begin{remark}
Our method as well as the works of Bombieri-Hejhal \cite{BH1995} and Selberg \cite{SCR} requires a strong zero density estimate for $L$-functions.
Unfortunately, the estimate has not been proved yet for many $L$-functions.
There have been other methods
to prove Selberg's central limit theorem without the strong zero density estimate such as Laurin\v{c}ikas \cite{L1987}
and Radziwi\l\l-Soundararajan \cite{RS2017}.
However, their methods do not apply to the central limit theorems for $\Im \log \zeta(s)$.
Nevertheless, Hsu-Wong \cite{HW2020} proved a joint central limit theorem (for fixed $V_{j}$)
for Dirichlet $L$-functions and certain $GL(2)$ $L$-functions twisted by Dirichlet characters by using the method in \cite{RS2017}.
The fact that the coefficients of these $L$-functions satisfy $|\chi(n)| \leq 1$ plays an important role in their proofs, which make it not applicable to more general $L$-functions.
\end{remark}
\subsection{Results for Dirichlet polynomials}
To prove the above theorems, we consider the Dirichlet polynomials associated with $F$.
We need some notation before stating our results.
Let $\Lambda_{F}(n)$ be the von Mangoldt function associated with $F$ defined by $\Lambda_{F}(n) = b_{F}(n) \log{n}$.
Let $\bm{x} = (x_{1}, \dots, x_{r}) \in \mathbb{R}^{r}$, $\bm{z} = (z_{1}, \dots, z_{r}) \in \mathbb{C}^{r}$,
and $\bm{F} = (F_{1}, \dots, F_{r})$ be an $r$-tuple of Dirichlet series,
and let $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in \mathbb{R}^r$.
When $\bm{F}$ satisfies (S4), we define
\begin{align} \label{def_D_P}
P_{F}(s, X)
:= \sum_{p^{\ell} \leq X}\frac{b_{F}(p^{\ell})}{p^{\ell s}}
= \sum_{2 \leq n \leq X}\frac{\Lambda_{F}(n)}{n^{s}\log{n}},
\end{align}
\begin{align} \label{def_var}
\sigma_{F}(X)
:= \sqrt{\frac{1}{2}\sum_{p \leq X} \sum_{\ell = 1}^{\infty}\frac{|b_{F}(p^{\ell})|^2}{p^{\ell}}},
\end{align}
\begin{align} \label{def_tau_F}
\tau_{j_{1}, j_{2}}(X) = \tau_{j_{1}, j_{2}}(X; \bm{F}, \bm{\theta})
:= \frac{1}{2}\sum_{p \leq X}\sum_{\ell = 1}^{\infty}
\frac{\Re e^{-i\theta_{j_{1}}}b_{F_{j_{1}}}(p^{\ell})\overline{e^{-i\theta_{j_{2}}} b_{F_{j_{2}}}(p^{\ell})}}{p^{\ell}},
\end{align}
\begin{align} \label{def_K_F}
K_{\bm{F}, \bm{\theta}}(p, \bm{z})
:= \sum_{1 \leq j_{1}, j_{2} \leq r}z_{j_{1}} z_{j_{2}}
\sum_{\ell = 1}^{\infty}\frac{\Re e^{-i\theta_{j_{1}}}b_{F_{j_{1}}}(p^{\ell})\overline{e^{-i\theta_{j_{2}}} b_{F_{j_{2}}}(p^{\ell})}}{p^{\ell}},
\end{align}
and
\begin{align}\label{S_X def}
\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta})
:= \set{t \in [T, 2T]}
{\frac{\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2}+it, X)} {\sigma_{F_{j}}(X)}
> V_{j} \text{ for } j = 1, \dots, r}.
\end{align}
Let $\{ \mathcal{X}(p) \}_{p \in \mathcal{P}}$ be a sequence of independent random variables
on a probability space $(\Omega, \mathscr{A}, \mathbb{P})$
with uniformly distributed on the unit circle in $\mathbb{C}$, where $\mathcal{P}$ is the set of prime numbers.
Denote
\begin{align} \label{def_RP}
P_{F}(\sigma, \mathcal{X}, X)
:= \sum_{p \leq X}\sum_{\ell = 1}^{\infty}\frac{b_{F}(p^{\ell}) \mathcal{X}(p)^{\ell}}{p^{\ell \sigma}},
\end{align}
\begin{align} \label{def_RDP_MGF}
M_{p, \sigma}(\bm{z}) = M_{p, \sigma}(\bm{z}; \bm{F}, \bm{\theta})
:= \EXP{\exp\l(\sum_{j = 1}^{r}z_{j} \Re e^{-i\theta_{j}}
\sum_{\ell = 1}^{\infty}\frac{b_{F_{j}}(p^{\ell}) \mathcal{X}(p)^{\ell}}{p^{\ell \sigma}}\r)},
\end{align}
where $\EXP{}$ is the expectation.
Finally, when $\bm{F}$ satisfies (S4) and (S5), we define
\begin{align} \label{def_Xi}
\Xi_{X}(\bm{x})
= \Xi_{X}(\bm{x}; \bm{F}, \bm{\theta})
:= \exp\l( \sum_{1 \leq j_{1} < j_{2} \leq r}x_{j_{1}} x_{j_{2}}\tau_{j_{1}, j_{2}}(X) \r)
\prod_{p}\frac{M_{p, \frac{1}{2}}(\bm{x})}{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{x})/4 \r)}.
\end{align}
The convergence of the infinite product of \eqref{def_Xi} is shown in Lemma \ref{Prop_Psi_FPP}.
We have the following the joint large deviations results for Dirichlet polynomials.
\begin{proposition} \label{Main_Prop_JVD}
Let $\bm{F} = (F_{1}, \dots, F_{r})$ be an $r$-tuple of Dirichlet series and $\bm{\theta} \in \mathbb{R}^{r}$
satisfy (S4), (S5), (A1), and (A2).
Let $T$, $X$ be large numbers with $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then there exists some positive constant $a_{7} = a_{7}(\bm{F})$ such that
for $\bm{V} = (V_{1}, \dots, V_{r}) \in \mathbb{R}^{r}$ with $|V_{j}| \leq a_{7} \sigma_{F_{j}}(X)$,
\begin{align} \label{MPJVD1}
& \frac{1}{T}\operatorname{meas}(\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& = \l( 1 + O_{\bm{F}}\l( \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^{2}}{\log{\log{X}}}\r) \r)\prod_{j = 1}^{r}
\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align}
\end{proposition}
We could improve the range $V_{j}$ in Propositions \ref{Main_Prop_JVD} with a weaker error term.
\begin{proposition} \label{Main_Prop_JVD3}
Let $\bm{F} = (F_{1}, \dots, F_{r})$ be an $r$-tuple of Dirichlet series and $\bm{\theta} \in \mathbb{R}^{r}$
satisfy (S4), (S5), (A1), and (A2).
Let $T$, $X$ be large numbers satisfying $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then for any $\bm{V} = (V_{1}, \dots, V_{r}) \in (\mathbb{R}_{\geq 0})^{r}$ with $\| \bm{V} \| \leq (\log{\log{X}})^{2r}$, we have
\begin{align} \label{MPJVD2}
& \frac{1}{T}\operatorname{meas}(\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta}))
= (1 + E) \times \Xi_{X}\l(\tfrac{V_{1}}{\sigma_{F_{1}}(X)}, \dots, \tfrac{V_{r}}{\sigma_{F_{r}}(X)}\r)
\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}},
\end{align}
where $E$ satisfies
\begin{align} \label{ERMPJVD}
E \ll_{\bm{F}}
\exp\l( C \l(\frac{\| \bm{V} \|}{\sqrt{\log{\log{X}}}}\r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\l\{\frac{\prod_{k = 1}^{r}(1 + V_{k})}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1}{\sqrt{\log{\log{X}}}}\r\}.
\end{align}
\end{proposition}
\begin{comment}
\begin{proposition} \label{Main_Prop_JVD2}
Let $\bm{F} = (F_{1}, \dots, F_{r})$ be an $r$-tuple of $L$-functions and $\bm{\theta} \in [0, 2\pi]^{r}$
satisfying (S4), (A1), and (A2).
Let $T$, $X$ be large numbers with $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then, for any $\bm{V} = (V_{1}, \dots, V_{r}) \in (\mathbb{R}_{\geq 0})^{r}$ with $\norm[]{\bm{V}} \leq (\log{\log{X}})^{2r}$,
\begin{align} \label{Main_Prop_JVD2_1}
& \frac{1}{T}\operatorname{meas}(\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta}))\ll_{\bm{F}} \\
& \l(\prod_{k = 1}^{r}\frac{1}{1 + V_{k}} + \frac{1}{(\log{\log{X}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
+ O_{\bm{F}}\l( \l( \frac{\norm[]{\bm{V}}}{\sqrt{\log{\log{X}}}} \r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
\end{align}
\end{proposition}
\end{comment}
\begin{remark}
In contrast to Proposition \ref{Main_Prop_JVD}, we allow $V_j$ to be of size $C\sqrt{\log\log X}$ for arbitrarily large $C$,
which is important in the proof of Theorem \ref{New_MVT_RH}.
We can prove an estimate similar to \eqref{MPJVD2} for larger $V_{j}$,
where we need to change the value of $X$ suitably in this case.
However, our main purpose is to prove Theorems \ref{Main_Thm_LD_JVD}, \ref{GJu}, and \ref{RH_LD_JVD},
and the case of larger $V_{j}$ is not required in their proofs.
For this reason, we give only the case $\norm[]{\bm{V}} \leq (\log{\log{X}})^{2r}$ for simplicity.
\end{remark}
Applying Proposition \ref{Main_Prop_JVD3}, we can prove the following theorem.
\begin{theorem} \label{GMDP}
Let $\bm{F} = (F_{1}, \dots, F_{r})$ be an $r$-tuple of Dirichlet series, and $\bm{\theta} \in \mathbb{R}^{r}$
satisfy (S4), (S5), (A1), and (A2). Denote $\vartheta_{\bm{F}}^{*}=\min_{1\leq j\leq r}\vartheta_{F_j}$.
Then for any fixed $k > 0$
there exist positive constants $c_1=c_1(k, \bm F)$, $c_2=c_2(k, \bm F)$ and $X_{0} = X_{0}(\bm{F}, k)$ such that
for any large $T$ and
$X_{0} \leq X \leq \max\{ c_{1}(\log{T} \log{\log{T}})^{\frac{2}{1 + 2\vartheta_{\bm{F}}^{*}}},
c_{2}\frac{(\log{T})^2 \log{\log{T}}}{\log_{3}{T}} \}$,
we have
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\exp\l(2k\min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2} + it, X)\r)dt \\
& = \frac{\exp\l( k^2 H_{\bm{F}}(X) \r) \prod_{j = 1}^{r}\sigma_{F_{j}}(X)}{\l( \sqrt{2\pi} k H_{\bm{F}}(X) \r)^{r-1}
\sqrt{\frac{1}{2}H_{\bm{F}}(X)}}
\Xi_{X}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r)\l( 1 + E \r),
\end{align*}
where $H_{\bm{F}}(X) = 2\l(\sum_{j = 1}^{r}\sigma_{F_{j}}(X)^{-2}\r)^{-1}$, and
\begin{align*}
E
\ll_{\bm{F}, k}
\frac{1}{(\log{\log{X}})^{\a_{\bm{F}} - \frac{r - 1}{2}}}
+ \frac{\log_{3}{X}}{\sqrt{\log{\log{X}}}}.
\end{align*}
\end{theorem}
\begin{remark}
When $r=1$, these type of results can be obtained via mean value theorems of Dirichlet series (see Gonek-Hughes-Keating \cite[Theorem 2]{GHK2007} for $\zeta(s)$ and Heap \cite[Theorem 2]{Heap13} for Dedekind zeta-functions associated with Galois extension number fields over $\mathbb{Q}$). Our method has the advantage to handle more than one $L$-functions, though the error term is weaker.
\end{remark}
\begin{remark}
The exact dependency of $c_1, c_2$ on $\bm F$ and $k$ are determined explicitly in the proof.
When $\vartheta_{\bm{F}}^{*} = 0$, the upper bound becomes $X \leq c_1(\log T\log\log T)^2$ where $c_1 \ll_{\bm F} \frac{1}{k^2}$.
\end{remark}
The asymptotic behavior can be further evaluated under some additional assumptions.
\begin{corollary}\label{jointDirichlet}
We use the same notation and assumptions as in Theorem \ref{GMDP}. We further assume $\vartheta_{\bm{F}} < \frac{1}{r + 1}$ and the strong Selberg Orthonormality Conjecture, that is,
\begin{align} \label{SSOC}
\sum_{n \leq X}\frac{a_{F_{i}}(p) \overline{a_{F_{j}}(p)}}{p}
= \delta_{F_{i}, F_{j}} n_{F_{j}}\log{\log{X}} + e_{i, j} + o(1), \quad X \rightarrow + \infty
\end{align}
for some constant $e_{i, j}$.
Here, $\delta_{F_{i}, F_{j}}$ is the Kronecker delta function, that is, $\delta_{F_{i}, F_{j}} = 1$ if $F_{i} = F_{j}$,
and $\delta_{F_{i}, F_{j}} = 0$ otherwise.
Then, as $X \rightarrow + \infty$,
\begin{align} \label{GGHK}
\hspace{-3mm}
\frac{1}{T}\int_{T}^{2T}
\exp\l( 2k \min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \sum_{2 \leq n \leq X}\frac{\Lambda_{F_{j}}(n)}{n^{1/2+it} \log{n}} \r) dt
\sim C(\bm{F}, k, \bm{\theta})\frac{(\log{X})^{k^2 / h_{\bm{F}}}}{(\log{\log{X}})^{(r - 1)/2}}
\end{align}
for some positive constant $C(\bm{F}, k, \bm{\theta})$.
\end{corollary}
\begin{remark}
When $r = 1$, $F_{1} = \zeta$, and $\theta_{1} = 0 \text{ or } \pi $ Corollary \ref{jointDirichlet} recovers a result of Heap \cite[Proposition 1]{He2021} for real moments of partial Euler product with $X<(\frac{1}{16k^2}-\epsilon )(\log T\log\log T)^2$
\end{remark}
\begin{comment}
Moreover, we can also prove the following proposition,
which has a consequence for Ramachandra's denseness problem.
Put
\begin{align*}
\tilde{\sigma}_{F}(X)
:= \sqrt{\frac{1}{2}\sum_{p \leq X}\sum_{k = 1}^{\infty}\frac{|b_{F}(p^{k})|^2}{p^{k}}},
\end{align*}
and for $\bm{V} = (V_{1}, V_{2})$,
\begin{align*}
\mathscr{T}_{X}(T, \bm{V}; F, \theta)
:= \set{t \in [T, 2T]}{\frac{\Re e^{-i\theta} P_{F}(\tfrac{1}{2} + it, X)}{\tilde{\sigma}_{F}(X)} > V_{1} \; \text{ and } \;
\frac{\Im e^{-i\theta} P_{F}(\tfrac{1}{2} + it, X)}{\tilde{\sigma}_{F}(X)} > V_{2}}.
\end{align*}
\begin{proposition} \label{EED2}
Let $F$ and $\theta \in \mathbb{R}$
satisfy (S4), (S5), (A1).
Let $T$, $X$ be large numbers with $X^{(\log{\log{X}})^{12}} \leq T$.
Then, there exists some positive constant $a_{10} = a_{10}(F)$ such that
for $\bm{V} = (V_{1}, V_{2}) \in \mathbb{R}^{r}$ with $|V_{j}| \leq a_{10} \sigma_{F}(X)$,
\begin{align*}
& \frac{1}{T}\operatorname{meas}(\mathscr{T}_{X}(T, \bm{V}; F, \theta)) \\
& = \l( 1 + O_{F}\l( \frac{(1 + |V_{1}|)(1 + |V_{2}|)}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \|\bm{V}\|^{3}}{(\log{\log{X}})^{3/2}}\r) \r)
\int_{V_{2}}^{\infty}\int_{V_{1}}^{\infty}e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi}.
\end{align*}
\end{proposition}
\begin{corollary} \label{DLDP}
Let $F$ satisfy (S4), (S5), and (A1), and the estimate $|a_{F}(p)| \ll_{F} p^{\vartheta_{F}}$
for some $\vartheta_{F} \in [0, 1/3)$.
Define $R(z, r) = \set{u + iv \in \mathbb{C}}{\max\{|\Re z - u|, |\Im z - v|\} < r}$.
For any $0 < \varepsilon \leq 1$, $z \in \mathbb{C}$, and any numbers $T$, $X$ with $X^{(\log{\log{X}})^{12}} \leq T$,
we have
\begin{align} \label{DLDP1}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{P_{F}(\tfrac{1}{2} + it, X) \in R(z, \varepsilon)}
\sim \iint_{R(z / \tilde{\sigma}_{F}(X), \varepsilon / \tilde{\sigma}_{F}(X))}
e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi},
\end{align}
as $X \rightarrow + \infty$.
Moreover, if $\vartheta_{F} \in [0, \tfrac{1}{4})$, we have
\begin{align} \label{DLDP2}
\hspace{-4mm}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\sum_{p \leq X}\log{F_{p}(\tfrac{1}{2} + it)} \in R(z, \varepsilon)}
\sim \iint_{R(z / \tilde{\sigma}_{F}(X), \varepsilon / \tilde{\sigma}_{F}(X))}
e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi},
\end{align}
and
\begin{align} \label{DLDP3}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\prod_{p \leq X}F_{p}(\tfrac{1}{2} + it) \in R(z, \varepsilon)}
\geq (1 + o(1))\frac{2 \varepsilon^2 + O(\varepsilon^4) }{n_{F}\log{\log{X}}}
\end{align}
as $X \rightarrow + \infty$.
\end{corollary}
\begin{remark}
Corollary \ref{DLDP} is related to Ramachandra's denseness problem, asking whether
the set of values of $\zeta(\tfrac{1}{2}+it)$, $t \in \mathbb{R}$ is dense in $\mathbb{C}$ or not.
Bohr and Courant \cite{B1916}, \cite{BC1914} showed that
the sets of values of $\zeta(\sigma+it)$ and $\log{\zeta(\sigma+it)}$ are dense in $\mathbb{C}$ when $\frac{1}{2} < \sigma \leq 1$.
Roughly speaking, the proofs are mainly divided into the following two parts,
and we succeed in proving Step 1 in the case $\sigma = \frac{1}{2}$ by Corollary \ref{DLDP}.
\begin{itemize}
\item[Step 1.] (Denseness lemma)
\textit{The corresponding Dirichlet polynomial to $\zeta(\sigma+it)$ and $\log{\zeta(\sigma+it)}$ can approximate to any complex numbers.}
\item[Step 2.]
\textit{For ``almost all" $t$, $\zeta(\sigma+it)$ and $\log{\zeta(\sigma+it)}$ equal to the corresponding Dirichlet polynomial. }
\end{itemize}
When $\frac{1}{2} < \sigma \leq 1$, these assertions were shown by Bohr and Courant.
On the other hand, when $\sigma = \frac{1}{2}$, neither Step 1 nor Step 2 was not proven yet.
Hence, it seems that Corollary \ref{DLDP} advances Ramachandra's problem.
However, it is difficult to prove Step 2 in our method.
To prove Step 2, we need to obtain a ``good estimate" for the mean value theorem for the gap of zeta-function and Dirichlet polynomials
on the critical line.
For this problem, there is a work due to Goldston \cite{G1987}, which relates the mean value theorem to the pair correlation function.
Hence, we may expect that some strong assumption for the pair correlation of zeros leads to Ramachandra's problem.
For more recent works on this problem, see \cite{EI2020}, \cite{GS2014}, \cite{KN2012}.
\end{remark}
\end{comment}
\section{\textbf{Approximate formulas for $L$-functions}}\label{appL}
\subsection{Approximate formulas for $L$-functions}
In this section, we give an approximate formula for $\log{F(s)}$.
Here, we choose the branch of $\log{F(\sigma + it)}$ as follows.
If $t$ is equal to neither imaginary parts of zeros nor poles of $F$, then we choose the branch by the integral
$\log{F(\sigma + it)} = \int_{\infty + it}^{\sigma + it}\frac{F'}{F}(z)dz$.
If $t \not= 0$ is equal to an imaginary part of a zero or a pole of $F$,
then we take $\log{F(\sigma + it)} = \lim_{\varepsilon \downarrow 0}\log{F(\sigma + i(t - \operatorname{sgn}(t)\varepsilon))}$.
Here, $\operatorname{sgn}$ is the signum function.
If there exists a pole or a zero of $F$ such that the imaginary part of zero, then we take $\log{F(\sigma)} = \lim_{\varepsilon \downarrow 0} \log{F(\sigma - \varepsilon)}$.
Next we introduce some notation.
\begin{notation*}
Let $H \geq 1$ be a real parameter.
The function $f: \mathbb{R} \rightarrow [0, +\infty)$ is mass one and supported on $[0, 1]$, and further $f$ is a
$C^{1}([0, 1])$-function, or $f$ belongs to $C^{d-2}(\mathbb{R})$ and is a $C^{d}([0, 1])$-function for some $d \geq 2$.
For such $f$'s, we define the number $D(f)$, and functions $u_{f, H}$, $v_{f, H}$ by
\begin{align} \label{def_D}
D(f) = \max\set{ d \in \mathbb{Z}_{\geq 1} \cup \{+\infty\}}{\text{$f$ is a $C^{d}([0, 1])$-function}},
\end{align}
$u_{f, H}(x) = Hf(H\log(x/e))/x$, and
\begin{gather}
\label{def_v}
v_{f, H}(y) = \int_{y}^{\infty}u_{f, H}(x)dx.
\end{gather}
Further, the function $U_{0}(z)$ is defined by
\begin{gather}
\label{def_U}
U_{0}(z) = \int_{0}^{\infty}u_{f, H}(x)E_{1}(z\log{x})dx
\end{gather}
for $\Im(z) \not= 0$.
Here, $E_{1}(z) = E_{1}(x+iy)$ is the exponential integral defined by
\begin{align*}
E_{1}(z)
:= \int_{x+iy}^{+\infty+iy}\frac{e^{-w}}{w} dw
= \int_{z}^{\infty}\frac{e^{-w}}{w} dw.
\end{align*}
When $\Im(z) = 0$, then $U_{0}(x) = \lim_{\varepsilon \uparrow 0}U_{0}(x + i\varepsilon)$.
Let $X \geq 3$ be a real parameter.
Let $\rho_{F} = \b_{F} + i\gamma_{F}$ be a nontrivial zero of $F$ with $\b_F$, $\gamma_F$ real numbers.
Here, nontrivial zeros refer to the zeros of $\Phi_F(s)$ that do not come from the Gamma factors $\gamma(s)$.
We also define $\sigma_{X, t}(F)$ for $F \not= 1$ and $w_{X}(y)$ by
\begin{gather}
\label{def_s_X_F}
\sigma_{X, t}(F)
= \frac{1}{2} + 2\max_{|t - \gamma_{F}| \leq \frac{X^{3|\b_{F} - 1/2|}}{\log{X}}}
\l\{ \b_{F} - \frac{1}{2}, \frac{2}{\log{X}} \r\},\\
\label{def_w_X}
w_{X}(y) = \l\{
\begin{array}{cl}
1 & \text{if\, $1 \leq y \leq X$,} \\[2mm]
\frac{(\log(X^3/y))^2 - 2(\log(X^2 / y))^2}{2(\log{X})^2} & \text{if\, $X \leq y \leq X^2$,} \\[2mm]
\frac{(\log(X^{3}/y))^2}{2(\log{X})^2} & \text{if\, $X^{2} \leq y \leq X^3$.}
\end{array}
\r.
\end{gather}
\end{notation*}
\begin{remark}
The number $\sigma_{X, t}(F)$ is well-defined as a finite value for every $X \geq 3$, $t \in \mathbb{R}$, $F \in \mathcal{S}^{\dagger} \setminus \{1\}$.
Actually, by the same method as Conrey-Ghosh \cite[Theorem 2]{CG1993}, we see that $d_{F} \geq 1$ or $F = 1$ for any $F \in \mathcal{S}^{\dagger}$.
We also find that there are infinitely many zeros of $F \in \mathcal{S}^{\dagger} \setminus \{1\}$ in the region $0 \leq \sigma \leq 1$
by the standard argument in the Riemann-von Mangoldt formula, and that $F$ does not have zeros for $\sigma > 1$ from (S1) and (S5).
Hence, we have $\frac{1}{2} + \frac{4}{\log{X}} \leq \sigma_{X, t}(F) \leq \frac{3}{2}$.
\end{remark}
Then we have the following theorem, which is a generalization of \cite[Theorem 1]{II2019} in the case when $F$ is the Riemann zeta-function $\zeta(s)$.
\begin{theorem} \label{Main_F_S}
Assume $F \in \mathcal{S}^{\dagger}$. Let $ D(f)\geq 2$, and $H$, $X$ be real parameters with $H \geq 1$, $X \geq 3$.
Then, for any $\sigma \geq 1/2$, $t \geq 14$, we have
\begin{align}
\label{EQ_F_S}
& \log{F(s)} = \\
& \sum_{2 \leq n \leq X^{1+1/H}}\frac{\Lambda_{F}(n)v_{f, H}\l( e^{\log{n}/\log{X}} \r)}{n^{s}\log{n}}
+\sum_{|s - \rho_{F}| \leq \frac{1}{\log{X}}}\log((s - \rho_{F})\log{X}) + R_{F}(s, X, H),
\end{align}
where the error term $R_{F}(s, X, H)$ satisfies \begin{align}
\label{ESR_S2}
& R_{F}(s, X, H)
\ll_{f} m_{F}\frac{X^{2(1-\sigma)} + X^{1-\sigma}}{t\log{X}} \\
& +H^{3}(\sigma_{X, t}(F) - 1/2)(X^{2(\sigma_{X, t}(F) - \sigma)} + X^{\sigma_{X, t}(F) - \sigma})
\l( \Bigg| \sum_{n \leq X^3}\frac{\Lambda_{F}(n)w_{X}(n)}{n^{\sigma_{X, t}(F) + it}} \Bigg| + d_{F}\log{t} \r)
\end{align}
for $|t| \geq t_{0}(F)$ with $t_{0}(F)$ a sufficiently large constant depending on $F$.
\end{theorem}
\begin{remark}
Note that in the above theorem we choose the branch of $\log{(s - \rho_{F})}$ as follows.
If $t \not= \gamma_{F}$, then $-\pi < \arg{(s - \rho_{F})} < \pi$,
and if $t = \gamma_{F}$, then $\arg{(s - \rho_{F})} = \lim_{\varepsilon \uparrow 0}\arg{(\sigma - \b_{F} + i\varepsilon)}$.
\end{remark}
\begin{remark}
Theorem \ref{Main_F_S} is a modification and generalization of the hybrid formula by Gonek, Hughes, and Keating \cite{GHK2007} to apply
the method of Selberg-Tsang \cite{KTDT}, where a different formula \cite[Lemma 5.4]{KTDT} was used.
With our formula, we can find the sign of the contribution from zeros to close $s$ by the form
$
\sum_{|s - \rho_{F}|\leq \frac{1}{\log{X}}}\log((s - \rho_{F})\log{X})
$.
This fact plays an important role in the proof of the theorems in Section \ref{General} in a fashion similar to the work of Soundararajan \cite{SM2009}.
\end{remark}
Theorem \ref{Main_F_S} can be obtained in the same method as the proof of \cite[Theorem 1]{II2019},
where we need the following proposition instead of \cite[Proposition 1]{II2019}.
\begin{proposition} \label{GSEF}
Let $F \in \mathcal{S}^{\dagger}$.
Let $X \geq 3$, $H \geq 1$ be real parameters.
Then, for any $s \in \mathbb{C}$, we have
\begin{align*}
\log{F(s)}
= & \sum_{2 \leq n \leq X^{1 + 1/H}}\frac{\Lambda_{F}(n)v_{f, H}(e^{\log{n}/\log{X}})}{n^{s} \log{n}}
+ m_{F}^{*}(U_{0}((s - 1)\log{X}) + U_{0}(s \log{X})) \\
& - \sum_{\substack{\rho_{F} \\ \rho_{F} \not= 0, 1}}U_{0}((s - \rho_{F})\log{X})
- \sum_{n = 0}^{\infty}\sum_{j = 1}^{k}U_{0}((s + (n+\mu_{j})/\lambda_{j})\log{X}),
\end{align*}
where the number $m_{F}^{*}$ is the integer such that the function $(s - 1)^{m_{F}^{*}}F(s)$ is entire and not equal to zero at $s = 1$.
\end{proposition}
Using Theorem \ref{Main_F_S}, we obtain the following propositions.
\begin{proposition} \label{KLI}
Let $F \in \mathcal{S}^{\dagger} $ satisfying \eqref{SNC} and (A3).
Let $\sigma \geq 1/2$, and $T$ be large.
Then, there exist positive constants $\delta_{F}$, $A_{1} = A_{1}(F)$ such that
for any $k \in \mathbb{Z}_{\geq 1}$, $3 \leq X \leq Y := T^{1/k}$,
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\bigg| \log{F(\sigma + it)} - P_{F}(\sigma + it, X)
- \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log((\sigma + it - \rho_{F})\log{Y}) \bigg|^{2k}dt \\
& \leq A_{1}^{k} k^{2k}T^{\delta_{F}(1 - 2\sigma)}
+ A_{1}^{k} k^{k} \l( \sum_{X < p \leq Y}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \r)^{k}.
\end{align*}
\end{proposition}
\begin{proposition} \label{KLST}
Suppose the same situation as Proposition \ref{KLI}.
Then, there exist positive constants $\delta_{F}$, $A_{2} = A_{2}(F)$ such that
for any $k \in \mathbb{Z}_{\geq 1}$, $3 \leq X \leq Y := T^{1/k}$,
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\l| \log{F(\sigma + it)} - P_{F}(\sigma + it, X) \r|^{2k}dt
\leq A_{2}^{k} k^{4k}T^{\delta_{F}(1 - 2\sigma)}
+ A_{2}^{k} k^{k} \l( \sum_{X < p \leq Y}\frac{|a_{F}(p)|^{2}}{p^{2\sigma}} \r)^{k}.
\end{align*}
\end{proposition}
Proposition \ref{KLST} has been essentially proved by Selberg \cite[Theorem 1]{S1992}.
Our range of $X$ and his are different, but this difference is filled by the argument in \cite[Sections 3, 5]{KTDT}.
On the other hand, we do not assume the Ramanujan Conjecture (S5).
Hence, we give the proof of Proposition \ref{KLST} for completeness.
\subsection{\textbf{Proof of Proposition \ref{GSEF} and Theorem \ref{Main_F_S}}}
We give the proofs of Proposition \ref{GSEF} and Theorem \ref{Main_F_S}, but the proofs are almost the same as
the proofs of \cite[Proposition 2]{II2019} and \cite[Theorem 1]{II2019}. Therefore, we give the sketches only.
\begin{lemma} \label{GHF}
Let $F \in \mathcal{S}^{\dagger} \setminus \{ 1 \}$. For all $s \in \mathbb{C}$ neither equaling to a pole nor a zero of $F$, we have
\begin{align} \label{GLF}
\frac{F'}{F}(\sigma + it)
= & \sum_{\substack{\rho_{F} \\ \rho_{F} \not= 0, 1}}\l( \frac{1}{s - \rho_{F}} + \frac{1}{\rho_{F}} \r)
+ \eta_F - m_{F}\l(\frac{1}{s - 1} + \frac{1}{s}\r) - \log{Q} \\
& - \sum_{n = 0}^{\infty}\sum_{j = 1}^{k}\lambda_{j}\frac{\Gamma'}{\Gamma}\l( \lambda_{j}s + \mu_{j} \r),
\end{align}
where $\eta_F$ is a complex number and satisfies $\Re(\eta_{F}) = -\Re\sum_{\rho_{F}}(1/\rho_{F})$.
In particular, for $|t| \geq t_{0}(F)$, we have
\begin{align} \label{GLFE}
\frac{F'}{F}(\sigma + it)
= \sum_{\substack{\rho_{F} \\ \rho_{F} \not= 0, 1}}\l( \frac{1}{s - \rho_{F}} + \frac{1}{\rho_{F}} \r)
+ O(d_{F} \log{|t|}).
\end{align}
\end{lemma}
\begin{proof}
We obtain equation \eqref{GLF} in the same way as the proof of \cite[eq. (10.29)]{MV}.
Moreover, by applying the Stirling formula to equation \eqref{GLF}, we can also obtain equation \eqref{GLFE}.
\end{proof}
\begin{lemma} \label{lem_far_zero}
Let $F \in \mathcal{S}^{\dagger}$.
For $|t| \geq t_0(F)$, $1 \leq H \leq \frac{|t|}{2}$, we have
\begin{gather}
\label{SIZDS}
\sum_{|t - \gamma_{F}| \leq H}1 \ll d_{F} H \log{|t|},\\
\label{EEZS}
\sum_{|t - \gamma_{F}| \geq H}\frac{1}{(t - \gamma_{F})^2}
\ll \frac{d_{F} \log{|t|}}{H}.
\end{gather}
\end{lemma}
\begin{proof}
Applying the Stirling formula to Lemma \ref{GHF},
for $|t| \geq t_{0}(F)$, we have
\begin{align*}
\Re\l(\frac{F'}{F}(H + it)\r)
= \sum_{\rho}\frac{H - \b_{F}}{(H - \b_{F})^2 + (t - \gamma_{F})^2} + O\l(d_{F} \log{|t|}\r).
\end{align*}
On the other hand, it holds that
\begin{gather*}
\sum_{\rho_{F}}\frac{H - \b_{F}}{(H - \b_{F})^2 + (t - \gamma_{F})^2}
\gg \sum_{|t - \gamma_{F}| \leq H}\frac{1}{H},\\
\sum_{\rho_{F}}\frac{H - \b_{F}}{(H - \b_{F})^2 + (t - \gamma_{F})^2}
\gg \sum_{|t - \gamma_{F}| \geq H}\frac{H}{(t - \gamma_{F})^2}.
\end{gather*}
Since $b_{F}(n)\log{n} \ll_{F} n^{1/2}$ from (S4), we find that
\begin{align*}
|(F'/F)(H + it)| = \bigg|\sum_{n = 2}^{\infty}b_{F}(n)\log{n}/n^{H + it}\bigg|
\ll_{F} \zeta(H - 1/2) - 1 \ll 2^{-H}.
\end{align*}
Hence we obtain \eqref{SIZDS} and \eqref{EEZS} for $H \geq 2$.
In addition, we immediately obtain these inequality the case
$1 \leq H \leq 2$ from \eqref{SIZDS} and \eqref{EEZS} in the case $H = 2$.
\end{proof}
\begin{lemma} \label{ZAL}
Let $F \in \mathcal{S}^{\dagger}$. For any $T \geq t_{0}(F)$, there exists some $t \in [T, T + 1]$ such that,
uniformly for $1/2 \leq \sigma \leq 2$,
\begin{align*}
\frac{F'}{F}(\sigma + it)
\ll_{F} (\log{T})^{2}.
\end{align*}
\end{lemma}
\begin{proof}
Using Lemma \ref{lem_far_zero}, we obtain this lemma in the same argument as the proof of \cite[Lemma 12.2]{MV}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{GSEF}]
By using Lemma \ref{ZAL}, we obtain Proposition \ref{GSEF}
in the same method as the proof of \cite[Proposition 1]{II2019}.
\end{proof}
\begin{lemma} \label{GUE}
Let $d$ be a nonnegative integer with $d < D = D(f) + 1$.
Let $s = \sigma + it$ be a complex number.
Set $X \geq 3$ be a real parameter.
Then, for any $z = a + ib$ with $a \in \mathbb{R}$, $b \in \mathbb{R} \setminus \{ t \}$, we have
\begin{align*}
U_{0}((s - z)\log{X})
\ll_{f, d} \frac{X^{(1+1/H)(a - \sigma)} + X^{a - \sigma}}{|t - b|\log{X}}
\min_{0 \leq l \leq d}\l\{\l(\frac{H}{|t - b|\log{X}} \r)^{l}\r\}.
\end{align*}
\end{lemma}
\begin{proof}
This lemma can be proved in the same way as \cite[Lemma 2]{II2019}.
\end{proof}
\begin{lemma} \label{EUE}
Let $s = \sigma + it$ be a complex number.
Set $X \geq 3$ be a real parameter.
Then, for any complex number $z = a + ib$ with $|t - b| \leq 1 / \log{X}$, we have
\begin{align}
U_{0}((s - z)\log{X})
= \l\{
\begin{array}{cl}
-\log((s - z)\log{X}) + O(1) & \text{if \; $|s - z| \leq 1/\log{X}$, } \\[5mm]
O\l(X^{(1+1/H)(a - \sigma)} + X^{a - \sigma} \r) & \text{if \; $|s - z| > 1 / \log{X}$.}
\end{array}
\r.
\end{align}
\end{lemma}
\begin{proof}
This lemma can be proved in the same way as \cite[Lemma 3]{II2019}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{Main_F_S}]
It follows from Proposition \ref{GSEF}, Lemma \ref{GUE}, and Lemma \ref{EUE} that
\begin{align*}
\log{F(s)} =
\sum_{2 \leq n \leq X^{1+1/H}}\frac{\Lambda_{F}(n)v_{f, H}\l( e^{\log{n}/\log{X}} \r)}{n^{s}\log{n}}
+\sum_{|s - \rho_{F}| \leq \frac{1}{\log{X}}}\log((s - \rho_{F})\log{X}) + R_{F}(s, X, H),
\end{align*}
where
\begin{multline}
R_{F}(s, X, H)
\ll_{f, d} \frac{m_{F}(X^{2(1-\sigma)} + X^{1-\sigma})}{t\log{X}}
+ \sum_{|t - \gamma_{F}| \leq \frac{1}{\log{X}}}(X^{2(\b_{F} - \sigma)} + X^{\b_{F} - \sigma})\\
+ \frac{1}{\log{X}}\sum_{|t - \gamma_{F}| > \frac{1}{\log{X}}}
\frac{X^{2(\b_{F} - \sigma)} + X^{\b_{F} - \sigma}}{|t - \gamma_{F}|}
\min_{0 \leq l \leq d}\l\{\l( \frac{H}{|t - \gamma_{F}|\log{X}} \r)^{l}\r\}.
\end{multline}
Hence, it suffices to show estimate \eqref{ESR_S2} on the range $|t| \geq t_{0}(F)$.
Following the proof of \cite[Proposition 2]{II2019}, we see that it suffices to check
\begin{align} \label{KSIE}
\sum_{\rho_{F}}\frac{\sigma_{X, t}(F) - 1/2}{(\sigma_{X, t}(F) - \b_{F})^2 + (t - \gamma_{F})^2}
\ll \l| \sum_{n \leq X^3}\frac{\Lambda_{F}(n)w_{X}(n)}{n^{\sigma_{X, t}(F) + it}} \r| + d_{F} \log{|t|},
\end{align}
which can be shown by the same proofs as in \cite[eq. (4.4); eq. (4.7)]{SCR} by using equation \eqref{GLFE} instead of \cite[Lemma 11]{SCR}.
\end{proof}
\subsection{Proof of Propositions \ref{KLI} and Proposition \ref{KLST}}
\begin{lemma} \label{SLL}
Let $T \geq 5$, and let $X \geq 3$. Let $k$ be a positive integer such that $X^{k} \leq T$.
Then, for any complex numbers $a(p)$, we have
\begin{align*}
\int_{T}^{2T}\bigg| \sum_{p \leq X}\frac{a(p)}{p^{it}} \bigg|^{2k}dt
\ll T k! \l( \sum_{p \leq X}|a(p)|^2 \r)^{k}.
\end{align*}
Here, the above sums run over prime numbers.
\end{lemma}
\begin{proof}
By \cite[Lemma 3.3]{KTDT}, we have
\begin{align*}
\int_{T}^{2T}\bigg| \sum_{p \leq X}\frac{a(p)}{p^{it}} \bigg|^{2k}dt
&\leq T k! \l( \sum_{p \leq X}|a(p)|^2 \r)^{k}
+ O\l( k! \l( \sum_{p \leq X}p|a(p)|^2 \r)^{k} \r)\\
&\leq T k! \l( \sum_{p \leq X}|a(p)|^2 \r)^{k}
+ O\l( k! X^{k}\l( \sum_{p \leq X}|a(p)|^2 \r)^{k} \r).
\end{align*}
Hence, we obtain this lemma when $X^{k} \leq T$.
\end{proof}
The next lemma is an analogue and a generalization of \cite[Lemma 5.2]{KTDT}.
\begin{lemma} \label{SL_s_X}
Let $F \in \mathcal{S}^{\dagger} \setminus \{ 1 \}$
be an $L$-function satisfying (A3).
Let $T$ be large, and $\kappa_{F}$ be the positive constant in \eqref{ZDC1}.
For $k \in \mathbb{Z}_{\geq 1}$, ${3 \leq X \leq T^{2/3}}$, $\xi \geq 1$
with $X \xi \leq T^{\kappa_{F} / 4}$, we have
\begin{align*}
\int_{T}^{2T}\l( \sigma_{X, t}(F) - \frac{1}{2} \r)^{k}\xi^{\sigma_{X, t}(F) - 1/2}dt
\ll_{F} T \l( \frac{4^{k} \xi^{\frac{4}{\log{X}}}}{(\log{X})^{k}} + \frac{C^{k} k!}{\log{X}(\log{T})^{k-1}} \r),
\end{align*}
where $C = C(F)$ is a positive constant.
\end{lemma}
\begin{proof}
By definition \eqref{def_s_X_F} of $\sigma_{X, t}(F)$, we obtain
\begin{align} \label{SL_INE1}
& \int_{T}^{2T}\l( \sigma_{X, t}(F) - \frac{1}{2} \r)^{k}\xi^{\sigma_{X, t}(F) - 1/2}dt \\
& \leq T \xi^{\frac{4}{\log{X}}}\l( \frac{4}{\log{X}} \r)^{k}
+ \frac{2^{k}}{\log{X}}
\sum_{\substack{T - \frac{X^{3|\b_{F} - 1/2|}}{\log{X}} \leq \gamma_{F} \leq 2T + \frac{X^{3|\b_{F} - \frac{1}{2}|}}{\log{X}}
\\ \b_{F} \geq 1/2}}
\l(\b_{F} - \frac{1}{2} \r)^{k}(X^3 \xi^2)^{\b_{F} - \frac{1}{2}}.
\end{align}
From the equation
\begin{align*}
& \l(\b_{F} - \frac{1}{2} \r)^{k} (X^3 \xi^2)^{\b_{F} - \frac{1}{2}}
= \int_{\frac{1}{2}}^{\b_{F}}
\l\{ (\sigma - 1/2)^{k}\log{(X^3 \xi^2)} + k \l(\sigma - 1/2 \r)^{k-1} \r\}(X^3 \xi^2)^{\sigma - \frac{1}{2}}d\sigma,
\end{align*}
we find that
\begin{align*}
& \sum_{\substack{T - \frac{X^{3|\b_{F} - 1/2|}}{\log{X}} \leq \gamma_{F} \leq 2T + \frac{X^{3|\b_{F} - \frac{1}{2}|}}{\log{X}}
\\ \b_{F} \geq 1/2}}
\l(\b_{F} - \frac{1}{2} \r)^{k}(X^3 \xi^2)^{\b_{F} - \frac{1}{2}} \\
& \leq \sum_{\substack{0 \leq \gamma_{F} \leq 3T \\ \b_{F} \geq 1/2}}\int_{\frac{1}{2}}^{\b_{F}}
\l\{ (\sigma - 1/2)^{k}\log{(X^3 \xi^2)} + k \l(\sigma - 1/2 \r)^{k-1} \r\}(X^3 \xi^2)^{\sigma - \frac{1}{2}}d\sigma \\
& \leq \int_{\frac{1}{2}}^{1}\l\{ (\sigma - 1/2)^{k}\log{(X^3 \xi^2)} + k \l(\sigma - 1/2 \r)^{k-1} \r\}(X^3 \xi^2)^{\sigma - \frac{1}{2}}
\sum_{\substack{0 \leq \gamma_{F} \leq 3T \\ \b_{F} \geq \sigma}}1 d\sigma\\
& = \int_{\frac{1}{2}}^{1}\l\{ (\sigma - 1/2)^{k}\log{(X^3 \xi^2)} + k \l(\sigma - 1/2 \r)^{k-1} \r\}(X^3 \xi^2)^{\sigma - \frac{1}{2}}
N_{F}(\sigma, 3T)d\sigma.
\end{align*}
By assumption (A3), we can use the estimate $N_{F}(\sigma, T) \ll_{F} T^{1 - \kappa_{F}(\sigma - 1/2)}\log{T}$,
and so, for $X \xi \leq T^{\kappa_{F}/4}$, the above most right hand side is
\begin{align*}
& \ll_{F} T \log{T}\int_{\frac{1}{2}}^{1}\l\{ (\sigma - 1/2)^{k}\log{(X^3 \xi^2)} + k \l(\sigma - 1/2 \r)^{k-1} \r\}
\l(\frac{X^3 \xi^2}{T^{\kappa_{F}}}\r)^{\sigma - \frac{1}{2}}d\sigma
\ll T \frac{C^{k} k!}{(\log{T})^{k-1}}
\end{align*}
for some $C = C(F) > 0$.
Hence, by this estimate and inequality \eqref{SL_INE1}, we obtain this lemma.
\end{proof}
\begin{lemma}\label{MomentLambdab}
Let $F$ be a Dirichlet series satisfying (S4), (S5), and \eqref{SNC}.
Suppose there exists some positive constant $A$ such that $|w(n)|\leq (\log n)^A$.
Then for any $c \in \mathbb{R}$, $\sigma\geq \frac{1}{2}$, $k \in \mathbb{Z}_{\geq 1}$ and any $X \geq Z \geq 2$ with $X^{k} \leq T$, we have
\begin{align}
& \int_T^{2T}\left| \sum_{Z \leq n \leq X}\frac{b_{F}(n)w(n) (\log(X n))^{c}}{n^{\sigma + it }} \right| ^{2k}dt \\
& \leq T C^{k} k^{k} \l((\log{X})^{2c} \sum_{Z \leq p \leq X}\frac{|b_{F}(p) w(p)|^2}{p^{2\sigma}}\r)^k
+ T C^{k} k^{k} Z^{k(1 - 2\sigma)}(\log{X})^{2(A + c)k},
\end{align}
where $C=C(F)$ is some positive constant.
\end{lemma}
\begin{proof}
Set $K_{1} = 2(1/2-\vartheta_F)^{-1}$.
Using $|w(n)| \leq (\log n)^A$, we find that for any $\sigma \geq \frac{1}{2}$
\begin{align}
\Bigg|\sum_{\ell > K_{1}}\sum_{Z\leq p^{\ell} \leq X}
\frac{b_{F}(p^{\ell}) w(p^{\ell}) (\log(X p^{\ell}))^{c}}{p^{\ell(\sigma + it)}}\Bigg|
& \ll_{F} (\log{X})^{c}\sum_{p>Z}\sum_{\ell > K_{1}}\frac{(\log (p^\ell))^A}{p^{\ell(\sigma-1/2)}p^{(1/2-\theta_F) \ell}} \\
& \ll_{F} (\log{X})^{c} Z^{1/2-\sigma}\sum_{p}\frac{1}{p^{2-\epsilon}}
\ll_{F} (\log{X})^{c} Z^{1/2-\sigma}.\label{LargePrimespower}
\end{align}
Therefore, by Lemma \ref{SLL}, we have
\begin{align}
& \int_T^{2T}\left|\sum_{Z\leq n\leq X}\frac{b_F(n)w(n)(\log(Xn))^{c}}{n^{\sigma+it}}\right|^{2k}dt \\
& \leq T C^k k^k \sum_{1 \leq \ell \leq K_1}
\l(\sum_{Z^{1/\ell}\leq p\leq X^{1/\ell}}\frac{|b_F(p^\ell) w(p^\ell) (\log(X p^{\ell}))^{c}|^2}{p^{2\ell \sigma}}\r)^k
+ T C^{k} Z^{k(1 - 2\sigma)} (\log{X})^{2ck}
\end{align}
for some constant $C=C(F)>0$.
From assumption (S5), we see that, for $2 \leq \ell \leq K_{1}$,
\begin{align*}
\sum_{Z^{1/\ell}\leq p\leq X^{1/\ell}}\frac{|b_F(p^\ell) w(p^\ell) (\log(X p^{\ell}))^{c}|^2}{p^{2\ell \sigma}}
\leq (\log{X})^{2(A + c)}Z^{1 - 2\sigma}\sum_{p}\frac{|b_{F}(p^{\ell})|^{2}}{p^{\ell}}
\ll_{F} Z^{1 - 2\sigma} (\log{X})^{2(A + c)},
\end{align*}
which completes the proof of this lemma.
\end{proof}
\begin{lemma}\label{SumbLargeP}
Let $F$ be a Dirichlet series satisfying (S4), (S5), and \eqref{SNC}.
Then for any $x \geq 3$ and $c > 1$,
\begin{gather}
\label{SumbLargeP1}
2\sigma_{F}(x)^{2}
= \sum_{n\leq x}\frac{|b_F(n)|^2}{n} + O_{F}(1)
=\sum_{p\leq x}\frac{|b_F(p)|^2}{p} + O_{F}(1)
=n_F\log \log x+O_{F}(1),\\
\label{SumbLargeP2}
\sum_{n \geq x }\frac{|b_F(n)|^2}{n^c}\ll x^{1-c}, \quad x \rightarrow \infty,
\end{gather}
where $\sigma_{F}(x)$ is defined by \eqref{def_var}.
\end{lemma}
\begin{proof}
Set $K_1=2(1/2-\vartheta_F)^{-1}$.
Similar to \eqref{LargePrimespower}, we have
\begin{align}
\sum_{\ell \geq K_1}\sum_{p}\frac{|b_F(p^{\ell})|^2}{p^{\ell}}
\ll_{F} \sum_{p}\sum_{\ell}\frac{1}{p^{2}}
\ll 1.
\end{align}
From assumption (S5), we also have
\begin{align}
\sum_{2 \leq \ell \leq K_1} \sum_{p}\frac{|b_F(p^\ell)|^2}{p^\ell}
\ll_{F} 1,
\end{align}
and thus
\begin{align*}
\sum_{\ell = 2}^{\infty} \sum_{p}\frac{|b_F(p^\ell)|^2}{p^\ell}
\ll_{F} 1.
\end{align*}
Therefore, by assumption \eqref{SNC}
\begin{align}
2\sigma_{F}(x)^2
&= \sum_{p \leq x}\sum_{\ell = 1}^{\infty}\frac{|b_F(p^{\ell})|^2}{p^{\ell}}
= \sum_{n\leq x}\frac{|b_F(n)|^2}{n} + O_{F}(1)\\
&= \sum_{p\leq x}\frac{b_F(p)}{p} + O_{F}(1)
= n_F\log\log x + O_{F}(1),
\end{align}
which is the assertion of \eqref{SumbLargeP1}.
From partial summation and \eqref{SumbLargeP1}, we also find that
\begin{align}
\sum_{n\geq x}\frac{|b_F(n)|^2}{n^c}
& =\int_{x}^{\infty}\frac{1}{\xi^{c-1}}d\sum_{x \leq n \leq \xi}\frac{|b_F(n)|^2}{n} \\
& = c\int_{x}^{\infty}\frac{n_{F}\log{\log{\xi} - n_{F}\log{\log{x}} + O_{F}(1)}}{\xi^{c}}
\ll_{F} x^{1-c}.
\end{align}
\end{proof}
\begin{lemma} \label{ESMDP}
Let $F \in \mathcal{S}^{\dagger}$ be an $L$-function satisfying \eqref{SNC} and (A3).
Let $T$ be large. Put $\delta_{F} = \min\{ \frac{1}{4}, \frac{\kappa_{F}}{20} \}$
with $\kappa_{F}$ the positive constant in \eqref{ZDC1}.
For any $k \in \mathbb{Z}_{\geq 1}$, $X \geq 3$ with $X \leq T^{\delta_{F} / k}$, we have
\begin{align*}
\int_{T}^{2T}\Bigg| \sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{\sigma_{X, t}(F) + it}} \Bigg|^{2k}dt
\ll T C^{k} k^{k} (\log{X})^{2k},
\end{align*}
where $w_{X}$ is the smoothing function defined by \eqref{def_w_X}, and $C = C(F)$ is a positive constant.
\end{lemma}
\begin{proof}
For brevity, we write $\sigma_{X, t}(F)$ as $\sigma_{X, t}$.
\begin{align*}
\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{\sigma_{X, t} + it}}
& = \sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{1/2 + it}}
- \sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{1/2 + it}}(1 - n^{1/2 - \sigma_{X, t}}) \\
& = \sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{1/2 + it}}
-\int_{1/2}^{\sigma_{X, t}}\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{n}}{n^{\a' + it}}d\a',
\end{align*}
Note that for $1/2 \leq \a' \leq \sigma_{X, t}$,
\begin{align*}
\bigg| \sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{n}}{n^{\a' + it}} \bigg|
& = \Bigg|X^{\a' - 1/2} \int_{\a'}^{\infty}X^{1/2 - \a}
\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}{n^{\a + it}}d\a\Bigg| \\
& \leq X^{\sigma_{X, t} - 1/2}\int_{1/2}^{\infty}X^{1/2 - \a}
\bigg|\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{(Xn)} \log{n}}{n^{\a + it}}\bigg|d\a.
\end{align*}
Thus
\begin{align}\label{S1+S2}
& \frac{1}{T}\int_T^{2T}\left|\sum_{n\leq X^3}\frac{\Lambda_F(n)w_X(n)}{n^{\sigma_{X,t}+it}} \right|^{2k}dt\ll C_1^k (\mathcal{S}_1+\mathcal{S}_2),
\end{align}
where
\begin{align}
& \mathcal{S}_1=\frac{1}{T}\int_{T}^{2T}\left|\sum_{n\leq X^3}\frac{\Lambda_F(n)w_X(n)}{n^{1/2+it}}\right|^{2k}dt, \\
& \mathcal{S}_2=\frac{1}{T}\int_T^{2T}\left| (\sigma_{X,t}-\frac{1}{2})X^{\sigma_{X, t} - 1/2}\int_{1/2}^{\infty}X^{1/2 - \a}
\bigg|\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{(Xn)} \log{n}}{n^{\a + it}}
\bigg|d\a\right|^{2k}dt.
\end{align}
Since $\Lambda_F(n)=b_F(n)\log n$ and $b_F(p)=a_F(p)$, we have
\begin{align}
\sum_{n \leq X^{3}}\frac{|\Lambda_{F}(n)|^2}{n}
& = (\log{X^{3}})^{2}\sum_{p \leq X^{3}}\frac{|a_{F}(p)|^2}{p}
- \int_{2}^{X^{3}}\frac{2\log{\xi}}{\xi}\sum_{p \leq \xi}\frac{|a_{F}(p)|^2}{p} d\xi+O(1) \\ \nonumber
& = n_{F}(\log{X^{3}})^{2}\log{\log{X^3}}
- 2n_{F}\int_{2}^{X^{3}}\frac{\log{\xi} \times \log{\log{\xi}}}{\xi}d\xi
+ O_{F}\l( (\log{X})^{2} \r) \\ \nonumber
& = n_{F}(\log{X^{3}})^{2}\log{\log{X^3}}
- n_{F}(\log{X^{3}})^{2}\log{\log{X^3}} + \frac{n_{F}}{2}(\log{X^3})^2
+ O_{F}\l( (\log{X})^{2} \r) \\
\label{Lambda^2/p}
& \ll_{F} (\log{X})^{2}.
\end{align}Combining Lemma \ref{MomentLambdab} with \eqref{Lambda^2/p}, we obtain
\begin{align}\label{S1}
\mathcal{S}_1\ll C_2^k k^k \left(\sum_{n\leq X^3}\frac{|b_F(n)|^2\log^2 n}{n}\right)^k\ll C_3^k k^k (\log X)^{2k}.
\end{align}
By the Cauchy-Schwarz inequality and Lemma \ref{SL_s_X}, when $\delta_{F} \leq \kappa_{F} / 20$,
it holds that
\begin{align}
\mathcal{S}_2
& \leq \Bigg(\int_{T}^{2T}(\sigma_{X, t}-1/2)^{4k}X^{4k(\sigma_{X, t} - 1/2)}dt\Bigg)^{1/2} \times \\
\nonumber
& \qquad \quad \times \l(\int_{T}^{2T}\l( \int_{1/2}^{\infty}X^{1/2 - \a}
\bigg|\sum_{n \leq X^{3}}
\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}{n^{\a + it}}\bigg|d\a\r)^{4k}dt
\r)^{\frac{1}{2}} \\
& \ll\frac{T^{\frac{1}{2}} C_{4}^{k}}{(\log{X})^{2k}}
\l(\int_{T}^{2T}\l( \int_{1/2}^{\infty}X^{1/2 - \a}
\bigg|\sum_{n \leq X^{3}}
\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}{n^{\a + it}}\bigg|d\a\r)^{4k}dt\r)^{\frac{1}{2}}\label{S21},
\end{align}
for some constant $C_{4} = C_{4}(F)>0$.
We apply H\"older's inequality to the innermost integral to obtain
\begin{align*}
& \l( \int_{1/2}^{\infty}X^{\frac{1}{2} - \a}
\bigg|\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}{n^{\a + it}}
\bigg|d\a\r)^{4k} \\
& \leq \l(\int_{1/2}^{\infty}X^{\frac{1}{2} - \a}d\a\r)^{4k-1} \times
\l( \int_{1/2}^{\infty} X^{\frac{1}{2} - \a}\bigg|
\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}
{n^{\a + it}}\bigg|^{4k}d\a \r) \\
& = \frac{1}{(\log{X})^{4k-1}}
\int_{1/2}^{\infty} X^{1/2 - \a}\bigg|\sum_{n \leq X^{3}}
\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}{n^{\a + it}}\bigg|^{4k}d\a.
\end{align*}
Therefore, by Lemma \ref{MomentLambdab} and \eqref{Lambda^2/p}, we find that
\begin{align}
& \int_{T}^{2T}\l( \int_{1/2}^{\infty}X^{\frac{1}{2} - \a}
\bigg|\sum_{n \leq X^{3}}\frac{\Lambda_{F}(n) w_{X}(n) \log{(X n)} \log{n}}{n^{\a + it}}
\bigg|d\a\r)^{4k}dt\nonumber \\
& \leq \frac{1}{(\log{X})^{4k-1}}\int_{1/2}^{\infty} X^{1/2-\a}\l(
\int_{T}^{2T}\bigg|\sum_{n \leq X^{3}}\frac{b_{F}(n) \log{(X n)} \log{n}}{n^{\a + it}}
\bigg|^{4k}dt\r)d\a\nonumber \\
& \leq \frac{T C^{k} k^{2k}}{(\log{X})^{4k-1}}\int_{1/2}^{\infty}X^{1/2-\a}
\l\{\l( (\log{X})^{2}\sum_{n \leq X^{3}}\frac{|\Lambda_{F}(n)|^2 (\log{n})^2}{n^{2 \a}} \r)^{2k} + (\log{X})^{8k} \r\}d\a \\
& \ll \frac{T C^{k} k^{2k}}{(\log{X})^{4k}}
\l( \l((\log{X})^{4} \sum_{n \leq X^{3 }}\frac{|\Lambda_{F}(n)|^2}{n} \r)^{2k} + (\log{X})^{8k} \r) \\
& \ll T C_{5}^{k} k^{2k} (\log{X})^{8k}.\label{S22}
\end{align}
Combining \eqref{S1+S2}, \eqref{S1}, \eqref{S21} and \eqref{S22}, we complete the proof of Lemma \ref{ESMDP}.
\end{proof}
\begin{lemma} \label{ESRSIZ}
Let $F \in \mathcal{S}^{\dagger}$ be an $L$-function satisfying \eqref{SNC} and (A3).
Let $\sigma \geq 1/2$, $T$ be large.
Let $\kappa_{F}$, $\delta_{F}$ be the same constants as in Lemma \ref{ESMDP}.
There exists a positive constant $C = C(F)$ such that
for any $k \in \mathbb{Z}_{\geq 1}$, $3 \leq X \leq T^{\delta_{F}/k}$,
\begin{align} \label{ESRSIZ1}
\int_{T}^{2T}\l( \sum_{|\sigma + it - \rho_F| \leq \frac{1}{\log{X}}}1 \r)^{2k}dt
\leq C^{k} T^{1 - (2\sigma - 1)\delta_{F} + \frac{8\delta_{F}}{\log{X}}}\l( \frac{\log{T}}{\log{X}} \r)^{2k},
\end{align}
and
\begin{align} \label{ESRSIZ2}
\begin{aligned}
& \int_{T}^{2T}\Bigg| \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}}\log((\sigma + it - \rho_{F}) \log{X}) \Bigg|^{k} dt \\
& \leq (C k)^{k} T^{1 - (\sigma - 1/2)\delta_{F} + \frac{4\delta_{F}}{\log{X}}}
\l( \frac{\log{T}}{\log{X}} \r)^{k+\frac{1}{2}}.
\end{aligned}
\end{align}
\end{lemma}
\begin{proof}
From \eqref{def_s_X_F},
there are no zeros of $F$ with $|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}$ when $\sigma \geq \sigma_{X, t}(F)$.
Put $\xi := T^{\delta_{F} / k}$.
Note that $\xi \geq 1$ because we suppose that $3 \leq X \leq T^{\delta_{F}/k} = \xi$.
From these facts, we have
\begin{align*}
\sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}}1
\leq \xi^{\sigma_{X, t}(F) - \sigma}\sum_{|t - \gamma_{F}| \leq \frac{1}{\log{X}}}1
\end{align*}
for $\sigma \geq 1/2$.
By definition \eqref{def_s_X_F} and the line symmetry of nontrivial zeros of $F$, we find that
\begin{align*}
\sum_{|t - \gamma_{F}| \leq \frac{1}{\log{X}}}1
\leq 2\sum_{\substack{|t - \gamma_{F}| \leq \frac{1}{\log{X}} \\ \b_{F} \geq 1/2}}1
\ll \sum_{\substack{|t - \gamma_{F}| \leq \frac{1}{\log{X}} \\ \b_{F} \geq 1/2}}
\frac{(\sigma_{X, t}(F) - 1/2)^2}{(\sigma_{X, t}(F) - \b_{F})^2 + (t - \gamma_{F})^2}.
\end{align*}
Applying estimate \eqref{KSIE} to the above right hand side, we obtain
\begin{align} \label{NZVDP}
\sum_{|t - \gamma_{F}| \leq \frac{1}{\log{X}}}1
\ll (\sigma_{X, t}(F) - 1/2)\l(\bigg| \sum_{n \leq X^3}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{\sigma_{X, t}(F) + it}}\bigg| + d_{F}\log{T} \r)
\end{align}
for $t \in [T, 2T]$.
Noting $X \xi^{2k} \leq T^{\kappa_{F} / 4}$ and using Lemmas \ref{SL_s_X}, \ref{ESMDP},
we have
\begin{align}
& \int_{T}^{2T}(\sigma_{X, t}(F) - 1/2)^{2k}\xi^{2k(\sigma_{X, t}(F) - \sigma)}
\l(\bigg| \sum_{n \leq X^3}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{\sigma_{X, t}(F) + it}}\bigg| + d_{F}\log{T} \r)^{2k}dt \\
& \leq C^{k}\xi^{2k(1/2 - \sigma)}\Bigg\{(\log{T})^{2k}
\int_{T}^{2T}(\sigma_{X, t}(F) - 1/2)^{2k}\xi^{2k(\sigma_{X, t}(F)-1/2)}dt \\
& + \l(\int_{T}^{2T}(\sigma_{X, t}(F) - 1/2)^{2k}\xi^{2k(\sigma_{X, t}(F)-1/2)}dt\r)^{1/2}
\l( \int_{T}^{2T}\bigg| \sum_{n \leq X^3}\frac{\Lambda_{F}(n) w_{X}(n)}{n^{\sigma_{X, t}(F) + it}}\bigg|^{2k}dt \r)^{1/2}
\Bigg\} \\
& \leq C^{k} \xi^{2k(1/2 - \sigma)} \l(T \xi^{\frac{8k}{\log{X}}} \l( \frac{\log{T}}{\log{X}} \r)^{2k}
+ T \xi^{\frac{4k}{\log{X}}} k^{k/2}\r) \\
\label{PESRSIZ1}
& \leq C_{2}^{k} T^{1 - (2\sigma - 1)\delta_{F} + \frac{8\delta_{F}}{\log{X}}}\l( \frac{\log{T}}{\log{X}} \r)^{2k}
\end{align}
for some constant $C_{2} = C_{2}(F) > 0$.
Hence, we obtain estimate \eqref{ESRSIZ1}.
Next, we show estimate \eqref{ESRSIZ2}.
We find that
\begin{align*}
\Bigg|\sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}}\log\l( (\sigma + it - \rho_{F})\log{X} \r)\Bigg|
\leq \l(g_{X}(s) + \pi\r)
\times \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}}1,
\end{align*}
where $g_{X}(s) = g_{X}(\sigma+it)$ is the function defined by
\begin{align*}
g_{X}(s) = \l\{
\begin{array}{cl}
\log\l(\frac{1}{|(\sigma + it - \rho_{s})\log{X}|}\r)
& \text{if there exists a zero $\rho_{F}$ with $|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}$,} \\
0 & \text{otherwise.}
\end{array}
\r.
\end{align*}
Here, $\rho_{s}$ indicates the zero of $F$ nearest from $s = \sigma + it$.
By using the Cauchy-Schwarz inequality and estimate \eqref{ESRSIZ1}, we obtain
\begin{align}
\label{PESRSIZ2}
& \int_{T}^{2T}\Bigg| \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}}\log((\sigma + it - \rho_{F}) \log{X}) \Bigg|^{k} dt \\
& \leq C_{3}^{k}\l(\int_{T}^{2T} g_{X}(\sigma+it)^{2k}dt + \pi^{2k} T\r)^{1/2}
\times T^{\frac{1}{2} - (\sigma - 1/2)\delta_{F} + \frac{4\delta_{F}}{\log{X}}}
\l( \frac{\log{T}}{\log{X}} \r)^{k}
\end{align}
for some constant $C_{3} = C_{3}(F) > 0$.
Moreover, we find that
\begin{align*}
\int_{T}^{2T}g_{X}(s)^{2k}dt
& \leq \int_{T}^{2T}\sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{X}}}
\l( \log{\l( \frac{1}{|\sigma+it-\rho_{F}|\log{X}} \r)} \r)^{2k}dt \\
& \leq \sum_{T-\frac{1}{\log{X}} \leq \gamma_{F} \leq 2T + \frac{1}{\log{X}}}
\int_{\gamma_{F} - \frac{1}{\log{X}}}^{\gamma_{F} + \frac{1}{\log{X}}}
\l(\log\l( \frac{1}{|t - \gamma_{F}|\log{X}} \r)\r)^{2k}dt \\
& \ll \frac{1}{\log{X}}\sum_{T-1 \leq \gamma_{F} \leq 2T + 1}
\int_{0}^{1}\l(\log\l( \frac{1}{\ell} \r)\r)^{2k}d\ell \\
& \ll_{F} T\frac{\log{T}}{\log{X}}\int_{0}^{1}\l(\log\l( \frac{1}{\ell} \r)\r)^{2k}d\ell.
\end{align*}
By induction, we can easily confirm that the last integral is equal to $(2k)!$. Hence, we obtain
\begin{align*}
\int_{T}^{2T}g_{X}(s)^{2k}dt
\ll_{F} (2k)!T\frac{\log{T}}{\log{X}}.
\end{align*}
By substituting this estimate to inequality \eqref{PESRSIZ2}, we obtain this lemma.
\end{proof}
\begin{proof}[Proof of Proposition \ref{KLI}]
Let $f$ be a fixed function satisfy the condition of this paper (see Notation) and $D(f) \geq 2$.
Let $\sigma \geq 1/2$, $k \in \mathbb{Z}_{\geq 1}$, and $3 \leq Z := T^{\delta_{F} / k}$,
where $\delta_{F} = \min\{ \frac{\kappa_{F}}{20}, \frac{1}{4} \}$.
It holds that
\begin{align*}
Z^{2(\sigma_{Z, t}(F) - \sigma)} + Z^{\sigma_{Z, t}(F) - \sigma}
&= Z^{2(1/2-\sigma)} \cdot Z^{2(\sigma_{Z, t}(F) - 1/2)} + Z^{1/2-\sigma} \cdot Z^{\sigma_{Z, t}(F) - 1/2}\\
&\leq 2Z^{1/2 - \sigma} \cdot Z^{2(\sigma_{Z, t}(F) - 1/2)}
\end{align*}
for $\sigma \geq 1/2$.
Using this inequality and estimate \eqref{ESR_S2} as $H = 1$, we find that
there exists a positive constant $C_{1} = C_{1}(F)$ such that
\begin{align} \label{PKLI0}
& \bigg| \log{F(\sigma + it)} - P_{F}(\sigma + it, Z)
- \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z}) \bigg|^{2k} \\
& \leq C_{1}^{k}\bigg| \sum_{Z < n \leq Z^{2}}
\frac{\Lambda_{F}(n) v_{f, 1}(e^{\log{n} / \log{Z}})}{n^{\sigma+it}\log{n}} \bigg|^{2k} + \\ \nonumber
& \quad + C_{1}^{k}Z^{2k(1/2-\sigma)}(\sigma_{Z, t}(F) - 1/2)^{2k}Z^{4k(\sigma_{Z, t}(F) - 1/2)}
\l( \bigg| \sum_{n \leq Z^3}\frac{\Lambda_{F}(n) w_{Z}(n)}{n^{\sigma_{Z, t}(F) + it}} \bigg| + \log{T} \r)^{2k}
\end{align}
for $t \in [T, 2T]$.
Since $|v_{f, 1}(e^{\log{p^{\ell}} / \log{Z}})| \leq 1$, $\Lambda_{F}(n) = b_{F}(n)\log n$, and $a_{F}(p) = b_{F}(p)$,
we can apply Lemma \ref{MomentLambdab} and Lemma \ref{SumbLargeP} to bound the first sum in \eqref{PKLI0} so that
\begin{align}
\hspace{-5mm}
\label{PKLI3}
\frac{1}{T}\int_{T}^{2T}\bigg| \sum_{Z < n \leq Z^{2}}
\frac{\Lambda_{F}(n) v_{f, 1}(e^{\log{n} / \log{Z}})}{n^{\sigma+it}\log{n}} \bigg|^{2k}dt
\leq C_3^k k^k\l( \sum_{Z< n \leq Z^2}\frac{|a_{F}(p)|^2}{p^{2\sigma}}\r)^{k}+C_3^k k^{k} Z^{k(1 - 2\sigma)}.
\end{align}
For the second sum in \eqref{PKLI0} , we can follow the proof of estimate \eqref{PESRSIZ1} and obtain
\begin{align}
\nonumber
& Z^{2k(1/2 - \sigma)}\int_{T}^{2T} (\sigma_{Z, t}(F) - 1/2)^{2k}Z^{4k(\sigma_{Z, t}(F) - 1/2)}
\l( \bigg| \sum_{n \leq Z^3}\frac{\Lambda_{F}(n) w_{Z}(n)}{n^{\sigma_{Z, t}(F) + it}} \bigg| + \log{T} \r)^{2k} dt \\
\label{PKLI1}
& \ll C^{k} T^{1 - \delta_{F}(2\sigma - 1) + \frac{16\delta_F}{\log{Z}}} \l( \frac{\log{T}}{\log{Z}} \r)^{2k}
\ll T^{1 - \delta_{F}(2\sigma - 1)} C_{2}^{k} k^{2k}
\end{align}
for some positive constant $C_{2} = C_{2}(F)$.
Combining \eqref{PKLI0}, \eqref{PKLI1} and \eqref{PKLI3}, we have
\begin{align} \label{PKLI4}
& \frac{1}{T}\int_{T}^{2T}\bigg| \log{F(\sigma + it)} - P_{F}(\sigma + it, Z)
- \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z}) \bigg|^{2k}dt \\
& \leq C^{k} k^{2k}T^{\delta_{F}(1 - 2\sigma)}
+ C^{k} k^{k} \l( \sum_{Z< p \leq Z^{2}}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \r)^{k}
\end{align}
for some constant $C = C(F) > 0$.
Now for any $3 \leq X \leq Y = T^{1/k}$,
we write
\begin{align*}
&\log{F(\sigma + it)} - P_{F}(\sigma + it, X)
- \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log((\sigma + it - \rho_{F})\log{Y})\\
&= \log{F(\sigma + it)} - P_{F}(\sigma + it, Z) - \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})\\
&- (P_{F}(\sigma + it, X) - P_{F}(\sigma + it, Z))
+ \sum_{\frac{1}{\log{Y}} < |\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})\\
&\qquad \qquad \qquad \qquad \qquad \qquad \qquad - \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log\l( \frac{\log{Y}}{\log{Z}} \r).
\end{align*}
By Lemma \ref{MomentLambdab}, we have
\begin{align*}
\int_{T}^{2T}|P_{F}(\sigma + it, X) - P_{F}(\sigma + it, Z)|^{2k}dt
\leq T C^{k} k^{k} \l( \sum_{\min\{ X, Z \} < p \leq \max\{ X, Z \}}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \r)^{k}.
\end{align*}
for some constant $C = C(F) > 0$.
If $Z < X \leq Y$, we see, using \eqref{SNC}, that the sum on the right hand side is $\ll_{F} Z^{1 - 2\sigma} = T^{\delta_{F}(1 - 2\sigma)/k}$.
If $X < Z$, it holds that the sum is $\leq \sum_{X < p \leq Y}\frac{|a_{F}(p)|^2}{p^{2\sigma}}$.
Hence, we have
\begin{align}\label{XToZ}
\hspace{-5mm}
\int_{T}^{2T}|P_{F}(\sigma + it, X) - P_{F}(\sigma + it, Z)|^{2k}dt
\leq T C^{k} k^{k} T^{\delta_{F}(1 - 2\sigma)} + T C^{k} k^{k} \l( \sum_{X < p \leq Y}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \r)^{k}.
\end{align}
Also, it holds that
\begin{align*}
\sum_{\frac{1}{\log{Y}} < |\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})
- \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log\l( \frac{\log{Y}}{\log{Z}} \r)
\ll \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}1.
\end{align*}
Therefore, by this inequality and Lemma \ref{ESRSIZ}, we have
\begin{align}
&\int_{T}^{2T}\l\{ \sum_{\frac{1}{\log{Y}} < |\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})
+ \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log\l( \frac{\log{Y}}{\log{Z}} \r) \r\}^{2k} dt\\
&\leq T^{1 - \delta_{F}(2\sigma - 1)} C^{k} k^{2k}.\label{zerosYZ}
\end{align}
Combining \eqref{PKLI4}, \eqref{XToZ} and \eqref{zerosYZ}, we obtain
\begin{align*}
&\frac{1}{T}\int_{T}^{2T}\bigg|\log{F(\sigma + it)} - P_{F}(\sigma + it, X)
- \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log((\sigma + it - \rho_{F})\log{Y})\bigg|^{2k}\\
&\leq C^{k} k^{2k} T^{\delta_{F}(1 - 2\sigma)} + C^{k} k^{k} \l( \sum_{Z < p \leq Z^{2}}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \r)^{k}
+ C^{k} k^{k} \l( \sum_{X < p \leq Y}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \r)^{k}.
\end{align*}
Using \eqref{SNC}, we see that $\sum_{Z < p \leq Z^{2}}\frac{|a_{F}(p)|^2}{p^{2\sigma}} \ll_{F} Z^{1 - 2\sigma} = T^{\delta_{F}(1 - 2\sigma)/k}$.
Therefore, the second term is absorbed into the first term. Thus, we complete the proof of Proposition \ref{KLI}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{KLST}]
By Proposition \ref{KLI}, it suffices to show that there exists a positive constant $A_{3} = A_{3}(F)$ such that
\begin{align} \label{PKLST1}
\int_{T}^{2T}\Bigg| \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log((\sigma+it - \rho_{F})\log{Y}) \Bigg|^{2k}
\leq T^{1 - (2\sigma - 1)\delta_{F}} A_{3}^{k} k^{4k}
\end{align}
with $Y = T^{1 / k}$.
Put $Z = T^{\delta_{F} / k}$.
We write as
\begin{align*}
&\sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log((\sigma+it - \rho_{F})\log{Y})\\
&= \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})
-\sum_{\frac{1}{\log{Y}} < |\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})\\
&\qquad \qquad \qquad \qquad \qquad \qquad \qquad + \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log\l( \frac{\log{Y}}{\log{Z}} \r).
\end{align*}
It holds that
\begin{align*}
-\sum_{\frac{1}{\log{Y}} < |\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}}\log((\sigma + it - \rho_{F})\log{Z})
+ \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Y}}}\log\l( \frac{\log{Y}}{\log{Z}} \r)
\ll \sum_{|\sigma + it - \rho_{F}| \leq \frac{1}{\log{Z}}} 1.
\end{align*}
Hence, by using Lemma \ref{ESRSIZ}, we obtain \eqref{PKLST1}.
\end{proof}
\section{\textbf{Proofs of results for Dirichlet polynomials}}\label{dpoly}
In this section, we prove Propositions \ref{Main_Prop_JVD}, \ref{Main_Prop_JVD3}, Theorem \ref{GMDP} and Corollary \ref{jointDirichlet}.
Throughout this section, we assume that $\bm{F} = (F_1, \dots, F_{r})$ is an $r$-tuple of Dirichlet series,
and $\bm{\theta} = (\theta_{1}, \dots, \theta_{r}) \in \mathbb{R}^{r}$.
We also assume that $\{ \mathcal{X}(p) \}_{p \in \mathcal{P}}$ is a sequence of independent random variables
on a probability space $(\Omega, \mathscr{A}, \mathbb{P})$ with uniformly distributed on the unit circle in $\mathbb{C}$.
\subsection{Approximate formulas for moment generating functions}
We will relate the moment generating function of $(P_{F_{j}}(\tfrac{1}{2}+it, X))_{j = 1}^{r}$
to the moment generating function of random $L$-series.
Recall that the Dirichlet polynomial $P_{F_{j}}(s, X)$ is defined by \eqref{def_D_P}.
To do this, we work with a subset of $[T,2T]$ such that the Dirichlet polynomials do not obtain large values.
More precisely, define the set $\mathcal{A} = \mathcal{A}(T, X, \bm{F})$ by
\begin{align} \label{def_sA_JVDPP}
\mathcal{A}
= \bigcap_{j = 1}^{r}\set{t \in [T, 2T]}{\frac{|P_{F_{j}}(\tfrac{1}{2}+it, X)|}{\sigma_{F_{j}}(X)} \leq (\log{\log{X}})^{2(r + 1)}}.
\end{align}
We show that the measure of $\mathcal A$ is sufficiently close to $T$,
and thus it enough to consider the moment generating function of $(P_{F_{j}}(\tfrac{1}{2}+it, X))_{j = 1}^{r}$ on $\mathcal A$.
We first show that the measure of $\mathcal A$ is close to $T$.
\begin{lemma} \label{ESAEG_JVDPP}
Assume that $\bm{F}$ satisfies (S4), (S5), (A1).
Let $T, X$ be large with $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then there exists a positive constant $b_{0} = b_{0}(\bm{F})$ such that
\begin{align*}
\frac{1}{T}\operatorname{meas}([T, 2T] \setminus \mathcal{A})
\leq \exp\l( -b_{0}(\log{\log{X}})^{4(r + 1)} \r).
\end{align*}
\end{lemma}
\begin{proof}
By Lemmas \ref{MomentLambdab}, \ref{SumbLargeP}, there exists a constant $C_{j}$ depending only on $F_{j}$ such that
\begin{align} \label{MVEDPCLPP}
\int_{T}^{2T}| P_{F_{j}}(\tfrac{1}{2}+it, X)|^{2k}dt
\leq T C_{j}^k k^k \l( \sum_{n\leq X}\frac{|b_{F_{j}}(n)|^2}{n} \r)^{k}
\leq T (2C_{j} k \sigma_{F_{j}}(X)^2)^{k}
\end{align}
for $1 \leq k \leq 2^{-1}(\log{\log{X}})^{4(r + 1)}$.
This implies that
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\frac{|P_{F_{j}}(\tfrac{1}{2}+it, X)|}{\sigma_{F_{j}}(X)} > (\log{\log{X}})^{2(r + 1)}}
\ll \l( \frac{2C_{j} k}{(\log{\log{X}})^{4(r + 1)}} \r)^{k}.
\end{align*}
Hence, there exists some constant $C = C(\bm{F}) > 0$ such that
\begin{align*}
\frac{1}{T}\operatorname{meas}([T, 2T] \setminus \mathcal{A})
\leq \l( \frac{2Ck}{(\log{\log{X}})^{4(r + 1)}} \r)^{k}.
\end{align*}
Choosing $k = \lfloor (2 e C)^{-1} (\log{\log{X}})^{4(r + 1)} \rfloor$, we obtain this lemma.
\end{proof}
Next we consider the moment generating function of $(P_{F_{j}}(\tfrac{1}{2}+it, X))_{j = 1}^{r}$ on $\mathcal A$.
\begin{proposition} \label{RKLJVDPP}
Assume that $\bm{F}$, $\bm{\theta}$ satisfy (S4), (S5), (A1), and (A2).
Let $T$, $X$ be large numbers with $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
For any $\bm{z} = (z_1, \dots, z_r) \in \mathbb{C}^{r}$ with $\norm[]{\bm{z}} \leq 2(\log{\log{X}})^{2r}$,
\begin{align} \label{RKLJVDPP1}
& \frac{1}{T}\int_{\mathcal{A}}\exp\l( \sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \r)dt\\
&= \prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z})
+ O\l(\exp\l( -b_{1}(\log{\log{X}})^{4(r + 1)} \r)\r),
\end{align}
where
$M_{p, \frac{1}{2}}$ is the function defined by \eqref{def_RDP_MGF}.
Here, $b_{1}$ is a positive constant depending only on $\bm{F}$.
\end{proposition}
We prepare some lemmas for the proof of Proposition \ref{RKLJVDPP}.
\begin{lemma} \label{GRLR}
Let $\{b_{1}(m)\}, \dots, \{ b_{n}(m) \}$ be complex sequences.
For any $q_{1}, \dots, q_{s}$ distinct prime numbers and any $k_{1, 1}, \dots k_{1, n}, \dots, k_{s, 1}, \dots, k_{s, n} \in \mathbb{Z}_{\geq 1}$,
we have
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\prod_{\ell = 1}^{s}
\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) q_{\ell}^{-i t k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) q_{\ell}^{-i t k_{n, \ell}}\r)dt \\
& = \EXP{\prod_{\ell = 1}^{s}\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) \mathcal{X}(q_{\ell})^{k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) \mathcal{X}(q_{\ell})^{k_{n, \ell}}\r)}
+ O\l( \frac{1}{T}\prod_{\ell = 1}^{s}\prod_{j = 1}^{n}q_{\ell}^{k_{j, \ell}}|b_{j}(q_{\ell}^{k_{j, \ell}})| \r).
\end{align*}
\end{lemma}
\begin{proof}
Denote $\psi_{h, \ell} = \arg b_{h}(q_{\ell}^{k_{h, \ell}})$.
Then we can write
\begin{align}
& \l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) q_{\ell}^{-i t k_{1, \ell}}\r)
\cdots (\Re b_{n}(q_{\ell}^{k_{n, \ell}}) q_{\ell}^{-i t k_{n, \ell}}) \\
& = \frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \}}e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})}
q_{\ell}^{-i(\varepsilon_{1} k_{1, \ell} + \cdots \varepsilon_{n} k_{n, \ell})t} \\
\label{pGRLR1}
& = \frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1, \ell} + \cdots + \varepsilon_{n}k_{n, \ell} = 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})} \\
& \qquad + \frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1, \ell} + \cdots + \varepsilon_{n}k_{n, \ell} \not= 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})}e^{-i(\varepsilon_{1} k_{1, \ell} + \cdots + \varepsilon_{n} k_{n, \ell})t \log{q_{\ell}}}.
\end{align}
Thus, we obtain
\begin{align*}
& \prod_{\ell = 1}^{s}\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) q_{\ell}^{-i t k_{1, \ell}}\r)
\cdots (\Re b_{n}(q_{\ell}^{k_{n, \ell}}) q_{\ell}^{-i t k_{n, \ell}}) \\
& = \prod_{\ell = 1}^{s}\frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1} + \cdots + \varepsilon_{n}k_{n} = 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})}
+ E(t),
\end{align*}
where $E(t)$ is the sum whose the number of terms is less than $2^{n s}$,
and the form of each term is $\delta'' e^{it(\b_{1}\log{q_{1}} + \cdots + \b_{s}\log{q_{s}})}$
with $0 \leq |\b_{\ell}| \leq \a_{\ell} := k_{1, \ell} + \cdots k_{n, \ell}$ and $\b_{u} \not= 0$ for some $1 \leq u \leq s$.
Here, $\delta''$ is a complex number independent of $t$, and satisfies
$|\delta''| \leq 2^{-n s}\prod_{\ell = 1}^{s}\prod_{j = 1}^{n}|b_{j}(q_{\ell}^{k_{j, \ell}})|$.
Since $|\b_{1}\log{q_{1}} + \cdots + \b_{s}\log{q_{s}}| \gg \prod_{\ell = 1}^{s}\prod_{j = 1}^{n}q_{\ell}^{-k_{j, \ell}}$,
the integral of each term of $E(t)$ is bounded by
$2^{-n s}\prod_{\ell = 1}^{s}\prod_{j = 1}^{n}q_{\ell}^{k_{j, \ell}}|b_{j}(q_{\ell}^{k_{j, \ell}})|$.
Hence, by this bound of $E$ and the bound for the number of terms of $E$,
we have
\begin{align*}
\int_{T}^{2T}E(t) dt
\ll \prod_{\ell = 1}^{s}\prod_{j = 1}^{n}q_{\ell}^{k_{j, \ell}}|b_{j}(q_{\ell}^{k_{j, \ell}})|.
\end{align*}
Therefore, we have
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\prod_{\ell = 1}^{s}\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) q_{\ell}^{-i t k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) q_{\ell}^{-i t k_{n, \ell}}\r)dt \\
& = \prod_{\ell = 1}^{s}\frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1} + \cdots + \varepsilon_{n}k_{n} = 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})}
+ O\l( \frac{1}{T}\prod_{\ell = 1}^{s}\prod_{j = 1}^{n}q_{\ell}^{k_{j, \ell}}|b_{j}(q_{\ell}^{k_{j, \ell}})| \r).
\end{align*}
On the other hand, by the independence of $\mathcal{X}(p)$'s, it follows that
\begin{align*}
& \EXP{\prod_{\ell = 1}^{s}\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) \mathcal{X}(q_{\ell})^{k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) \mathcal{X}(q_{\ell})^{k_{n, \ell}}\r)} \\
& = \prod_{\ell = 1}^{s}\EXP{\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) \mathcal{X}(q_{\ell})^{k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) \mathcal{X}(q_{\ell})^{k_{n, \ell}}\r)}.
\end{align*}
As in \eqref{pGRLR1}, we can write
\begin{align*}
& \l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) \mathcal{X}(q_{\ell})^{k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) \mathcal{X}(q_{\ell})^{k_{n, \ell}}\r) \\
& = \frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1, \ell} + \cdots + \varepsilon_{n}k_{n, \ell} = 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})} \\
& \qquad + \frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1, \ell} + \cdots + \varepsilon_{n}k_{n, \ell} \not= 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})}\mathcal{X}(q_{\ell})^{-i(\varepsilon_{1} k_{1, \ell} + \cdots + \varepsilon_{n} k_{n, \ell})}.
\end{align*}
Since $\mathcal{X}(p)$ is uniformly distributed on the unit circle in $\mathbb{C}$, we have
\begin{align} \label{BEQXp}
\EXP{\mathcal{X}(p)^{a}}
= \l\{
\begin{array}{cc}
1 & \text{if \; $a = 0$,} \\
0 & \text{otherwise}
\end{array}
\r.
\end{align}
for any $a \in \mathbb{Z}$.
Hence, we obtain
\begin{align*}
& \prod_{\ell = 1}^{s}\EXP{\l(\Re b_{1}(q_{\ell}^{k_{1, \ell}}) \mathcal{X}(q_{\ell})^{k_{1, \ell}}\r)
\cdots \l(\Re b_{n}(q_{\ell}^{k_{n, \ell}}) \mathcal{X}(q_{\ell})^{k_{n, \ell}}\r)} \\
& = \prod_{\ell = 1}^{s}\frac{|b_{1}(q_{\ell}^{k_{1, \ell}}) \cdots b_{n}(q_{\ell}^{k_{n, \ell}})|}{2^{n}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}k_{1, \ell} + \cdots + \varepsilon_{n}k_{n, \ell} = 0}}
e^{i(\varepsilon_{1}\psi_{1, \ell} + \cdots + \varepsilon_{n}\psi_{n, \ell})},
\end{align*}
which completes the proof of the lemma.
\end{proof}
\begin{lemma} \label{UBDRP}
Let $\{a(p)\}_{p \in \mathcal{P}}$ be a complex sequence.
Then, for any $k \in \mathbb{Z}_{\geq 1}$, $X \geq 3$, we have
\begin{align*}
\EXP{\bigg| \sum_{p \leq X}a(p)\mathcal{X}(p) \bigg|^{2k}}
\leq k! \l( \sum_{p \leq X}|a(p)|^2 \r)^{k}.
\end{align*}
\end{lemma}
\begin{proof}
Since $\mathcal{X}(p)$'s are independent and uniformly distributed on the unit circle in $\mathbb{C}$, it holds that
\begin{align} \label{BEUIR}
\EXP{ \frac{\mathcal{X}(p_1)^{a_{1}} \cdots \mathcal{X}(p_{k})^{a_{k}}}{ \mathcal{X}(q_1)^{b_1} \cdots \mathcal{X}(q_{\ell})^{b_{\ell}} } }
=
\begin{cases}
1 & \text{if $p_{1}^{a_1} \cdots p_{k}^{a_{k}} = q_{1}^{b_1} \cdots q_{\ell}^{b_\ell}$}, \\
0 & \text{if $p_{1}^{a_1} \cdots p_{k}^{a_{k}} \neq q_{1}^{b_1} \cdots q_{\ell}^{b_\ell}$}.
\end{cases}
\end{align}
It follows that
\begin{align*}
\EXP{ \bigg| \sum_{p \leq X}a(p) \mathcal{X}(p) \bigg|^{2k} }
& = \sum_{\substack{p_1, \dots, p_{k} \leq X \\ q_{1}, \dots, q_{k} \leq X}}
a(p_1) \cdots a(p_{k}) \overline{a(q_{1}) \cdots a(q_{k})}
\EXP{\frac{\mathcal{X}(p_1) \cdots \mathcal{X}(p_{k}) }{ \mathcal{X}(q_{1}) \cdots \mathcal{X}(q_{k})} } \\
& \leq k!\sum_{p_{1}, \dots, p_{k} \leq X}|a(p_{1})|^2 \cdots |a(p_{k})|^{2}
\leq k! \l( \sum_{p \leq X}|a(p)|^2 \r)^{k},
\end{align*}
which completes the proof of the lemma.
\end{proof}
\begin{lemma} \label{UBMVRDP}
Assume that $F$ is a Dirichlet series satisfying (S4), (S5), and \eqref{SNC}.
Let $P_{F}(\sigma, \mathcal{X}, X)$ be the random Dirichlet polynomial defined by \eqref{def_RP}.
There exists a positive constant $C = C(F)$ such that for any $k \in \mathbb{Z}_{\geq 1}$, and any large $X$, we have
\begin{align*}
\EXP{| P_{F}(\tfrac{1}{2}, \mathcal{X}, X)|^{2k}}
\ll (C k \sigma_{F}(X)^{2})^{k}.
\end{align*}
\end{lemma}
\begin{proof}
Using Lemma \ref{UBDRP}, we can prove this lemma in the same way as Lemma \ref{MomentLambdab}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{RKLJVDPP}]
Let $T$, $X$ be large numbers such that $X^{(\log{\log{X}})^{4(r+1)}} \leq T$.
Let $\bm{z} = (z_{1}, \dots, z_{r}) \in \mathbb{C}^{r}$ with $\norm[]{\bm{z}} \leq 2(\log{\log{X}})^{2r}$.
From \eqref{def_sA_JVDPP}, we have
\begin{align*}
& \frac{1}{T}\int_{\mathcal{A}}\exp\l( \sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \r)dt \\
& = \frac{1}{T}\sum_{0 \leq k \leq Y}\frac{1}{k!}\int_{\mathcal{A}}
\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg)^{k}dt \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad+ O\l(\sum_{k > Y}\frac{1}{k!}
\l( C (\log{\log{X}})^{2r + 5/2} \norm[]{\bm{z}} \r)^{k} \r)
\end{align*}
with $Y = \frac{1}{4}(\log{\log{X}})^{4(r+1)}$.
Here, $C = C(\bm{F})$ is some positive constant.
We see that this $O$-term is $\ll \exp\l( - (\log{\log{X}})^{4(r+1)} \r)$ by the bound for $\|\bm z\|$.
Using the Cauchy-Schwarz inequality, we find that
\begin{align*}
& \frac{1}{T}\int_{\mathcal{A}}\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg)^{k}dt \\
& = \frac{1}{T}\int_{T}^{2T}\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg)^{k}dt \\
& +O\l( \frac{1}{T}(\operatorname{meas}([T, 2T] \setminus \mathcal{A}))^{1/2}
\l(\int_{T}^{2T} \bigg|\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg|^{2k}dt\r)^{1/2} \r).
\end{align*}
By Lemma \ref{ESAEG_JVDPP}, estimate \eqref{MVEDPCLPP}, and the bound for $\|\bm{z}\|$, this $O$-term is
\begin{align*}
\ll \exp\l( -(2 b_{0})^{-1} (\log{\log{X}})^{4(r+1)} \r) \l( C_{1} k^{1/2}(\log{\log{X}})^{2r + 1/2} \r)^{k}
\end{align*}
for $0 \leq k \leq Y$, where $C_{1} = C_{1}(\bm{F}) > 0$ is a positive constant.
It holds from this estimate and the Stirling formula that
\begin{align} \label{GRKLG5}
& \frac{1}{T}\int_{\mathcal{A}}\exp\l( \sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \r)dt \\
& = \frac{1}{T}\sum_{0 \leq k \leq Y}\frac{1}{k!}
\int_{T}^{2T}\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg)^{k}dt \\
& \qquad \qquad \qquad+ O\l(\exp\l( -(2 b_{0})^{-1} (\log{\log{X}})^{4(r + 1)} \r)
\sum_{0 \leq k \leq Y}\frac{(C_{1} e (\log{\log{X}})^{2r+1/2})^{k}}{k^{k/2}}\r).
\end{align}
When $X$ is sufficiently large, it follows that
\begin{align*}
\sum_{0 \leq k \leq Y}\frac{(C_{1} e (\log{\log{X}})^{2r+1/2})^{k}}{k^{k/2}}
& = \sum_{0 \leq k \leq (\log{\log{X}})^{4r + 2}}\frac{(C_{1} e (\log{\log{X}})^{2r+1/2})^{k}}{k^{k/2}} + O(1) \\
& \ll \exp\l( (\log{\log{X}})^{4r + 3} \r).
\end{align*}
Hence, the $O$-term on the right hand side of \eqref{GRKLG5} is $\ll \exp\l( -(3c_{1})^{-1}(\log{\log{X}})^{4(r + 1)} \r)$.
Now, we can write
\begin{align*}
& \int_{T}^{2T}\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg)^{k}dt \\
& = \sum_{1 \leq j_{1}, \dots, j_{k} \leq r}z_{j_{1}} \cdots z_{j_{k}}
\sum_{p_{1}^{\ell_{1}}, \dots, p_{k}^{\ell_{k}} \leq X}\frac{1}{p_{1}^{\ell_{1}/2} \cdots p_{k}^{\ell_{k}/2}} \\
& \qquad \qquad \qquad \times \int_{T}^{2T}\l(\Re e^{-i\theta_{j_{1}}} b_{F_{j_{1}}}(p^{\ell_{1}}) p_{1}^{-i t \ell_{1}} \r)
\cdots \l(\Re e^{-i\theta_{j_{k}}} b_{F_{j_{k}}}(p^{\ell_{k}}) p_{k}^{-i t \ell_{k}} \r)dt.
\end{align*}
From this equation and Lemma \ref{GRLR}, we have
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \bigg)^{k}dt \\
&= \EXP{\bigg(\sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \bigg)^{k}} \\
&\quad+O\l( \frac{1}{T}\sum_{1 \leq j_{1}, \dots, j_{k} \leq r}|z_{j_{1}} \cdots z_{j_{k}}|
\sum_{p_{1}^{\ell_{1}}, \dots, p_{k}^{\ell_{k}} \leq X}|b_{F_{j_{1}}}(p_{1}^{\ell_{1}}) \cdots b_{F_{j_{k}}}(p_{k}^{\ell_{k}})|
p_{1}^{\ell_{1}/2} \cdots p_{k}^{\ell_{k}/2} \r).
\end{align*}
Additionally, we see that this $O$-term is
\begin{align*}
\ll \frac{1}{T}\l( \sum_{j = 1}^{r}|z_{j}|\sum_{n \leq X}|b_{F_{j}}(n)|n^{1/2} \r)^{k}
\leq \frac{(C X)^{2k}}{T}
\leq \frac{C^{2k}}{T^{1/2}}
\leq \exp\l( -(3b_{0})^{-1} (\log{\log{X}})^{4(r + 1)} \r)
\end{align*}
for $0 \leq k \leq Y$ when $T$ is sufficiently large.
Therefore, we have
\begin{align*}
& \frac{1}{T}\int_{\mathcal{A}}\exp\l( \sum_{j = 1}^{r} z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \r)dt \\
& = \sum_{0 \leq k \leq Y}\frac{1}{k!}\EXP{\l( \sum_{j = 1}^{r}z_{j}\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \r)^{k}}
+ O\l( \exp\l( -(3c_{1})^{-1}(\log{\log{X}})^{4(r + 1)} \r) \r) \\
& = \EXP{\exp\l( \sum_{j = 1}^{r}z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \r)}
- \sum_{k > Y}\frac{1}{k!}\EXP{\l( \sum_{j = 1}^{r}z_{j}\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \r)^{k}} \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad + O\l( \exp\l( -(3b_{0})^{-1}(\log{\log{X}})^{4(r + 1)} \r) \r).
\end{align*}
The independence of $\mathcal{X}(p)$'s yields that
\begin{align*}
\EXP{\exp\l( \sum_{j = 1}^{r}z_{j} \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \r)}
= \prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z}).
\end{align*}
Using Lemmas \ref{SumbLargeP}, \ref{UBMVRDP}, the Cauchy-Schwarz inequality, and the bound for $\| \bm{z} \|$, we obtain
\begin{align}
\EXP{\l( \sum_{j = 1}^{r}z_{j}\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \r)^{k}}
\leq k^{k/2}\l(C (\log{\log{X}})^{2r + 1/2}\r)^{k}
\end{align}
for some constant $C = C(\bm{F}) > 0$.
Therefore, it holds that
\begin{align*}
\sum_{k > Y}\frac{1}{k!}\EXP{\l( \sum_{j = 1}^{r}z_{j}\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2}, \mathcal{X}, X) \r)^{k}}
& \leq \sum_{k > Y}\frac{\l(C (\log{\log{X}})^{2r + 1/2}\r)^{k}}{k^{k/2}}
\leq \sum_{k > Y}\l( \frac{2C}{(\log{\log{X}})^{1/2}} \r)^{k} \\
& \leq \exp\l( -(3b_{0})^{-1}(\log{\log{X}})^{4(r + 1)} \r).
\end{align*}
Thus, we complete the proof of Proposition \ref{RKLJVDPP}.
\end{proof}
\subsection{Estimates for the main term of the moment generating function.}
We give some lemmas to estimate the main term in Proposition \ref{RKLJVDPP}.
\begin{lemma} \label{EQTM}
Assume that $\bm{F}$, $\bm{\theta}$ satisfy (S4), (S5).
For any $\bm{z} = (z_{1}, \dots, z_{r}) \in \mathbb{C}^{r}$, $\sigma \geq 1/2$, we have
\begin{align*}
M_{p, \sigma}(\bm{z})
& = 1 + \frac{1}{4}\sum_{1 \leq j_{1}, j_{2} \leq r}z_{j_{1}} z_{j_{2}}\sum_{\ell = 1}^{\infty}
\frac{\Re e^{-i\theta_{j_{1}}}b_{F_{j_{1}}}(p^{\ell})\overline{e^{-i\theta_{j_{2}}}b_{F_{j_{2}}}(p^{\ell})}}{p^{2 \ell \sigma}} \\
& \qquad \qquad \qquad \qquad
+ O\l( \sum_{n = 3}^{\infty}\frac{1}{n!}\l(\sum_{j = 1}^{r} |z_{j}|
\sum_{\ell = 1}^{\infty}\frac{|b_{F_{j}}(p^{\ell})|}{p^{\ell \sigma}} \r)^{n} \r).
\end{align*}
\end{lemma}
\begin{proof}
We write
\begin{align*}
& \exp\l( \sum_{j = 1}^{r}z_{j} \Re e^{-i\theta_{j}}\sum_{\ell = 1}^{\infty}\frac{b_{F_{j}}(p^{\ell})\mathcal{X}(p)^{\ell}}{p^{\ell \sigma}} \r) \\
& =1 + \sum_{n = 1}^{\infty}\frac{1}{n!}\sum_{1 \leq j_{1}, \dots, j_{n} \leq r}z_{j_{1}} \cdots z_{j_{n}}
\sum_{\ell_{1}, \dots, \ell_{n} = 1}^{\infty}
\frac{\Re e^{-i\theta_{j_{1}}} b_{F_{j_{1}}}(p^{\ell_{1}}) \mathcal{X}(p)^{\ell_{1}}}{p^{\ell_{1} \sigma}}
\cdots
\frac{\Re e^{-i\theta_{j_{n}}} b_{F_{j_{n}}}(p^{\ell_{n}}) \mathcal{X}(p)^{\ell_{n}}}{p^{\ell_{n} \sigma}}.
\end{align*}
Let $\psi_{h} = \arg b_{F_{h}}(p^{\ell_{h}}) - \theta_{j_{h}}$.
Then, using equation \eqref{BEQXp}, we obtain
\begin{align*}
& \EXP{\exp\l( \sum_{j = 1}^{r}z_{j} \Re e^{-i\theta_{j}}
\sum_{\ell = 1}^{\infty}\frac{b_{F_{j}}(p^{\ell})\mathcal{X}(p)^{\ell}}{p^{\ell \sigma}} \r)} = \\
& 1 + \sum_{n = 1}^{\infty}\frac{1}{2^{n} n!}\sum_{1 \leq j_{1}, \dots, j_{n} \leq r}z_{j_{1}} \cdots z_{j_{n}}
\sum_{\ell_{1}, \dots, \ell_{n} = 1}^{\infty}\frac{|b_{F_{j_{1}}}(p^{\ell_{1}}) \cdots b_{F_{j_{n}}}(p^{\ell_{n}})|}
{p^{(\ell_{1} + \cdots + \ell_{n})\sigma}}
\sum_{\substack{\varepsilon_{1}, \dots, \varepsilon_{n} \in \{ -1, 1 \} \\ \varepsilon_{1}\ell_{1} + \cdots + \varepsilon_{n}\ell_{n} = 0}}
e^{i(\varepsilon_{1} \psi_{1} + \cdots + \varepsilon_{n} \psi_{n})}.
\end{align*}
If $n = 1$, then the equation $\varepsilon_{1}\ell_{1} = 0$ is always false, so the term corresponds to $n = 1$ vanishes.
The term becomes corresponds to $n=2$ is
\begin{align*}
& \frac{1}{4}\sum_{1 \leq j_{1}, j_{2} \leq r}z_{j_{1}} z_{j_{2}}\sum_{\ell = 1}^{\infty}
\frac{\Re e^{-i\theta_{j_{1}}}b_{F_{j_{1}}}(p^{\ell})\overline{e^{-i\theta_{j_{2}}}b_{F_{j_{2}}}(p^{\ell})}}{p^{2 \ell \sigma}}.
\end{align*}
The term corresponds to $n\geq 3$ can be bounded by
\begin{align*}
& \leq \frac{1}{n!}\sum_{1 \leq j_{1}, \dots, j_{n} \leq r}|z_{j_{1}} \cdots z_{j_{n}}|
\sum_{\ell_{1}, \dots, \ell_{n} = 1}^{\infty}\frac{|b_{F_{j_{1}}}(p^{\ell_{1}}) \cdots b_{F_{j_{n}}}(p^{\ell_{n}})|}
{p^{(\ell_{1} + \cdots + \ell_{n})\sigma}}
= \frac{1}{n!}\l( \sum_{j = 1}^{r} |z_{j}| \sum_{\ell = 1}^{\infty}\frac{|b_{F_{j}}(p^{\ell})|}{p^{\ell \sigma}} \r)^{n}.
\end{align*}
This completes the proof of this lemma.
\end{proof}
\begin{lemma} \label{Prop_Psi_FPP}
Assume that $\bm{F}$ satisfies (S4), (S5), (A1), and (A2).
Put
\begin{align} \label{def_Psi_F}
\Psi(\bm{z})
= \Psi(\bm{z}; \bm{F}, \bm{\theta})
:= \prod_{p}\frac{M_{p, \frac{1}{2}}(\bm{z})}
{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)}.
\end{align}
Then, the infinity product is uniformly convergent on any compact set in $\mathbb{C}^{r}$.
In particular, $\Psi$ is analytic on $\mathbb{C}^{r}$.
Moreover, it holds that
\begin{align} \label{Psi_F_1}
| \Psi(\bm{z})|
\leq \bigg|\prod_{j = 1}^{r}\exp\l( -\frac{z_{j}^2}{2}\sigma_{F_{j}}\l( |z_{j}| \r)^2
+ O_{\bm{F}}\l( |z_{j}|^{2} + |z_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)\bigg|
\end{align}
for any $\bm{z} = (z_{1}, \dots, z_{r}) \in \mathbb{C}$, and that
\begin{align} \label{Psi_F_3}
\Psi(\bm{x})
= \prod_{j = 1}^{r}\exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}\l( |x_{j}| \r)^2
+ O_{\bm{F}}\l( x_{j}^{2} + |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)
\end{align}
for $\bm{x} \in \mathbb{R}^{r}$.
Furthermore, for any $\bm{z} = (x_{1} + iu_{1}, \dots, x_{r} + iu_{r}) \in \mathbb{C}^{r}$
satisfying $x_{j}, u_{j} \in \mathbb{R}$ with $\| \bm{u} \| \leq 1$,
we have
\begin{align} \label{Psi_F_2}
\Psi(\bm{z})
=\Psi(x_{1}, \dots, x_{r})\prod_{j = 1}^{r}\l(1
+ O_{\bm{F}}\l(|u_{j}| \exp\l( D_{1} \|\bm{x}\|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)\r).
\end{align}
Here, $D_{1} = D_{1}(\bm{F})$ is a positive constant.
\end{lemma}
\begin{proof}
By the definition of $K_{\bm{F}, \bm{\theta}}$ (see \eqref{def_K_F}), it holds that
\begin{align} \label{ES_sInp2}
|K_{\bm{F}, \bm{\theta}}(p, \bm{z})|
\leq 2 r \| \bm{z} \|^{2} \sum_{\ell = 1}^{\infty}\frac{\|b_{\bm{F}}(p^{\ell})\|^2}{p^{\ell}}.
\end{align}
Here, $b_{\bm{F}}(p^{\ell})$ indicates the vector $(b_{F_{1}}(p^{\ell}), \dots, b_{F_{r}}(p^{\ell}))$.
Using this estimate and Lemma \ref{EQTM},
we find that
\begin{align}
& \frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp(K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4)} \\
& = \l\{ 1 + \frac{1}{4}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
+ O\l( \sum_{n = 3}^{\infty}\frac{1}{n!}\l(r \| \bm{z} \|
\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} \r)^{n} \r) \r\} \\
& \qquad \qquad \qquad \qquad \times \l\{ 1 - \frac{1}{4}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
+ O\l( \sum_{n = 3}^{\infty}\frac{1}{n!}\l(\frac{r}{2} \| \bm{z} \|^2
\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}} \r)^{n} \r) \r\} \\
\label{ES_sI_new_p1}
& = 1 + O\l( \sum_{n = 3}^{\infty}\frac{(r \| \bm{z} \|)^{n}}{n!}
\l( \sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} \r)^{n}
+ \sum_{n = 2}^{\infty}\frac{(r \|\bm{z}\|^2)^{n}}{n!}
\l(\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}} \r)^{n} \r).
\end{align}
Let $Y \geq C_{1} (\| \bm{z} \|)^{\frac{2}{1 - 2\vartheta_{\bm{F}}}}$ with $C_{1} = C_{1}(\bm{F}) > 0$ a suitably large constant.
Then, by estimate \eqref{ES_sI_new_p1} and the definition of $\vartheta_{\bm{F}}$, it holds that
\begin{align*}
\bigg|\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp(K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4)} - 1\bigg|
\leq \frac{1}{2}
\end{align*}
for any $p \geq Y$.
From this inequality and estimate \eqref{ES_sI_new_p1}, we obtain
\begin{align}
& \frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp(K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4)}
= \exp\l(\log \frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp(K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4)}\r) \\
\label{ES_sI_PP_p1}
& = \exp\l\{ O\l( \sum_{n = 3}^{\infty}\frac{(r \| \bm{z} \|)^{n}}{n!}
\l( \sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} \r)^{n}
+ \sum_{n = 2}^{\infty}\frac{(r \|\bm{z}\|^2)^{n}}{n!}
\l(\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}} \r)^{n} \r) \r\}
\end{align}
for any $p \geq Y$.
Set $K_{1} = 2(\frac{1}{2} - \vartheta_{\bm{F}})^{-1}$.
Then, it holds that
$$
\sum_{\ell > K_{1}}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}}
\ll_{\bm{F}} p^{-2},\text{ and }
\sum_{\ell > K_{1}}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}}
\ll_{\bm{F}} p^{-4}.
$$
Therefore, there exists some $C = C(F)>0$ such that
\begin{align} \label{ES_sI_PP_p5}
\l( \sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} \r)^{n}
& \leq C^{n}\l(\max_{1 \leq \ell \leq K_{1}} \frac{\| b_{\bm{F}}(p^{\ell}) \|^{n}}{p^{\ell n/2}} + p^{-2n} \r)
\leq C^{n}\l(\sum_{1 \leq \ell \leq K_{1}} \frac{\| b_{\bm{F}}(p^{\ell}) \|^{n}}{p^{\ell n/2}} + p^{-2n} \r) \\
& \leq C^{n}\l(\sum_{j = 1}^{r} \sum_{1 \leq \ell \leq K_{1}} \frac{\| b_{F_{j}}(p^{\ell}) \|^{n}}{p^{\ell n/2}} + p^{-2n} \r)
\end{align}
and similarly that
\begin{align} \label{ES_sI_PP_p51}
\l(\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}} \r)^{n}
\leq C^{n}\l(\sum_{j = 1}^{r} \sum_{1 \leq \ell \leq K_{1}} \frac{\| b_{F_{j}}(p^{\ell}) \|^{2 n}}{p^{\ell n}} + p^{-4n} \r).
\end{align}
Using these inequalities and Lemma \ref{SumbLargeP}, we have for any $Z \geq 2, n\in \mathbb{Z}$
\begin{align} \label{ES_sI_PP_p3}
\sum_{p > Z}\l( \sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} \r)^{n}
& \leq C^{n}\sum_{j = 1}^{r}\sum_{m > Z}\frac{|b_{F_{j}}(m)|^2}{m^{1 + (1/2 - \vartheta_{\bm{F}})(n - 2)}} + C^{n} Z^{1 - 2n} \\
& \ll_{\bm{F}} C^{n} Z^{-(\frac{1}{2} - \vartheta_{\bm{F}})(n-2)}, \quad n \geq 3
\end{align}
and
\begin{align} \label{ES_sI_PP_p4}
\sum_{p > Z} \l(\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}} \r)^{n}
& \leq C^{n}\sum_{j = 1}^{r}\sum_{m > Z}\frac{|b_{F_{j}}(m)|^2}{m^{1 + (1 - 2\vartheta_{\bm{F}})(n - 1)}} + C^{n} Z^{1 - 4n} \\
& \ll_{\bm{F}} C^{n} Z^{-(1 - 2\vartheta_{\bm{F}})(n-1)}, \quad n \geq 2.
\end{align}
%
Applying these estimates to \eqref{ES_sI_PP_p1}, we obtain
\begin{align}
\prod_{p > Y}\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp(K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4)}
& = \exp\l( O_{\bm{F}}\l( \frac{\| \bm{z} \|^3}{Y^{\frac{1}{2} - \vartheta_{\bm{F}}}} \r) \r) \\
\label{ES_sI_PP_p2}
& = \prod_{j = 1}^{r}\exp\l( O_{\bm{F}}\l( \frac{|z_{j}|^3}{Y^{\frac{1}{2} - \vartheta_{\bm{F}}}} \r) \r)
\end{align}
for any $Y \geq C_{1} (\| \bm{z} \|)^{\frac{2}{1 - 2\vartheta_{\bm{F}}}}$ with $C_{1} = C_{1}(\bm{F})$ a suitably large constant.
In particular, we see that this estimate implies the infinite product
$
\prod_{p}\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp(K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4)}
$
is uniformly convergent for $\bm{z} \in D$ with $D$ an arbitrary compact set in $\mathbb{C}^{r}$.
Hence, $\Psi$ is analytic on $\mathbb{C}^{r}$.
Next, we prove \eqref{Psi_F_1}.
Put $L = C_{1} (\| \bm{z} \|)^{\frac{2}{1 - 2\vartheta_{\bm{F}}}}$.
Then we divide the range of the product as
\begin{align} \label{p_Psi_F_1}
\prod_{p}\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)}
= \l(\prod_{p \leq L} \times \prod_{p > L}\r)
\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)}.
\end{align}
If $L < 2$, then
$
\prod_{p \leq L}M_{p, \frac{1}{2}}(\bm{z})
= 1
= \exp\l( O\l( \| \bm{z} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
$
Next, we suppose $L \geq 2$.
Using Lemma \ref{EQTM}, we see that
\begin{align*}
|M_{p, \frac{1}{2}}(\bm{z})|
\leq \exp\l( \sum_{j = 1}^{r} |z_{j}| \sum_{\ell = 1}^{\infty}\frac{|b_{F_{j}}(p^{\ell})|}{p^{\ell/2}} \r)
\leq \exp\l( r \| \bm{z} \| \sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} \r).
\end{align*}
By \eqref{ES_sI_PP_p5} and the Cauchy-Schwarz inequality, we also find that
\begin{align*}
\sum_{p \leq L}\sum_{\ell = 1}^{\infty}\frac{\|b_{\bm{F}}(p^{\ell})\|}{p^{\ell/2}}
\ll_{\bm{F}} \max_{1 \leq \ell \leq K_{1}}\sum_{p \leq L}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}} + 1
\leq \max_{1 \leq \ell \leq K_{1}}\l(\sum_{p \leq L}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}}\r)^{1/2}\l( \sum_{p \leq L} 1 \r)^{1/2} + 1.
\end{align*}
Assumptions (S5), (A1), and the prime number theorem yield that
\begin{align*}
\max_{1 \leq \ell \leq K_{1}}\l(\sum_{p \leq L}\frac{\| b_{\bm{F}}(p^{\ell}) \|^2}{p^{\ell}}\r)^{1/2}\l( \sum_{p \leq L} 1 \r)^{1/2}
\ll_{\bm{F}} L^{1/2}
\ll_{\bm{F}} \| \bm{z} \|^{\frac{1}{1 - 2\vartheta_{\bm{F}}}}.
\end{align*}
Therefore, we obtain
\begin{align*}
\prod_{p \leq L}M_{p, \frac{1}{2}}(\bm{z})
\leq \exp\l( O_{\bm{F}}\l( \| \bm{z} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)
= \prod_{j = 1}^{r} \exp\l( O_{\bm{F}}\l( |z_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)
\end{align*}
for $L \geq 2$.
Combing this estimate with the estimate in the case $L < 2$, we have
\begin{align} \label{p_Psi_F_5}
\prod_{p \leq L}M_{p, \frac{1}{2}}(\bm{z})
\leq \prod_{j = 1}^{r} \exp\l( O_{\bm{F}}\l( |z_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)
\end{align}
for all $L \geq 0$.
Also, it follows from the definitions of $\sigma_{F}(X)$, $\tau_{j_{1}, j_{2}}(X)$, and $K_{\bm{F}, \bm{\theta}}(p, \bm{z})$ that
\begin{align} \label{ES_sI_PP_p6}
\sum_{p \leq L}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
= 2\sum_{j = 1}^{r}z_{j}^2 \sigma_{F_{j}}(L)^2
+ 4\sum_{1 \leq j_{1} < j_{2} \leq r} z_{j_{1}} z_{j_{2}} \tau_{j_{1}, j_{2}}(L).
\end{align}
Using (A1), (A2), we find that
\begin{align*}
\tau_{j_{1}, j_{2}}(Z)
= \frac{1}{2}\sum_{p \leq Z}\sum_{\ell = 2}^{\infty}
\frac{\Re e^{-i\theta_{j_{1}}}b_{F_{j_{1}}}(p^{\ell}) \overline{e^{-i\theta_{j_{2}}}b_{F_{j_{2}}}(p^{\ell})}}{p^{\ell}}
+ O_{\bm{F}}(1)
\end{align*}
for any $Z \geq 2$ when $j_{1} \not= j_{2}$.
Similarly to \eqref{ES_sI_PP_p51}, we obtain
\begin{align*}
\frac{1}{2}\sum_{p \leq Z}\sum_{\ell = 2}^{\infty}
\frac{\Re e^{-i\theta_{j_{1}}}b_{F_{j_{1}}}(p^{\ell}) \overline{e^{-i\theta_{j_{2}}}b_{F_{j_{2}}}(p^{\ell})}}{p^{\ell}}
\ll_{\bm{F}} \sum_{p}\sum_{2 \leq \ell \leq K_{1}}\frac{\|b_{\bm{F}}(p^{\ell})\|^2}{p^{\ell}} + 1,
\end{align*}
and by (S5), this is $\ll_{\bm{F}} 1$.
Hence, assumptions (S4), (S5), (A1), and (A2) yield that
\begin{align} \label{BDtau}
\tau_{j_{1}, j_{2}}(Z)
\ll_{\bm{F}} 1
\end{align}
for any $Z \geq 2$ when $j_{1} \not= j_{2}$.
By Lemma \ref{SumbLargeP},
\begin{align} \label{ESTsigma}
\sigma_{F_{j}}(Z)^2 = \frac{n_{F_{j}}}{2} \log{\log{Z}} + O_{F_{j}}(1)
\end{align}
for any $Z \geq 2$.
If $|z_{j}| \leq \|\bm{z}\|^{1/2}$, then $z_{j}^{2} \sigma_{F_{j}}(L)^2 \ll_{\bm{F}} \| \bm{z} \|^{2}$.
If $|z_{j}| > \|\bm{z}\|^{1/2}$, then we use \eqref{ESTsigma} to obtain
\begin{align*}
\frac{z_{j}^2}{2}\sigma_{F_{j}}(L)^2
= \frac{z_{j}^{2}}{2}\sigma_{F_{j}}(|z_{j}|)^2 + \frac{z_{j}^{2}}{4} \sum_{|z_{j}| < p \leq L}\frac{|a_{F_{j}}|^2}{p}
= \frac{z_{j}^{2}}{2}\sigma_{F_{j}}(|z_{j}|)^2 + O_{F_{j}}\l(\| \bm{z} \|^{2}\r).
\end{align*}
From this observation and \eqref{ES_sI_PP_p6}, we find that
\begin{align}
\prod_{p \leq L}\exp\l( - K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)
\label{p_Psi_F_6}
& = \prod_{j = 1}^{r}\exp\l( -\frac{z_{j}^2}{2}\sigma_{F_{j}}(|z_{j}|)^2 + O_{\bm{F}}(|z_{j}|^2) \r).
\end{align}
Hence, we obtain
\begin{align} \label{p_Psi_F_2}
\bigg|\prod_{p \leq L}\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)}\bigg|
\leq \bigg|\prod_{j = 1}^{r}\exp\l( -\frac{z_{j}^2}{2}\sigma_{F_{j}}\l( |z_{j}| \r)^2
+ O_{\bm{F}}\l( |z_{j}|^{2} + |z_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)\bigg|.
\end{align}
We also see that
\begin{align} \label{ES_sI_PP_p7}
\prod_{p > L}\frac{M_{p, \frac{1}{2}}(\bm{z})}{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z})/4 \r)}
= \exp\l( O_{\bm{F}}\l(\| \bm{z} \|^2\r) \r)
= \prod_{j = 1}^{r}\exp\l( O_{\bm{F}}\l( |z_{j}|^2 \r) \r)
\end{align}
from \eqref{ES_sI_PP_p2}.
These estimates yield inequality \eqref{Psi_F_1}.
Next, we show \eqref{Psi_F_3}.
It holds from the definition of $M_{p, \frac{1}{2}}(\bm{z})$ that
\begin{align*}
\exp\l( -\sum_{j = 1}^{r}|x_{j}|\sum_{k = 1}^{\infty}\frac{|b_{F_{j}}(p^{k})|}{p^{k/2}} \r)
\leq M_{p, \frac{1}{2}}(\bm{x})
\leq \exp\l( \sum_{j = 1}^{r}|x_{j}|\sum_{k = 1}^{\infty}\frac{|b_{F_{j}}(p^{k})|}{p^{k/2}} \r)
\end{align*}
for any $\bm{x} = (x_{1}, \dots, x_{r}) \in \mathbb{R}^{r}$.
Therefore, we have
\begin{align*}
\prod_{p \leq L}M_{p, \frac{1}{2}}(\bm{x})
= \exp\l( O_{\bm{F}}\l( \| \bm{x} \| \sum_{p \leq L}\sum_{k = 1}^{\infty}\frac{\|b_{\bm{F}}(p^{k})\|}{p^{k/2}} \r) \r).
\end{align*}
Similarly to \eqref{p_Psi_F_5} and by this equation, we have
\begin{align*}
\prod_{p \leq L}M_{p, \frac{1}{2}}(\bm{x})
= \prod_{j = 1}^{r}\exp\l( O_{\bm{F}}\l(|x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
\end{align*}
We can calculate the other parts similarly to the proof of \eqref{Psi_F_1}, and obtain \eqref{Psi_F_3}.
Finally, we prove equation \eqref{Psi_F_2}.
Since $\Psi$ is analytic on $\mathbb{C}^{r}$, we can write
\begin{align*}
\Psi(x_{1} + iu_{1}, \dots, x_{r} + iu_{r})
& = \sum_{n = 0}^{\infty}\sum_{\substack{k_{1} + \dots + k_{r} = n \\ k_{1}, \dots, k_{r} \geq 0}}
\frac{1}{k_{1}! \cdots k_{r}!}
\frac{\partial^{n} \Psi(x_{1}, \dots, x_{r})}{\partial z_{1}^{k_{1}} \cdots \partial z_{r}^{k_{r}}}
(iu_{1})^{k_{1}} \cdots (iu_{r})^{k_{r}}.
\end{align*}
It follows from estimates \eqref{Psi_F_1} and \eqref{Psi_F_3} that
\begin{align*}
|\Psi(z_{1}, \dots, z_{r})|
\leq \Psi(x_{1}, \dots, x_{r})\exp\l( C\l( \| \bm{x} \|^2 + \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}}\r) \r)
\end{align*}
for some $C = C(\bm{F}) > 0$ when $|z_{1} - x_{1}| = \cdots |z_{r} - x_{r}| = 2$.
Using this estimate and Cauchy's integral formula, we find that
\begin{align*}
\frac{\partial^{n} \Psi(x_{1}, \dots, x_{r})}{\partial z_{1}^{k_{1}} \cdots \partial z_{r}^{k_{r}}}
& = \frac{k_{1}! \cdots k_{r}!}{(2\pi i)^{r}}\int_{|z_{r} - x_{r}| = 2} \cdots \int_{|z_{1} - x_{1}| = 2}
\frac{\Psi(z_{1}, \dots, z_{r})}{(z_{1} - x_{1})^{k_{1}} \cdots (z_{r} - x_{r})^{k_{r}}}dz_{1} \cdots dz_{r} \\
& \leq 2^{-(k_{1} + \cdots + k_{r})}k_{1}! \cdots k_{r}! \Psi(x_{1}, \dots, x_{r})
\exp\l(C \l( \| \bm{x} \|^2 + \|\bm{x}\|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}}\r) \r).
\end{align*}
Hence, when $\| \bm{u} \| \leq 1$, we have
\begin{align*}
\Psi(x_{1} + iu_{1}, \dots, x_{r} + iu_{r})
& = \Psi(x_{1}, \dots, x_{r})\l(1
+ O_{\bm{F}}\l(\| \bm{u} \| \exp\l( C\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\theta_{\bm{F}}}} \r) \r)\r) \\
& = \Psi(x_{1}, \dots, x_{r})\prod_{j = 1}^{r}\l(1
+ O_{\bm{F}}\l(|u_{j}| \exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\theta_{\bm{F}}}} \r) \r)\r),
\end{align*}
which completes the proof of \eqref{Psi_F_2}.
\end{proof}
\begin{lemma} \label{EST_Xi_lem}
Let $\Xi_{X}$ be the function defined by \eqref{def_Xi}.
For $\bm{x} = (x_{1}, \dots, x_{r}) \in (\mathbb{R}_{\geq 0})^{r}$, we have
\begin{align} \label{EST_Xi}
\Xi_{X}(\bm{x})
= \prod_{j = 1}^{r}\exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}(|x_{j}|)^{2}
+ O_{\bm{F}}\l( x_{j}^2 + |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
\end{align}
\end{lemma}
\begin{proof}
From \eqref{BDtau}, formula \eqref{Psi_F_3}, and the definition of $\Xi_{X}$, we have
\begin{align*}
\Xi_{X}(\bm{x})
& = \prod_{j = 1}^{r}\exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}(|x_{j}|)^{2}
+ O_{\bm{F}}\l( \| \bm{x} \|^{2} + x_{j}^2 + |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r) \\
& = \prod_{j = 1}^{r}\exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}(|x_{j}|)^{2}
+ O_{\bm{F}}\l( x_{j}^2 + |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r),
\end{align*}
which completes this lemma.
\end{proof}
\begin{lemma} \label{ES_sI}
Assume that $\bm{F}$ satisfies (S4), (A1), and (A2).
For $\bm{z} = (z_1, \dots, z_r) \in \mathbb{C}^{r}$, $X \geq C \|\bm{z}\|^{\frac{2}{1-2\vartheta_{\bm{F}}}} + 3$
with $C = C(\bm{F})$ a sufficiently large positive constant, we have
\begin{align}
\label{ES_sI1}
\bigg|\prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z})\bigg|
\leq \bigg|\prod_{j = 1}^{r}
\exp\l(\frac{z_{j}^2}{2} \l(\sigma_{F_{j}}(X)^2 - \sigma_{F_{j}}(|z_{j}| \r)^2
+ O_{\bm{F}}\l( |z_{j}|^{2} + |z_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)\bigg|,
\end{align}
where $\sigma_{F_{j}}(X)$ is defined by \eqref{def_var}.
Moreover, there exists a positive constant $b_{2} = b_{2}(\bm{F})$ such that,
for any $X \geq 3$ and any $\bm{z} = (z_{1}, \dots, z_{r}) \in \mathbb{C}^{r}$ with $\norm[]{\bm{z}} \leq b_{2}$, we have
\begin{align} \label{ES_sI2}
\prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z})
= \prod_{j = 1}^{r}\l( 1 + O_{\bm{F}}\l( |z_{j}|^2 \r) \r)\exp\l( \frac{z_{j}^2}{2} \sigma_{F_{j}}(X)^2 \r).
\end{align}
Furthermore, for any $\bm{z} = (x_{1} + iu_{1}, \dots, x_{r} + iu_{r}) \in \mathbb{C}^{r}$ with $x_{j}, u_{j} \in \mathbb{R}$ and $\| \bm{u} \| \leq 1$,
and any $X \geq C \|\bm{x} \|^{\frac{2}{1-2\vartheta_{\bm{F}}}} + 3$
with $C = C(\bm{F})$ a sufficiently large positive constant, we have
\begin{align} \label{ES_sI3}
& \prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z}) \\
& = \Xi_{X}(\bm{x})
\prod_{j = 1}^{r}\l( 1 + O_{\bm{F}}\l( |u_{j}|\exp\l( D_{1}\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
+ \frac{|z_{j}|^{4}}{\log{X}} \r) \r)\exp\l( \frac{z_{j}^2}{2}\sigma_{F_{j}}(X)^2 \r),
\end{align}
where $\Xi_{X}$ is the function defined by \eqref{def_Xi}, and $D_{1}$ is the same constant as in Lemma \ref{Prop_Psi_FPP}.
\end{lemma}
\begin{proof}
First, we prove \eqref{ES_sI1}.
It holds that
\begin{align} \label{ES_sI_PP1}
& \prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z})
= \Psi(\bm{z})\prod_{p \leq X}\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)
\times \prod_{p > X}\frac{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)}{M_{p, \frac{1}{2}}(\bm{z})}.
\end{align}
Using \eqref{ES_sI_PP_p2}, we have
\begin{align} \label{ES_sI_PP2}
\prod_{p > X}\frac{\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)}{M_{p, \frac{1}{2}}(\bm{z})}
& = \prod_{j = 1}^{r}\exp\l( O_{\bm{F}}\l( \frac{|z_{j}|^3}{X^{\frac{1}{2} - \vartheta_{\bm{F}}}} \r) \r) \\
& = \prod_{j = 1}^{r}\exp\l( O_{\bm{F}}\l( |z_{j}|^2 \r) \r)
\end{align}
when $X \geq C \| \bm{z} \|^{\frac{2}{1 - 2\vartheta_{\bm{F}}}}$ with $C$ a suitably large constant.
Also, as in the proof of \eqref{p_Psi_F_6}, we find that
\begin{align} \label{ES_sI_PP3}
\prod_{p \leq X}\exp\l( K_{\bm{F}, \bm{\theta}}(p, \bm{z}) / 4 \r)
= \prod_{j = 1}^{r}\exp\l( \frac{z_{j}^2}{2}\sigma_{F_{j}}(X)^{2} + O_{\bm{F}}\l( |z_{j}|^{2} \r) \r).
\end{align}
Combing the above two estimates and inequality \eqref{Psi_F_1}, we have estimate \eqref{ES_sI1}.
Next, we prove \eqref{ES_sI2}.
When $\| \bm{z} \| \leq b_{2}$ with $b_{2} = b_{2}(\bm{F})$ sufficiently small, it holds from Lemma \ref{EQTM} that
\begin{align*}
|M_{p, \frac{1}{2}}(\bm{z}) - 1|
\leq \frac{1}{2},
\end{align*}
and that
\begin{align*}
M_{p, \frac{1}{2}}(\bm{z})
=1 + \frac{1}{4}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
+ O_{\bm{F}}\l( \|\bm{z}\|^3 \l(\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}}\r)^{3} \r)
\end{align*}
with $b_{\bm{F}}(p^{\ell}) = (b_{F_{1}}(p^{\ell}), \dots, b_{F_{r}}(p^{\ell}))$.
Using these and estimate \eqref{ES_sI_PP_p3},
we obtain
\begin{align*}
\sum_{p \leq X}\log{M_{p, \frac{1}{2}}(\bm{z})}
& = \sum_{p \leq X}\l(\frac{1}{4}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
+ O_{\bm{F}}\l( \| \bm{z} \|^{3} \l(\sum_{\ell = 1}^{\infty}\frac{\| b_{\bm{F}}(p^{\ell}) \|}{p^{\ell/2}}\r)^{3} \r) \r) \\
& = \frac{1}{4}\sum_{p \leq X}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
+ O_{\bm{F}}\l(\norm{\bm{z}}^{3}\r).
\end{align*}
Similarly to \eqref{ES_sI_PP_p6}, we also have
\begin{align}
\label{EQKF2}
& \frac{1}{4}\sum_{p \leq X}K_{\bm{F}, \bm{\theta}}(p, \bm{z})
= \sum_{j=1}^{r}\frac{z_{j}^2}{2}\sigma_{F_{j}}(X)^2
+ O_{\bm{F}}\l( \| \bm{z} \|^2 \r).
\end{align}
Hence, it holds that
\begin{align*}
\sum_{p \leq X}\log{M_{p, \frac{1}{2}}(\bm{z})}
& = \sum_{j = 1}^{r}\frac{z_{j}^2}{2}\sigma_{F_{j}}(X)^2
+ O_{\bm{F}}\l( \|\bm{z}\|^2 \r) \\
& = \sum_{j = 1}^{r}\frac{z_{j}^2}{2}\l(\sigma_{F_{j}}(X)^2 + O_{\bm{F}}\l( 1 \r)\r),
\end{align*}
which completes the proof of \eqref{ES_sI2}.
Finally, we prove \eqref{ES_sI3}.
It follows from equations \eqref{ES_sI_PP1} and \eqref{ES_sI_PP2} that
\begin{align*}
\prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z})
= \Psi(\bm{z})\exp\l( \frac{1}{4}\sum_{p \leq X}K_{\bm{F}, \bm{\theta}}(p, \bm{z}) \r)
\prod_{j = 1}^{r}\l( 1 + O_{\bm{F}}\l( \frac{| z_{j} |^3}{X^{\frac{1}{2} - \vartheta_{\bm{F}}}} \r) \r).
\end{align*}
for $X \geq C \|\bm x\|^{\frac{2}{1 - 2\vartheta_{\bm{F}}}} + 3$ with $C = C(\bm{F})$ sufficiently large.
Moreover, by equation \eqref{ES_sI_PP_p6}, it holds that
\begin{align}
\exp\l( \frac{1}{4}\sum_{p \leq X}K_{\bm{F}, \bm{\theta}}(p, \bm{z}) \r)
= \exp\l( \sum_{j = 1}^{r}\frac{z_{j}^2}{2}\sigma_{F_{j}}(X)
+ \sum_{1 \leq j_{1} < j_{2} \leq r}z_{j_{1}} z_{j_{2}} \tau_{j_{1}, j_{2}}(X) \r).
\end{align}
In particular, from \eqref{BDtau}, the estimate $\tau_{{l_{1}}, {l_{2}}}(X) \ll_{\bm{F}} 1$ holds
for all $1 \leq j_{1} < j_{2} \leq r$ by assumptions (S4), (S5), (A1), and (A2),
and so the above is equal to
\begin{align}
\exp\l(\sum_{1 \leq l_{1} < l_{2} \leq r}x_{l_{1}} x_{l_{2}} \tau_{l_{1}, l_{2}}(X) \r)
\prod_{j = 1}^{r}(1 + O_{\bm{F}}( |u_{j}| \cdot \| \bm{x} \| + u_{j}^2 ))\exp\l( \frac{z_{j}^2}{2}\sigma_{F_{j}}(X) \r).
\end{align}
Additionally, we have
$
\Psi(\bm{z}) = \Psi(\bm{x})\prod_{j = 1}^{r}\l(1 + O_{\bm{F}}\l(|u_{j}|
\exp\l( D_{1} \| \bm{x} \|^{(2-2\vartheta_{\bm{F}}) / (1 - 2\vartheta_{\bm{F}})} \r) \r)\r)
$
by \eqref{Psi_F_2}.
From the above estimates and the definition of $\Xi_{X}$ \eqref{def_Xi}, we also obtain formula \eqref{ES_sI3}.
Thus, we complete the proof of this lemma.
\end{proof}
\subsection{Proofs of Propositions \ref{Main_Prop_JVD} and \ref{Main_Prop_JVD3}} \label{Proof_Props_JVD}
In this section, we prove Propositions \ref{Main_Prop_JVD} and Proposition \ref{Main_Prop_JVD3}.
Proposition \ref{Main_Prop_JVD3} is a generalization of Radziwi\l\l's result\footnote{That there are some incomplete arguments
when truncating the integral in his proof, but we can fix his proof by a nontrivial modification.} \cite[Proposition 2]{Ra2011}.
Radziwi\l\l \ adopted Tenenbaum's formula \cite{Te1988} which relates the measure of the large values of the Dirichlet polynomial
to a line integral.
If we follow the same approach, we will need to assume Ramanujan's conjecture.
The reason is that
we have to use formula \eqref{RKLJVDPP1} in a larger range of $\bm{z}$ when truncating the integral and
the estimate \eqref{ES_sI1} in Lemma \ref{ES_sI} is not enough when $\vartheta_{\bm{F}} > 0$.
We avoid these obstacles by probabilistic methods in large deviation theory (see Cramér \cite{Cramer}, Hwang \cite{HH1996})).
We consider the associated distribution $\nu_{T, \bm{F}, \bm{x}}$ (see below)
and use the Selberg-Beurling functions to approximate regions in $\mathbb R^r$.
The associated distribution $\nu_{T, \bm{F}, \bm{x}}$ plays the role of localizing the measure by changing $\bm{x}$.
Eventually, we choose a suitable $\bm{x}$ in the proof of Propositions \ref{Main_Prop_JVD}, \ref{Main_Prop_JVD3},
and the $\bm{x}$ is the saddle point of the integrand in Lemma \ref{SPKLH}.
The Selberg-Beurling formula has the advantage that it only involves finite integrals.
We introduce some notation.
Define the $\mathbb{R}^{r}$-valued function $\bm{F}_{\bm{\theta}, X}(t)$ by
\begin{align*}
\bm{F}_{\bm{\theta}, X}(t)
= (\Re e^{-i\theta_{1}} P_{F_{1}}(\tfrac{1}{2}+it, X), \dots, \Re e^{-i\theta_{r}} P_{F_{r}}(\tfrac{1}{2}+it, X)),
\end{align*}
and $\mu_{T, \bm{F}}$ the measure on $\mathbb{R}^{r}$ by
$
\mu_{T, \bm{F}}(B)
:= \frac{1}{T}\operatorname{meas}(\bm{F}_{\bm{\theta}, X}^{-1}(B) \cap \mathcal{A})
$
for $B \in \mathcal{B}(\mathbb{R}^{r})$.
Put $y_{j} = V_{j} \sigma_{F_{j}}(X)$.
Then we find that
\begin{align}
\label{RTmuS}
& \frac{1}{T}\operatorname{meas}(\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& = \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty))
+ O_{\bm{F}}\l(\exp\l( -b_{0}(\log{\log{X}})^{4(r + 1)} \r)\r)
\end{align}
since $\operatorname{meas}([T, 2T] \setminus \mathcal{A}) \ll_{\bm{F}} T\exp( -b_{0}(\log{\log{X}})^{4(r + 1)} )$ by Lemma \ref{ESAEG_JVDPP}.
For $\bm{x} = (x_{1}, \dots, x_{r}) \in \mathbb{R}^{r}$, set
\begin{align*}
\nu_{T, \bm{F}, \bm{x}}(B)
:= \int_{B}e^{x_{1} \xi_{1} + \cdots + x_{r} \xi_{r}}d\mu_{T, \bm{F}}(\bm{\xi})
\end{align*}
for $B \in \mathcal{B}(\mathbb{R}^{r})$.
Note that $\nu_{T, \bm{F}, \bm{x}}$ is a measure on $\mathbb{R}^{r}$, and has a finite value
for every $B \in \mathcal{B}(\mathbb{R}^{r})$, $\bm{x} \in \mathbb{R}^{r}$, $X \geq 3$
in the sense
\begin{align*}
\nu_{T, \bm{F}, \bm{x}}(B)
\leq \nu_{T, \bm{F}, \bm{x}}(\mathbb{R}^{r})
= \frac{1}{T}\int_{\mathcal{A}}\exp\l( \sum_{j = 1}^{r}x_{j}\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \r)dt
< +\infty.
\end{align*}
This is an analogue of the associated distribution function of $\tilde{M}_{n}(w)$ in Hwang \cite[page 300]{HH1996}.
Under the above notation, we state and prove three lemmas.
\begin{lemma} \label{SPKLH}
For $x_{1}, \dots, x_{r} > 0$, we have
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& = \int_{x_{r} y_{r}}^{\infty} \cdots \int_{x_{1} y_{1}}^{\infty}e^{-(\tau_{1} + \cdots + \tau_{r})}
\nu_{T, \bm{F}, \bm{x}}((y_{1}, \tau_{1}/x_{1}) \times \cdots \times (y_{r}, \tau_{r}/x_{r}))d\tau_{1} \cdots d\tau_{r}.
\end{align*}
\end{lemma}
\begin{proof}
For every $B \in \mathcal{B}(\mathbb{R}^{r})$, it holds that
\begin{align}
\mu_{T, \bm{F}}(B)
& = \int_{B}e^{-(x_{1}v_{1} + \cdots + x_{r}v_{r})}d\nu_{T, \bm{F}, \bm{x}}(\bm{v}).
\end{align}
By Fubini's theorem, we find that
\begin{align*}
& \int_{(y_{1}, \infty) \times \cdots \times (y_{r}, \infty)}e^{-(x_{1}v_{1} + \cdots + x_{r}v_{r})}
d\nu_{T, \bm{F}, \bm{x}}(\bm{v}) \\
& = \int_{(y_{1}, \infty) \times \cdots \times (y_{r}, \infty)}
\l(\int_{x_{r} v_{r}}^{\infty} \cdots \int_{x_{1} v_{1}}^{\infty}
e^{-(\tau_{1} + \cdots + \tau_{r})}d\tau_{1} \cdots d\tau_{r}\r)
d\nu_{T, \bm{F}, \bm{x}}(\bm{v}) \\
& = \int_{x_{r} y_{r}}^{\infty} \cdots \int_{x_{1} y_{1}}^{\infty}e^{-(\tau_{1} + \cdots + \tau_{r})}
\l(\int_{(y_{1}, \tau_{1}/x_{1}) \times \cdots \times (y_{r}, \tau_{r}/x_{r})}1d\nu_{T, \bm{F}, \bm{x}}(\bm{v})\r)
d\tau_{1} \cdots d\tau_{r} \\
& = \int_{x_{r} y_{r}}^{\infty} \cdots \int_{x_{1} y_{1}}^{\infty}e^{-(\tau_{1} + \cdots + \tau_{r})}
\nu_{T, \bm{F}, \bm{x}}((y_{1}, \tau_{1}/x_{1}) \times \cdots \times (y_{r}, \tau_{r}/x_{r}))d\tau_{1} \cdots d\tau_{r}.
\end{align*}
\end{proof}
The next lemma is a generalization of \cite[Lemma 6.2]{LLR2019} in multidimensions.
Define
\begin{gather}
G(u) = \frac{2u}{\pi} + \frac{2 (1 - u)u}{\tan{\pi u}}, \ \
f_{c, d}(u) = \frac{e^{-2\pi i c u} - e^{-2 \pi i d u}}{2}.
\end{gather}
For a set $A$, we denote the indicator function of $A$ by $\bm{1}_{A}$.
\begin{lemma} \label{Multi_BSF}
Let $L$ be a positive number.
Let $c_{1}, \dots, c_{r}, d_{1}, \dots, d_{r}$ be real numbers with $c_{j} < d_{j}$.
Put $\mathscr{R} = (c_{1}, d_{1}) \times \cdots \times (c_{r}, d_{r}) \subset \mathbb{R}^{r}$.
For any $\bm{\xi} = (\xi_{1}, \dots, \xi_{r}) \in \mathbb{R}^{r}$, we have
\begin{align*}
\bm{1}_{\mathscr{R}}(\bm{\xi})
& = W_{L, \mathscr{R}}(\bm{\xi})
+ O_{r}\l( \sum_{j = 1}^{r}\l\{ \l( \frac{\sin(\pi L(\xi_{j} - c_{j}))}{\pi L(\xi_{j} - c_{j})} \r)^2
+ \l( \frac{\sin(\pi L(\xi_{j} - d_{j}))}{\pi L(\xi_{j} - d_{j})} \r)^2 \r\} \r),
\end{align*}
where $W_{L, \mathscr{R}}(\bm{\xi})$ is defined as if $r$ is even,
\begin{align*}
\frac{i^{r}}{2^{r-1}}\sum_{j = 1}^{r}(-1)^{j-1}
\Re \prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i \varepsilon_{j}(h) u \xi_{h}}f_{c_{h}, d_{h}}(\varepsilon_{j}(h)u)\frac{du}{u},
\end{align*}
if $r$ is odd,
\begin{align*}
\frac{i^{r+1}}{2^{r-1}}\sum_{j = 1}^{r}(-1)^{j-1}
\Im \prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i \varepsilon_{j}(h) u \xi_{h}}f_{c_{h}, d_{h}}(\varepsilon_{j}(h)u)\frac{du}{u}.
\end{align*}
Here, $\varepsilon_{j}(h) = 1$ if $1 \leq h \leq j-1$, and $\varepsilon_{j}(h) = -1$ otherwise.
\end{lemma}
\begin{proof}
We use the following formula (cf. \cite[equation (6.1)]{LLR2019})
\begin{align*}
\bm{1}_{(c_{h}, d_{h})}(\xi_{h})
& = \Im \int_{0}^{L}G\l( \frac{u}{L} \r)e^{2\pi i u \xi_{h}}f_{c_{h}, d_{h}}(u)\frac{du}{u} \\
& \quad + O\l( \l( \frac{\sin(\pi L(\xi_{h} - c_{h}))}{\pi L(\xi_{h} - c_{h})} \r)^2
+ \l( \frac{\sin(\pi L(\xi_{h} - d_{h}))}{\pi L(\xi_{h} - d_{h})} \r)^2 \r),
\end{align*}
which leads to
the estimate $\Im \int_{0}^{L}G\l( \frac{u}{L} \r)e^{2\pi i u \xi_{j}}f_{c_{j}, d_{j}}(u)\frac{du}{u} \ll 1$.
Therefore, we obtain
\begin{align}
\label{Multi_BSF_1}
\bm{1}_{\mathscr{R}}(\bm{\xi})
& = \prod_{h = 1}^{r}\Im \int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i u \xi_{h}}f_{c_{h}, d_{h}}(u) \frac{du}{u} \\
& \quad + O_{r}\l( \sum_{j = 1}^{r}\l\{ \l( \frac{\sin(\pi L(\xi_{j} - c_{j}))}{\pi L(\xi_{j} - c_{j})} \r)^2
+ \l( \frac{\sin(\pi L(\xi_{j} - d_{j}))}{\pi L(\xi_{j} - d_{j})} \r)^2 \r\} \r).
\end{align}
For any complex numbers $w_{1}, \dots, w_{r}$, we observe that
\begin{align*}
\Im(w_{1}) \cdots \Im(w_{r})
= \frac{i^{r}}{2^{r}}\sum_{j = 1}^{r}(-1)^{j - 1}
\l(w_{1} \cdots w_{j - 1}\overline{w_{j} \cdots w_{r}} + (-1)^{r} \overline{w_{1} \cdots \overline{w_{j} \cdots w_{r}}}\r).
\end{align*}
In particular, if $r$ is even, then
\begin{align*}
\Im(w_{1}) \cdots \Im(w_{r})
= \frac{i^{r}}{2^{r-1}}\Re\sum_{j = 1}^{r}(-1)^{j - 1}w_{1} \cdots w_{j - 1}\overline{w_{j} \cdots w_{r}},
\end{align*}
and if $r$ is odd, then
\begin{align*}
\Im(w_{1}) \cdots \Im(w_{r})
= \frac{i^{r+1}}{2^{r-1}}\Im\sum_{j = 1}^{r}(-1)^{j - 1}w_{1} \cdots w_{j - 1}\overline{w_{j} \cdots w_{r}}.
\end{align*}
Substituting these to \eqref{Multi_BSF_1}, we obtain Lemma \ref{Multi_BSF}.
\end{proof}
\begin{lemma} \label{FMnu}
Assume that $\bm{F}$, $\bm{\theta}$ satisfy (S4), (A1), and (A2).
Let $c_{1}, \dots, c_{r}, d_{1}, \dots, d_{r}$ be real numbers with $c_{j} < d_{j}$.
Put $\mathscr{R} = (c_{1}, d_{1}) \times \cdots \times (c_{r}, d_{r})$.
Let $T$, $X$ be large numbers depending on $\bm{F}$ and satisfying $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then for any $\bm{x} = (x_{1}, \dots, x_{r}) \in \mathbb{R}^{r}$ satisfying $\|\bm{x}\| \leq (\log{\log{X}})^{2r}$, we have
\begin{align} \label{FMnu1}
\nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
& = \Xi_{X}(\bm{x})\l( \prod_{h = 1}^{r}e^{\frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2} \r)
\times \l\{\prod_{j = 1}^{r}\int_{x_{j} \sigma_{F_{j}}(X) - \frac{d_{j}}{\sigma_{F_{j}}(X)}}^{
x_{j}\sigma_{F_{j}}(X) - \frac{c_{j}}{\sigma_{F_{j}}(X)}}e^{-v^2/2}\frac{dv}{\sqrt{2\pi}} + E_{1}\r\},
\end{align}
where the error term $E_{1}$ satisfies
\begin{align*}
& E_{1}
\ll_{\bm{F}}
\frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{\sqrt{\log{\log{X}}}}
\prod_{h = 1}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}
\end{align*}
for some constant $C = C(\bm{F}) > 0$.
Moreover, if $\|\bm{x}\| \leq b_{3}$ with $b_{3} = b_{3}(\bm{F}) > 0$ sufficiently small,
we have
\begin{align} \label{FMnu2}
\nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
=\l( \prod_{h = 1}^{r}e^{\frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2} \r)
\times \l\{\prod_{j = 1}^{r}\int_{x_{j} \sigma_{F_{j}}(X) - \frac{d_{j}}{\sigma_{F_{j}}(X)}}^{
x_{j}\sigma_{F_{j}}(X) - \frac{c_{j}}{\sigma_{F_{j}}(X)}}e^{-v^2/2}\frac{dv}{\sqrt{2\pi}} + E_{2}\r\},
\end{align}
where the error term $E_{2}$ satisfies
\begin{align*}
& E_{2}
\ll_{\bm{F}} \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \sum_{k = 1}^{r}\l(\frac{x_{k}^2 (d_{k} - c_{k})}{\sigma_{F_{k}}(X)} + \frac{1}{\sigma_{F_{k}}(X)^2} \r)
\prod_{\substack{h = 1 \\ h \not= k}}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}.
\end{align*}
\end{lemma}
\begin{proof}
We show formula \eqref{FMnu1}.
Put $L = b_{4} (\log{\log{X}})^{\a_{\bm{F}}}$ with $b_{4} = b_{4}(\bm{F})$ a small positive constant to be chosen later.
Recall that $\a_{\bm{F}} = \min\l\{ 2r, \frac{1 - 2\vartheta_{\bm{F}}}{2\vartheta_{\bm{F}}} \r\}$.
It follows from Lemma \ref{Multi_BSF} that
\begin{align}
\label{MPnp1}
& \nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
= \int_{\mathbb{R}^{r}}W_{L, \mathscr{R}}(\bm{\xi})e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi}) + E,
\end{align}
where the error term $E$ satisfies the estimate
\begin{align*}
E
\ll \sum_{j = 1}^{r}\int_{\mathbb{R}^{r}}
\l\{ \l( \frac{\sin(\pi L(\xi_{j} - c_{j}))}{\pi L(\xi_{j} - c_{j})} \r)^2
+ \l( \frac{\sin(\pi L(\xi_{j} - d_{j}))}{\pi L(\xi_{j} - d_{j})} \r)^2 \r\}
e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi}).
\end{align*}
First, we estimate $E$.
For $\bm{z} = (z_{1}, \dots, z_{r}) \in \mathbb{C}^{r}$, define
\begin{align*}
\tilde{M}_{T}(\bm{z})
= \int_{\mathbb{R}^{r}}e^{z_{1} \xi_{1} + \cdots + z_{r} \xi_{r}}d\mu_{T, \bm{F}}(\bm{\xi}).
\end{align*}
Then, it holds that
\begin{align*}
\tilde{M}_{T}(\bm{z})
=\frac{1}{T}\int_{\mathcal{A}}\exp\l( \sum_{j = 1}^{r}z_{j}\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X) \r)dt.
\end{align*}
Put $\bm{w} = (w_{1}, \dots, w_{r}) = (x_{1} + i u_{1}, \dots, x_{r} + i u_{r})$ with $u_{j} \in \mathbb{R}$.
When $\norm[]{(u_{1}, \dots, u_{r})} \leq L$ holds,
we have
\begin{align*}
& |\tilde{M}_{T}(\bm{w})| \\
& \leq \bigg| \prod_{j = 1}^{r}\exp\l( \frac{(x_{j} + i u_{j})^2}{2}
\l( \sigma_{F_{j}}(X)^2 - \sigma_{F_{j}}(|w_{j}|)^2 \r)
+ O_{\bm{F}}\l(|x_{j} + i u_{j}|^2 + |x_{j} + i u_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}}\r) \r) \bigg| \\
& \qquad \qquad \qquad \qquad + O_{\bm{F}}\l( \exp\l( -b_{1}(\log{\log{X}})^{4(r + 1)} \r) \r) \\
& \ll_{\bm{F}} \exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\prod_{j = 1}^{r}\exp\l( \frac{x_{j}^2}{2}\l(\sigma_{F_{j}}(X)^2 - \sigma_{F_{j}}(|w_{j}|)^{2} \r) - \frac{u_{j}^2}{3}\sigma_{F_{j}}(X)^2
+ O_{\bm{F}}\l(u_{j}^2L^{\frac{2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}}\r) \r)
\end{align*}
by Proposition \ref{RKLJVDPP} and \eqref{ES_sI1}, where $C = C(\bm{F})$ is some positive constant.
Additionally, by \eqref{EST_Xi}, we find that
\begin{align*}
\prod_{j = 1}^{r}\exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}(|w_{j}|)^{2}\r)
\leq \prod_{j = 1}^{r}\exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}(|x_{j}|)^{2} \r)
\ll_{\bm{F}} \Xi_{X}(\bm{x})\exp\l( C\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\end{align*}
for some $C = C(\bm{F}) > 0$.
By the definition of $\a_{\bm{F}}$, the inequality
$
L^{\frac{2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}}
\leq (2 b_{4})^{^{\frac{2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}}}\log{\log{X}}
$ holds.
Therefore, when $b_{4}$ is sufficiently small, we have
\begin{align} \label{ESTMT1}
& |\tilde{M}_{T}(x_{1} + iu_{1}, \dots, x_{r} + iu_{r})| \\
& \ll_{\bm{F}} \Xi_{X}(\bm{x})\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\prod_{j = 1}^{r}\exp\l( \l(\frac{x_{j}^2}{2} - \frac{u_{j}^2}{4} \r)\sigma_{F_{j}}(X)^2 \r)
\end{align}
for $\|(u_{1}, \dots, u_{r})\| \leq L$.
For any $\ell, \xi \in \mathbb{R}$, we can write
\begin{align}
\l( \frac{\sin(\pi L (\xi - \ell))}{\pi L (\xi - \ell)} \r)^{2}
& = \frac{2}{L^2}\int_{0}^{L}(L - u)\cos(2\pi (\xi - \ell) u)du \\
\label{STTBS}
& = \frac{2}{L^2}\Re \int_{0}^{L}(L - u)e^{2\pi i (\xi - \ell) u}du.
\end{align}
Thus
\begin{align*}
& \int_{\mathbb{R}^{r}}\l( \frac{\sin(\pi L(\xi_{j} - \ell))}{\pi L(\xi_{j} - \ell)} \r)^2
e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}d\mu_{T, \bm{F}}(\bm{\xi}) \\
& = \frac{2}{L^2}\Re \int_{0}^{L}(L - u)\int_{\mathbb{R}^{r}}e^{2\pi i (\xi_{j} - \ell) u}e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi}) du \\
& = \frac{2}{L^2}\Re \int_{0}^{L}e^{-2\pi i \ell u}
(L - u)\tilde{M}_{T}(x_{1}, \dots, x_{j-1}, x_{j} + 2\pi i u, x_{j+1}, \dots, x_{r})du,
\end{align*}
which, by \eqref{ESTMT1}, is
\begin{align*}
& \ll_{\bm{F}} \Xi_{X}(\bm{x})\frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{L^2}\l(\prod_{k = 1}^{r}
\exp\l( \frac{x_{k}^2}{2}\sigma_{F_{k}}(X) \r)\r)
\int_{0}^{L}(L - u)\exp\l( - (\pi \sigma_{F_{j}}(X) u)^2 \r)du \\
& \ll \Xi_{X}(\bm{x})\frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{L\sigma_{F_{j}}(X)}
\prod_{k = 1}^{r}\exp\l( \frac{x_{k}^2}{2}\sigma_{F_{k}}(X)^2 \r).
\end{align*}
It then follows that equation \eqref{MPnp1} satisfies
\begin{align}
\label{EqnubS}
\nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
& = \int_{\mathbb{R}^{r}}W_{L, \mathscr{R}}(\bm{\xi})e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi})+E
\end{align}
with
\begin{align*}
E
\ll_{\bm{F}} \Xi_{X}(\bm{x})\frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{L\sqrt{\log{\log{X}}}}
\prod_{k = 1}^{r}\exp\l( \frac{x_{k}^2}{2}\sigma_{F_{k}}(X)^2 \r).
\end{align*}
For the main term in \eqref{EqnubS}, it is enough to calculate
\begin{align}
\label{MPnp2}
\int_{\mathbb{R}^{r}}\l(\prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i \varepsilon_{j}(h) u \xi_{h}}f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u)\frac{du}{u}\r)
e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi})
\end{align}
for every fixed $1 \leq j \leq r$.
Using Fubini's theorem, we find that \eqref{MPnp2} is equal to
\begin{multline*}
\int_{0}^{L} \cdots \int_{0}^{L}
\l(\prod_{h = 1}^{r}G\l( \frac{u_{h}}{L} \r)\frac{f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u_{h})}{u_{h}}\r)\\
\times \tilde{M}_{T}\l( x_{1} + 2\pi i \varepsilon_{j}(1) u_{1}, \dots, x_{r} + 2 \pi i \varepsilon_{j}(r) u_{r} \r)
du_{1} \cdots du_{r}.
\end{multline*}
Here we divide the range of this integral as
\begin{align*}
\int_{0}^{L} \cdots \int_{0}^{L}
= \int_{0}^{1} \cdots \int_{0}^{1} +
\sum_{k = 0}^{r-1} \int \cdots \int_{D_{k}},
\end{align*}
where
\begin{align*}
\int \cdots \int_{D_{k}}
= \int_{0}^{1} \cdots \int_{0}^{1} \int_{1}^{L}
\overbrace{\int_{0}^{L} \cdots \int_{0}^{L}}^{k}.
\end{align*}
By estimate \eqref{ESTMT1} and the estimates $\frac{f_{c, d}(\pm u)}{u} \ll d - c$,
$G(u / L) \ll 1$ for $0 \leq u \leq L$, the integral over $D_{r-k}$ for $1 \leq k \leq r$ is
\begin{align*}
& \ll_{\bm{F}} \Xi_{X}(\bm{x})\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\l(\prod_{h = 1}^{k-1}(d_{h} - c_{h})\int_{0}^{1}\exp\l( \l(\frac{x_{h}^2}{2} - (\pi u)^2\r)\sigma_{F_{h}}(X)^2 \r)du\r) \\
& \qquad \qquad \qquad \times (d_{k} - c_{k})\int_{1}^{L}\exp\l( \l(\frac{x_{k}^2}{2} - (\pi u)^2 \r)\sigma_{F_{k}}(X)^2 \r)du \\
& \qquad \qquad \qquad \times \l(\prod_{h = k+1}^{r}(d_{h} - c_{h})\int_{0}^{L}
\exp\l( \l(\frac{x_{h}^2}{2} - (\pi u)^2 \r)\sigma_{F_{h}}(X)^2 \r)du\r) \\
& \ll_{\bm{F}} \Xi_{X}(\bm{x})\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
e^{-\sigma_{F_{k}}(X)^2}\prod_{h = 1}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}
\exp\l( \frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r).
\end{align*}
Hence, integral \eqref{MPnp2} is equal to
\begin{multline} \label{MPnp3}
\int_{0}^{1} \cdots \int_{0}^{1}
\l(\prod_{h = 1}^{r}G\l( \frac{u_{h}}{L} \r)\frac{f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u_{h})}{u_{h}}\r)
\tilde{M}_{T}\l( x_{1} + 2\pi i \varepsilon_{j}(1) u_{1}, \dots, x_{r} + 2 \pi i \varepsilon_{j}(r) u_{r} \r)du_{1} \cdots du_{r}\\
+ O_{\bm{F}}\l(\sum_{k=1}^r \Xi_{X}(\bm{x})\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
e^{-\sigma_{F_{k}}(X)^2}\prod_{h = 1}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}
\exp\l( \frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r) \r).
\end{multline}
When $\|(u_{1}, \dots, u_{r})\| \leq 1$, it follows from Proposition \ref{RKLJVDPP} and equation \eqref{ES_sI3} that
\begin{align*}
& \tilde{M}_{T}\l( x_{1} + 2\pi i \varepsilon_{j}(1) u_{1}, \dots, x_{r} + 2 \pi i \varepsilon_{j}(r) u_{r} \r) \\
& = \Xi_{X}(\bm{x})\prod_{h = 1}^{r}\l(1
+ O_{\bm{F}}\l(|u_{h}|\exp\l( D_{1}\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
+ \frac{x_{j}^{4} + u_{j}^{4}}{\log{X}}\r)\r) \\
& \qquad \times\exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u_{h} - 4 \pi^2 u_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r)
+ O_{\bm{F}}\l( \exp\l( -b_{1}(\log{\log{X}})^{4(r + 1)} \r) \r).
\end{align*}
Therefore, the integral of \eqref{MPnp3} is equal to
\begin{multline}
\label{MPnp4}
\Xi_{X}(\bm{x})\prod_{h = 1}^{r}\int_{0}^{1}\l(1
+ O_{\bm{F}}\l(u\exp\l( D_{1}\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
+ \frac{x_{h}^{4} + u_{h}^{4}}{\log{X}} \r)\r)
G\l( \frac{u}{L} \r)f_{c_{h}, d_{h}}(\varepsilon_{j}(h)u)\\
\times \exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u - 4 \pi^2 u^2}{2}\sigma_{F_{h}}(X)^2 \r) \frac{du}{u}\\
+ O_{\bm{F}}\l( \exp\l( -c_{1}(\log{\log{X}})^{4(r + 1)} \r)
\prod_{h = 1}^{r}\int_{0}^{1} \bigg|G\l( \frac{u}{L} \r) \frac{f_{c_{h}, d_{h}}(u)}{u}\bigg|du \r).
\end{multline}
Since $G(u / L) \ll 1$
and $\frac{f_{c_{h}, d_{h}}(\pm u)}{u} \ll d_{h} - c_{h}$,
we find that
\begin{align*}
& \int_{0}^{1}\l(u\exp\l( D_{1}\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
+ \frac{x_{h}^{4} + u^{4}}{\log{X}} \r)G\l( \frac{u}{L} \r)f_{c_{h}, d_{h}}(\varepsilon_{j}(h)u) \\
& \qquad \qquad \qquad \times \exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u - 4 \pi^2 u^2}{2}\sigma_{F_{h}}(X)^2 \r) \frac{du}{u} \\
& \ll \exp\l( \frac{x_{h}^{2}}{2}\sigma_{F_{h}}(X)^2 \r)(d_{h} - c_{h}) \\
& \qquad \qquad \qquad \times \int_{0}^{1} \l(u \exp\l( D_{1}\| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
+ \frac{x_{h}^{4} + u^{4}}{\log{X}} \r)\exp\l(- 2 \pi^2 u^2\sigma_{F_{h}}(X)^2 \r)du \\
& \ll_{\bm{F}} \exp\l( D_{1} \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\times \exp\l( \frac{x_{h}^{2}}{2}\sigma_{F_{h}}(X)^2 \r)\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)^{2}},
\end{align*}
and that
\begin{align*}
& \int_{0}^{1}G\l( \frac{u}{L} \r)f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u)
\exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u - 4 \pi^2 u^2}{2}\sigma_{F_{h}}(X)^2 \r) \frac{du}{u} \\
& \ll (d_{h} - c_{h})\exp\l( \frac{x_{h}^{2}}{2}\sigma_{F_{h}}(X)^2 \r)
\int_{0}^{1} \exp\l(- 2 \pi^2 u^2\sigma_{F_{h}}(X)^2 \r)du \\
& \ll \exp\l( \frac{x_{h}^{2}}{2}\sigma_{F_{h}}(X)^2 \r)
\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}.
\end{align*}
Moreover, we find that
\begin{align*}
& \int_{1}^{L}G\l( \frac{u}{L} \r)f_{c_{h}, d_{h}}( \varepsilon_{j}(h) u)
\exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u - 4 \pi^2 u^2}{2}\sigma_{F_{h}}(X)^2 \r) \frac{du}{u} \\
& \ll_{\bm{F}} \exp\l( \frac{x_{h}^{2}}{2}\sigma_{F_{h}}(X)^{2} \r)\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}e^{-\sigma_{F_{h}}(X)^2},
\end{align*}
and that
\begin{align*}
\int_{0}^{1}\bigg|G\l( \frac{u}{L} \r)\frac{f_{c_{h}, d_{h}}( \varepsilon_{j}(h) u)}{u}\bigg|du
\ll d_{h} - c_{h}.
\end{align*}
Also, by \eqref{EST_Xi}, it holds that
\begin{align*}
&\Xi_{X}(\bm{x})\exp\l( D_{2} \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\prod_{j = 1}^{r} \exp\l( \frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 \r) \times \frac{1}{\sqrt{\log{\log{X}}}}\prod_{h = 1}^{r}\frac{1}{\sigma_{F_{h}}(X)}\\
&\geq \frac{1}{\sqrt{\log{\log{X}}}}\prod_{h = 1}^{r}\frac{1}{\sigma_{F_{h}}(X)}
\geq \exp\l( -b_{1}(\log{\log{X}})^{4(r + 1)} \r)
\end{align*}
for some constant $D_{2} = D_{2}(\bm{F}) > 0$.
From these estimates and \eqref{MPnp4}, integral \eqref{MPnp3} is equal to
\begin{multline*}
\Xi_{X}(\bm{x})\prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)f_{c_{h}, d_{h}}(\varepsilon_{j}(h)u)
\exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u - 4 \pi^2 u^2}{2}\sigma_{F_{h}}(X)^2 \r) \frac{du}{u}\\
+ O_{\bm{F}}\l( \Xi_{X}(\bm{x})\exp\l( D_{3} \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\prod_{j = 1}^{r} \exp\l( \frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 \r) \times
\frac{1}{\sqrt{\log{\log{X}}}}\prod_{h = 1}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}\r)
\end{multline*}
with $D_{3} = \max\{ D_{1}, D_{2} \}$.
Using the well known formula
\begin{align}
\label{FTGF}
\frac{1}{\sqrt{2\pi}}\int_{\mathbb{R}}e^{-iv\xi}e^{-\eta v^2}dv
= \frac{1}{\sqrt{2\eta}}\exp\l( -\frac{\xi^2}{4\eta} \r),
\end{align}
we can rewrite the above main term as
\begin{multline*}
\Xi_{X}(\bm{x})\l(\prod_{h = 1}^{r} \frac{\exp\l( \frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r)}{\sqrt{2\pi}} \r)\\
\times \int_{\mathbb{R}^{r}} e^{-(v_{1}^2 + \cdots + v_{r}^2)/2}
\bigg\{
\prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i \varepsilon_{j}(h) u (x_{h}\sigma_{F_{h}}(X)^2 - v_{h} \sigma_{F_{h}}(X))}f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u)\frac{du}{u}\bigg\}
d\bm{v}.
\end{multline*}
Combining this with \eqref{MPnp3}, we see that integral \eqref{MPnp2} is equal to
\begin{align*}
& \Xi_{X}(\bm{x})\l(\prod_{h = 1}^{r} \frac{\exp\l( \frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r)}{\sqrt{2\pi}} \r) \\
& \qquad \times \int_{\mathbb{R}^{r}} e^{-(v_{1}^2 + \cdots + v_{r}^2)/2}
\bigg\{
\prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i \varepsilon_{j}(h) u (x_{h}\sigma_{F_{h}}(X)^2 - v_{h} \sigma_{F_{h}}(X))}f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u)\frac{du}{u}\bigg\}
d\bm{v} \\
& + O_{\bm{F}}\l( \Xi_{X}(\bm{x})\exp\l( D_{1} \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\prod_{j = 1}^{r} \exp\l( \frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 \r) \times \frac{1}{\sqrt{\log{\log{X}}}}
\prod_{h = 1}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)} \r).
\end{align*}
Substituting this equation to the definition of $W_{L, \mathscr{R}}$ and using Lemma \ref{Multi_BSF} and equation \eqref{EqnubS},
we obtain
\begin{align}
\label{MPnp5}
& \nu_{T, \bm{F}, \bm{x}}(\mathscr{R}) \\
& = \Xi_{X}(\bm{x})\l(\prod_{h = 1}^{r} \frac{\exp\l( \frac{x_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r)}{\sqrt{2\pi}} \r) \\ \nonumber
& \qquad\times\bigg\{\int_{\mathbb{R}^{r}} e^{-\frac{v_{1}^2 + \cdots + v_{r}^2}{2}}
\bm{1}_{\mathscr{R}}\l( x_{1}\sigma_{F_{1}}(X)^2 - v_{1}\sigma_{F_{1}}(X), \dots, x_{r}\sigma_{F_{r}}(X)^2 - v_{r}\sigma_{F_{r}}(X) \r)
d\bm{v}
+ E_{3} + E_{4}\bigg\},
\end{align}
where $E_{3}$ and $E_{4}$ satisfy
\begin{align*}
& E_{3} \ll_{\bm{F}}
\sum_{j = 1}^{r}\int_{\mathbb{R}^{r}}\bigg\{ \l(
\frac{\sin(\pi L (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - c_{j}))}
{\pi L (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - c_{j})} \r)^{2} \\
& \qquad \qquad \qquad \qquad+ \l(\frac{\sin(\pi L (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - d_{j}))}
{\pi L (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - d_{j})} \r)^{2} \bigg\}
\times e^{-(v_{1}^2 + \cdots + v_{r}^2)/2}d\bm{v},
\end{align*}
and
\begin{align*}
E_{4}
\ll_{\bm{F}}
\frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{\exp\l( C \| \bm{x} \|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)}{\sqrt{\log{\log{X}}}}
\prod_{h = 1}^{r}\frac{d_{h} - c_{h}}{\sigma_{F_{h}}(X)}
\end{align*}
for some constant $C = C(\bm{F}) > 0$.
By equation \eqref{STTBS}, it holds that, for any $\ell \in \mathbb{R}$,
\begin{align*}
& \int_{\mathbb{R}^{r}}\l(
\frac{\sin(\pi L (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - \ell))}
{\pi L (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - \ell)} \r)^{2}
\times e^{-(v_{1}^2 + \cdots + v_{r}^2)/2}d\bm{v} \\
& = \frac{2}{L^2}\Re\int_{0}^{L}(L - \a)\l(\int_{\mathbb{R}^{r}}e^{2\pi i (x_{j}\sigma_{F_{j}}(X)^2 - v_{j}\sigma_{F_{j}}(X) - \ell)\a}
e^{-(v_{1}^2 + \cdots + v_{r}^2)/2}d\bm{v}\r) d\a \\
& = \frac{2(2\pi)^{(r-1)/2}}{L^2}\Re\int_{0}^{L}(L - \a)e^{2\pi i (x_{j}\sigma_{F_{j}}(X)^2 - \ell)\a}
\l(\int_{\mathbb{R}}e^{-2\pi i v\sigma_{F_{j}}(X) \a}e^{-v^2/2}dv \r)d\a,
\end{align*}
which, by \eqref{FTGF}, becomes
\begin{align*}
& = \frac{2(2\pi)^{r/2}}{L^2}\Re\int_{0}^{L}(L - \a)e^{2\pi i (x_{j}\sigma_{F_{j}}(X)^2 - \ell)\a}
\exp\l( -2 \pi^2 \a^2 \sigma_{F_{j}}(X)^{2} \r)d\a \\
& \ll_{\bm{F}} \frac{1}{L \sigma_{F_{j}}(X)}
\ll_{\bm{F}} \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}.
\end{align*}
Hence, we have $E_{3} \ll_{\bm{F}} \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}$.
Finally, by simple calculations, we can write
\begin{align*}
& \int_{\mathbb{R}^{r}} e^{-\frac{v_{1}^2 + \cdots + v_{r}^2}{2}}
\bm{1}_{\mathscr{R}}\l( x_{1}\sigma_{F_{1}}(X)^2 - v_{1}\sigma_{F_{1}}(X), \dots, x_{r}\sigma_{F_{r}}(X)^2 - v_{r}\sigma_{F_{r}}(X) \r)
d\bm{v} \\
& =\prod_{j = 1}^{r}\int_{x_{j}\sigma_{F_{j}}(X) - \frac{d_{j}}{\sigma_{F_{j}}(X)}}^{x_{j} \sigma_{F_{j}}(X) - \frac{c_{j}}{\sigma_{F_{j}}(X)}}
e^{-v^2/2}\frac{dv}{\sqrt{2\pi}}
\end{align*}
and this completes the proof of \eqref{FMnu1}.
Next, we consider \eqref{FMnu2}.
Using Proposition \ref{RKLJVDPP} and equation \eqref{ES_sI2}, we have
\begin{align*}
& M_{T}\l( x_{1} + 2\pi i \varepsilon_{j}(1) u_{1}, \dots, x_{r} + 2 \pi i \varepsilon_{j}(r) u_{r} \r) \\
& = \prod_{h = 1}^{r}\l(1 + O_{\bm{F}}\l(|x_{h} + iu_{h}|^2\r)\r)
\exp\l( \frac{x_{h}^2 + 4 \pi i \varepsilon_{j}(h) x_{h} u_{h} - 4 \pi^2 u_{h}^2}{2}\sigma_{F_{h}}(X)^2 \r) \\
& \qquad \qquad \qquad \qquad+ O_{\bm{F}}\l( \exp\l( -b_{1}(\log{\log{X}})^{4(r + 1)} \r) \r)
\end{align*}
when $\| \bm{x} \|$, $\| \bm{u} \|$ are sufficiently small.
By using this equation, we can prove \eqref{FMnu2} similarly to the proof of \eqref{FMnu1}.
\end{proof}
\begin{comment}
\begin{lemma} \label{FMnu4}
Assume that $\bm{F}$, $\bm{\theta}$ satisfy (S4), (A1), and (A2).
Let $c_{1}, \dots, c_{r}, d_{1}, \dots, d_{r}$ be real numbers with $c_{j} < d_{j}$.
Put $\mathscr{R} = (c_{1}, d_{1}) \times \cdots \times (c_{r}, d_{r})$.
Let $T$, $X$ be large numbers with $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then, there exists a positive constant $b_{3} = b_{3}(\bm{F})$ such that
for any $\bm{x} = (x_{1}, \dots, x_{r}) \in \mathbb{R}^{r}$ with
$\norm[]{\bm{x}} \leq (\log{\log{X}})^{2r}$, we have
\begin{align*}
& \nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
\ll_{\bm{F}} \l(\prod_{k = 1}^{r}\frac{d_{k} - c_{k}}{\sigma_{F_{k}}(X)} + \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}} \r)
\times \prod_{j = 1}^{r}\exp\l( \frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2
+ O_{\bm{F}}\l( |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
\end{align*}
\end{lemma}
\begin{proof}
We give the sketch of the proof only because this lemma can be shown in the same way as Lemma \ref{FMnu}.
Similarly to \eqref{ESTMT1}, we can show that
\begin{align}
\label{ESTMT1m}
& |M_{T}(x_{1} + iu_{1}, \dots, x_{r} + iu_{r})|
\ll_{\bm{F}} \bigg| \prod_{j = 1}^{r}\exp\l( \l(\frac{x_{j}^2}{2} - \frac{u_{j}^2}{4} \r) \sigma_{F_{j}}(X)^2
+ O_{\bm{F}}\l( |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r) \bigg|
\end{align}
for $\norm[]{(u_{1}, \dots, u_{r})} \leq b_{3} (\log{\log{X}})^{\a_{\bm{F}}} =: L$,
where $b_{3} > 0$ is the same constant as in the proof of Lemma \ref{FMnu}.
By using this inequality, we obtain
\begin{align*}
\nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
& = \int_{\mathbb{R}^{r}}W_{L, \mathscr{R}}(\bm{\xi})e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi}) \\
& \quad+ O_{\bm{F}}\l( \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
\prod_{k = 1}^{r}\exp\l( \frac{x_{k}^2}{2}\sigma_{F_{k}}(X)^2
+ O_{\bm{F}}\l( |x_{j}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r) \r)
\end{align*}
similarly to \eqref{EqnubS}.
Also, from Fubini's theorem, it holds that
\begin{align}
\label{pFMnu21}
& \int_{\mathbb{R}^{r}}\l(\prod_{h = 1}^{r}\int_{0}^{L}G\l( \frac{u}{L} \r)
e^{2\pi i \varepsilon_{j}(h) u \xi_{h}}f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u)\frac{du}{u}\r)
e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi}) \\
& = \int_{0}^{L} \cdots \int_{0}^{L}
\l(\prod_{h = 1}^{r}G\l( \frac{u_{h}}{L} \r)\frac{f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u_{h})}{u_{h}}\r) \\
& \qquad \qquad \qquad \qquad \times M_{T}\l( x_{1} + 2\pi i \varepsilon_{j}(1) u_{1}, \dots, x_{r} + 2 \pi i \varepsilon_{j}(r) u_{r} \r)
du_{1} \cdots du_{r}.
\end{align}
Using the estimates $G(u / L) \ll 1$, $f_{c_{h}, d_{h}}(\pm u) \ll 1$, $\frac{f_{c_{h}, d_{h}}(\pm u)}{u} \ll d_{h} - c_{h}$,
and inequality \eqref{ESTMT1m}, we find that
\begin{align*}
& \int_{0}^{L} \cdots \int_{0}^{L}
\l(\prod_{h = 1}^{r}G\l( \frac{u_{h}}{L} \r)\frac{f_{c_{h}, d_{h}}(\varepsilon_{j}(h) u_{h})}{u_{h}}\r) \\
& \qquad \qquad \qquad \qquad \times M_{T}\l( x_{1} + 2\pi i \varepsilon_{j}(1) u_{1}, \dots, x_{r} + 2 \pi i \varepsilon_{j}(r) u_{r} \r)
du_{1} \cdots du_{r} \\
& \ll_{\bm{F}} \prod_{k = 1}^{r}\frac{d_{k} - c_{k}}{\sigma_{F_{k}}(X)}\exp\l( \frac{x_{k}^2}{2}\sigma_{F_{k}}(X)^2
+ O_{\bm{F}}\l( |x_{k}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
\end{align*}
From this estimate, equation \eqref{pFMnu21}, and the definition of $W_{L, \mathscr{R}}$, we have
\begin{align*}
& \int_{\mathbb{R}^{r}}W_{L, \mathscr{R}}(\bm{\xi})e^{x_{1}\xi_{1} + \cdots + x_{r}\xi_{r}}
d\mu_{T, \bm{F}}(\bm{\xi}) \\
& \ll_{\bm{F}} \prod_{k = 1}^{r}\frac{d_{k} - c_{k}}{\sigma_{F_{k}}(X)}\exp\l( \frac{x_{k}^2}{2}\sigma_{F_{k}}(X)^2
+ O_{\bm{F}}\l( |x_{k}|^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r).
\end{align*}
Thus, we obtain this lemma.
\end{proof}
\end{comment}
\begin{proof}[Proof of Proposition \ref{Main_Prop_JVD}]
We firstly prove Proposition \ref{Main_Prop_JVD} in the case $V_{j}$'s are nonnegative.
Let $\bm{x} = (x_{1}, \dots, x_{r}) \in (\mathbb{R}_{> 0})^{r}$ satisfying
$\|\bm{x}\| \leq b_{4}$ with $b_{4}$ the same number as in Lemma \ref{FMnu}.
By Lemma \ref{SPKLH} and equation \eqref{FMnu2}, we have
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& = \prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}\int_{x_{j} y_{j}}^{\infty}e^{-\tau}
\int_{x_{j}\sigma_{F_{j}}(X) - \frac{\tau / x_{j}}{\sigma_{F_{j}}(X)}}^{x_{j}\sigma_{F_{j}}(X) - \frac{y_{j}}{\sigma_{F_{j}}(X)}}
e^{-v^2 / 2}\frac{dv}{\sqrt{2\pi}} d\tau \\
& \qquad \qquad \qquad+ O_{\bm{F}}\l( \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
\prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 -x_{j} y_{j}}
+ E \times \prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2} \r),
\end{align*}
where
\begin{align*}
& E =
\sum_{k = 1}^{r}\int_{x_{r} y_{r}}^{\infty} \cdots \int_{x_{1} y_{1}}^{\infty}
e^{-(\tau_{1} + \cdots + \tau_{r})}
\l(\frac{x_{k}^2 (\tau_{k}/x_{k} - y_{k})}{\sigma_{F_{k}}(X)} + \frac{1}{\sigma_{F_{k}}(X)^2} \r)
\prod_{\substack{h = 1 \\ h \not= k}}^{r}\frac{\tau_{h} / x_{h} - y_{h}}{\sigma_{F_{h}}(X)}d\tau_{1} \cdots d\tau_{r}.
\end{align*}
Now, simple calculations lead that
\begin{align*}
\int_{x_{j} y_{j}}^{\infty}e^{-\tau}
\int_{x_{j}\sigma_{F_{j}}(X) - \frac{\tau / x_{j}}{\sigma_{F_{j}}(X)}}^{x_{j}\sigma_{F_{j}}(X) - \frac{y_{j}}{\sigma_{F_{j}}(X)}}
e^{-v^2 / 2}\frac{dv}{\sqrt{2\pi}}d\tau
= \exp\l( -\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 \r)\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{align*}
since $y_{j} = V_{j} \sigma_{F_{j}}(X)$.
Therefore, we obtain
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& =\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2 / 2}\frac{du}{\sqrt{2\pi}}
+ O_{\bm{F}}\l( \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
\prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 -x_{j} y_{j}}
+ E \prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2} \r).
\end{align*}
Here, we decide $x_{j}$'s as $x_{j} = \max\{ 1, V_{j} \} / \sigma_{F_{j}}(X)$, where
$V_{j}$'s must satisfy the inequality $V_{j} \leq R \sigma_{F_{j}}(X) $.
Then, we see that
\begin{align}
\label{pMPJVD1}
e^{-x_{j} y_{j}}
= e^{-\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 - x_{j} y_{j}}
\ll e^{-\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2} e^{-V_{j}^2/2}.
\end{align}
This estimate leads that
\begin{align*}
\frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
\prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 -x_{j} y_{j}}
& \ll_{r} \frac{e^{-(V_{1}^2 + \cdots + V_{r}^2)/2}}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}} \\
& \ll_{r} \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
\prod_{j = 1}^{r}(1 + V_{j})\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
Moreover, since it holds that
\begin{align} \label{BFpte}
\int_{x_{j}y_{j}}^{\infty}\l( \frac{\tau}{x_{j}} - y_{j} \r)e^{-\tau}d\tau
= \frac{e^{-x_{j}y_{j}}}{x_{j}},
\end{align}
we have
\begin{align}
& \int_{x_{r} y_{r}}^{\infty} \cdots \int_{x_{1} y_{1}}^{\infty}e^{-(\tau_{1} + \cdots + \tau_{r})}
\l(\frac{x_{k}^2 (d_{k} - c_{k})}{\sigma_{F_{k}}(X)} + \frac{1}{\sigma_{F_{k}}(X)^2} \r)
\prod_{\substack{h = 1 \\ h \not= k}}^{r}\frac{\tau_{h} / x_{h} - y_{h}}{\sigma_{F_{h}}(X)}
d\tau_{1} \cdots d\tau_{r} \\ \nonumber
& = x_{k}^{2}\prod_{j = 1}^{r}
\frac{1}{\sigma_{F_{j}}(X)}\int_{x_{j} y_{j}}^{\infty}\l( \frac{\tau}{x_{j}} - y_{j} \r)e^{-\tau}d\tau
+ \frac{e^{-x_{k} y_{k}}}{\sigma_{F_{k}}(X)^2}\prod_{\substack{j = 1 \\ j \not= k}}^{r}\frac{1}{\sigma_{F_{j}}(X)}
\int_{x_{j} y_{j}}^{\infty}\l( \frac{\tau}{x_{j}} - y_{j} \r)e^{-\tau}d\tau \\
\label{pMPJVD2}
& = x_{k}^2\prod_{j = 1}^{r}\frac{e^{-x_{j} y_{j}}}{x_{j}\sigma_{F_{j}}(X)}
+ \frac{x_{k}}{\sigma_{F_{k}}(X)}
\prod_{j = 1}^{r}\frac{e^{-x_{j} y_{j}}}{x_{j} \sigma_{F_{j}}(X)}
\end{align}
for every $1 \leq k \leq r$.
By estimate \eqref{pMPJVD1} and $x_{j}\sigma_{F_{j}}(X) \asymp 1 + V_{j}$,
we can write
\begin{align*}
\frac{e^{-x_{j} y_{j}}}{x_{j}\sigma_{F_{j}}(X)}
\ll e^{-\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}\frac{e^{-V_{j}^2 / 2}}{1 + V_{j}}
\ll e^{-\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
By this observation, \eqref{pMPJVD2} is
\begin{align*}
& \ll_{r} \l(x_{k}^2 + \frac{x_{k}}{\sigma_{F_{k}}(X)}\r)
\prod_{j = 1}^{r}e^{-\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \\
& \ll_{\bm{F}} \frac{1 + V_{k}^2}{\log{\log{X}}}
\prod_{j = 1}^{r}e^{-\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
Hence, we have
\begin{align*}
E \prod_{j = 1}^{r}e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}
\ll_{\bm{F}} \frac{1 + \norm[]{\bm{V}}^2}{\log{\log{X}}}\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
From the above estimations, we obtain
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& = \prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
+ O_{\bm{F}}\l( \l\{\frac{\prod_{k = 1}^{r}(1 + V_{k})}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^2}{\log{\log{X}}}\r\}
\times \prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \r)
\end{align*}
for $0 \leq V_{j} \leq b_{2} \sigma_{F_{j}}(X)$.
Thus, by this formula and \eqref{RTmuS}, we complete the proof of Proposition \ref{Main_Prop_JVD}
in the case $V_{j}$'s are nonnegative.
In order to finish the proof of Proposition \ref{Main_Prop_JVD}, we consider the negative cases.
It suffices to show that,
for the case $- b\sigma_{F_{1}}(X) \leq V_{1} \leq 0$ and $0 \leq V_{j} \leq b \sigma_{F_{j}}(X)$,
\begin{align*}
& \frac{1}{T}\operatorname{meas}(\S_{X}(T, (-V_{1}, V_2, \dots, V_{r}); \bm{F}, \bm{\theta})) \\
& = \l( 1 + O_{\bm{F}}\l( \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^2}{\log{\log{X}}} \r) \r)
\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}
\end{align*}
since other cases can be shown similarly by induction.
By the definition of the set $\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta})$, it holds that
\begin{multline*}
\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta})
= \S_{X}(T, (V_{2}, \dots, V_{r}); (F_{2}, \dots, F_{r}), (\theta_{2}, \dots, \theta_{r}))\\
\setminus \S_{X}(T, (-V_{1}-0, V_{2}, \dots, V_{r}); \bm{F}, (\pi - \theta_{1}, \theta_2, \dots, \theta_{r})),
\end{multline*}
where we regard that if $r = 1$, the first set on the right hand side is $[T, 2T]$.
Therefore, from the nonnegative cases, we have
\begin{align}
\label{PNMPJVD1}
& \frac{1}{T}\operatorname{meas}(\S_{X}(T, (-V_{1}, V_2, \dots, V_{r}); \bm{F}, \bm{\theta})) \\
& = \frac{1}{T}\operatorname{meas}(\S_{X}(T, (V_{2}, \dots, V_{r}); (F_{2}, \dots, F_{r}), (\theta_{2}, \dots, \theta_{r}))) \\
& \qquad \qquad \qquad \qquad
-\frac{1}{T}\operatorname{meas}(\S_{X}(T, (-V_{1} - 0, V_{2}, \dots, V_{r}); \bm{F}, (\pi - \theta_{1}, \theta_2, \dots, \theta_{r}))) \\
& = \l( 1 + E_{1} \r)\prod_{j = 2}^{r}\int_{V_{j}}^{\infty}e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}
- (1 + E_{2})\prod_{j = 1}^{r}
\int_{|V_{j}|}^{\infty}e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}.
\end{align}
Here, $E_{1}$ and $E_{2}$ satisfy
\begin{gather*}
E_{1}
\ll_{\bm{F}} \frac{\prod_{k = 2}^{r}(1 + V_{k})}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{(V_{2}, \dots, V_{r})}^2}{\log{\log{X}}},\\
E_{2}
\ll_{\bm{F}} \frac{\prod_{k = 1}^{r}(1 + V_{k})}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{(V_{1}, \dots, V_{r})}^2}{\log{\log{X}}}.
\end{gather*}
Hence, we find that \eqref{PNMPJVD1} is equal to
\begin{align*}
& \l(1 - \int_{-V_{1}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}\r)
\prod_{j = 2}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad+ E_{1}\prod_{j = 2}^{r}\int_{V_{j}}^{\infty}e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}
+ E_{2}\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}} \\
& = \l( 1 + O_{\bm{F}}\l( \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^2}{\log{\log{X}}} \r) \r)\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
Thus, we also obtain the negative cases of Proposition \ref{Main_Prop_JVD}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{Main_Prop_JVD3}]
Let $\bm{V} = (V_{1}, \dots, V_{r}) \in (\mathbb{R}_{\geq 0})^{r}$ satisfying
$\| \bm{V} \| \leq (\log{\log{X}})^{2r}$,
and put $x_{j} = \max\{ 1, V_{j} \} / \sigma_{F_{j}}(X)$.
Similarly to the proof of Proposition \ref{Main_Prop_JVD} by using \eqref{FMnu1} instead of \eqref{FMnu2}, we obtain
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& = \Xi_{X}(\bm{x})\Bigg\{\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2 / 2}\frac{du}{\sqrt{2\pi}} \\
& \qquad \qquad \qquad+ O_{\bm{F}}\l( \exp\l( C \l(\frac{\| \bm{V} \|}{\sqrt{\log{\log{X}}}}\r)^{\frac{2 - 2\vartheta{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\frac{\prod_{k =1}^{r}(1 + V_{k})}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2 / 2}\frac{du}{\sqrt{2\pi}} + E\r)\Bigg\}
\end{align*}
for $0 \leq V_{j} \leq (\log{\log{X}})^{2r}$, where
\begin{align*}
& E =
\exp\l( C \l(\frac{\| \bm{V} \|}{\sqrt{\log{\log{X}}}}\r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\frac{1}{\sqrt{\log{\log{X}}}} \times \prod_{j = 1}^{r}\frac{e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}}{\sigma_{F_{j}}(X)}
\int_{x_{j}}^{y_{j}}\l( \frac{\tau}{x_{j}} - y_{j} \r)e^{-\tau}d\tau.
\end{align*}
Here, $C = C(\bm{F})$ is a positive constant.
Moreover, using \eqref{BFpte} we have
\begin{align*}
E
= \exp\l( C \l(\frac{\| \bm{V} \|}{\sqrt{\log{\log{X}}}}\r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\frac{1}{\sqrt{\log{\log{X}}}} \prod_{j = 1}^{r}\frac{e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2 - x_{j} y_{j}}}{x_{j}\sigma_{F_{j}}(X)}
\end{align*}
By estimate \eqref{pMPJVD1} and $x_{j}\sigma_{F_{j}}(X) \asymp 1 + V_{j}$,
we can write
\begin{align*}
\frac{e^{\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2-x_{j} y_{j}}}{x_{j}\sigma_{F_{j}}(X)}
\ll \frac{e^{-V_{j}^2 / 2}}{1 + V_{j}}
\ll \int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
By this observation, we have
\begin{align*}
E
\ll_{r} \exp\l( C \l(\frac{\| \bm{V} \|}{\sqrt{\log{\log{X}}}}\r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\frac{1}{\sqrt{\log{\log{X}}}}\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
From the above estimations, we obtain
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& = \Xi_{X}(\bm{x})\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \\
& \qquad \qquad \qquad\times \Bigg\{ 1
+ O_{\bm{F}}\l( \exp\l( C \l(\frac{\| \bm{V} \|}{\sqrt{\log{\log{X}}}}\r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r)
\l\{\frac{\prod_{k = 1}^{r}(1 + V_{k})}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1}{\sqrt{\log{\log{X}}}}\r\} \r) \Bigg\}.
\end{align*}
In particular, by the definition of $\Xi_{X}$ \eqref{def_Xi}, assumptions (A1), (A2),
and the analyticity of $\Xi_{X}$ coming from Lemma \ref{Prop_Psi_FPP}, it holds that
\begin{align*}
\Xi_{X}(\bm{x})
= \l( 1 + O_{\bm{F}}\l( \frac{1}{\sqrt{\log{\log{X}}}} \r) \r)
\Xi_{X}\l( \tfrac{V_{1}}{\sigma_{F_{1}}(X)}, \dots, \tfrac{V_{r}}{\sigma_{F_{r}}(X)} \r).
\end{align*}
Thus, by these formulas and \eqref{RTmuS}, we obtain Proposition \ref{Main_Prop_JVD3} when $\| \bm{V} \| \leq (\log{\log{X}})^{2r}$.
\end{proof}
\subsection{Proof of Theorem \ref{GMDP}}
We prove the following lemma as a preparation.
\begin{lemma} \label{LDESRDP}
Let $F$ be a Dirichlet series satisfying (S4), (S5), and \eqref{SNC}.
There exists a positive constant $c_{0} = c_{0}(F)$ such that for any large numbers $X$, $T$
with $X \leq T^{1/2}$
and for any $\Delta > 0$, $10 \Delta \sigma_{F}(X)^2 + \Delta^{-1} \leq V \leq \frac{\log{(T / \log{T})}}{\Delta \log{X}}$
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{P_{F}(\tfrac{1}{2} + it, X) > V}
\ll \exp\l( -\Delta V \log\l(\frac{c_{0} V}{\Delta \sigma_{F}(X)^2}\r) \r).
\end{align*}
\end{lemma}
\begin{proof}
Similarly to the proof of \eqref{MVEDPCLPP}, we can obtain
\begin{align*}
\frac{1}{T}\int_{T}^{2T}|P_{F}(\tfrac{1}{2} + it, X)|^{2k}dt
\ll (C k \sigma_{F}(X)^2)^{k}
\end{align*}
for any $1 \leq k \leq \frac{\log{(T / \log{T})}}{\log{X}}$ with $C = C(F)$ a positive constant.
It follows from this estimate that
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{P_{F}(\tfrac{1}{2} + it, X) > V}
\ll \l( \frac{C k \sigma_{F}(X)^2}{V^2} \r)^{k}.
\end{align*}
Hence, choosing $k = \lfloor \Delta V \rfloor$, we obtain
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{P_{F}(\tfrac{1}{2} + it, X) > V}
\ll \exp\l( -\Delta V \log\l( \frac{V}{4 \Delta \sigma_{F}(X)^2} \r) \r)
\end{align*}
for any $V \leq \frac{\log{(T / \log{T})}}{\Delta \log{X}}$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{GMDP}]
To make the dependency of the limit on $X$ and $T$ explicit,
we introduce three more parameters $n_{\bm{F}}^{*}$, $\vartheta_{\bm{F}}^{*}$, and $b_{\bm{F}}^{*}$, where
\begin{align*}
n_{\bm{F}}^{*} = \min_{1 \leq j \leq r}n_{F_{j}},\
\vartheta_{\bm{F}}^{*} = \min_{1 \leq j \leq r}\vartheta_{F_{j}}, \
b_{\bm{F}}^{*}
=\min_{\substack{1\leq i\leq r\\\vartheta_{F_i}=\vartheta_{\bm F}^*} }\sup_{p \text{ prime}}\{\frac{|b_{F_i}(p)|}{p^{\vartheta_{F_i}}} \}.
\end{align*}
Let $\varepsilon > 0$ be any small fixed constant and let
$ X
\leq \max\{c_{1}(\log{T} \log{\log{T}})^{\frac{2}{1 + 2\vartheta_{\bm{F}}^{*}}}, c_{2}\frac{(\log{T})^2 \log{\log{T}}}{\log_{3}{T}}\}$,
where $c_{1} = c_{1}(\bm{F}, k, \varepsilon)$ and $c_{2} = c_{2}(\bm{F}, k, \varepsilon)$ are some positive constants to be chosen later.
Denote $\phi_{\bm{F}}(t, X) = \min_{1 \leq j \leq r}\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2} + it, X)$ and
\begin{align}
\Phi_{\bm{F}}(T, V, X)
= \operatorname{meas}\set{t \in [T, 2T]}{\phi_{\bm{F}}(t, X) > V}.
\end{align}
Then we have
\begin{align*}
\int_{T}^{2T}\exp\l( 2k \phi_{\bm{F}}(t, X) \r)dt
= \int_{-\infty}^{\infty}2k e^{2k V} \Phi_{\bm{F}}(T, V, X)dV.
\end{align*}
We divide the range of the integral on the right hand side as
\begin{align*}
& I_{1} + I_{2} + I_{3} + I_{4} + I_{5} \\
& := \bigg(\int_{-\infty}^{\sqrt{\log{\log{X}}}}
+ \int_{\sqrt{\log{\log{X}}}}^{C k \log{\log{X}}}
+ \int_{C k \log{\log{X}}}^{\frac{\log{T}}{\log{\log{T}}}} + \int_{\frac{\log{T}}{\log{\log{T}}}}^{c_{3} \log{T}}
+ \int_{c_{3} \log{T}}^{\infty}\bigg) 2k e^{2k V} \Phi_{\bm{F}}(T, V, X)dV
\end{align*}
with $c_{3} = \frac{1}{4k + \sqrt{\varepsilon} / 4}$ and $C = C(\bm{F})$ a suitably large constant.
The trivial bound $\Phi_{\bm{F}}(T, V, X) \leq T$ yields the inequality $I_{1} \leq T e^{2k\sqrt{\log{\log{X}}}}$.
Applying Lemma \ref{LDESRDP} with $\Delta = 4k$, we have $I_{3} \ll T$ when $C$ is suitably large depending only on $\bm{F}$.
Similarly, applying Lemma \ref{LDESRDP} with $\Delta = \frac{2k + \sqrt{\varepsilon}/10}{\log{\log{T}}}$, we have $I_{4} \ll_{k, \varepsilon} T$.
Next, we consider $I_{5}$.
We have
\begin{align*}
\phi_{\bm F}(t, X)\leq \min_{1\leq j\leq r} |P_{F_{j}}(\tfrac{1}{2} + it, X)|
\leq\min_{1\leq j\leq r} \sum_{p \leq X}\frac{|a_{F_{j}}(p)|}{p^{1/2}}
+ \max_{1\leq j\leq r} \sum_{2 \leq \ell \leq \frac{\log{X}}{\log{2}}}\sum_{p \leq X^{1/\ell}}\frac{|b_{F_{j}}(p^{\ell})|}{p^{\ell/2}}.
\end{align*}
By the Cauchy-Schwarz inequality, the prime number theorem, and assumption \eqref{SNC}, the first sum is
\begin{align*}
\leq \min_{1\leq j\leq r} \l( \sum_{p \leq X}1 \r)^{1/2} \l( \sum_{p \leq X}\frac{|a_{F_{j}}(p)|^2}{p} \r)^{1/2}
= (1 + o(1))\sqrt{\frac{X}{\log{X}} \times n_{\bm F}^* \log{\log{X}}}.
\end{align*}
On the other hand, we also obtain, using the prime number theorem and partial summation, that the first sum is
\begin{align*}
\leq \min_{1\leq j\leq r}\sup_{p}\frac{|b_{F_{j}}(p)|}{p^{\vartheta_{F_j}}}\sum_{p \leq X}\frac{p^{\vartheta_{F_{j}}}}{p^{1/2}}
= (1 + o(1))\frac{2b_{\bm F}^*}{1 + 2\vartheta_{\bm F}^*}\frac{X^{1/2 + \vartheta_{\bm F}^*}}{\log{X}}
\end{align*}
By the assumption (S5) and the Cauchy-Schwarz inequality, the second sum is
\begin{align*}
\ll_{F_{j}} \sum_{2 \leq \ell \leq \frac{\log{X}}{\log{2}}}\l( \sum_{p \leq X^{1/\ell}}\frac{1}{(\log{p^{\ell}})^2} \r)^{1/2}
\l( \sum_{p \leq X^{1/\ell}}\frac{|b_{F_{j}}(p^{\ell}) \log{p^{\ell}}|^2}{p^{\ell}} \r)^{1/2}
\ll X^{1/4}.
\end{align*}
Therefore, we have
\begin{align*}
\phi_{\bm{F}}(t, X)
\leq (1 + o(1))
\min\l\{ \frac{2b_{\bm F}^*}{1 + 2\vartheta_{\bm F}^*}\frac{X^{1/2 + \vartheta_{\bm{F}}^{*}}}{\log{X}}, \sqrt{n_{\bm{F}}^{*} \frac{X \log{\log{X}}}{\log{X}}} \r\}.
\end{align*}
Choosing $c_{1} = (b_{\bm{F}}^{*}(4k + \sqrt{\varepsilon}/2)^{-2 / (1 + 2\vartheta_{\bm{F}}^{*})}$
and $c_{2} = ( (8k^2 + \varepsilon)n_{\bm{F}}^{*})^{-1}$,
we have $\Phi_{\bm{F}}(T, V, X) = 0$ for $V \geq c_{3}\log{T}$ and thus $I_{5} = 0$.
Finally, we consider $I_{2}$.
From Proposition \ref{Main_Prop_JVD3}, it holds that
\begin{align*}
\Phi_{\bm{F}}(T, V, X)
= T(1 + E_{1}) \times \Xi_{X}\l( \tfrac{V}{\sigma_{F_{1}}(X)^2}, \dots, \tfrac{V}{\sigma_{F_{r}}(X)^2} \r)
\prod_{j = 1}^{r}\int_{V / \sigma_{F_{j}}(X)}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{align*}
for $\sqrt{\log{\log{X}}} \leq V \leq C k \log{\log{X}}$.
Here, $E_{1}$ satisfies the estimate
\begin{align*}
E_{1} \ll_{\bm{F}, k}
\frac{V^{r}}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{r + 1}{2}}}
+ \frac{1}{\sqrt{\log{\log{X}}}}.
\end{align*}
We also have
\begin{align*}
\int_{x}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
= \frac{e^{-x^{2}/2}}{\sqrt{2\pi}x}\l(1 + O_{\varepsilon}\l(\frac{1}{x^2} \r)\r)
\end{align*}
for $x \geq \varepsilon$ with any $\varepsilon > 0$.
Therefore, we obtain
\begin{align*}
& \Phi_{\bm{F}}(T, V, X) \\
& = T\l(\prod_{j = 1}^{r}\sigma_{F_{j}}(X)\r)\frac{1 + E_{1} + O_{\bm{F}}(\log{\log{X}} / V^2)}{(2\pi)^{r/2}V^{r}}
\times \Xi_{X}\l( \tfrac{V}{\sigma_{F_{1}}(X)^2}, \dots, \tfrac{V}{\sigma_{F_{r}}(X)^2} \r)
\exp\l( -H_{\bm{F}}(X)^{-1}V^2 \r)
\end{align*}
for $\sqrt{\log{\log{X}}} \leq V \leq C k \log{\log{X}}$.
Hence, we have
\begin{align} \label{PGMDP1}
& 2k e^{2kV}\Phi_{\bm{F}}(T, V, X) \\
& = T 2k \l( \prod_{j = 1}^{r}\sigma_{F_{j}}(X) \r)\frac{1 + E_{1} + O(\log{\log{X}} / V^2)}{(2\pi)^{r/2} V^{r}}
\exp\l( k^{2}H_{\bm{F}}(X) \r) \\
& \qquad \qquad \qquad \qquad \times \Xi_{X}\l( \tfrac{V}{\sigma_{F_{1}}(X)^2}, \dots, \tfrac{V}{\sigma_{F_{r}}(X)^2} \r)
\exp\l( -\frac{1}{H_{\bm{F}}(X)}\l( V - k H_{\bm{F}}(X) \r)^2 \r)
\end{align}
for $\sqrt{\log{\log{X}}} \leq V \leq C k \log{\log{X}}$.
Put $H = D\sqrt{\log{\log{X}} \log_{3}{X}}$ with $D = D(\bm{F})$ a suitably large constant.
Here, we choose $X_{0}(\bm{F}, k)$ large such that $k H_{\bm{F}}(X) \geq 2H$ if necessary.
We write
\begin{align*}
I_{2}
= \l(\int_{\sqrt{\log{\log{X}}}}^{kH_{\bm{F}}(X) - H} + \int_{kH_{\bm{F}}(X) - H}^{kH_{\bm{F}}(X) + H}
+ \int_{kH_{\bm{F}}(X) + H}^{C k \log{\log{X}}} \r)2ke^{2kV}\Phi_{\bm{F}}(T, V, X)dV
=: I_{2, 1} + I_{2, 2} + I_{2, 3}.
\end{align*}
Applying \eqref{PGMDP1}, we find that
\begin{align*}
I_{2, 1}
& \ll_{\bm{F}, k} T \exp\l( k^2 H_{\bm{F}}(X) \r)
\int_{\sqrt{\log{\log{X}}}}^{kH_{\bm{F}}(X) - H} \exp\l( -\frac{1}{H_{\bm{F}}(X)}\l( V - kH_{\bm{F}}(X) \r)^2 \r)dV \\
& \leq T \exp\l( k^2 H_{\bm{F}}(X) \r)\int_{H}^{\infty}\exp\l( -\frac{V^2}{H_{\bm{F}}(X)} \r)dV
\ll T \frac{H_{\bm{F}}(X)}{H}\exp\l( k^2 H_{\bm{F}}(X) - \frac{H^2}{H_{\bm{F}}(X)} \r),
\end{align*}
and that
\begin{align*}
I_{2, 3}
& \ll_{\bm{F}, k} T \frac{\exp(k^2 H_{\bm{F}}(X))}{(\log{\log{X}})^{r/2}}
\int_{kH_{\bm{F}}(X) + H}^{C k \log{\log{X}}}\exp\l( -\frac{1}{H_{\bm{F}}(X)}\l( V - k H_{\bm{F}}(X) \r)^2 \r) dV \\
& \leq T \frac{\exp(k^2 H_{\bm{F}}(X))}{(\log{\log{X}})^{r/2}}
\int_{H}^{\infty}\exp\l( -\frac{V^2}{H_{\bm{F}}(X)} \r)dV
\ll T \frac{H_{\bm{F}}(X)}{H (\log{\log{X}})^{r/2}}\exp\l( k^2 H_{\bm{F}}(X) - \frac{H^2}{H_{\bm{F}}(X)} \r).
\end{align*}
Hence, we have
\begin{align*}
I_{2, 1} + I_{2, 3}
\ll_{\bm{F}, k} T\frac{\exp\l( k^2 H_{\bm{F}}(X)\r)}{(\log{\log{X}})^{2r}}
\end{align*}
Our main term comes from $I_{2, 2}$, which is equal to
\begin{multline*}
2k \frac{\exp\l( k^2 H_{\bm{F}}(X) \r)}{(\sqrt{2\pi} k H_{\bm{F}}(X))^{r}} \l( \prod_{j = 1}^{r}\sigma_{F_{j}}(X)\r)
\l( 1 + O_{\bm{F}}\l( \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} - \frac{r - 1}{2}}}
+ \frac{1}{\sqrt{\log{\log{X}}}} + \frac{H}{H_{\bm{F}}(X)} \r) \r)\\
\times \Xi_{X}\l( \frac{k H_{\bm{F}}(X) + O(H)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X) + O(H)}{\sigma_{F_{r}}(X)^2} \r)
\int_{k H_{\bm{F}}(X) - H}^{k H_{\bm{F}}(X) + H}\exp\l( -\frac{1}{H_{\bm{F}}(X)}(V - kH_{\bm{F}}(X))^2 \r).
\end{multline*}
The analyticity of $\Xi_{X}$ yields that
\begin{align*}
& \Xi_{X}\l( \frac{k H_{\bm{F}}(X) + O(H)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X) + O(H)}{\sigma_{F_{r}}(X)^2} \r) \\
& = \l( 1 + O_{k, \bm{F}}\l( \frac{H}{\log{\log{X}}} \r) \r)
\Xi_{X}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r).
\end{align*}
We also see that
\begin{align*}
\int_{k H_{\bm{F}}(X) - H}^{k H_{\bm{F}}(X) + H}\exp\l( -\frac{1}{H_{\bm{F}}(X)}(V - kH_{\bm{F}}(X))^2 \r)
& = \sqrt{\pi H_{\bm{F}}(X)} + O\l( \int_{H}^{\infty}\exp\l( -\frac{V^2}{H_{\bm{F}}(X)} \r) \r) \\
& = \sqrt{\pi H_{\bm{F}}(X)} + O\l( \frac{H_{\bm{F}}(X)}{H}\exp\l( -\frac{H^2}{H_{\bm{F}}(X)} \r) \r).
\end{align*}
Therefore, we have
\begin{align*}
I_{2, 2}
= T\frac{\exp\l( k^2 H_{\bm{F}}(X) \r) \prod_{j = 1}^{r}\sigma_{F_{j}}(X)}{\l( \sqrt{2\pi} k H_{\bm{F}}(X) \r)^{r-1}
\sqrt{\frac{1}{2}H_{\bm{F}}(X)}}
\Xi_{X}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r)\l( 1 + E_{2} \r),
\end{align*}
where
\begin{align*}
E_{2}
\ll_{\bm{F}, k}
\frac{1}{(\log{\log{X}})^{\a_{\bm{F}} - \frac{r - 1}{2}}}
+ \frac{\log_{3}{X}}{\sqrt{\log{\log{X}}}}.
\end{align*}
From the above estimates of $I_{1}$, $I_{2}$, and $I_{3}$, we obtain Theorem \ref{GMDP}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{jointDirichlet}]
Assume the Selberg Orthonormality Conjecture, \eqref{SSOC} and $\vartheta_{\bm{F}} < \frac{1}{r + 1}$.
Note that the error term $E$ in Theorem \ref{GMDP} is $= o(1)$ as $X \rightarrow + \infty$ when $\vartheta_{\bm{F}} < \frac{1}{r + 1}$.
Hence, it suffices to show that for $k > 0$
\begin{align*}
&\frac{\exp\l( k^2 H_{\bm{F}}(X) \r) \prod_{j = 1}^{r}\sigma_{F_{j}}(X)}{\l( \sqrt{2\pi} k H_{\bm{F}}(X) \r)^{r-1}
\sqrt{\frac{1}{2}H_{\bm{F}}(X)}}
\Xi_{X}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r)
\sim C(\bm{F}, k, \bm{\theta})\frac{(\log{X})^{k^2 / h_{\bm{F}}}}{(\log{\log{X}})^{(r - 1)/2}}
\end{align*}
as $X \rightarrow + \infty$.
Using \eqref{SSOC}, (S4), (S5), we find that
\begin{align} \label{pGGHK1}
\sigma_{F_{j}}(X)^2
= \frac{n_{F_{j}}}{2}\log{\log{X}} + c_{j} + o(1)
\end{align}
as $X \rightarrow + \infty$. Here, $c_{j}$ is a constant depending only on $\bm{F}$.
From this equation, we also see that $\sigma_{F_{j}}(X) = (1 + o(1))\sqrt{\frac{n_{F_{j}}}{2}\log{\log{X}}}$.
Moreover, it holds that as $X \rightarrow + \infty$
\begin{align} \label{pGGHK2}
H_{\bm{F}}(X)
= \frac{\prod_{j = 1}^{r}(n_{F_{j}}\log{\log{X}} + 2c_{j})}{\sum_{h = 1}^{r}\underset{j \not= h}{\prod_{j = 1}^{r}}(n_{F_{j}}\log{\log{X}} + 2c_{j})}
+ o(1)
= \frac{\log{\log{X}}}{h_{\bm{F}}} + g_{\bm{F}} + o(1)
\end{align}
for a constant $g_{\bm{F}}$.
Hence, we have
\begin{align*}
\frac{\exp\l( k^2 H_{\bm{F}}(X) \r) \prod_{j = 1}^{r}\sigma_{F_{j}}(X)}{\l( \sqrt{2\pi} k H_{\bm{F}}(X) \r)^{r-1}
\sqrt{\frac{1}{2}H_{\bm{F}}(X)}}
\sim C_{1}(\bm{F}, k)\frac{(\log{X})^{k^2 / h_{\bm{F}}}}{(\log{\log{X}})^{\frac{r - 1}{2}}}
\end{align*}
as $X \rightarrow + \infty$.
Also, by \eqref{SSOC}, (S4), and (S5), we can prove that, for $j_{1} \not= j_{2}$,
\begin{align*}
\tau_{j_{1}, j_{2}}(X) = c_{j_{1}, j_{2}} + o(1)
\end{align*}
as $X \rightarrow + \infty$, where $c_{j_{1}, j_{2}}$ is a constant depending only on $\bm{F}$, $\bm{\theta}$.
From this equation and the definition of $\Psi_{\bm{F}}$ \eqref{def_Psi_F}, we can write
\begin{align*}
&\Xi_{X}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r)\\
&= (1 + o(1))\exp\l( k^2\sum_{1 \leq j_{1} < j_{2} \leq r}\frac{H_{\bm{F}}(X)^2}{(\sigma_{F_{j_{1}}}(X) \sigma_{F_{j_{2}}}(X))^2}c_{j_{1}, j_{2}} \r)
\Psi_{\bm{F}}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r)
\end{align*}
as $X \rightarrow + \infty$.
Equations \eqref{pGGHK1}, \eqref{pGGHK2} yield that for some constant $C_{2}(\bm{F}, k, \bm{\theta})$
\begin{align*}
\exp\l( k^2\sum_{1 \leq j_{1} < j_{2} \leq r}\frac{H_{\bm{F}}(X)^2}{(\sigma_{F_{j_{1}}}(X) \sigma_{F_{j_{2}}}(X))^2}c_{j_{1}, j_{2}} \r)
\sim C_{2}(\bm{F}, k, \bm{\theta}).
\end{align*}
Moreover, it follows from the analyticity shown in Lemma \ref{Prop_Psi_FPP} and equations \eqref{pGGHK1}, \eqref{pGGHK2} that
\begin{align*}
\Psi_{\bm{F}}\l( \frac{k H_{\bm{F}}(X)}{\sigma_{F_{1}}(X)^2}, \dots, \frac{k H_{\bm{F}}(X)}{\sigma_{F_{r}}(X)^2} \r)
&= \Psi_{\bm{F}}\l( \frac{2k}{n_{F_{1}}h_{\bm{F}}} + o(1), \dots, \frac{2k}{n_{F_{r}}h_{\bm{F}}} + o(1) \r)\\
&\sim \Psi_{\bm{F}}\l( \frac{2k}{n_{F_{1}}h_{\bm{F}}}, \dots, \frac{2k}{n_{F_{r}}h_{\bm{F}}} \r).
\end{align*}
Thus, taking
$C(\bm{F}, k, \bm{\theta}) = C_{1}(\bm{F}, k) \cdot C_{2}(\bm{F}, k, \bm{\theta})
\cdot \Psi_{\bm{F}}\l( \frac{2k}{n_{F_{1}}h_{\bm{F}}}, \dots, \frac{2k}{n_{F_{r}}h_{\bm{F}}} \r)$, we obtain \eqref{GGHK}.
\end{proof}
\begin{comment}
\begin{proof}[Proof of \eqref{GGHK}]
Let $\chi$ a primitive Dirichlet character.
When $k = 0$, asymptotic formula \eqref{GGHK} is clear.
Using Theorem \ref{GMDP} with $\theta = 0$ if $k > 0$, and with $\theta = \pi$ if $k < 0$, we have
\begin{align*}
\frac{1}{T}\int_{T}^{2T}\exp\l( 2k \sum_{n \leq X}\frac{\Lambda(n) \chi(n)}{n^{1/2+it}} \r)dt
= \exp\l( k^{2}\sum_{p \leq X}\sum_{\ell = 1}^{\infty}\frac{|\chi(p^{\ell})|^2}{\ell^2 p^{\ell}} \r)
\Xi_{X}(2k)\l( 1 + O_{\chi, k}\l( \frac{\log_{3}{X}}{\sqrt{\log{\log{X}}}} \r) \r)
\end{align*}
for any $X_{0}(\chi, k) \leq X \leq c_{0}(\chi, k)(\log{T})^2$.
The well known formula
\begin{align*}
\sum_{p \leq X}\frac{1}{p}
= \log{\log{X}} + b + O\l( \exp\l(-c \sqrt{\log{X}}\r) \r)
\end{align*}
with $b = \gamma - \sum_{p}\sum_{\ell = 2}^{\infty}\frac{1}{\ell p^{\ell}}$
yields that
\begin{align*}
\sum_{p \leq X}\frac{|\chi_{j}(p)|^2}{p}
= \log{\log{X}} + b - \sum_{p | q}\frac{1}{p} + O\l( \exp\l(-c\sqrt{\log{X}}\r) \r),
\end{align*}
where $q$ is the conductor of $\chi$.
Also, we see that
\begin{align*}
\sum_{p > X}\sum_{\ell = 2}^{\infty}\frac{|\chi(p^{\ell})|^2}{\ell^2 p^{\ell}}
\ll \sum_{p > X}p^{-2}
\ll X^{-1}.
\end{align*}
Therefore, we have
\begin{align*}
& \exp\l( k^{2}\sum_{p \leq X}\sum_{\ell = 1}^{\infty}\frac{|\chi(p^{\ell})|^2}{\ell^2 p^{\ell}} \r) \\
& = (e^{\gamma}{\log{X}})^{k^2}
\exp\l( k^2 \sum_{p}\sum_{\ell = 2}^{\infty}\frac{|\chi(p^{\ell})|^{2}}{\ell^{2}p^{\ell}}
-k^2\sum_{p}\sum_{\ell = 2}^{\infty}\frac{1}{\ell p^{\ell}}
-k^2\sum_{p | q}\frac{1}{p} \r)
\l( 1 + O\l( \frac{1}{\log{X}} \r) \r)
\end{align*}
Also, we can write as
\begin{align*}
\Xi_{X}(2k)
= \prod_{p}\frac{\EXP{|1 - \chi(p)p^{-1/2}\mathcal{X}(p)|^{-2k}}}{
\exp\l(k^2 \sum_{\ell=1}^{\infty}\frac{|\chi(p^{\ell})|^2}{\ell^2 p^{\ell}}\r)}.
\end{align*}
Hence, we obtain
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\exp\l( 2k \sum_{n \leq X}\frac{\Lambda(n) \chi(n)}{n^{1/2+it}} \r)dt \\
& = (e^{\gamma} \log{X})^{k^2}
\l\{ \prod_{p}\frac{\EXP{|1 - \chi(p)p^{-1/2}\mathcal{X}(p)|^{-2k}}}{
\exp\l(k^2 \sum_{\ell=1}^{\infty}\frac{1}{\ell p^{\ell}}\r)} \r\}\l( 1 + O_{\chi, k}\l( \frac{\log_{3}{X}}{\sqrt{\log{\log{X}}}} \r) \r).
\end{align*}
It follows that $\exp\l( k^2 \sum_{\ell=1}^{\infty}\frac{1}{\ell p^{\ell}} \r) = (1 - p^{-1})^{-k^2}$.
Moreover, we find that
\begin{align*}
|1 - \chi(p)p^{-1/2}\mathcal{X}(p)|^{-2k}
& = \l\{(1 - \chi(p)\mathcal{X}(p) p^{-1/2}) (1 - \overline{\chi(p) \mathcal{X}(p)} p^{-1/2}) \r\}^{-k} \\
& = \l(\sum_{\ell = 0}^{\infty}\frac{d_{k}(p^{\ell}; \chi) \mathcal{X}(p)^{\ell}}{p^{\ell/2}}\r)
\times \l(\sum_{m = 0}^{\infty}\frac{d_{k}(p^{m}; \overline{\chi}) \mathcal{X}(p)^{-m}}{p^{m/2}}\r) \\
& = \sum_{\ell, m \geq 0}\frac{d_{k}(p^{\ell}; \chi) d_{k}(p^{m}; \overline{\chi})\mathcal{X}(p)^{\ell - m}}{p^{(\ell + m)/2}}.
\end{align*}
Hence, by equation \eqref{BEUIR}, we have
\begin{align*}
\EXP{|1 - \chi(p)p^{-1/2}\mathcal{X}(p)|^{-2k}}
= \sum_{\ell = 0}^{\infty}\frac{|d_{k}(p^{\ell}; \chi)|^2}{p^{\ell}}.
\end{align*}
Thus, we obtain
\begin{align*}
\frac{1}{T}\int_{T}^{2T}\exp\l( 2k \sum_{n \leq X}\frac{\Lambda(n) \chi(n)}{n^{1/2+it}} \r)dt
\sim \l\{ \prod_{p}\l( 1 - \frac{1}{p} \r)^{k^2}\sum_{\ell = 0}^{\infty}\frac{|d_{k}(p^{\ell}; \chi)|^2}{p^{\ell}} \r\}
(e^{\gamma} \log{X})^{k^2}
\end{align*}
as $X \rightarrow + \infty$.
\end{proof}
\end{comment}
\begin{comment}
\begin{proof}[Proof of Proposition \ref{Main_Prop_JVD2}]
Let $T$, $X$ be large numbers such that $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Let $\bm{x} = (x_{1}, \dots, x_{r}) \in (\mathbb{R}_{> 0})^{r}$
with $\norm[]{\bm{x}} \leq (\log{\log{X}})^{2r}$.
By Lemmas \ref{SPKLH} and \ref{FMnu3}, we obtain
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& \ll_{\bm{F}} \frac{1}{(\log{\log{X}})^{\a_{\bm{F}}+\frac{1}{2}}}\prod_{j = 1}^{r}
\exp\l({\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2} - x_{j}y_{j}
+ O_{\bm{F}}\l( x_{j}^{\frac{2-2\vartheta_{\bm{F}}}{1-2\vartheta_{\bm{F}}}} \r)\r) \\
& \qquad \qquad \qquad+ \prod_{j = 1}^{r}\frac{1}{\sigma_{F_{j}}(X)}\exp\l({\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2}
+ O_{\bm{F}}\l( x_{j}^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)
\int_{x_{j}y_{j}}^{\infty}\l( y_{j} - \frac{\tau}{x_{j}} \r)e^{-\tau}d\tau.
\end{align*}
Using \eqref{BFpte}, we have
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty)) \\
& \ll_{\bm{F}} \l(\prod_{k = 1}^{r}\frac{1}{x_{k}\sigma_{F_{k}}(X)} + \frac{1}{(\log{\log{X}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\times \prod_{j = 1}^{r}\exp\l({\frac{x_{j}^2}{2}\sigma_{F_{j}}(X)^2} - x_{j}y_{j}
+ O_{\bm{F}}\l( x_{j}^{\frac{2-2\vartheta_{\bm{F}}}{1-2\vartheta_{\bm{F}}}} \r)\r).
\end{align*}
Putting $x_{j} = \max\{ 1, V_{j} \} / \sigma_{F_{j}}(X)$, we obtain
\begin{align*}
& \mu_{T, \bm{F}}((y_{1}, \infty) \times \cdots \times (y_{r}, \infty))
\ll_{\bm{F}} \\
& \l(\prod_{k = 1}^{r}\frac{1}{1 + V_{k}} + \frac{1}{(\log{\log{X}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
+ O_{\bm{F}}\l( \l( \frac{\norm[]{\bm{V}}}{\sqrt{\log{\log{X}}}} \r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \r)
\end{align*}
for $0 \leq V_{j} \leq (\log{\log{X}})^{2r}$.
Combining this with \eqref{RTmuS}, we complete the proof of Proposition \ref{Main_Prop_JVD2}.
\end{proof}
\end{comment}
\begin{comment}
\subsection{Proof of Proposition \ref{EED2}}
In this section, we prove Proposition \ref{EED2}.
The proof and the proofs of some lemmas are written roughly because those many points are similar to
the proofs of Proposition \ref{Main_Prop_JVD}.
When $\bm{F} = (F, F) \in (\mathcal{S}^{\dagger} \setminus \{1\})^{2}$, $\bm{\theta} = (\theta_{1}, \theta_{2}) \in \mathbb{R}^{2}$ satisfy (A2)
(i.e. $|\theta_{1} - \theta_{2}| = \pi / 2$ in this case),
we can write
\begin{align} \label{EQKJF}
K_{\bm{F}, \bm{\theta}}(p, \bm{z})
= (z_{1}^2 + z_{2}^2)\sum_{k = 1}^{\infty}\frac{|b_{F}(p^{k})|^2}{p^{k}}.
\end{align}
Thanks to this equation, we can improve formula \eqref{ES_sI2} and Lemma \ref{FMnu} to the following lemmas.
We omit the proofs of those because the lemmas can be shown similarly the proofs of formula \eqref{ES_sI2} and Lemma \ref{FMnu}
just by using equation \eqref{EQKJF}.
\begin{lemma}
Let $\bm{F} = (F, F) \in (\mathcal{S}^{\dagger} \setminus \{1\})^{2}$
and $\bm{\theta} = (\theta_{1}, \theta_{2}) \in \mathbb{R}^{2}$
satisfying (S4), (S5), (A1), and (A2).
There exists a positive $b = b(F)$ such that for any $\bm{z} = (z_{1}, z_{2}) \in \mathbb{C}^{2}$ with $\norm[]{\bm{z}} \leq b$
we have
\begin{align*}
\prod_{p \leq X}M_{p, \frac{1}{2}}(\bm{z})
= \exp\l(\frac{z_{1}^2}{2}\tilde{\sigma}_{F}(X)^2 + O_{F}(|z_{1}|^{3})\r)\exp\l(\frac{z_{2}^2}{2}\tilde{\sigma}_{F}(X)^2 + O_{F}(|z_{2}|^{3})\r).
\end{align*}
\end{lemma}
\begin{lemma} \label{FMnu3}
Suppose that $\bm{F} = (F, F) \in (\mathcal{S}^{\dagger} \setminus \{1\})^{2}$ and $\bm{\theta} \in \mathbb{R}^{2}$ satisfy (S4), (S5), (A1), and (A2).
Let $c_{1}, c_{2}, d_{1}, d_{2}$ be real numbers with $c_{j} < d_{j}$.
Put $\mathscr{R} = (c_{1}, d_{1}) \times (c_{2}, d_{2}) \subset \mathbb{R}^2$.
Let $T$, $X$ be large numbers with $X^{(\log{\log{X}})^{4(r + 1)}} \leq T$.
Then, there exists a positive constant $b_{2} = b_{2}(\bm{F})$ such that for $\bm{x} = (x_{1}, x_{2}) \in \mathbb{R}^{2}$ with
$\norm[]{\bm{x}} \leq b_{2}$, we have
\begin{align*}
\nu_{T, \bm{F}, \bm{x}}(\mathscr{R})
& = e^{\frac{x_{1}^2 + x_{2}^2}{2}\sigma_{F}(X)^2}
\times \l\{\prod_{j = 1}^{2}\int_{x_{j} \sigma_{F}(X) - \frac{d_{j}}{\sigma_{F}(X)}}^{
x_{j}\sigma_{F}(X) - \frac{c_{j}}{\sigma_{F}(X)}}e^{-v^2/2}\frac{dv}{\sqrt{2\pi}} + E\r\},
\end{align*}
where the error term $E$ satisfies
\begin{align*}
& E
\ll_{F} \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \sum_{k = 1}^{2}\l(\frac{x_{k}^3 (d_{k} - c_{k})}{\sigma_{F}(X)} + \frac{1}{\sigma_{F}(X)^3} \r)
\prod_{\substack{h = 1 \\ h \not= k}}^{2}\frac{d_{h} - c_{h}}{\sigma_{F}(X)}.
\end{align*}
\end{lemma}
\begin{proof}[Proof of Proposition \ref{EED2}]
Using Lemma \ref{FMnu3}, we can prove Proposition \ref{EED2} in the same way as Proposition \ref{Main_Prop_JVD}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{DLDP}]
Let $0 < \varepsilon \leq 1$ and $z \in \mathbb{C}$.
We may assume $|z| + 2 \leq a_{7}\sigma_{F}(X)$, where $a_{7}$ is the same constant as in Theorem \ref{EED2}.
and Proposition \ref{EED2} in the case $\theta_{1} = 0$ and $\theta_{2} = \frac{\pi}{2}$, we obtain
\begin{align*}
& \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{P_{F}(\tfrac{1}{2}+it, X) \in R(z, \varepsilon)} \\
& = \iint_{R(z / \tilde{\sigma}_{F}(X), \varepsilon / \tilde{\sigma}_{F}(X))}e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi}
+ O_{F}\l( \frac{1}{(\log{\log{X}})^{\a_{\bm{F}} + 1/2}} + \frac{1}{(\log{\log{X}})^{3/2}} \r).
\end{align*}
Since we assume that $|a_{F}(p)| \ll_{F} p^{\vartheta_{F}}$ for some $\vartheta_{F} \in [0, 1/3)$,
the inequality $\a_{\bm{F}} \geq 1/2 + c_{0}$ holds for a constant $0 < c_{0} < 1$.
Therefore, the above $O$-term is $\leq \frac{C}{(\log{\log{X}})^{1 + c_{0}}}$ for some $C = C(F) > 0$.
Also, we see that the main term is $\gg_{F, z, \varepsilon} \frac{1}{\log{\log{X}}}$.
Thus, we obtain \eqref{DLDP1}.
Next, we prove \eqref{DLDP2}. Assume that $\vartheta_{F} \in [0, \tfrac{1}{4})$.
We see that
\begin{align*}
\sum_{p \leq X}\log{F_{p}(\tfrac{1}{2} + it)}
& = P_{F}(\tfrac{1}{2} + it, X) + \sum_{p \leq X}\sum_{k > \frac{\log{X}}{\log{p}}}\frac{b_{F}(p^{k})}{p^{k(1/2+it)}}.
\end{align*}
Put $K_{1} = 2(\tfrac{1}{2} - \vartheta_{F})^{-1}$.
Then it holds that
\begin{align*}
\sum_{p \leq X}\sum_{k > K_{1}\frac{\log{X}}{\log{p}}}\frac{b_{F}(p^{k})}{p^{k(1/2+it)}}
\ll_{F} \sum_{p \leq X}\sum_{k > K_{1}\frac{\log{X}}{\log{p}}}p^{-k(\frac{1}{2} - \vartheta_{F})}
\ll_{F} X^{-1}.
\end{align*}
Therefore, we have
\begin{align} \label{pDLDP1}
\sum_{p \leq X}\log{F_{p}(\tfrac{1}{2} + it)}
& = P_{F}(\tfrac{1}{2} + it, X)
+ \sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}\frac{b_{F}(p^{k})}{p^{k(1/2+it)}}
+ O_{F}(X^{-1}).
\end{align}
Here, we can write
\begin{multline*}
\bigg| \sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}\frac{b_{F}(p^{k})}{p^{k(1/2+it)}} \bigg|^2
= \sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}
\frac{|b_{F}(p^{k})|^2}{p^{k}}\\
+ \underset{(p_{1}, k_{1}) \not= (p_{2}, k_{2})}
{\sum_{p_{1} \leq X}\sum_{p_{2} \leq X}\sum_{\frac{\log{X}}{\log{p_{1}}} < k_{1} \leq K_{1}\frac{\log{X}}{\log{p_{1}}}}
\sum_{\frac{\log{X}}{\log{p_{2}}} < k_{2} \leq K_{1}\frac{\log{X}}{\log{p_{2}}}}}
\frac{b_{F}(p_{1}^{k_{1}}) \overline{b_{F}(p_{2}^{k_{2}})}}{p_{1}^{k_{1} / 2} p_{2}^{k_{2} / 2}}\l( \frac{p_{2}^{k_{2}}}{p_{1}^{k_{1}}} \r)^{it}.
\end{multline*}
Hence, it holds that
\begin{align*}
& \int_{T}^{2T}
\bigg| \sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}\frac{b_{F}(p^{k})}{p^{k(1/2+it)}} \bigg|^{2} dt \\
& = T\sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}
\frac{|b_{F}(p^{k})|^2}{p^{k}}
+ O\l( X^{K_{1}}\l( \sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}
\frac{|b_{F}(p^{k})|}{p^{k/2}} \r)^{2} \r)
\end{align*}
We see that this $O$-term is $\ll_{F} X^{K_{1} + 3} \ll T^{1/2}$ for large $X$.
Also, the main term is
\begin{align*}
= T\sum_{2 \leq k \leq K_{1}\frac{\log{X}}{\log{2}}}\sum_{X^{\frac{1}{k}} < p \leq X}\frac{|b_{F}(p^{k})|^2}{p^{k}}
& \ll_{F} T\sum_{2 \leq k \leq K_{1}\frac{\log{X}}{\log{2}}}\sum_{X^{\frac{1}{k}} < p \leq X}p^{-k(1 - 2\vartheta_{F})} \\
& \ll T\sum_{2 \leq k \leq K_{1}\frac{\log{X}}{\log{2}}} k\frac{X^{\frac{1}{k} - (1 - 2\vartheta_{F})}}{\log{X}} \\
& \ll T X^{-(\frac{1}{2} - 2\vartheta)}\log{X}
\ll T X^{-c}
\end{align*}
for some constant $c = c(F) > 0$.
Hence, we obtain
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}
\bigg| \sum_{p \leq X}\sum_{\frac{\log{X}}{\log{p}} < k \leq K_{1}\frac{\log{X}}{\log{p}}}\frac{b_{F}(p^{k})}{p^{k(1/2+it)}} \bigg|^{2} dt]
\ll_{F} X^{-c}.
\end{align*}
Combing this estimate with equation \eqref{pDLDP1}, we have
\begin{align*}
\frac{1}{T}\int_{T}^{2T}\bigg| \sum_{p \leq X}\log{F_{p}(\tfrac{1}{2} + it)} - P_{F}(\tfrac{1}{2} + it, X) \bigg|^{2}dt
\ll_{F} X^{-c}.
\end{align*}
It follows from this estimate and \eqref{DLDP1} that
\begin{align*}
& \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\sum_{p \leq X}\log{F_{p}(\tfrac{1}{2} + it)} \in R(z, \varepsilon)} \\
& = \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{P_{F}(\tfrac{1}{2} + it, X) \in R(z, \varepsilon + O(X^{-c/4}))} + O_{F}(X^{-c/2}) \\
& = (1 + o(1))\iint_{R(z / \tilde{\sigma}_{F}(X), \varepsilon / \tilde{\sigma}_{F}(X) + O(X^{-c/4}))}e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi}
+ O_{F}(X^{-c/2}).
\end{align*}
Additionally, when $X \geq X_{0}(z, \varepsilon)$ with $X_{0}(z, \varepsilon)$ sufficiently large, it holds that
\begin{align*}
\iint_{R(z / \tilde{\sigma}_{F}(X), \varepsilon / \tilde{\sigma}_{F}(X) + O(X^{-c/4}))}e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi}
= \iint_{R(z / \tilde{\sigma}_{F}(X), \varepsilon / \tilde{\sigma}_{F}(X))}e^{-\frac{u^2 + v^2}{2}}\frac{du dv}{2\pi}
+ O(X^{-c/4}).
\end{align*}
Thus, we complete the proof of \eqref{DLDP2}.
Finally, we prove \eqref{DLDP3}.
We see that... (In the following, I will write later)
\end{proof}
\end{comment}
\section{\textbf{Proof of Theorems \ref{Main_Thm_LD_JVD}--Theorem \ref{New_MVT}}} \label{uncond}
\begin{lemma}{\label{diff}}
Under the same situation as in Proposition \ref{KLI}.
Let $r \in \mathbb{Z}_{\geq 1}$ be given.
There exists a positive constant $A_{4} = A_{4}(F, r)$ such that
for $X = T^{1 / (\log{\log{T}})^{4(r + 1)}}$, $Y = T^{1 / k}$, $k \in \mathbb{Z}_{\geq 1}$
with $k \leq (\log{\log{T}})^{{4}(r + 1)}$,
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\bigg| \log{F(\tfrac{1}{2} +it)} - P_{F}(\tfrac{1}{2}+it, X)
- \sum_{|\frac{1}{2}+it-\rho_{F}| \leq \frac{1}{\log{Y}}}\log{((\tfrac{1}{2}+it-\rho_{F})\log{Y})} \bigg|^{2k}dt \\
& \leq A_{4}^{k} k^{2k} + A_{4}^{k}k!(\log_3T)^{k},
\end{align*}
and
\begin{align*}
\frac{1}{T}\int_{T}^{2T}\bigg| \log{F(\tfrac{1}{2} +it)} - P_{F}(\tfrac{1}{2}+it, X) \bigg|^{2k}dt
\leq A_{4}^{k} k^{4k} + A_{4}^{k}k!(\log_3T)^{k}.
\end{align*}
\end{lemma}
\begin{proof}
By Proposition \ref{KLI}, it suffices to show that
\begin{align} \label{diffp1}
\sum_{X < p \leq Y}\frac{|a_{F}(p)|^2}{p}
\ll_{F, r} \log_{3}T,
\end{align}
where $Y = T^{1/ k}$.
Using formula \eqref{SNC}, we find that
\begin{align*}
\sum_{X < p < Y}\frac{|a_{F}(p)|^2}{p}
= n_{F}\l( \log{\log{Y^2}} - \log{\log{X}} \r) + O_{F}(1)
\ll_{F, r} \log_3T.
\end{align*}
Thus, we obtain estimate \eqref{diffp1}.
\end{proof}
\noindent
\textit{Proof of Theorem \ref{Main_Thm_LD_JVD}}.
We consider \eqref{Main_Thm_LD_JVD1} and \eqref{Main_Thm_LD_JVD2}--\eqref{Main_Thm_LD_JVD3} separately.
\begin{proof}[Proof of \eqref{Main_Thm_LD_JVD1}]
Let $T$ be large.
Put $X = T^{1 / (\log{\log{T}})^{4(r + 1)}}$.
Let $A \geq 1$ be a fixed arbitrary constant.
Let the set $\mathcal{E}_j$ be
\begin{align*}
\mathcal{E}_j:=\set{t \in [T, 2T]}{\bigg|\log F_{j}(\tfrac{1}{2} + it) - P_{F_{j}}(\tfrac{1}{2} + it, X)\bigg|
\geq \mathcal{L}}.
\end{align*}
From Lemma \ref{diff}, we have
\begin{align}
\operatorname{meas}(\mathcal{E}_j)
\ll T\mathcal{L}_{j}^{-2k}A_{5}^{k}(k^{4k} + k^{k}(\log_3 T)^k)
\end{align}
for all $j$ with $A_{5} := A_{5}(\bm{F}) = \max_{1 \leq j \leq r}A_{4}(F_{j}, r)+1$,
where $A_{4}(F_{j}, r)$ has the same meaning as in Lemma \ref{diff}.
Here the parameter $\mathcal{L}$ satisfying $\mathcal{L} \geq (2A_5\log_{3}{T})^{2/3}$ will be chosen later.
Set $k=\lfloor \mathcal{L}^{1/2}/e A_{5}^{1/4}\rfloor$ so that
$\operatorname{meas}(\mathcal{E}_j)\ll T \exp(-c_{1}\mathcal{L}^{1/2})$ for some $c_1>0$.
Therefore except on the set $\mathcal{E}:=\bigcup_{j=1}^r\mathcal{E}_j$ with measure $O_r(T \exp(-c_{1}\mathcal{L}^{1/2}))$, we have
\begin{align}\label{logFjVj}
\Re e^{-i\theta_j}\log F_j(\tfrac{1}{2}+it)
=\Re e^{-i\theta_j}P_{F_{j}}(\tfrac{1}{2} + it, X) + \beta_j(t)\mathcal{L}
\end{align}
with $|\beta_j(t)|\leq 1$ for all $j=1, \dots, r$.
By \eqref{logFjVj} and Proposition \ref{Main_Prop_JVD}, the measure of $t\in[T, 2T]\backslash \mathcal{E}$ such that
\begin{align}\label{large dev}
\Re e^{-i\theta_j}\log F_j(\tfrac{1}{2}+it)
\geq V_{j} \sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}, \; \forall j=1, \dots, r
\end{align}
is at least (since $\beta_j(t)\leq 1$
\begin{multline}\label{lower meas}
T\l( 1 + O_{\bm{F}}\l(
\frac{\prod_{j = 1}^{r}(1 + |V_{k}| + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}})}{ (\log{\log{T}})^{\alpha_F+\frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^{2}+\frac{\mathcal{L}^2}{\log{\log{T}}}}{\log{\log{T}}}\r) \r)\\
\times \prod_{j = 1}^{r}\int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} + \mathcal{L})}^{\infty}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{multline}
for $\frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}, \norm[]{\bm{V}}
\leq c\sqrt{\log{\log{T}}}$ with $c$ sufficiently small.
Similarly, the measure of $t\in [T, 2T]\backslash \mathcal{E}$ such that \eqref{large dev} holds is at most
\begin{multline}\label{upper meas}
T\l( 1 + O_{\bm{F}}\l(
\frac{\prod_{k = 1}^{r}(1 + |V_{k}| + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}})}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^{2}+\frac{\mathcal{L}^2}{\log{\log{T}}}}{\log{\log{T}}}\r) \r)\\
\times \prod_{j = 1}^{r}\int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} - \mathcal{L})}^{\infty}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{multline}
for $\frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}, \norm[]{\bm{V}}
\leq c\sqrt{\log{\log{T}}}$.
By using equation \eqref{SNC}, we find that
\begin{align*}
\sigma_{F_{j}}(X)
= \sqrt{\frac{n_{F_{j}}}{2} \log{\log{T}}} + O_{r, F_{j}}\l( \frac{\log_{3}{T}}{\sqrt{\log{\log{T}}}} \r),
\end{align*}
and so we also have
\begin{align}
\label{sigaminv}
\sigma_{F_{j}}(X)^{-1}
= \frac{1}{\sqrt{\frac{n_{F_{j}}}{2} \log{\log{T}}}} \l(1+ O_{r, F_{j}}\l( \frac{\log_{3}{T}}{\log{\log{T}}} \r)\r).
\end{align}
Therefore, when $|V_{j}| \leq \sqrt{\frac{\log{\log{T}}}{\log_{3}{T}}}$,
$(|V_{j}| + 1)\mathcal{L} \leq B_{1}\sqrt{\log{\log{T}}}$ with $B_{1} > 0$ a constant to be chosen later,
we find that
\begin{align*}
& \int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} \pm \mathcal{L})}^{\infty}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \\
& = \int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
+ \int_{V_{j}+ O_{r, F_{j}}( |V_{j}|\frac{\log_{3}{T}}{\log{\log{T}}} + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}})}^{V_{j}}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \\
& = \int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
+ O_{r, F_{j}, B_{1}}\l( \l(\frac{|V_{j}| \log_{3}{T}}{\log{\log{T}}} + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}} \r)
e^{-V_{j}^2/2}\r) \\
& = \l( 1 + O_{r, F_{j}, B_{1}}
\l( \frac{|V_{j}|(|V_{j}| + 1)\log_{3}{T} + \mathcal{L}(|V_{j}| + 1)\sqrt{\log{\log{T}}}}{\log{\log{T}}} \r) \r)
\int_{V_{j}}^{\infty}e^{-u^2 / 2}\frac{du}{\sqrt{2\pi}}.
\end{align*}
Hence, choosing $\mathcal{L} = 2c_{1}^{-2}r^{2}(\norm[]{\bm{V}}^{4} + (\log_{3}{T})^{2})$,
we find that \eqref{lower meas} and \eqref{upper meas} become
\begin{align*}
T \l( 1 + O_{\bm{F}, B_{1}}\l(\frac{(\norm[]{\bm{V}}^{4} + (\log_{3}{T})^2)(\norm[]{\bm{V}} + 1)}{\sqrt{\log{\log{T}}}}
+ \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}\r)\r)
\prod_{j = 1}^{r}\int_{V_{j}}^\infty e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}},
\end{align*}
and
\begin{align*}
T \exp(-c_{1} \mathcal{L}^{1/2})
\ll T \exp(-r(\|\bm V\|^2+\log_3T))\ll T \frac{1}{\log{\log{T}}}\prod_{j = 1}^{r}\int_{V_{j}}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{align*}
when $\norm[]{\bm{V}} \leq c_{1}^{2}r^{-2}B_{1}(\log{\log{T}})^{1/10}$.
Choosing $B_{1} = A c_{1}^{-2} r^{2}$, we have
\begin{align*}
& \frac{1}{T}\operatorname{meas}\left( \bigcap_{j = 1}^{r}\set{t \in [T, 2T]}
{\frac{\Re e^{-i\theta_{j}}\log{F_{j}(1/2 + it)}}{\sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}} \geq V_j}\right) \\
& =\l( 1 + O_{\bm{F}, A}\l(\frac{(\norm[]{\bm{V}}^{4} + (\log_{3}{T})^2)(\norm[]{\bm{V}} + 1)}{\sqrt{\log{\log{T}}}}
+ \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{T}})^{\alpha_F+1/2}}\r)\r)\prod_{j = 1}^{r}
\int_{V_j}^{\infty} e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}}
\end{align*}
for $\norm[]{\bm{V}} \leq A(\log{\log{T}})^{1/10}$.
Thus, we complete the proof of \eqref{Main_Thm_LD_JVD1}.
\end{proof}
\begin{proof}[Proof of \eqref{Main_Thm_LD_JVD2} and \eqref{Main_Thm_LD_JVD3}]
Let $X = T^{1 / (\log{\log{T}})^{{4}(r + 1)}}$, $Y=T^{1/k}$ and
let $\mathcal{B}_j$ be the set of $t\in [T, 2T]$ such that
\begin{align}
\left| \log F_j(\tfrac{1}{2}+it)-P_{F_{j}}(\tfrac{1}{2} + it, X)
- \sum_{|1/2+it-\rho_{F_j}| \leq \frac{1}{\log Y}}\log ((\tfrac{1}{2}+it-\rho_{F_j})\log{Y})\right|
\geq \mathcal{L}.
\end{align}
By Lemma \ref{diff}, we know
\begin{align*}
\operatorname{meas}(\mathcal{B}_j)
\leq T\mathcal{L}^{-2k}A_{5}^k (k^{2k}+k^k(\log_{3}T)^k),
\end{align*}
where $A_5=\max_{1\leq j\leq r}{A_4(F_j,r)}+2$ and $A_{4}(F_{j}, r)$ has the same meaning as in Lemma \ref{diff}.
By taking $k=\lfloor \mathcal{L}/\sqrt{A_{5}e}\rfloor$, we have that
$\operatorname{meas}(\mathcal{B}_j)\leq T \exp(-c_2 \mathcal{L} )$ for some $c_2>0$ as long as
$\mathcal{L} \geq 2A_5\log_3T$.
Therefore, it follows that for $t \in [T, 2T] \setminus \bigcup_{j=1}^r\mathcal{B}_j$
\begin{align}
& \Re e^{-i\theta_{j}} \log F_j(\tfrac{1}{2}+it) \\
& =\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2} + it, X)
+ \sum_{|1/2+it-\rho_{F_j}|\leq \frac{1}{\log Y}}
\Re e^{-i\theta_{j}} \log((\tfrac{1}{2}+it-\rho_{F_j})\log{Y}) + \b_j(t)\mathcal{L}
\end{align}
holds for all $1\leq j\leq r$ with some $|\beta_j(t)|\leq 1$.
Let $\mathcal{C}_j$ be the set of $t\in [T, 2T]$ such that
\begin{align}
\sum_{|1/2-it-\rho_{F_j}|\leq \frac{1}{\log{Y}}}\Re e^{-\theta_{j}} \log((\tfrac{1}{2}+it-\rho_{F_j})\log{Y})
\geq \mathcal{L}.
\end{align}
When $\theta_{j} \in [-\frac{\pi}{2}, \frac{\pi}{2}]$ and $|1/2+it-\rho_{F_j}| \leq \frac{1}{\log{Y}}$,
we find that
\begin{align}
& \Re e^{-i\theta_{j}} \log((\tfrac{1}{2}+it-\rho_{F_j})\log{Y}) \\
\label{KPlt}
& = \cos{\theta_{j}}\log|(\tfrac{1}{2}+it-\rho_{F_j})\log{Y}| + \sin{\theta_{j}} \arg((\tfrac{1}{2}+it-\rho_{F_j})\log{Y})
\leq \pi.
\end{align}
Hence, we have
\begin{align*}
\sum_{|1/2+it-\rho_{F_j}|\leq \frac{1}{\log{Y}}}\Re e^{-i\theta_{j}} \log((\tfrac{1}{2}+it-\rho_{F_j})\log{Y})
\leq \pi\sum_{|1/2+it-\rho_{F_j}|\leq \frac{1}{\log{Y}}}1,
\end{align*}
and thus by Lemma \ref{ESRSIZ}
\begin{align*}
\operatorname{meas}(\mathcal{C}_j)
\leq C^k k^{2k} T \mathcal{L}^{-2k}
\end{align*}
for some constant $C = C(F_{j}) > 0$.
By choosing $k = \lfloor \mathcal{L} / \sqrt{C e} \rfloor$, we have $\operatorname{meas}(\mathcal{C}_j)\leq T \exp(-c_3\mathcal{L})$ for some $c_3>0$. Now we have that the measure of
$t\in[T, 2T] \setminus \bigcup_{j=1}^r(\mathcal{B}_j\cup \mathcal{C}_j)$ such that
\begin{align*}
\Re e^{-\theta_j}\log{F_{j}(\tfrac{1}{2}+it)}
\geq V_j \sqrt{\frac{n_{F_j}}{2}\log\log T}
\end{align*}
is bounded by the measure of the set $t\in[T, 2T]$ such that
\begin{align}\label{lower V-L-D}
\Re e^{-i\theta_{j}} P_{F_{j}}(\tfrac{1}{2}+it, X)
\geq V_j\sqrt{\frac{n_{F_j}}{2}\log\log T}- 2\mathcal{L}.
\end{align}
From Proposition \ref{Main_Prop_JVD}, we know \eqref{lower V-L-D} holds with measure
\begin{multline}\label{upper bound V-L-D}
T\l( 1 + O_{\bm{F}}\l(
\frac{\prod_{k = 1}^{r}(1 + |V_{k}| + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}})}{ (\log{\log{T}})^{\alpha_F+\frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^{2}+\frac{\mathcal{L}^2}{\log{\log{T}}}}{\log{\log{T}}}\r) \r)\\
\times \prod_{j = 1}^{r}\int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} - 2\mathcal{L})}^{\infty}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{multline}
for $\frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}, \norm[]{\bm{V}}
\leq c\sqrt{\log{\log{T}}}$ with $c$ sufficiently small. Since we have $\operatorname{meas}(\bigcup_{j=1}^r(\mathcal{B}_j\cup \mathcal{C}_j))\ll T\exp(-c_4\mathcal L)$ for some $c_4>0$, we choose $\mathcal{L} =c_4^{-1}r( \norm[]{\bm{V}}^{2} + 2A_5\log_{3}{T})$ so that that \eqref{upper bound V-L-D} becomes
\begin{multline}
T \l( 1 + O_{\bm{F}, A}\l(\frac{(\norm[]{\bm{V}}^{2} + \log_{3}{T})(\norm[]{\bm{V}} + 1)}{\sqrt{\log{\log{T}}}}
+ \frac{\prod_{k = 1}^{r}(1 + |V_{k}|)}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}\r)\r)
\prod_{j = 1}^{r}\int_{V_{j}}^\infty e^{-\frac{u^2}{2}}\frac{du}{\sqrt{2\pi}},
\end{multline}
for $\| \bm V\|\leq A (\log \log T)^{1/6}$.
This completes the proof of \eqref{Main_Thm_LD_JVD2}.
The proof of \eqref{Main_Thm_LD_JVD3} is similar by noting that
when $\theta_{j} \in [\frac{\pi}{2}, \frac{3\pi}{2}]$ and $|\tfrac{1}{2}+it-\rho_{F_j}| \leq \frac{1}{\log{Y}}$,
\begin{align*}
& \Re e^{-i\theta_{j}} \log((\tfrac{1}{2}+it-\rho_{F_j})\log{Y}) \\
& = \cos{\theta_{j}}\log|(\tfrac{1}{2}+it-\rho_{F_j})\log{Y}| + \sin{\theta_{j}} \arg((\tfrac{1}{2}+it-\rho_{F_j})\log{Y})
\geq -\pi.
\end{align*}
and thus the set of $t\in[T, 2T]$ such that
\begin{align}
\sum_{|1/2+it-\rho_{F_j}|\leq \frac{1}{\log{Y}}}\Re e^{-\theta_{j}} \log((\tfrac{1}{2}+it-\rho_{F_j})\log{Y})
\leq - \mathcal{L}
\end{align}
has measure bounded by $C^k k^{2k} T \mathcal{L}^{-2k}$ for some constant $C=C(F_j)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{GJu}]
Let $X = T^{1 / (\log{\log{T}})^{4(r + 1)}}$.
Let $a_{1} = a_{1}(\bm{F}) > 0$ be a sufficiently small constant to be chosen later,
Let $\bm{V} = (V_{1}, \dots, V_{r}) \in (\mathbb{R}_{\geq 0})^{r}$ such that
$\norm[]{\bm{V}} \leq a_{1}(1 + V_{m}^{{1/2}})(\log{\log{T}})^{{1/4}}$ with $V_{m} := \min_{1 \leq j \leq r}V_{j}$.
We consider the case when $\bm \theta\in[-\frac{\pi}{2}, \frac{\pi}{2}]^r$ first.
Similarly to the proof of \eqref{Main_Thm_LD_JVD2} (see \eqref{lower V-L-D}),
we find that the measure of the set of $t \in [T, 2T]$ except for a set of measure
$T\exp(-c_{4}\mathcal{L})$ ($\mathcal{L} \gg \log_{3}{T}$) such that
$
\Re e^{-i\theta_{j}} \log{F_{j}\l( \frac{1}{2} + it \r)}
\geq V_{j} \sqrt{\frac{n_{F_{j}}}{2} \log{\log{T}}}
$
is at most the measure of the set $t \in [T, 2T]$ such that
\begin{align*}
\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2} + it, X)
\geq V_{j} \sqrt{\frac{n_{F_{j}}}{2} \log{\log{T}}} - 2\mathcal{L}.
\end{align*}
From Proposition \ref{Main_Prop_JVD}, the measure of $t \in [T, 2T]$ satisfying this inequality for all $j = 1, \dots, r$
is equal to
\begin{multline}
\label{pGJu1}
T\l( 1 + O_{\bm{F}}\l(
\frac{\prod_{k = 1}^{r}(1 + V_{k} + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}})}{ (\log{\log{T}})^{\alpha_F+\frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^{2}+\frac{\mathcal{L}^2}{\log{\log{T}}}}{\log{\log{T}}}\r) \r)\\
\times \prod_{j = 1}^{r}\int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} - 2\mathcal{L})}^{\infty}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{multline}
for $\frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}, \norm[]{\bm{V}} \leq c \sqrt{\log{\log{T}}}$ with $c$ sufficiently small.
Now, we choose $\mathcal{L} = 2 r c_{4}^{-1} \norm[]{\bm{V}}^2 + \log_{3}{T}$
and $a_{1}$ small enough so that we have the inequalities $\norm[]{\bm{V}} \leq a_{9} \sqrt{\log{\log{T}}}$ and
$4\mathcal{L} \leq (1 + V_{j})\sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}$
for all $j = 1, \dots, r$, where $a_{9}$ is the same constant as in Proposition \ref{Main_Prop_JVD}.
Then, by equation \eqref{sigaminv} and the estimate $\int_{V}^{\infty}e^{-u^2/2}du \ll \frac{1}{1 + V}e^{-V^2/2}$
for $V \geq 0$,
we obtain
\begin{align*}
& \int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} - 2\mathcal{L})}^{\infty}
e^{-u^2/2}\frac{du}{\sqrt{2\pi}} \\
& \ll_{r, F_{j}} \frac{1}{1 + V_{j}}
\exp\l( -\frac{1}{2}\l( V_{j}
+ O_{r, F_{j}}\l( \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}} + V_{j}\frac{\log_{3}{T}}{\log{\log{T}}} \r) \r)^2 \r) \\
& \ll_{\bm F} \frac{1}{1 + V_{j}}
\exp\l( -\frac{V_{j}^2}{2} + O_{\bm{F}}\l( \frac{V_{j}\norm[]{\bm{V}}^2}{\sqrt{\log{\log{T}}}}
+ \frac{\norm[]{\bm{V}}^{4}}{\log{\log{T}}} \r) \r).
\end{align*}
Hence, when $0 \leq V_{1}, \dots, V_{r} \leq a \sqrt{\log{\log{T}}}$ with $a$ sufficiently small,
\eqref{pGJu1} is
\begin{align*}
\ll_{\bm{F}} T \l\{\l(\prod_{j = 1}^{r}\frac{1}{1 + V_{j}}\r) + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r\}
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
+ O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}}} \r) \r).
\end{align*}
Moreover, we have
\begin{align*}
T \exp(-c_{4} \mathcal{L})
\leq T \exp(-2 r \norm[]{\bm{V}}^2)
& \leq T \prod_{j = 1}^{r}\exp\l( -2(V_{1}^2 + \cdots + V_{r}^2) \r) \\
& \ll T \prod_{j = 1}^{r}\frac{1}{1 + V_{j}}\exp\l(-\frac{V_{j}^2}{2}\r).
\end{align*}
Similarly when $\bm \theta\in[\frac{\pi}{2}, \frac{3\pi}{2}]^r$, except for a set of measure $T \exp(-c_4\mathcal L)$ ($\mathcal L \gg \log_3T$), the measure of $t\in[T, 2T]$ such that
\begin{align}
\Re e^{-i\theta_j} \log F_j(\tfrac{1}{2}+it) \geq V_j \sqrt{\tfrac{n_{F_j}}{2}\log\log T}
\end{align}
is at least the measure of $t\in[T, 2T]$ such that
\begin{align}
\Re e^{-i\theta_j} P_{F_j}(\tfrac{1}{2}+it) \geq V_j \sqrt{\tfrac{n_{F_j}}{2}\log\log T} + 2\mathcal L
\end{align}
When $4 \mathcal L\leq V_j\sqrt{\frac{n_{F_j}}{2}\log\log T}$, we have the measure of $t$ satisfying the above inequality for all $j = 1, \dots, r$ is (by Proposition \ref{Main_Prop_JVD})
\begin{multline}\label{lowden}
T\l( 1 + O_{\bm{F}}\l(
\frac{\prod_{k = 1}^{r}(1 + V_{k} + \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}})}{ (\log{\log{T}})^{\alpha_F+\frac{1}{2}}}
+ \frac{1 + \norm[]{\bm{V}}^{2}+\frac{\mathcal{L}^2}{\log{\log{T}}}}{\log{\log{T}}}\r) \r)\\
\times \prod_{j=1}^r\int_{\sigma_{F_{j}}(X)^{-1}(V_{j}\sqrt{(n_{F_{j}}/2)\log{\log{T}}} + 2\mathcal{L})}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}},
\end{multline}
which can be bounded by
\begin{align}
& \gg_{\bm{F}} T \l(\prod_{j = 1}^{r}\frac{1}{1 + V_{j}}\r)
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2} + O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}\mathcal L}{\sqrt{\log{\log{T}}}} \r) \r)
\end{align}
when $\prod_{j = 1}^{r}(1 + V_{j}) \leq c(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}$
with $c = c(\bm{F}) > 0$ a suitably small constant.
Choose $\mathcal{L}=2r\|\bm V\|^2+ \log_3 T$ and $a_1$ small enough so that $\|\bm V\|\leq a_6 \sqrt{\log\log T}$ and $4\mathcal L \leq V_j\sqrt{\frac{n_{F_j}}{2}\log\log T}$ hold for all $j=1, \dots, r$, where $a_6$ is the same constant as in Proposition \ref{Main_Prop_JVD}. Then \eqref{lowden} is
\begin{align}
& \gg_{\bm{F}} T \l(\prod_{j = 1}^{r}\frac{1}{1 + V_{j}}\r)
\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2} + O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}}} \r) \r),
\end{align}
which completes the proof of Theorem \ref{GJu}.
\end{proof}
We prepare a lemma to prove Theorem \ref{New_MVT}.
\begin{lemma} \label{LLDSL}
Let $\theta \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]$, and $F \in \mathcal{S}^{\dagger}$ satisfying \eqref{SNC} and (A3).
There exist positive constants $a_{8} = a_{8}(F)$ such that for any large $V$,
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\Re e^{-i\theta} \log{F(\tfrac{1}{2}+it)} > V}
\leq \exp\l(-a_{8} \frac{V^2}{\log{\log{T}}}\r)
+ \exp\l( -a_{8} V \r).
\end{align*}
\end{lemma}
\begin{proof}
We can show that, for $t \in [T, 2T]$, the inequality $\Re e^{-i\theta} \log F(1/2+it) \leq C_{1}\log{T}$
with $C_{1} = C_{1}(F) > 0$ a suitable constant by using
Theorem \ref{Main_F_S} in the case $X = 3$, $H = 1$ and estimate \eqref{SIZDS}.
Hence, this lemma holds when $V \geq C_{1} \log{T}$ with $C_{1} = C_{1}(F) > 0$.
In the following, we consider the case $V \leq C_{1}\log{T}$.
Similarly to the proof of Lemma \ref{diff}, we obtain
\begin{align*}
& \frac{1}{T}\int_{T}^{2T}\bigg| \log{F(\tfrac{1}{2} +it)} - P_{F}(\tfrac{1}{2}+it, X)
- \sum_{|\frac{1}{2}+it-\rho_{F}| \leq \frac{1}{\log{X}}}\log{((\tfrac{1}{2}+it-\rho_{F})\log{X})} \bigg|^{2k}dt \\
& \leq A_{4}^{k}k^{2k} + A_{4}^{k} k! (\log\log T)^{k}
\end{align*}
for $X = T^{\delta_F/ k}$, where $\delta_F$ has the same meaning as in Lemma \ref{ESRSIZ}.
Moreover, by Lemma \ref{SLL} and Lemma \ref{ESRSIZ}, we obtain
\begin{align*}
\frac{1}{T}\int_{T}^{2T}|P_{F}(\tfrac{1}{2}+it, X)|^{2k} dt
\leq \l( Ck \log{\log{T}} \r)^{k},
\end{align*}
and
\begin{align*}
\frac{1}{T}\int_{T}^{2T}\l( \sum_{|\frac{1}{2}+it-\rho_{F}| \leq \frac{1}{\log{X}}} 1\r)^{2k}dt
\leq C^{k} k^{2k}.
\end{align*}
When $V \leq \log{\log{T}}$, we choose $k = \lfloor c V^2 / \log{\log{T}} \rfloor$,
and when $V \geq \log{\log{T}}$, we choose $k = \lfloor c V \rfloor$.
Here, $c$ is a suitably small constant depending only on $F$.
Then, by the above calculations and the inequality $\Re e^{-i\theta}\log{((1/2+it-\rho_{F})\log{X})} \leq \pi$,
we obtain
\begin{align*}
\frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\Re e^{-i\theta} \log{F(\tfrac{1}{2}+it)} > V}
\leq \exp\l(-a_{8} \frac{V^2}{\log{\log{T}}}\r)
+ \exp\l( -a_{8} V \r),
\end{align*}
which completes the proof of Lemma \ref{LLDSL}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{New_MVT}]
Let $0 \leq k \leq a_{6}$ with $a_{6} = a_{6}(\bm{F}) > 0$ suitably small to be chosen later.
Put $\phi_{\bm{F}}(t) = \min_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}(1/2+it)}$ and
\begin{align*}
\Phi_{\bm{F}}(T, V)
:= \operatorname{meas}\set{t \in [T, 2T]}{\phi_{\bm{F}}(t) > V}.
\end{align*}
Then we have
\begin{align}
\label{MINEF}
\int_{T}^{2T}\exp\l( 2k \phi_{\bm{F}}(t) \r)dt
= \int_{-\infty}^{\infty}2k e^{2k V}\Phi_{\bm F}(T, V)dV.
\end{align}
We consider the case when $\bm \theta \in [-\frac{\pi}{2},\frac{\pi}{2}]^r$ first.
From Theorem \ref{GJu}, it follows that, for any $0 \leq V \leq a_{3}\log{\log{T}}$ with $a_{3} = a_{3}(\bm{F})$ a suitable constant,
\begin{align}
\label{APGJu}
& \Phi_{\bm{F}}(T, V) \\
& \ll_{\bm{F}}
T\l(\frac{1}{1 + (V/\sqrt{\log{\log{T}}})^{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r)
\exp\l( -h_{\bm{F}}\frac{V^2}{\log{\log{T}}} + C_{1}\frac{V^3}{(\log{\log{T}})^2} \r)
\end{align}
for some constant $C_{1} = C_{1}(\bm{F}) > 0$
Moreover, by Lemma \ref{LLDSL}, it holds that
\begin{align}
\label{TBPhiF}
\Phi_{\bm{F}}(T, V)
\leq T\exp\l(-a_{8} \frac{V^2}{\log{\log{T}}}\r) + T\exp\l( -a_{8} V \r)
\end{align}
for any large $V$.
Now we choose $a_{6} = \min\{a_{3}a_{8} / 4, a_{8}/4\}$.
Put $D_{1} = 4a_{8}^{-1}$.
We divide the integral on the right hand side of \eqref{MINEF} to
\begin{align*}
\l(\int_{-\infty}^{0} + \int_{0}^{D_{1}k\log{\log{T}}} + \int_{D_{1}k\log{\log{T}}}^{\infty}\r)2k e^{2k V}\Phi_{\bm{F}}(T, V)dV
=: I_{1} + I_{2} + I_{3},
\end{align*}
say. We use the trivial bound $\Phi_{\bm{F}}(T, V) \leq T$ to obtain $I_{1} \leq T$.
Also, by inequality \eqref{TBPhiF}, it follows that
\begin{align*}
I_{3}
& \leq T\int_{D_{1} k \log{\log{T}}}^{\infty}2k \l\{\exp\l( \l(-a_{8}\frac{V}{\log{\log{T}}} + 2k\r)V \r)
+ e^{(-a_{8} + 2k)V}\r\}dV \\
& \leq T\int_{0}^{\infty}4k e^{-2k V}dV
\leq 2T.
\end{align*}
Moreover, using inequality \eqref{APGJu}, we find that
\begin{align*}
I_{2}
& \ll_{\bm{F}} T\int_{0}^{D_{1} k \log{\log{T}}}
\l( E_{1} + E_{2}\r)
\exp\l(2kV - h_{\bm{F}}\frac{V^2}{\log{\log{T}}} + C_{1}\frac{V^3}{(\log{\log{T}})^2} \r)dV \\
& \ll T(\log{T})^{k^2/h_{\bm{F}} + C_{1}D_{1}^{3}k^{3}}
\int_{0}^{D_{1} k \log{\log{T}}}(E_{1} + E_{2})
\exp\l( - \frac{h_{\bm{F}}}{\log{\log{T}}}\l( V - \frac{k}{h_{\bm{F}}}\log{\log{T}} \r)^2 \r)dV,
\end{align*}
where $E_{1} = \frac{k}{1 + (V/\sqrt{\log{\log{T}}})^r}$ and $E_{2} = \frac{k}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}$.
We see that
\begin{align*}
& \int_{0}^{D_{1} k \log{\log{T}}}E_{2}
\exp\l( - \frac{h_{\bm{F}}}{\log{\log{T}}}\l( V - \frac{k}{h_{\bm{F}}}\log{\log{T}} \r)^2 \r)dV \\
& \leq \frac{k}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}}
\int_{-\infty}^{\infty}\exp\l( - \frac{h_{\bm{F}}}{\log{\log{T}}}V^2\r)dV
\ll_{\bm{F}} \frac{k}{(\log{\log{T}})^{\a_{\bm{F}}}}.
\end{align*}
Also, we write
\begin{align*}
& \int_{0}^{D_{1} k \log{\log{T}}}E_{1}
\exp\l( - \frac{h_{\bm{F}}}{\log{\log{T}}}\l( V - \frac{k}{h_{\bm{F}}}\log{\log{T}} \r)^2 \r)dV \\
& = \l( \int_{0}^{\frac{k}{2h_{\bm{F}}}\log{\log{T}}} + \int_{\frac{k}{2h_{\bm{F}}}\log{\log{T}}}^{D_{1}k\log{\log{T}}} \r)
k\frac{\exp\l( - \frac{h_{\bm{F}}}{\log{\log{T}}}\l( V - \frac{k}{h_{\bm{F}}}\log{\log{T}} \r)^2 \r)}{
1 + (V/\sqrt{\log{\log{T}}})^{r}}dV
=: I_{2, 1} + I_{2, 2},
\end{align*}
say.
We find that
\begin{align*}
I_{2, 2}
\ll_{\bm{F}} \frac{k}{1 + (k\sqrt{\log{\log{T}}})^{r}}
\int_{-\infty}^{\infty}\exp\l( -\frac{h_{\bm{F}}}{\log{\log{T}}}V^2 \r)dV
\ll_{\bm{F}} \frac{k \sqrt{\log{\log{T}}}}{1 + (k\sqrt{\log{\log{T}}})^{r}},
\end{align*}
and that
\begin{align*}
I_{2, 1}
\leq \int_{\frac{k}{2h_{\bm{F}}}\log{\log{T}}}^{\frac{k}{h_{\bm{F}}}\log{\log{T}}}
k\exp\l( -\frac{h_{\bm{F}}}{\log{\log{T}}}V^2 \r)dV
\leq \sqrt{\frac{\log{\log{T}}}{h_{\bm{F}}}}
\int_{\frac{k}{2\sqrt{h_{\bm{F}}}}\sqrt{\log{\log{T}}}}^{\infty}ke^{-u^2}du.
\end{align*}
If $k \leq (\log{\log{T}})^{-1/2}$, the last is clearly $\ll_{\bm{F}} 1$.
If $k \geq (\log{\log{T}})^{-1/2}$, we apply the estimate $\int_{x}^{\infty}e^{-u^2}du \ll x^{-1}e^{-x^2}$ to obtain
\begin{align*}
I_{2, 1}
\leq \sqrt{\frac{\log{\log{T}}}{h_{\bm{F}}}}
\int_{\frac{k}{2\sqrt{h_{\bm{F}}}}\sqrt{\log{\log{T}}}}^{\infty}ke^{-u^2}du
\ll 1.
\end{align*}
Hence, we obtain
\begin{align}
\label{pNew_MVT2_1_I2}
I_{2}
\ll_{\bm{F}} T + kT(\log{T})^{k^2/h_{\bm{F}} + C_{1}D_{1}^{3}k^{3}}
\l( \frac{\sqrt{\log{\log{T}}}}{1 + (k\sqrt{\log{\log{T}}})^{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r).
\end{align}
Combing this estimate and the estimates for $I_{1}$, $I_{3}$,
we complete the proof of \eqref{New_MVT1}.
Next, we consider the case $\bm \theta \in [\frac{\pi}{2},\frac{3\pi}{2}]^r$.
By equation \eqref{MINEF}, estimate \eqref{NNGJu2}, and positivity of $\Phi_{\bm{F}}$, we have
\begin{align}
\label{pNew_MVT2_1}
& \int_{T}^{2T}\exp\l( 2k \phi_{\bm{F}}(t) \r)dt \\
& \geq \int_{0}^{1}2k e^{2kV} \Phi_{\bm{F}}(T, V)dV
+ \int_{\frac{k}{h_{\bm F}}\log\log{T}}^{\frac{k}{h_{\bm{F}}} \log{\log{T}} + \sqrt{\log{\log{T}}}}
2k e^{2k V}\Phi_{\bm F}(T, V)dV.
\end{align}
By estimate \eqref{NNGJu2}, the first integral on the right hand side is $\gg_{\bm{F}} T$,
and the second integral on the right hand side is
\begin{align}
& \gg_{\bm{F}}
\frac{kT}{1 + (k\sqrt{\log\log T})^{r}}
\int_{\frac{k}{h_{\bm F}}\log\log T}^{\frac{k}{h_{\bm{F}}} \log{\log{T}} + \sqrt{\log{\log{T}}}}
\exp\l(2kV - h_{\bm{F}}\frac{V^2}{\log{\log{T}}} - C_{1}\frac{V^3}{(\log{\log{T}})^2} \r)dV \\
& \geq \frac{kT (\log{T})^{\frac{k^2}{h_{\bm{F}}} - C_{2}k^{3}}}{1 + (k\sqrt{\log\log T})^{r}}
\int_{\frac{k}{h_{\bm F}}\log\log T}^{\frac{k}{h_{\bm{F}}} \log{\log{T}} + \sqrt{\log{\log{T}}}}
\exp\l( -\frac{h_{\bm{F}}}{\log{\log{T}}}\l( V - \frac{k}{h_{\bm{F}}}\log{\log{T}} \r)^2 \r) \\
& \gg _{\bm{F}} kT (\log{T})^{\frac{k^2}{h_{\bm{F}}} - C_{2}k^{3}}\frac{\sqrt{\log{\log{T}}}}{1 + (k\sqrt{\log\log T})^{r}},
\end{align}
where $C_2 \geq 0$ is some constant depending on $\bm F$.
Hence, we also obtain Theorem \ref{New_MVT} in the case $\bm \theta \in [\frac{\pi}{2},\frac{3\pi}{2}]^r$.
\end{proof}
\section{\textbf{Proofs of Theorem \ref{RH_LD_JVD} and Theorem \ref{New_MVT_RH}}} \label{cond}
\begin{proof}[Proof of Theorem \ref{RH_LD_JVD}]
Let $\bm{F} \in (\mathcal{S}^{\dagger})^{r}$ and $\bm{\theta} \in [-\tfrac{\pi}{2}, \tfrac{3\pi}{2}]^{r}$ satisfying $\mathscr{A}$.
Let $T$ be a sufficiently large constant depending on $\bm F$.
Set $Y = T^{K_{1} / \mathcal{L}}$ where
$K_{1} = K_{1}(\bm{F}) > 0$ is a suitably large constant and $\mathcal{L} \geq (\log_{3}{T})^{2}$ is a large parameter to be chosen later.
Let $f$ be a fixed function satisfying the condition of this paper (see Notation) and $D(f) \geq 2$.
Assuming the Riemann Hypothesis for $F_{1}, \dots, F_{r}$, we apply Theorem \ref{Main_F_S} with $X = Y$, $H = 1$ to obtain
\begin{align*}
\log{F_{j}(\tfrac{1}{2}+it)}
& = \sum_{2 \leq n \leq Y^2}\frac{\Lambda_{F_{j}}(n) v_{f, 1}(e^{\log{n} / \log{Y}})}{n^{1/2+it} \log{n}} \\
& \qquad+ \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}\log((\tfrac{1}{2} + it -\rho_{F_{j}})\log{Y})
+ R_{F_{j}}(\tfrac{1}{2}+it, Y, 1),
\end{align*}
where
\begin{align*}
\l| R_{F_{j}}(\tfrac{1}{2}+it, Y, 1)\r|
\leq C_{0}\l(\frac{1}{\log{Y}}\bigg| \sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg|
+ \frac{d_{F_{j}}\log{T}}{\log{Y}}\r)
\end{align*}
for any $t \in [T, 2T]$.
Here $C_{0}$ is a positive constant depending only on $f$.
Moreover, when $\theta_{j} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]$, it holds that
\begin{align*}
\Re e^{-i\theta_{j}} \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}\log((\tfrac{1}{2} + it -\rho_{F_{j}})\log{Y})
\leq \pi \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}1,
\end{align*}
and when $\theta_{j} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]$, it holds that
\begin{align*}
\Re e^{-i\theta_{j}} \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}\log((\tfrac{1}{2} + it -\rho_{F_{j}})\log{Y})
\geq -\pi \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}1.
\end{align*}
Hence, there exists some positive constant $C_{1} > 0$ such that we have (by \eqref{NZVDP}),
\begin{align*}
& \Re e^{-i\theta_{j}} \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}\log((\tfrac{1}{2} + it -\rho_{F_{j}})\log{Y}) \\
& \leq C_{1} \l( \frac{1}{\log{Y}}\bigg|
\sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg|
+ \frac{d_{F_{j}}\log{T}}{\log{Y}} \r)
\end{align*}
when $\theta_{j} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]$,
and
\begin{align*}
& \Re e^{-i\theta_{j}} \sum_{|1/2+it-\rho_{F_{j}}| \leq \frac{1}{\log{Y}}}\log((\tfrac{1}{2} + it -\rho_{F_{j}})\log{Y}) \\
& \geq -C_{1} \l( \frac{1}{\log{Y}}\bigg|
\sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg|
+ \frac{d_{F_{j}}\log{T}}{\log{Y}} \r)
\end{align*}
when $\theta_{j} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]$.
Taking $K_{1} = 2(C_{0} + C_{1})\max_{1 \leq j \leq r}d_{F_{j}}$, we find that there exists some positive constant $C_{2}$ depending on $f$ such that for any $t \in [T, 2T]$ and all $j = 1, \dots, r$,
\begin{align} \label{upRH}
\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}
\leq & \Re e^{-i\theta_{j}}\sum_{2 \leq n \leq Y^2}\frac{\Lambda_{F_{j}}(n) v_{f, 1}(e^{\log{n} / \log{Y}})}{n^{1/2+it} \log{n}} \\
& + \frac{C_{2}}{\log{Y}}\bigg| \sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg|
+ \frac{\mathcal{L}}{2}
\end{align}
when $\theta_{j} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]$,
and
\begin{align}\label{loRH}
\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}
\geq & \Re e^{-i\theta_{j}}\sum_{2 \leq n \leq Y^2}\frac{\Lambda_{F_{j}}(n) v_{f, 1}(e^{\log{n} / \log{Y}})}{n^{1/2+it} \log{n}} \\
& - \frac{C_{2}}{\log{Y}}\bigg| \sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg|
- \frac{\mathcal{L}}{2}
\end{align}
when $\theta_{j} \in \l[ \frac{\pi}{2}, \frac{3\pi}{2} \r]$.
Put $X = Y^{1 / (\log{\log{T}})^{4(r + 1)}}$.
By Lemma \ref{SLL} and assumption (A1), we obtain
\begin{align*}
\int_{T}^{2T}\bigg| \sum_{X < p \leq Y^2}\frac{a_{F_{j}}(p) v_{f, 1}(e^{\log{p} / \log{Y}})}{p^{1/2+it}} \bigg|^{2k}dt
& \ll T k! \l( \sum_{X < p \leq Y^2}\frac{|a_{F_{j}}(p)|^2}{p} \r)^{k} \\
& \leq T C_{3}^{k}k^{k}\l(\log_{3}{T}\r)^{k}
\end{align*}
for some constant $C_{3} = C_{3}(F_{j}, r) > 0$.
Similarly to the proofs of estimates \eqref{PKLI3},
we can show that for any integer $k$ with $1 \leq k \leq \mathcal{L} / 4K_{1}$
\begin{align*}
\int_{T}^{2T}\bigg| \sum_{\substack{X < p^{\ell} \leq Y^2 \\ \ell \geq 2}}
\frac{\Lambda_{F_{j}}(p^{\ell}) v_{f, 1}(e^{\log{p^{\ell}} / \log{Y}})}{p^{\ell(1/2+it)} \log{p^{\ell}}} \bigg|^{2k}dt
\leq T C_{4}^k k^{k}
\end{align*}
for some constant $C_{4} = C_{4}(F_{j}) > 0$.
Moreover, by Lemma \ref{ESMDP}, we have
\begin{align*}
\int_{T}^{2T}\l( \frac{C_{1}}{\log{Y}}
\bigg| \sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg| \r)^{2k}dt
\leq T C_{5}^{k} k^{k}
\end{align*}
for any integer $k$ with $1 \leq k \leq \mathcal{L} / 4K_{1}$ and for some constant $C_{5} = C_{5}(F_{j}) > 0$.
Here the assumptions in Lemma \ref{ESMDP} is satisfied as we can take $\kappa_F$ arbitrarily large.
Therefore, the set of $t \in [T, 2T]$ such that for all $j = 1, \dots, r$,
\begin{align*}
\frac{\mathcal{L}}{2}
\leq \bigg| \sum_{X < n \leq Y^2}\frac{\Lambda_{F}(n) v_{f, 1}(e^{\log{n} / \log{Y}})}{n^{1/2+it}\log{n}} \bigg|
& + \frac{C_{2}}{\log{Y}}\bigg| \sum_{n \leq Y^3}\frac{\Lambda_{F_{j}}(n) w_{Y}(n)}{n^{\frac{1}{2}+\frac{4}{\log{Y}}+it}} \bigg|
\end{align*}
has a measure bounded by
$
T \mathcal{L}^{-2k} C_{6}^{k} k^{k} \l(\log_{3}{T}\r)^{k}
$
with $C_{6} = C_{6}(\bm{F}) > 0$ a suitably large constant.
Choosing $k = \lfloor c_{1} \mathcal{L} \rfloor$ with $c_{1}$ suitably small depending only on $\bm{F}$, we find that
there exists a set $\mathcal{X} \subset [T, 2T]$ with
\begin{align}\label{measX}
\operatorname{meas}(\mathcal{X}) \leq T\exp\l( -c_{1}\mathcal{L}\log\l( \frac{\mathcal{L}}{\log_{3}{T}} \r) \r)
\end{align}
such that for any $t \in [T, 2T] \setminus \mathcal{X}$ and any $j = 1, \dots, r$,
\begin{align}
\label{RH_LD_JVDp1}
\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}
\leq \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X)
+ \mathcal{L} \quad \text{ when } \theta_{j} \in [-\tfrac{\pi}{2}, \tfrac{\pi}{2}],
\end{align}
and
\begin{align}
\label{RH_LD_JVDp2}
\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)}
\geq \Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X)
- \mathcal{L} \quad \text{ when } \theta_{j} \in [\tfrac{\pi}{2}, \tfrac{3\pi}{2}].
\end{align}
Now we consider the case when
$\|\bm{V}\| \leq a_{5}V_{m}^{{1/2}}(\log{\log{T}})^{{1/4}} (\log_{3}{T})^{{1/2}}$, where $a_{5}$ is a sufficiently small positive constant.
Set $\mathcal{L} = 4 r c_{1}^{-1}\l(\frac{\norm[]{\bm{V}}^2}{\log{\norm[]{\bm{V}}}} + (\log_{3}{T})^{{4}}\r)$.
Then we can verify from \eqref{measX} that $\operatorname{meas}(\mathcal{X}) \ll_{\bm{F}} T\exp\l( -2 r \norm[]{\bm{V}}^{2} \r)$.
Moreover, when $\theta_i\in[-\frac{\pi}{2}, \frac{\pi}{2}]$, the measure of $t\in[T, 2T]\backslash \mathcal{X}$ such that
\begin{align*}
\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)} \geq V_{j} \sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}
\end{align*}
is bounded above by the measure of $t\in [T, 2T]$ such that
\begin{align*}
\frac{\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X)}{\sigma_{F_{j}}(X)}
\geq V_{j} - C_{\bm{F}}\l( \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}+ \frac{V_j\log_3T}{\log\log T} \r).
\end{align*}
where $C_{\bm F}$ is some positive constant and we used \eqref{sigaminv} for $\sigma_{F_j}(X)^{-1}$.
Similarly when $\theta_i\in[\frac{\pi}{2}, \frac{3\pi}{2}]$, the measure of $t\in[T, 2T]\backslash \mathcal{X}$ such that
\begin{align*}
\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)} \geq V_{j} \sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}
\end{align*}
is bounded below by the measure of $t\in [T, 2T]$ such that
\begin{align*}
\frac{\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2}+it, X)}{\sigma_{F_{j}}(X)}
\geq V_{j} + C_{\bm{F}}\l( \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}+ \frac{V_j\log_3T}{\log\log T} \r).
\end{align*}
Here, we choose $a_{3}$ so that
$C_{\bm{F}}\l( \frac{\mathcal{L}}{\sqrt{\log{\log{T}}}}+ \frac{V_j\log_3T}{\log\log T} \r)\leq \frac{V_{m}}{2}$.
From these observations, the estimate $\int_{V}^{\infty}e^{-u^2/2}du \asymp \frac{1}{1+V}e^{-V^2/2}$ for $V \geq 0$,
estimate \eqref{EST_Xi}, and Proposition \ref{Main_Prop_JVD3}, we find that if $\theta_{j} \in [-\frac{\pi}{2}, \frac{\pi}{2}]$,
\begin{align*}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \ll_{\bm{F}} \frac{1}{T}\operatorname{meas}(\mathcal{X}) \\
& + \l(\frac{1}{V_{1} \cdots V_{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\prod_{j = 1}^{r}\exp\Bigg( -\frac{V_{j}^2}{2}
- \frac{V_{j}^2}{2 \sigma_{F_{j}}(X)^2}\sigma_{F_{j}}\l( \frac{V_{j}^2}{\sigma_{F_{j}}(X)^2} \r)^2 + \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad + O_{\bm{F}}\l( \frac{\norm[]{\bm{V}} \mathcal{L}}{\sqrt{\log{\log{T}}}} + \frac{\mathcal{L}^2}{\log{\log{T}}}
+ \l( \frac{\norm[]{\bm{V}}}{\sqrt{\log{\log{T}}}} \r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \Bigg) \\
& \ll_{\bm{F}} \l(\frac{1}{V_{1} \cdots V_{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\prod_{j = 1}^{r}\exp\l( -\frac{V_{j}^2}{2}
+ O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}}} \r) \r)
\end{align*}
for $\|\bm{V}\| \leq a_{5} V_{m}^{{1/2}} (\log{\log{T}})^{{1/4}} (\log_{3}{T})^{{1/2}}$.
Hence, we obtain estimate \eqref{RH_LD_JVD1u}.
Similarly, we can also find that if $\theta_{j} \in [\frac{\pi}{2}, \frac{3\pi}{2}]$,
\begin{align*}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \gg_{\bm{F}}
\l(\frac{1}{V_{1} \cdots V_{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}}+\frac{1}{2}}}\r)
\prod_{j = 1}^{r}\exp\Bigg( -\frac{V_{j}^2}{2}
- \frac{V_{j}^2}{2 \sigma_{F_{j}}(X)^2}\sigma_{F_{j}}\l( \frac{V_{j}^2}{\sigma_{F_{j}}(X)^2} \r)^2 + \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad - O_{\bm{F}}\l( \frac{\norm[]{\bm{V}} \mathcal{L}}{\sqrt{\log{\log{T}}}} + \frac{\mathcal{L}^2}{\log{\log{T}}}
+ \l( \frac{\norm[]{\bm{V}}}{\sqrt{\log{\log{T}}}} \r)^{\frac{2 - 2\vartheta_{\bm{F}}}{1 - 2\vartheta_{\bm{F}}}} \r) \Bigg) \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad - \frac{1}{T}\operatorname{meas}(\mathcal{X}) \\
& \gg_{\bm{F}} \frac{1}{V_{1} \cdots V_{r}}\exp\l( -\frac{V_{1}^2 + \cdots + V_{r}^2}{2}
- O_{\bm{F}}\l( \frac{\norm[]{\bm{V}}^{3}}{\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}}} \r) \r)
\end{align*}
for $\|\bm{V}\| \leq a_{5} V_{m}^{{1/2}} (\log{\log{T}})^{{1/4}} (\log_{3}{T})^{{1/2}}$ satisfying
$\prod_{j = 1}^{r}V_{j} \leq a_{6}(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}$ with $a_{6} = a_{6}(\bm{F}) > 0$ a suitably small constant.
Hence, we also obtain \eqref{RH_LD_JVD1l}.
Now we consider \eqref{RH_LD_JVD2}, where $\theta_{j} \in \l[ -\frac{\pi}{2}, \frac{\pi}{2} \r]$.
Putting $\mathcal{L} = \frac{4K_{1}\log{T}}{\log{\log{T}}}$,
we see that $Y = (\log{T})^{1/4}$, and hence there exists a positive constant $A = A(\bm{F})$ such that
the right hand side of \eqref{upRH} is $\leq A \frac{\log{T}}{\sqrt{\log{\log{T}}}}\sqrt{\frac{n_{F_{j}}}{2}\log{\log{T}}}$
uniformly for any $t \in [T, 2T]$ and all $j = 1, \dots, r$.
Thus, we may assume $\norm[]{\bm{V}} \leq A\frac{\log{T}}{\sqrt{\log{\log{T}}}}$.
We first consider the case when $\sqrt{\log{\log{T}}} \leq \norm[]{\bm{V}} \leq A\frac{\log{T}}{\sqrt{\log{\log{T}}}}$.
Set $\mathcal{L} = b_{1}\frac{\norm[]{\bm{V}}}{2}\sqrt{\log{\log{T}}}$, where $b_{1}$ is some small positive constant such that the inequality $Y \geq 3$ holds.
Then we see \eqref{measX} becomes
$$
\operatorname{meas}(\mathcal{X})
\ll_{\bm{F}} T \exp\l( -c_{2} \norm[]{\bm{V}}\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}} \r)
$$
for some constant $c_{2} = c_{2}(\bm{F}) > 0$.
Using Lemma \ref{SLL}, we have, uniformly for any $j = 1, \dots, r$,
\begin{align}\label{PF_2k}
\int_{T}^{2T}|P_{F_{j}}(\tfrac{1}{2}+it, X)|^{2k}dt
\ll_{\bm{F}} T(C_{6} k \log{\log{T}})^{k}
\end{align}
for any integer $k$ with $1 \leq k \leq \mathcal{L}\log{\log{T}}$ and some $C_{6} = C_{6}(\bm{F})$.
Combing \eqref{RH_LD_JVDp1} and \eqref{PF_2k}, we obtain
\begin{align*}
& \frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta})) \\
& \ll \min_{1\leq j\leq r} \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\Re e^{-i\theta_{j}} \log{F_{j}(\tfrac{1}{2}+it)} > V_{j}} \\
& \ll_{\bm{F}}
\norm[]{\bm{V}}^{-2k} C_{7}^{k} k^{k}
+ \exp\l( -c_{2} \norm[]{\bm{V}}\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}} \r).
\end{align*}
When $\norm[]{\bm{V}} \leq \log{\log{T}}$, we choose $k = \lfloor c_{3} \norm[]{\bm{V}}^2 \rfloor$,
and when $\norm[]{\bm{V}} > \log{\log{T}}$, we choose $k = \lfloor c_{3} \norm[]{\bm{V}}\sqrt{\log{\log{T}}} \rfloor$,
where $c_{3} $ is a suitably small positive constant depending only on $\bm{F}$.
Then, it follows that
\begin{align*}
\frac{1}{T}\operatorname{meas}(\S(T, \bm{V}; \bm{F}, \bm{\theta}))
\ll_{\bm{F}}
\exp\l( -c_{4}\|\bm{V}\|^2 \r) + \exp\l( -c_{4} \|\bm{V}\|\sqrt{\log{\log{T}}} \log{\norm[]{\bm{V}}} \r),
\end{align*}
which completes the proof of \eqref{RH_LD_JVD2}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{New_MVT_RH}]
Let $T$ be large, and put $\varepsilon(T) = (\log_{3}{T})^{-1}$.
Let $k \geq 0$.
We recall equation \eqref{MINEF}, which is
\begin{align*}
\int_{T}^{2T}\exp\l( 2k \phi_{\bm{F}}(t) \r)dt
= \int_{-\infty}^{\infty}2k e^{2k V}\Phi_{\bm{F}}(T, V)dV.
\end{align*}
We divide the integral on the right hand side to
\begin{align*}
\l(\int_{-\infty}^{0}+ \int_{0}^{D_{2}k\log{\log{T}}}
+ \int_{D_{2}k\log{\log{T}}}^{\infty}\r)2k e^{2k V}\Phi_{\bm{F}}(T, V)dV
=: I_{4} + I_{5} + I_{6},
\end{align*}
say. Here, $D_{2} = D_{2}(\bm{F})$ is a suitably large positive constant.
Now we consider the case when $\theta_{j} \in [-\frac{\pi}{2}, \frac{\pi}{2}]$.
We use the trivial bound $\Phi_{\bm{F}}(T, V) \leq T$ to obtain $I_{4} \leq T$.
Applying estimate \eqref{RH_LD_JVD2}, we find that the estimate
\begin{align*}
\Phi_{\bm{F}}(T, V)
\ll_{\bm{F}} T\exp\l( - 4k V \r)
\end{align*}
holds for $V \geq D_{2} k \log{\log{T}}$ when $D_{2}$ is suitably large.
Therefore, we have
\begin{align*}
I_{6}
\ll_{\bm{F}} T \int_{D\log{\log{T}}}^{\infty}2k e^{-2kV}dV
\ll T.
\end{align*}
By estimate \eqref{RH_LD_JVD1u}, we find that
\begin{align*}
& \Phi_{\bm{F}}(T, V) \\
& \ll_{k, \bm{F}} T \l( \frac{1}{1 + (V / \sqrt{\log{\log{T}}})^{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r)
\exp\l( -\frac{h_{\bm{F}}V^{2}}{\log{\log{T}}} + \frac{C_{1} V^3}{(\log{\log{T}})^2 \log_{3}{T}} \r) \\
& \ll T\l( \frac{1}{1 + (V / \sqrt{\log{\log{T}}})^{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r)
(\log{T})^{C_{1} D_{2}^{3} k^{3} \varepsilon(T)}\exp\l( -h_{\bm{F}}\frac{V^{2}}{\log{\log{T}}} \r)
\end{align*}
for $(\log{\log{T}})^{2/3} \leq V \leq D_{2}k\log{\log{T}}$.
Here, $C_{1} = C_{1}(\bm{F})$ is some positive constant.
Similarly to the proof of \eqref{pNew_MVT2_1_I2} by using this estimate, we obtain
\begin{align*}
I_{5}
\ll_{\bm{F}} T + T(\log{T})^{k^{2}/h_{\bm{F}} + Bk^{3}\varepsilon(T)}
\l( \frac{k \sqrt{\log{\log{T}}}}{1 + (k \sqrt{\log{\log{T}}})^{r}} + \frac{1}{(\log{\log{T}})^{\a_{\bm{F}} + \frac{1}{2}}} \r).
\end{align*}
Hence, we obtain \eqref{New_MVT_RH1}.
For estimate \eqref{New_MVT_RH2}, it holds from the positivity of $\Phi_{\bm{F}}(T, V)$ and equation \eqref{MINEF} that
\begin{align*}
\int_{T}^{2T}\exp\l( 2k \phi_{\bm{F}}(t) \r)dt
\gg_{k} \int_{\frac{k}{h_{\bm{F}}}\log{\log{T}}}^{\frac{k}{h_{\bm{F}}}\log{\log{T}} + \sqrt{\log{\log{T}}}}
e^{2k V}\Phi_{\bm{F}}(T, V)dV.
\end{align*}
When $\theta_i\in[\frac{\pi}{2}, \frac{3\pi}{2}]$,
assuming $\vartheta_{\bm{F}} < \frac{1}{r + 1}$, we use \eqref{RH_LD_JVD1l} to obtain
\begin{align*}
\Phi_{\bm{F}}(T, V)
& \gg_{k, \bm{F}} \frac{T}{1 + (V / \sqrt{\log{\log{T}}})^{r}}
\exp\l( -h_{\bm{F}}\frac{V^{2}}{\log{\log{T}}} - \frac{CV^3}{(\log{\log{T}})^2 \log_{3}{T}} \r) \\
& \gg_{\bm{F}} \frac{T (\log{T})^{-C_{2} k^{3} \varepsilon(T)}}{1 + (V / \sqrt{\log{\log{T}}})^{r}}
\exp\l( -h_{\bm{F}}\frac{V^{2}}{\log{\log{T}}} \r)
\end{align*}
for $\frac{k}{h_{\bm{F}}}\log{\log{T}} \leq V \leq \frac{k}{h_{\bm{F}}}\log{\log{T}} + \sqrt{\log{\log{T}}}$.
Here, $C_{2} = C_{2}(\bm{F})$ is a positive constant.
Similarly to the proof of \eqref{New_MVT2} by using this estimate
and the bound $\Phi_{\bm{F}}(T, V) \gg_{\bm{F}} T$ for $0 \leq V \leq 1$,
we can also obtain \eqref{New_MVT_RH2}.
\end{proof}
\section{\textbf{Concluding remarks}}\label{finalremarks}
\subsection{Moments of max/min values}
Under the same conditions as in Theorem \ref{New_MVT}, we can apply Theorem \ref{GJu} to show that
for sufficiently small $k$, there exists some constant $B=B(\bm F)>0$ such that
\begin{gather}
\label{MAXLUC}
\int_{T}^{2T}\l(\max_{1 \leq j \leq r}|F_{j}(\tfrac{1}{2} + it)|\r)^{2k}dt
\ll_{\bm{F}} T (\log{T})^{n_{\bm{F}}k^2 + B k^{3}}, \\
\label{MINLUC}
\int_{T}^{2T}\l(\min_{1 \leq j \leq r}|F_{j}(\tfrac{1}{2} + it)|\r)^{-2k}dt
\gg_{\bm{F}} T (\log{T})^{n_{\bm{F}}k^2 - B k^{3}},
\end{gather}
where $n_{\bm{F}} =\max_{1 \leq j \leq r}n_{F_{j}}$.
Assuming the Riemann Hypothesis for these $L$-functions, we can replace the term $Bk^3$ in the exponent by any $\epsilon>0$ and the results hold for any $k>0$.
In fact, consider
\begin{align*}
S_{j}
= \set{t \in [T, 2T]}{\Re e^{-i\theta_{j}} \log{F_{j}}(\tfrac{1}{2} + it) > V}.
\end{align*}
Using the inclusion-exclusion principle, we see that
\begin{align}
\begin{aligned}
\label{IEPS_j}
&\sum_{j = 1}^{r}\operatorname{meas} S_{j} - \sum_{1 \leq \ell < m \leq r}\operatorname{meas}\l( S_{\ell} \cap S_{m} \r) \\
&\leq \operatorname{meas}\l(\bigcup_{j = 1}^{r}S_{j}\r)
= \operatorname{meas}\set{t \in [T, 2T]}{\max_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}}(\tfrac{1}{2} + it) > V} \\
& \leq \sum_{j = 1}^{r}\operatorname{meas} S_{j}.
\end{aligned}
\end{align}
Applying the second inequality of \eqref{IEPS_j} and Theorem \ref{GJu} with $r = 1$ and
$V_{1} = \frac{V}{\sqrt{(n_{F_{j}} / 2)\log{\log{T}}}}$ for $j = 1, \dots, r$, we obtain
\begin{align*}
& \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\max_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}}(\tfrac{1}{2} + it) > V} \\
& \ll_{\bm{F}} \frac{1}{1 + V / \sqrt{\log{\log{T}}}}
\exp\l( -\frac{V^2}{n_{\bm{F}}\log{\log{T}}}\l( 1 + O_{\bm{F}}\l( \frac{V}{\log{\log{T}}} \r) \r) \r)
\end{align*}
for any $\bm{\theta} \in [-\tfrac{\pi}{2}, \tfrac{\pi}{2}]^{r}$ and $0 \leq V \leq a \log{\log{T}}$ with $a = a(\bm{F}) > 0$ suitable small.
Similarly, applying the first inequality of \eqref{IEPS_j} and Theorem \ref{GJu} with $r = 1, 2$,
$V_{1} = \frac{V}{\sqrt{(n_{F_{j}}/2)\log{\log{T}}}}$ for $j = 1, \dots, r$,
and $(V_{1}, V_{2}) = \l( \frac{V}{\sqrt{(n_{F_{\ell}} / 2)\log{\log{T}}}}, \frac{V}{\sqrt{(n_{F_{m}} / 2)\log{\log{T}}}} \r)$
for $1 \leq \ell < m \leq r$, we obtain
\begin{align*}
& \frac{1}{T}\operatorname{meas}\set{t \in [T, 2T]}{\max_{1 \leq j \leq r}\Re e^{-i\theta_{j}} \log{F_{j}}(\tfrac{1}{2} + it) > V} \\
& \gg_{\bm{F}} \frac{1}{1 + V / \sqrt{\log{\log{T}}}}
\exp\l( -\frac{V^2}{n_{\bm{F}}\log{\log{T}}}\l( 1 + O_{\bm{F}}\l( \frac{V}{\log{\log{T}}} \r) \r) \r)
\end{align*}
for any $\bm{\theta} \in [\tfrac{\pi}{2}, \tfrac{3\pi}{2}]^{r}$
and $0 \leq V \leq a_{3}\min\{\log{\log{T}}, (\log \log T)^{\frac{\a_{\bm{F}}}{r} + \frac{1}{2} + \frac{1}{2r}}\}$.
The rest of the argument follows the same way as in Theorems \ref{New_MVT} and Theorem \ref{New_MVT_RH}.
Note that assumptions \eqref{SOC2}, (A2) are not needed for \eqref{MAXLUC}.
\subsection{Moments of product of $L$-functions}
Our method can be modified to study the moment of product of $L$-functions. Under the same assumptions as in Theorem \ref{New_MVT}, we can also show that for sufficiently small $k_1, \dots, k_r>0$ there exists some constant $B=B(\bm F)$ such that
\begin{align}
\label{PRDLUC}
&\int_{T}^{2T}\prod_{j = 1}^{r}|F_{j}(\tfrac{1}{2} + it)|^{2k_{j}}dt
\ll_{\bm{F}} T (\log{T})^{n_{F_{1}}k_{1}^2 + \cdots + n_{F_{r}}k_{r}^2 + B K^{3}},\\
\label{PRDLUC2}
&\int_{T}^{2T}\prod_{j = 1}^{r}|F_{j}(\tfrac{1}{2} + it)|^{-2k_{j}}dt
\gg_{\bm{F}} T (\log{T})^{n_{F_{1}}k_{1}^2 + \cdots + n_{F_{r}}k_{r}^2 - B K^{3}}, \quad \text{ if }\theta_{\bm F}\leq \frac{1}{r+1}
\end{align}
where $K=\max_{1\leq j\leq r}k_j$.
If we assume the Riemann Hypothesis for these $L$-functions, the term $BK^3$ in the exponent can be replaced by any $\epsilon>0$ and the results hold for any $k_1, \dots, k_r>0$.
To do this, we need to modify the definition of $\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta})$.
Let $X \geq 3$, $\bm{k} = (k_{1}, \dots, k_{r}) \in (\mathbb{R}_{\geq 0})^{r}$,
and let $\bm{F}$ be a $r$-tuple of Dirichlet series and $\bm{\theta} \in \mathbb{R}^{r}$ satisfying (S4), (S5), (A1), and (A2).
Let
\begin{align*}
\sigma_{\bm{F}}(X, \bm{k})
= \sqrt{\frac{1}{2}\sum_{j = 1}^{r}k_{j}\sum_{p \leq X} \sum_{\ell = 1}^{\infty}\frac{|b_{F_{j}}(p^{\ell})|^2}{p^{\ell}}}
\end{align*}
and
\begin{align*}
\S_{X}(T, \bm{V}; \bm{F}, \bm{\theta}, \bm k)=\frac{1}{T}\operatorname{meas}
\set{t \in [T, 2T]}{\frac{\sum_{j = 1}^{r}k_{j}\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2} + it, X)}{\sigma_{\bm{F}}(X; \bm{k})} > V}.
\end{align*}
Then, we can show that for any $\bm{k} \in (\mathbb{R}_{> 0})^{r}$
and $|V| \leq a \sigma_{\bm{F}}(X; \bm{k})$ with $a = a(\bm{F}, \bm{k}) > 0$ suitably small,
\begin{align*}
\frac{1}{T}\operatorname{meas}
\set{t \in [T, 2T]}{\frac{\sum_{j = 1}^{r}k_{j}\Re e^{-i\theta_{j}}P_{F_{j}}(\tfrac{1}{2} + it, X)}{\sigma_{\bm{F}}(X; \bm{k})} > V}
\sim \int_{V}^{\infty}e^{-u^2/2}\frac{du}{\sqrt{2\pi}}
\end{align*}
by modifying the proofs in Section \ref{Proof_Props_JVD}.
Based on this asymptotic formula, we can also prove an analogue of Theorem \ref{GJu} and thus \eqref{PRDLUC}.
Similarly, we can also obtain \eqref{PRDLUC2}.
\subsection{Other large deviation results}
We would like to mention that our method also recovers the work of Heuberger-Kropf \cite{HK2018}
for higher dimensional quasi-power theorem, and it is likely our method yields improvement to their work in the direction of large deviations.
|
\section{Introduction}
Polar codes \cite{b1}, invented by Ar{\i}kan, are a great break through in coding theory. As code length $N = 2^n$ approaches infinity, the synthesized channels become either noiseless or pure-noise, and the fraction of the noiseless channels approaches channel capacity. Thanks to channel polarization, efficient successive cancellation (SC) decoding algorithm can be implemented with a complexity of $O(NlogN)$. However, the performance of polar codes under SC decoding is poor at short to moderate block lengths.
In \cite{b2}, a successive cancellation list (SCL) decoding algorithm was proposed. As the list size $L$ increases, the performance of SCL decoding approaches that of ML decoding. But the ML performance of polar codes is still inferior due to low minimum distance. Consequently, concatenation of polar codes with CRC \cite{b3} and PC \cite{b4} were proposed to improve weight spectrum.
Recently, Ar{\i}kan proposed polarization-adjusted convolutional (PAC) codes \cite{b5}, which is shown to approach binary input additive white Gaussian noise (BIAWGN) dispersion bound \cite{b7} under large list decoding\cite{b8}.
CRC-Aided (CA) polar, PC-polar, and PAC codes can be viewed as pre-transformed polar codes with upper-triangular transformation matrices\cite{b8}. In\cite{b9}, it is proved that any pre-transformation with an upper-triangular matrix does not reduce the minimum Hamming weight, and a properly designed pre-transformation can reduce the number of minimum-weight codewords. In this paper, we propose an efficient method to calculate the average weight spectrum of pre-transformed polar codes. Moreover, the method holds for arbitrary information sub-channel selection criteria, thus covers polar codes, RM codes and is not constrained by "partial order"\cite{b6}. Our results confirm that the pre-transformation with an upper-triangular matrix can reduce the number of minimum-weight codewords significantly when the information set is properly chosen. In the meantime, it enhances error-correcting performance of SCL decoding.
In section II, we review polar codes and pre-transformed polar codes. In section III we propose a formula to calculate the average weight spectrum of pre-transformation polar codes. In section IV the simulation results are presented to verify the accuracy of the formula. Finally we draw some conclusions in section V.
\section{Background}
\subsection{Polar Code}
Given a B-DMC $W: \{ 0,1 \} \rightarrow \mathcal{Y}$, the channel transition
probabilities are defined as $W(y|x)$ where $y \in \mathcal{Y} , x \in \{ 0,1 \}$. $W$ is said to be symmetric if there is a permutation $\pi$, such that $ \forall$ $y \in \mathcal{Y}$, $W(y|1)=W(\pi(y)|0)$ and $\pi^2 = id$.
Then the symmetric capacity and the Bhattacharyya parameter of $W$ are defined as
\begin{equation}
I(W) \triangleq \sum_{y \in \mathcal{Y}} \sum_{x \in \mathcal{X}} \frac{1}{2} W(y \mid x) \log \frac{W(y \mid x)}{\frac{1}{2} W(y \mid 0)+\frac{1}{2} W(y \mid 1)},
\end{equation}
and
\begin{equation}
Z(W) \triangleq \sum_{y \in \mathcal{Y}} \sqrt{W(y \mid 0) W(y \mid 1)}.
\end{equation}
Let
$
F=\left[\begin{array}{ll}
1 & 0 \\
1 & 1
\end{array}\right]
$, $N=2^m$, and $F_N=F^{\otimes m}$. Starting from $N = 2^m$ independent channels $W$, we obtain $N$ polarized channels $W_N^{(i)}$, after channel combining and splitting operations \cite{b1}, where
\begin{equation}
W_{N}\left(y_{1}^{N}|u_1^N \right) \triangleq W^N \left(y_1^N|u_1^NF_N\right),
\end{equation}
\begin{equation}
W_{N}^{(i)}\left(y_{1}^{N}, u_{1}^{i-1} \mid u_{i}\right) \triangleq \sum_{u_{i+1}^{N} \in \mathcal{X}^{N-i}} \frac{1}{2^{N-1}} W_{N}\left(y_{1}^{N} \mid u_{1}^{N}\right).
\end{equation}
Polar codes can be constructed by selecting the indices of $K$ information sub-channels, denoted by the information set $\mathcal{A} = \left\{ I_1,I_2,\dots,I_K \right\}$. The optimal sub-channel selection criterion for SC decoding is reliability, i.e., selecting the $K$ most reliable sub-channel as information set. Density evolution (DE) algorithm\cite{b10}, Gaussian approximation (GA) algorithm\cite{b11} and the channel-independent polarization weight (PW) construction algorithm\cite{b12} are efficient methods to find reliable sub-channels. The optimal sub-channel selection criterion for SCL decoding is still an open problem. Some heuristic approaches cosider both reliability and row weight to improve minimum code distance.
After determining the information set $\mathcal{A}$, the complement set $\mathcal{A}^c$ is called the frozen set. Let
$u_1^N=(u_1,u_2, \dots , u_N)$ be the bit sequence to be encoded. The information bits are inserted into $u_{\mathcal{A}}$,
and all zeros are filled into $u_{\mathcal{A}^c}$. Then the codeword $x_1^N$ is obtained by $x_1^N=u_1^NF_N$.
\subsection{Weight Spectrum of Polar Codes}
There are several prior works to obtain the weight spectrum of polar codes. In \cite{b15}, the authors use SCL decoding with a large list size to decode an all-zeros codeword. Codewords within the list are enumerated to estimate the number of low-weight codewords. In \cite{r1}, this approach is improved in term of memory usage. The above methods only obtain partial weight spectrum. In \cite{r2}\cite{r3}, probabilistic computation methods are proposed to estimate the weight spectrum of polar codes.
\subsection{Weight Spectrum of Polar Cosets}
As in \cite{r4}, let $u_1^{i-1} \in \{0,1\}^{i-1}$, $u_i \in \{0,1\}$, define the polar coset $C_N^{(i)}\left(u_1^{i-1},u_i \right)$ as
$$C_N^{(i)}\left(u_1^{i-1},u_i \right)=\left\{ (u_1^i,u')F_N|u' \in \{0,1\}^{n-i} \right\}.$$
In \cite{r5}\cite{r6}, recursive formulas are proposed to efficiently compute the weight spectrum of $C_N^{(i)}\left(0_1^{i-1},1 \right)$. The weight spectrum of $C_N^{(i)}\left(0_1^{i-1},1 \right)$ is tightly associated with the performance of SC decoding, our analysis of average weight spectrum of pre-transformed polar codes is based on the polar coset spectrum as well.
\subsection{Pre-Transformed Polar Codes}
\begin{equation*}
T
=\begin{bmatrix}
1 & T_{12} & \cdots\ &T_{1N}\\
0 & 1 & \cdots\ & T_{2N}\\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots\ & 1\\
\end{bmatrix}
\end{equation*}
The above non-degenerate upper-triangular pre-transformation matrix $T$ has all ones on the main diagonal. Let $G_N = TF_N$ and $u_{\mathcal{A}^c}=\textbf{0}$, the codeword of the pre-transformed polar codes is given by $x_1^N=u_1^NG_N=u_1^NTF_N$. In original polar codes, the frozen bits are fixed to be zeros. While in pre-transformed polar codes, the frozen bits are linear combination of previous information bits.
\section{Average Code Spectrum
Analysis}
In this section, we propose a formula to compute the average weight spectrum of the pre-transformed polar codes, with focus on the number of low-weight codewords. The average number assumes that $T_{ij}$, $1\leq i < j \leq N$ are $i.i.d.$ $Bernoulli(\frac{1}{2})$ $r.v.$.
\subsection{Notations and Definitions}
$f_N^{(i)}$ is the $i$-$th$ row vector of $F_N$, and $g_N^{(i)}$ is the $i$-$th$ row vector of $G_N$. The number of codewords with Hamming weight $d$ of the pre-transformed polar codes is denoted by $N_d(\mathcal{A},T)$. The minimum distance of polar/RM codes and the pre-transformed codes are denoted by $d_{min}(\mathcal{A})$ and $d_{min}(\mathcal{A},T)$, respectively. The number of minimum-weight codewords of polar/RM codes and the pre-transformed codes are denoted by $N_{min}(\mathcal{A})$ and $N_{min}(\mathcal{A},T)$, respectively.
\subsection{Code Spectrum Analysis}
The expected number of codewords with Hamming weight $d$ is
\begin{align}
\label{eq:1}
&E\left[ N_d(\mathcal{A},T)\right] \notag \\
&= \sum_{u_{I_1},u_{I_2},\dots,u_{I_K}\in \{0,1\}^K} P\left(w \left(\oplus \sum_{i=1}^K u_{I_i} g_N^{(I_i)} \right)=d \right) \notag \\
&= \sum_{j=1}^K \sum_{\mbox{\tiny$\begin{array}{c}
u_1,\dots,u_{I_{j-1}}=0\\
u_{I_j}=1\\
u_{I_{j+1}},\dots,u_{I_K} \in \left\{0,1\right\}^{K-j}\end{array}$}} \notag \\
&P \left(w\left(g_N^{(I_j)}\oplus \sum_{i=j+1}^K u_{I_i} g_N^{(I_i)}\right)=d\right).
\end{align}
The expectation is with respect to $T$, so $E\left[ N_d(\mathcal{A},T)\right]$ is a function of $\mathcal{A}$.
\begin{Le}
${\forall}$ $u_{I_{j+1}},\dots,u_{I_K} \in \left\{0,1\right\}^{K-j}$,
$$P \left(w\left(g_N^{(I_j)}\oplus \sum_{i=j+1}^K u_{I_i} g_N^{(I_i)}\right)=d \right)=P \left(w\left(g_N^{(I_j)}\right)=d \right).$$
\end{Le}
\begin{proof}According to the pre-transformation matrix,
\begin{equation*}
g_N^{(I_j)}=f_N^{(I_j)}\oplus \sum_{i=I_j+1}^N T_{I_ji} f_N^{(i)},
\end{equation*}
$$g_N^{(I_j)}\oplus \sum_{i=j+1}^K u_{I_i} g_N^{(I_i)}=f_N^{(I_j)}\oplus \sum_{i=I_j+1}^N T'_{I_ji} f_N^{(i)}.$$
And
\begin{equation*} T'_{I_ji} \triangleq
\begin{cases}
\sum_{I_k < i,u_{I_k}=1}T_{I_ki} & i \notin \left[I_{j+1},\dots,I_K \right] \\
\sum_{I_k < i,u_{I_k}=1}T_{I_ki} \oplus u_i & i \in \left[I_{j+1},\dots,I_K \right].
\end{cases}
\end{equation*}
It is straightforward to see that when $T_{I_ji}$ are $i.i.d.$ $Bernoulli(\frac{1}{2})$ $r.v.$, $T'_{I_ji}$ are $i.i.d.$ $Bernoulli(\frac{1}{2})$ $r.v.$ as well.
As a result, $g_N^{(I_j)}$ and $g_N^{(I_j)}\oplus \sum\limits_{i=j+1}^K u_{I_i} g_N^{(I_i)}$ follow the same distribution too, ${\forall}$ $u_{I_{j+1}},\dots,u_{I_K} \in \left\{0,1\right\}^{K-j}$.
Therefore \textbf{Lemma 1} holds.
\end{proof}
\begin{Le}
If $w(f^{(I_j)}_N) > d$, $P \left(w(g_N^{(I_j)})=d \right)=0$.
\end{Le}
\begin{proof}
Recall that
$$g_N^{(I_j)}=f_N^{(I_j)}\oplus \sum_{i=I_j+1}^N T_{I_ji} f_N^{(i)}.$$
According to \cite[Corollary 1]{b9},
$$w\left(g_N^{(I_j)}\right) \geq w\left(f_N^{(I_j)} \right)>d,$$
therefore
$$P \left(w\left(g_N^{(I_j)}\right)=d \right)=0.$$
\end{proof}
According to \textbf{Lemma 1} and \textbf{Lemma 2}, (\ref{eq:1}) can be further simplified to
\begin{align}
\label{eq:2}
E\left[ N_d(\mathcal{A},T)\right] = \sum_{\mbox{\tiny$\begin{array}{c}
1 \leq j \leq K\\
w(f_{I_j}) \leq d\end{array}$}}2^{K-j}P \left(w\left(g_N^{(I_j)}\right)=d \right).
\end{align}
Let $P(m,i,d) \triangleq P\left(w\left(g_{2^m}^{(i)}\right)=d\right)$ , (\ref{eq:2}) can be rewritten as
\begin{align}
\label{eq:4}
E\left[ N_d(\mathcal{A},T)\right] = \sum_{\mbox{\tiny$\begin{array}{c}
1 \leq j \leq K\\
w(f_{I_j}) \leq d\end{array}$}}2^{K-j}P \left(m,I_j,d \right).
\end{align}
In particular, let $P(m,i) \triangleq P\left(w\left(g_{2^m}^{(i)}\right)=w\left(f_{2^m}^{(i)}\right)\right)$. So if $d=d_{min}$, (\ref{eq:2}) can be rewritten as\\
$E\left[ N_{min}(\mathcal{A},T)\right] =$
\begin{align}
\label{eq:5}
\sum_{\mbox{\tiny$\begin{array}{c}
1 \leq j \leq K\\
w(f_{I_j})=d_{min}(\mathcal{A})\end{array}$}}2^{K-j}P \left(m,I_j\right).
\end{align}
Let $A_{d}$ denote the number of codewords in $C_N^{(i)}\left(0_1^{i-1},1 \right)$ with Hamming weight $d$. Clearly, $2^{N-i}P(m,i)=A_{w\left(f_N^{(i)}\right)}$, $2^{N-i}P(m,i,d)=A_d$.
In \cite{r5}\cite{r6}, the authors propose recursive formulas to calculate the weight spectrum of polar cosets.
In \textbf{Theorem 1} and \textbf{Theorem 2}, we investigate the recursive fomulas for $P(m,i)$ and $P(m,i,d)$, which are similar to the formula in \cite{r6}. But instead of polar cosets, we are interested in the pre-transformed polar codes. For the completeness of the paper, the proofs are in the appendix.
\begin{theo}
\begin{align}
\label{eq:8}
P(m,i)=
\begin{cases}
\frac{2^{w(f_{2^m}^{(i)})}}{2^{2^{m-1}}}P(m-1,i)& 1 \leq i \leq 2^{m-1}\\
P(m-1,i-2^{m-1})& 2^{m-1} < i \leq 2^m,
\end{cases}
\end{align}
with the boundary conditions $P(1,1)=P(1,2)=1$.
\end{theo}
With (\ref{eq:5}) and (\ref{eq:8}), we can recursively calculate the average number of minimum-weight codewords. We are also interested in other low-weight codewords on the weight spectrum, since together they determine the ML performance at high SNR. The problem boils down to evaluating the more general formula of $P(m,i,d)$. As we will see in \textbf{Theorem 2}, the average weight spectrum can be calculated efficiently in the same recursive manner especially for codewords with small Hamming weight.
\begin{theo}
If $1 \leq i \leq 2^{m-1}$
\begin{align}
\label{eq:r1}
P(m,i,d)=\sum\limits_{\mbox{\tiny$\begin{array}{c}
d'=w\left(f_{2^m}^{(i)}\right)\\
d-d' is \ even \end{array}$}}^{d}P(m-1,i,d')\frac{2^{d'}\begin{pmatrix} 2^{m-1}-d' \\ \frac{d-d'}{2} \end{pmatrix}}{2^{2^{m-1}}}.
\end{align}
If \ $2^{m-1} < i \leq 2^m$
\begin{align}
\label{eq:r2}
P(m,i,d)=\begin{cases}P(m-1,i-2^{m-1},d/2) &d \ is \ even\\
0 &d \ is\ odd,
\end{cases}
\end{align}
with the boundary conditions $P(1,1,1)=P(1,2,2)=1$. And
\begin{align}
\label{eq:6}
\begin{cases}
P(m,1,d)=0,&\ if \ d \ is \ even \\
P(m,i,d)=0,&\ if \ i > 1 \ and \ d \ is \ odd.
\end{cases}
\end{align}
\end{theo}
\subsection{Complexity Analysis}
In this section, we consider the computational complexity of average weight spectrum of pre-transformed polar codes.
According to \textbf{Theorem 1} and (\ref{eq:5}), the computational complexity of average number of minimum-weight codewords is $O(NlogN)$.
Let $\chi(N)$ denote the worst case complexity of computing the average spectrum of pre-transformed polar codes with code length $N$. According to (\ref{eq:r1}) and (\ref{eq:r2}), $O(N)$ operations are required for computing each $P(m,i,d)$, after the computation of average spectrum with code length $\frac{N}{2}$. So the computation complexity of $P(m,i,d)$ over $1 \leq i \leq N$, $0 \leq d \leq N$ is $O(N^3)$. At last, (\ref{eq:4}) requires no more than $N$ calculation. In short, $\chi(N) \leq \chi(N/2) + O(N^3)$. Consequently, $\chi(N) = O(N^3)$.
\section{Simulation}
In this section, we verify the correctness of the recursive formula through simulations. In particular, we employ the "large list decoding" method described in \cite{b15} to collect low-weight codewords. Transmit all-zero codeword without noise, and use list decoding to decode the channel output. With sufficiently large list size $L$, the decoder collects all the low-weight codewords. At first, we randomly generate one thousand pre-transfom matrices for RM(128, 64), and set $L=5\times 10^3$ to count the number of minimum-weight codewords for each matrix, and obtain their average $N_{min}$. The result is shown in Fig.\;\ref{fig1}: $d_{min}=16$, $N_{min}^{simulation}=2768.1$, $N_{min}^{recursion}=2766.9$.
To show that our recursive formula is applicable for any sub-channel selection criterion we also construct polar code(128, 64) by the PW algorithm \cite{b12}. The simulation result is shown in Fig.\;\ref{fig2}: $d_{min}=8$, $N_{min}^{simulation}=272.64$, $N_{min}^{recursion}=272$.
Our recursive formula is also applicable for longer codes. We set $L=5\times 10^4$ to count the number of minimum-weight codewords for one thousand pre-transformed RM(512, 256). The result is shown in Fig.\;\ref{fig6}: $d_{min}=32$, $N_{min}^{simulation}=1.5933\times 10^4$, $N_{min}^{recursion}=1.5936\times 10^4$.
As seen, the recursively calculated minimum-weight codeword numbers are very close to ones obtained through simulation. Furthermore, they show that the variance of the number of minimum-weight codeword is small.
\begin{figure}[htbp]
\centerline{\includegraphics[width = .5\textwidth]{RMt.eps}}
\caption{RM(128,64), the black solid line is calculated, and the blue solid line is obtained from simulation.}
\label{fig1}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[width = .5\textwidth]{PWt.eps}}
\caption{ PW(128,64), the black solid line is calculated, and the blue solid line is obtained from simulation.}
\label{fig2}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[width = .5\textwidth]{RM512.eps}}
\caption{RM(512,256), the black solid line is calculated, and the blue solid line is obtained from simulation.}
\label{fig6}
\end{figure}
In Table.\;\ref{tab1}, we display the number of minimum codewords of the original RM/polar codes, and the average number is recursively calculated. It is shown that pre-transforming significantly reduces the number of minimum-weight code words, especially in RM(128, 64). The significant improvement of weight spectrum after pre-transformation explains why the CA-polar, PC-polar, and PAC codes outperform the original polar codes under list decoding with large list size.
\begin{table}[htbp]
\caption{comparsion between original polar codes and Pre-transfomed polar codes}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\ \ &\multicolumn{3}{|c|}{\textbf{Minimum-weight codewords}} \\
\cline{2-4}
\ \ & $d_{min}$ &\textit{Original}& \textit{Pre-trasformed} \\
\hline
RM(128,64)&16&94488 &2767 \\
\hline
PW(128,64)&8&304 &272 \\
\hline
\end{tabular}
\label{tab1}
\end{center}
\end{table}
The improvement can be observed under different code lengths and rates, as we can see from Fig.\;\ref{fig3}. In all cases, pre-transformation reduces the number of minimum codewords significantly.
\begin{figure}[htbp]
\centerline{\includegraphics[width = .5\textwidth]{ori_avet.eps}}
\caption{comparsion between original polar codes and pre-transformed polar codes under different code lengths and rates}
\label{fig3}
\end{figure}
In addition to minimum-weight codewords, we also simulate to verify the accuracy of the formula for other low-weight codewords. The simulation results are shown in Table.\;\ref{tab2} for RM(128, 64) and PW(128, 64) respectively, where
$N^{sim}$ is the simulation results, and $N^{recur}$ is the calculation results.
\begin{table}[htbp]
\caption{comparsion between simulation results of RM(128, 64) and PW(128, 64) with 500 realizations and the calculation results by the proposed recursive formula}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\multicolumn{3}{|c|}{\textbf{RM(128,64)}}& \multicolumn{3}{|c|}{\textbf{PW(128,64)}} \\
\hline
$d$ & $N_d^{sim}$ & $N_d^{recur}$ & $d$ & $N_d^{sim}$ & $N_d^{recur}$ \\
\hline
16&2764.5&2766.9 &8&272.2&272 \\
\hline
18&397.1&393.5 &12&896.6&896 \\
\hline
20&80251 &80182&16&76812.2&77111 \\
\hline
\multicolumn{6}{l}{Note that $N_{10}=N_{14}=0$ for PW(128, 64)}
\end{tabular}
\label{tab2}
\end{center}
\end{table}
In PC-polar codes \cite{b4}, both reliability and code distance are taken into consideration when selecting the information set. A coefficient $\alpha$ is used to control the tradeoff between reliability and code distance. The larger $\alpha$ is, the greater code distance is. A parity check pattern can be considered as a realization of the pre-transformation matrix. Take PC-polar codes(128, 64) ($\alpha=1.5$) as an example, we calculate the average number of low-weight codewords. The result implies that pre-transformtion can increase the minimum code distance when the information set is properly chosen, that is, reducing the number of original minimum-codewords to zero. The number of low-weight codewords of the original code, a realization of the pre-transformed code and the code ensemble average are shown in Table.\;\ref{tab3}. In this case, although some rows of $F_N$ with Hamming weight 8 are selected into the information set, PC-polar codes can increase the minimum distance from 8 to 12.
\begin{table}[htbp]
\caption{comparsion between the low-weight codewords number of the original codes, a realization of the pre-transformed codes and the ensemble average}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Hamming weight &\multicolumn{3}{|c|}{$N_d$} \\
\cline{2-4}
$d$ & Original & Pre-transformed & Average \\
\hline
8 &32 &0 &0.5 \\
\hline
10 &0 &0 &0.0547 \\
\hline
12 &0 &48 &39.5 \\
\hline
14 &128 &28 &27 \\
\hline
16 &57048 &5228 &5250 \\
\hline
\end{tabular}
\label{tab3}
\end{center}
\end{table}
In CA-polar codes \cite{b3}, $r$ CRC bits are attached to $K$ information bits and all the $K'=K+r$ bits are fed into the polar encoder. To construct CA-polar codes, $K'$ indices are selected, and the first $K$ of them are information bits, while the others are dynamic frozen bits\cite{b19}. We construct CA-polar code ($N=128, K=64, r=6$) by reliability sequence in \cite{b21}. The number of low-weight codewords of the original code, a CA-polar code with generator polynomial $g(D) = D^6+D+1$ and the code ensemble average are shown in Table.\;\ref{tab4}.
\begin{table}[htbp]
\caption{comparsion between the low-weight codewords number of the original codes, a CA-polar code and the ensemble average}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Hamming weight &\multicolumn{3}{|c|}{$N_d$} \\
\cline{2-4}
$d$ & Original & CA & Average \\
\hline
8 &529 &4 &10.75 \\
\hline
10 &0 &0 &0.0547 \\
\hline
12 &0 &145 &85.5 \\
\hline
14 &0 &0 &27.07 \\
\hline
16 &3.364$\times 10^5$ &12550 &4952.4 \\
\hline
\end{tabular}
\label{tab4}
\end{center}
\end{table}
Fig.\;\ref{fig4} and Fig.\;\ref{fig5} provide the BLER performances of various constructions under different list sizes, with reference to finite-length performance bounds such as normal approximation (NA), random-coding union (RCU) and meta-converse (MC) bounds \cite{b7} \cite{b16} \cite{b17}. PW pre, PC-Polar ($\alpha$ = 1.5), PC-Polar ($\alpha$ = 3.5) codes are specific realizations drawn from the code ensemble with different information set selections. The information sets of PAC and PC-Polar ($\alpha$ = 3.5) codes turn to be the same. In PAC codes, the transformation matrix $T$ is a upper-triangular Toeplitz matrix, while in PC-Polar codes ($\alpha$ = 3.5), $T$ is a randomly generated upper-triangular matrix. It is observed that reliability is the only contributing factor to decoding performance under SC decoding. Under SCL decoding with list size $L = 8$, the PC-polar codes ($\alpha = 1.5$) strike a good balance between reliability and distance, and shows the best decoding performance. When the list size is large enough, both PAC and PC-polar codes ($\alpha$ = 3.5) can approach NA bound with their ML performances.
\begin{figure}[htbp]
\centerline{\includegraphics[width = .5\textwidth]{SCr.eps}}
\caption{Performance comparison between different code constructions under SC and SCL, $L$=8}
\label{fig4}
\end{figure}
\begin{figure}[htbp]
\centerline{\includegraphics[width = .5\textwidth]{SCL256r.eps}}
\caption{Performance comparison between different code constructions under SCL L=256, with reference to finite-length performance bounds}
\label{fig5}
\end{figure}
\section{Conclusion}
In this paper, we propose recursive formulas to efficiently calculate the average weight spectrum of pre-transformed polar codes, which include CA-polar, PC-polar and PAC codes as special cases. It is worth mentioning that our formulas work for any sub-channel selection criteria. We found that, with pre-transformation, the average number of minimum codewords decreases significantly, therefore outperforming the original RM/polar codes under the ML decoding and SCL decoding with large list sizes. Furthermore, as in the instance of PC-polar codes ($\alpha=1.5$), the combination of a proper sub-channel selection and pre-transformation has the potential to increase minimum code distance by eliminating minimum-weight codewords.
\newpage
\begin{appendix}
\emph{A. Proof of Theorem 1}
\begin{proof}
A trivial examination can prove the correctness of the boundary conditions. Let us focus on deriving the recursive formula.
\emph{Case 1}: $1 \leq i \leq 2^{m-1}$\\
$$g_{2^m}^{(i)}=f_{2^m}^{(i)} \oplus \sum_{j=i+1}^{2^{m-1}}T_{ij}f_{2^m}^{(j)} \oplus \sum_{j=2^{m-1}+1}^{2^m}T_{ij}f_{2^m}^{(j)}.$$
Let $f_{2^m}^{(i)} \oplus \sum\limits_{j=i+1}^{2^{m-1}}T_{ij}f_{2^m}^{(j)} \triangleq [\textbf{X},\textbf{0}]$, $\sum\limits_{j=2^{m-1}+1}^{2^m}T_{ij}f_{2^m}^{(j)} \triangleq [\textbf{Y},\textbf{Y}]$, where $\textbf{0}$ is an all-zero row vector of length $ 2^{m-1}$, $\textbf{X}=\left(x_1,\dots,x_{2^{m-1}}\right)$ , $\textbf{Y}=\left(y_1,\dots,y_{2^{m-1}}\right)$.
Apparently, \textbf{X} and \textbf{Y} are independent, and ${\forall} \ \textbf{a}=(a_1,\dots,a_{2^{m-1}}) \in \{0,1\}^{2^{m-1}}$, $P(\textbf{Y}=\textbf{a})=2^{-2^{m-1}}$. Let $w(\textbf{X})=d_1$, $w(\textbf{Y})=d_2$, and $c$ be the number of positions where \textbf{X} and \textbf{Y} are both 1. We have
\begin{align*}
w(g_{2^m}^{(i)})&=w([\textbf{X} \oplus \textbf{Y},\textbf{Y}])\\
&=w(\textbf{X} \oplus \textbf{Y})+w(\textbf{Y})\\
&=d_1+2d_2-2c.
\end{align*}
Because $d_2 \geq c$ and $d_1 \geq w(f_{2^m}^{(i)})$ \cite[Corollary 1]{b9}, the equation $w(g_{2^m}^{(i)})=d_1+2d_2-2c=w(f_{2^m}^{(i)})$ holds if and only if $d_1=w(f_{2^m}^{(i)})$, $d_2=c$. In fact, \{$d_2=c$\} denotes the event that \textbf{X} \emph {covers} the first half of \textbf{Y}, i.e., if $x_i=0$ then $y_i =0$, for all $1 \leq i \leq 2^{m-1}$. Let $\{i_1,\dots,i_{d_1}\}$ denote the $d_1$ locations where $x_{i_1},\dots,x_{i_{d_1}}=1$, hence the recursive formula is
\begin{align*}
P(m,i)&=P(d_1=w(f_{2^m}^{(i)}))*P(d_2=c|d_1=w(f_{2^m}^{(i)}))\\
&=P(m-1,i)*\\
&P\left(y_{i_1},\dots,y_{i_{d_1}} \in \{0,1\}^{d_1},y_i=0 \ otherwise \right)\\
&=P(m-1,i)*\frac{2^{d_1}}{2^{2^{m-1}}}\\
&=P(m-1,i)*\frac{2^{w(f_{2^m}^{(i)})}}{2^{2^{m-1}}}.
\end{align*}
\emph{Case 2}: $2^{m-1} < i \leq 2^m$
\begin{align}
\label{eq:7}
g_{2^m}^{(i)}&=\left[f_{2^m}^{(i)} \oplus \sum_{j=i+1}^{2^m}T_{ij}f_{2^m}^{(j)} \right] \notag \\
&=\bigg[f_{2^{m-1}}^{(i-2^{m-1})} \oplus \sum_{j=i+1}^{2^m}T_{ij}f_{2^{m-1}}^{(j-2^{m-1})}, \notag \\
& \ \ \ \ f_{2^{m-1}}^{(i-2^{m-1})} \oplus \sum_{j=i+1}^{2^m}T_{ij}f_{2^{m-1}}^{(j-2^{m-1})} \bigg] \notag \\
&\backsim \left[g_{2^{m-1}}^{(i-2^{m-1})},g_{2^{m-1}}^{(i-2^{m-1})}\right],
\end{align}
where $X_1\backsim X_2$ means $X_1,X_2$ have the same distribution.
\end{proof}
\emph{B. Proof of Theorem 2}
\begin{proof}
(\ref{eq:6}) is obtained with the observation that $w\left(f_N^{(1)}\right)$ is odd and $\forall$ $i > 1$, $w\left(f_N^{(i)}\right)$ is even.
\emph{Case 1}: $1 \leq i \leq 2^{m-1}$
Similar to the proof of \textbf{Theorem 1}, let $w(\textbf{X})=d_1$, $w(\textbf{Y})=d_2$ and $c$ be the number of positions where \textbf{X} and \textbf{Y} are both 1. Denoted by $\mathcal{V}=\{v_1,\dots,v_c\}$ the set of positions where \textbf{X} and \textbf{Y} are both 1, and $\mathcal{V}^c$ its complement. Let $\textbf{Y}_{\mathcal{V}^c}$\ denote the corresponding subvector of \textbf{Y}, we have $w\left(\textbf{Y}_{\mathcal{V}^c}\right)=d_2-c$. Because
\begin{align*}
w(g_{2^m}^{(i)})&=w([\textbf{X} \oplus \textbf{Y},\textbf{Y}])\\
&=w(\textbf{X} \oplus \textbf{Y})+w(\textbf{Y})\\
&=d_1+2d_2-2c\\
&=d,
\end{align*}
then $d_2-c=\frac{d-d_1}{2}$, so $d-d_1$ must be even. No matter what $c$ is, the equation is satisfied if and only if $w\left(\textbf{Y}_{\mathcal{V}^c}\right)=\frac{d-d_1}{2}$. Based on the above observations, $P(m,i,d)$ can be formulated as
\begin{align*}
&P(m,i,d)\\
&=\sum\limits_{\mbox{\tiny$\begin{array}{c}
d'=w\left(f_{2^m}^{(i)}\right)\\
d-d' is \ even \end{array}$}}^{d}P(m,i,d|w\left(\textbf{X}\right)=d') \ast P(w\left(\textbf{X}\right)=d')\\
&=\sum\limits_{\mbox{\tiny$\begin{array}{c}
d'=w\left(f_{2^m}^{(i)}\right)\\
d-d' is \ even \end{array}$}}^{d}P(m,i,d|w\left(\textbf{X}\right)=d') \ast P(m-1,i,d').
\end{align*}
The last equality holds due to $\textbf{X}\backsim g_{2^{m-1}}^{(i)}$.
In particular
\begin{align*}
P(m,i,d|w\left(\textbf{X}\right)=d') &= P\left(w\left(\textbf{Y}_{\mathcal{V}^c}\right)=\frac{d-d_1}{2}\right)\\
&= \frac{2^{d'}\begin{pmatrix} 2^{m-1}-d' \\ \frac{d-d'}{2} \end{pmatrix}}{2^{2^{m-1}}}.
\end{align*}
Consequently, the recursive formula is
\begin{align*}
P(m,i,d)=\sum\limits_{\mbox{\tiny$\begin{array}{c}
d'=w\left(f_{2^m}^{(i)}\right)\\
d-d' is \ even \end{array}$}}^{d}P(m-1,i,d')\frac{2^{d'}\begin{pmatrix} 2^{m-1}-d' \\ \frac{d-d'}{2} \end{pmatrix}}{2^{2^{m-1}}}.
\end{align*}
\emph{Case 2}: $2^{m-1} < i \leq 2^m$, according to (\ref{eq:7})
\begin{align*}
g_{2^m}^{(i)}
&\backsim \left[g_{2^{m-1}}^{i-2^{m-1}},g_{2^{m-1}}^{i-2^{m-1}}\right].
\end{align*}
It is straightforward to obtain the recursive formula
$$P(m,i,d)=P(m-1,i-2^{m-1},d/2).$$
\end{proof}
\end{appendix}
|
\section{Background}
Currently, the most scalable technology with the potential to achieve the feeling of being immersed in a remote location is a panoramic camera streaming to a \ac{vr} \ac{hmd}; see Fig.~\ref{fig:teaser}. The immersion provided by an \ac{hmd} can facilitate more natural interaction and thus ease the difference between telepresence and physical presence \cite{slater1997framework}. Unfortunately, there are several concerns regarding the use of telepresence with an \ac{hmd} that do not typically occur in telepresence with a Fixed Naked-eye Display (FND, a traditional display such as on a computer or a phone). The velocities that the robot uses or the passing distances to objects and walls, which would otherwise be comfortable in FND telepresence, can be uncomfortable in immersive telepresence. Second, a serious consequence of using an \ac{hmd} is \textit{VR sickness} \cite{LaValle_bookVR}, which can result in the user experiencing nausea and vertigo.
Although there are known methods to reduce VR sickness \cite{Chang_Kim_Yoo_2020}, such as reducing the \ac{fov} \cite{Teixeira_Palmisano_2020},
these techniques may have a negative effect on the feeling of being present \cite{weech2019presence}. However, by planning autonomous trajectories for the robot while avoiding motions known for contributing to VR sickness, the feeling of sickness can be decreased without sacrificing the feeling of presence. Additionally, speeds and distances to objects can also be taken into account to make the experience comfortable for the user. Thus, to effectively adapt the robot's motion planning criteria priorities to each user, the impacts of these components on the user must be analyzed. For example, turns have been shown to play a role in VR sickness, \cite{hu1999systematic}, but there is evidence that the duration of the turn is more important than the total angle of rotation \cite{widdowson2019assessing}. This knowledge was used to plan piecewise linear trajectories for a robot, with rotations in place, which resulted in reduced VR sickness in an immersive telepresence experience \cite{becerra2020human}.
However, there are other aspects to consider besides reducing VR sickness; issues such as the path not feeling natural or qualities of the turns when rotating in place may detract from the experience. These issues are typically not a problem in telepresence through an FND, since it is the immersion experienced through the near-eye displays of an \ac{hmd} that allow the user to really feel them. These issues are also not frequently studied in pure VR research, since motions in VR are usually either directly controlled by the user, or teleportation is used. Although these methods for locomotion in VR may be more comfortable for the user, teleportation is infeasible in telepresence and manual control is perceived as bothersome in large environments \cite{rae2017robotic}. Thus, there is a need to research autonomous motions to overcome these obstacles in immersive telepresence.
Negative impacts, like VR sickness, are not the only concern when designing autonomous telepresence robot trajectories. An interesting, yet unexplored, positive element of the telepresence experience is the \textit{naturalness} of the robot's motions. Some work has discussed attempts to make robot motion more natural \cite{kwon2008natural} or deployed voiced commands to make robots perform more naturally \cite{laviers2019make}. However, to the authors' knowledge, there is no widely accepted definition of natural motions for mobile robots. Moreover, from the perspective of the user embodying the telepresence robot, there may not currently be any research on whether people even prefer motions that they would perceive as natural, or more generally what kind of trajectories are preferred for an autonomously moving telepresence robot to take.
\begin{figure}
\vspace{4mm}
\includegraphics[width=\columnwidth]{VAM-HRI_museum_image.png}
\caption{A screenshot inside the virtual museum, seen by participants from the point of view of the virtual telepresence robot. The entrance hallways is seen on top, and the main gallery in the museum is shown on the bottom.}
\Description{The top shows a hallway with paintings. One painting shows brightly colored balloons. The bottom shows a room with statues of a green bird, a red abstract design, a brown warthog, and a blue circle with legs.}
\label{fig:museum}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Naturals.png}
\caption{Conceptualization of the dimensions on which subjects consider movement to be natural. Horizontally is the time frame dimension. Things closer to the origin are more immediate (turns), whereas away from the origin, longer term, relates to things like the path. On the generality dimension, things closer to the origin are specific to a place (in this museum) whereas things away from the origin are more general (in any setting). The units of each dimension are not discrete and can vary by individual.}
\Description{Along the bottom is an arrow pointing to the right, which is labelled "Time Frame." Along the left side of the image is an arrow pointing up which is labelled "Generality." Various points are labeled inside of these axes, based on whether they are short term (to the left of the image) or long term (to the right of the image) and whether they are very specific to a certain context (near the bottom) or more universally applicable (towards the top).}
\label{fig:axes}
\end{figure}
\section{Methods}
We performed an experiment where the subjects wore an \ac{hmd} and watched three pre-recorded trajectories of a virtual telepresence robot moving through a virtual museum. A screenshot from inside the museum can be seen in Fig.~\ref{fig:museum}, and the videos (in 2D format) can be viewed at: https://youtu.be/FPbrIlw3OhY. The data was collected in February 2020, before coronavirus restrictions were enacted in Finland. Two of the trajectories were piecewise linear trajectories with 45$^{\circ}$ or 90$^{\circ}$ turns, optimized for the least amount of turns and for the shortest path (see \cite{becerra2020human} for the path planning), while the third path was a smooth path created with a Rapidly-exploring Random Trees (RRT) algorithm \cite{lavalle2001randomized}. We selected the RRT because it is commonly used in robot motion planning, it respects the differential drive robot system's dynamics while planning the trajectory in 5-dimensional space, and because there is evidence that RRT paths are considered to be human-like \cite{turnwald2019human}. We asked the subjects to rate aspects of the paths on a Likert-scale, and to select their preferred path, the most comfortable path, and the most natural path. We also asked them open-ended questions regarding why they made those selections.
The environments were created with the Unity 3D game engine and displayed to subjects as 360 degree videos played through the Virtual Desktop application into an Oculus Rift S. Subjects were seated in a stationary chair and able to look around while the videos were played. Thirty-six subjects from campus and the local community were run in a gender and path counterbalanced within-subjects design, with each subject viewing all three of the stimulus videos and completing a Simulator Sickness Questionnaire (SSQ, \cite{kennedy1993simulator}) and questionnaires regarding their experience on each path. Subjects signed a consent form before participating and were debriefed after the study was completed.
\section{Results}
\label{results}
We set out to examine the interplay between the preference, comfort, and perceived naturalness of robot motions for first-person experiences in VR telepresence. We found the most salient aspect regarding comfort was in relation to the turns. Comfort had a strong association with preference, but not with naturalness. We also found that, after comfort, preference for a particular path was influenced the most by forward speed and the turns. Preference was strongly associated with the users' perceived naturalness, which was primarily determined by the ability to see salient objects, the distance to the walls and objects, as well as the turns. Participants favored the paths that had a one meter per second forward speed and rated the path with the least amount of turns as the most comfortable. The piecewise linear paths with rotation in place were preferred and selected as the most natural more often than the smooth path where rotation and translation occurred simultaneously. Finally, we also found that the robot's speed and passing distances had a significant impact on the users, and must be carefully considered when implementing immersive telepresence.
Naturalness is a topic often mentioned as a goal in robotics publications, but there is no widely accepted definition of naturalness and it can mean different things to different people. In order to further elucidate how naturalness is thought of by people interacting with a robot, we asked participants open-ended questions regarding the naturalness of the robot's motions. We examined the answers on why people considered a certain path to be the most natural and found that there seem to be two dimensions (see Fig.~\ref{fig:axes}) on which people interpreted the given definition of naturalness, "how you would have moved through the museum if you were in control": time frame, and generality. On the time frame dimension, responses regarding either how the robot performed the turns (fifteen comments, twenty-five percent of the total comments on naturalness) or how close to walls and objects the robot moved (thirteen comments, twenty-two percent) can be considered to reflect something more immediate, while comments related to things further in time referred to aspects of the whole path through the museum (fifteen comments, twenty-five percent). Regarding the generality dimension that describes either context-specific or more general behavior, subjects specifically mentioned the museum setting in seven answers spread across all paths, but some comments regarding turning, for example, could be considered as not specific to the context.
The two dimensions in Fig.~\ref{fig:axes} depict a categorization of certain aspects of the answers regarding naturalness when subjects were queried on how they would have moved through the museum if they were in control. Answers about naturalness seemed to generally reflect these two dimensions, though how naturalness is conceptualized will include significant individual variation. Furthermore, it should be noted that these answers are a expression of what participants though of as natural in a particular context, the virtual museum. Whereas some answers were specific to this setting (\textit{"I would visit one side first and then go to the other side of the museum, slowly exploring. So the first one was more natural"}), others discussed naturalness more generally (\textit{"the walking speed was closet [sic] to my own"}). Although the situation in which these answers were generated was quite distinct, the authors believe that people think of naturalness on these two dimensions in many other cases and situations, which our subsequent work will examine.
\section{Broader Impact}
When attempting to make a robot move "naturally," researchers should take individual variation into account and consider the two dimensions on which it might be perceived. For example, researchers can consider that natural motion in a certain context might not be natural in other contexts, which we found in answers where subjects contrasted how they would move in a museum (\textit{"I think in real life I would move similarly to the 3rd path, because it would allow me to view the items better (closer). Also, it simulates the hesitation of a visitor when going in around a museum"}) versus how they consider themselves to move more generally (\textit{"Walking through a corridor in straight line felt more natural than it was in the first path where the movement was from one corner towards another"}). Researchers can also consider that the naturalness of parts of the path like the speed and how the turns are executed (\textit{"This speed is more natural for a scenario or location such as this. However, the turns were a bit sharp and you feel like you're too close to the objects"}), or the whole path that is taken (\textit{"It felt most like a path that an actual human would take, getting closer to inspect the paintings"}), can each be considered to exhibit naturalness or a lack thereof. All of the individual pieces of a trajectory may be construed as natural or not natural in conflicting ways, so there does not seem to be one universal method to categorize robot motion as natural or not. Therefore, care should be exercised in the third-party classification of the naturalness of robot motion; indeed what a roboticist designing the motion or a reviewer reading the paper might assess as natural may be quite distinct from what a general member of the public may appraise as being natural. Thus the best determination of naturalness is, therefore, attained through the evaluation of research participants.
\section{Future Work}
Future work will will test these same trajectories through the museum, except without any artwork present. Since many people regarded path qualities to be related to their ability to see objects in the museum, the artwork had a significant impact on the subjects' responses regarding preference, comfort, and naturalness. In contrast to the first study with the art, we hypothesize that the turns will still have the largest impact on comfort, and that the path with the least amount of turns will be the preferred path when moving through an empty room. We also postulate that the least turns path will be rated as the most natural because it is piecewise linear. The resulting comparisons will further elucidate how naturalness is conceptualized more broadly and will provide further clarification on user preferences from a first person perspective regarding turn speed, forward speed, trajectories, and comfort for the physical movements of immersive telepresence robots.
\newpage
\bibliographystyle{ACM-Reference-Format}
|
\section*{Acknowledgments}
We would like to thank C. J. Ballance, T. P. Harty, J. P. Gaebler, S. B. Libby, D. M. Lucas, V.~M.~Sch{\"a}fer, and T. R. Tan for helpful discussions. We thank M. Affolter and A.~L.~Collopy for helpful and insightful comments on the manuscript. At the time the work was performed, R.S., S.C.B., H.M.K., A.K., and D.T.C.A. were supported as Associates in the Professional Research Experience Program (PREP) operated jointly by NIST and the University of Colorado Boulder under Award No.\,70NANB18H006 from the U.S. Department of Commerce, National Institute of Standards and Technology.
{\bf Funding:} This work was supported by the NIST Quantum Information Program and ONR.
{\bf Author contributions:} R.S. and H.M.K. carried out the experiments, assisted by S.C.B., D.T.C.A, and D.H.S.; D.H.S., R.S., H.M.K., A.K., and R.T.S. analyzed the data and performed numerical simulations, with support from E.K. and S.G.; D.T.C.A., D.H.S., R.S., S.C.B., and H.M.K. built and maintained the experimental apparatus; R.S. wrote the manuscript with input from all authors; A.C.W., D.L., D.H.S., and D.J.W. secured funding for the work; and D.H.S. and D.T.C.A. supervised the work with support from A.C.W., D.L., S.G., E.K., and D.J.W.
{\bf Competing interests:} The authors declare no competing interests.
\noindent$^{\dagger}$Current address: Department of Physics, University of Oxford, Clarendon Laboratory, Parks Road, Oxford OX1 3PU, U.K.\\
\noindent$^{\xi}$Current address: Department of Physics, Stanford University, Stanford, CA 94305.
\bibliographystyle{Science}
\section*{Introduction}
In this file, we present some tips and sample mark-up to assure your
\LaTeX\ file of the smoothest possible journey from review manuscript
to published {\it Science\/} paper. We focus here particularly on
issues related to style files, citation, and math, tables, and
figures, as those tend to be the biggest sticking points. Please use
the source file for this document, \texttt{scifile.tex}, as a template
for your manuscript, cutting and pasting your content into the file at
the appropriate places.
{\it Science\/}'s publication workflow relies on Microsoft Word. To
translate \LaTeX\ files into Word, we use an intermediate MS-DOS
routine \cite{tth} that converts the \TeX\ source into HTML\@. The
routine is generally robust, but it works best if the source document
is clean \LaTeX\ without a significant freight of local macros or
\texttt{.sty} files. Use of the source file \texttt{scifile.tex} as a
template, and calling {\it only\/} the \texttt{.sty} and \texttt{.bst}
files specifically mentioned here, will generate a manuscript that
should be eminently reviewable, and yet will allow your paper to
proceed quickly into our production flow upon acceptance \cite{use2e}.
\section*{Formatting Citations}
Citations can be handled in one of three ways. The most
straightforward (albeit labor-intensive) would be to hardwire your
citations into your \LaTeX\ source, as you would if you were using an
ordinary word processor. Thus, your code might look something like
this:
\begin{quote}
\begin{verbatim}
However, this record of the solar nebula may have been
partly erased by the complex history of the meteorite
parent bodies, which includes collision-induced shock,
thermal metamorphism, and aqueous alteration
({\it 1, 2, 5--7\/}).
\end{verbatim}
\end{quote}
\noindent Compiled, the last two lines of the code above, of course, would give notecalls in {\it Science\/} style:
\begin{quote}
\ldots thermal metamorphism, and aqueous alteration ({\it 1, 2, 5--7\/}).
\end{quote}
Under the same logic, the author could set up his or her reference list as a simple enumeration,
\begin{quote}
\begin{verbatim}
{\bf References and Notes}
\begin{enumerate}
\item G. Gamow, {\it The Constitution of Atomic Nuclei
and Radioactivity\/} (Oxford Univ. Press, New York, 1931).
\item W. Heisenberg and W. Pauli, {\it Zeitschr.\ f.\
Physik\/} {\bf 56}, 1 (1929).
\end{enumerate}
\end{verbatim}
\end{quote}
\noindent yielding
\begin{quote}
{\bf References and Notes}
\begin{enumerate}
\item G. Gamow, {\it The Constitution of Atomic Nuclei and
Radioactivity\/} (Oxford Univ. Press, New York, 1931).
\item W. Heisenberg and W. Pauli, {\it Zeitschr.\ f.\ Physik} {\bf 56},
1 (1929).
\end{enumerate}
\end{quote}
That's not a solution that's likely to appeal to everyone, however ---
especially not to users of B{\small{IB}}\TeX\ \cite{inclme}. If you
are a B{\small{IB}}\TeX\ user, we suggest that you use the
\texttt{Science.bst} bibliography style file and the
\texttt{scicite.sty} package, both of which are downloadable from our author help site.
{\bf While you can use B{\small{IB}}\TeX\ to generate the reference list, please don't submit
your .bib and .bbl files; instead, paste the generated .bbl file into the .tex file, creating
\texttt{\{thebibliography\}} environment.}
You can also
generate your reference lists directly by using
\texttt{\{thebibliography\}} at the end of your source document; here
again, you may find the \texttt{scicite.sty} file useful.
Whatever you use, be
very careful about how you set up your in-text reference calls and
notecalls. In particular, observe the following requirements:
\begin{enumerate}
\item Please follow the style for references outlined at our author
help site and embodied in recent issues of {\it Science}. Each
citation number should refer to a single reference; please do not
concatenate several references under a single number.
\item The reference numbering continues from the
main text to the Supplementary Materials (e.g. this main
text has references 1-3; the numbering of references in the
Supplementary Materials should start with 4).
\item Please cite your references and notes in text {\it only\/} using
the standard \LaTeX\ \verb+\cite+ command, not another command
driven by outside macros.
\item Please separate multiple citations within a single \verb+\cite+
command using commas only; there should be {\it no space\/}
between reference keynames. That is, if you are citing two
papers whose bibliography keys are \texttt{keyname1} and
\texttt{keyname2}, the in-text cite should read
\verb+\cite{keyname1,keyname2}+, {\it not\/}
\verb+\cite{keyname1, keyname2}+.
\end{enumerate}
\noindent Failure to follow these guidelines could lead
to the omission of the references in an accepted paper when the source
file is translated to Word via HTML.
\section*{Handling Math, Tables, and Figures}
Following are a few things to keep in mind in coding equations,
tables, and figures for submission to {\it Science}.
\paragraph*{In-line math.} The utility that we use for converting
from \LaTeX\ to HTML handles in-line math relatively well. It is best
to avoid using built-up fractions in in-line equations, and going for
the more boring ``slash'' presentation whenever possible --- that is,
for \verb+$a/b$+ (which comes out as $a/b$) rather than
\verb+$\frac{a}{b}$+ (which compiles as $\frac{a}{b}$).
Please do not code arrays or matrices as
in-line math; display them instead. And please keep your coding as
\TeX-y as possible --- avoid using specialized math macro packages
like \texttt{amstex.sty}.
\paragraph*{Tables.} The HTML converter that we use seems to handle
reasonably well simple tables generated using the \LaTeX\
\texttt{\{tabular\}} environment. For very complicated tables, you
may want to consider generating them in a word processing program and
including them as a separate file.
\paragraph*{Figures.} Figure callouts within the text should not be
in the form of \LaTeX\ references, but should simply be typed in ---
that is, \verb+(Fig. 1)+ rather than \verb+\ref{fig1}+. For the
figures themselves, treatment can differ depending on whether the
manuscript is an initial submission or a final revision for acceptance
and publication. For an initial submission and review copy, you can
use the \LaTeX\ \verb+{figure}+ environment and the
\verb+\includegraphics+ command to include your PostScript figures at
the end of the compiled file. For the final revision,
however, the \verb+{figure}+ environment should {\it not\/} be used;
instead, the figure captions themselves should be typed in as regular
text at the end of the source file (an example is included here), and
the figures should be uploaded separately according to the Art
Department's instructions.
\section*{What to Send In}
What you should send to {\it Science\/} will depend on the stage your manuscript is in:
\begin{itemize}
\item {\bf Important:} If you're sending in the initial submission of
your manuscript (that is, the copy for evaluation and peer review),
please send in {\it only\/} a PDF version of the
compiled file (including figures). Please do not send in the \TeX\
source, \texttt{.sty}, \texttt{.bbl}, or other associated files with
your initial submission. (For more information, please see the
instructions at our Web submission site.)
\item When the time comes for you to send in your revised final
manuscript (i.e., after peer review), we require that you include
source files and generated files in your upload. {\bf The .tex file should include
the reference list as an itemized list (see "Formatting citations" for the various options). The bibliography should not be in a separate file.}
Thus, if the
name of your main source document is \texttt{ltxfile.tex}, you
need to include:
\begin{itemize}
\item \texttt{ltxfile.tex}.
\item \texttt{ltxfile.aux}, the auxilliary file generated by the
compilation.
\item A PDF file generated from
\texttt{ltxfile.tex}.
\end{itemize}
\end{itemize}
|
\section{Introduction}
\label{sec:introduction}
Online social networking platforms place very few constraints on the type and structure of the textual content posted online. The unstructured nature of these content results in the creation of textual content which could be far from the original grammatical, syntactic, or semantic rules of the language they are written in \cite{arora2019character, gautam2014sentiment}. One deviation that has been observed quite often is the use of words from more than one language in the text \cite{barik2019normalization}. Commonly known as “code-mixed” text. These texts are written in language A but include one or more words of language B (either written in the official alphabets of language B or transliterated to language A). \\
In this study, we investigate code-mixed Persian-English data and perform sentiment analysis on these texts. The reason why sentiment analysis of these texts would differ from a text written purely in the Persian language is that the words containing the emotional energies of the text could be written in English which would make the Persian-only sentiment analysis models unable to produce the correct outputs. Grammatical differences could also render such monolingual models useless \cite{xu2015proceedings}. The difficulties of this task have been shown in other language pairs \cite{mandal2018preparing, banerjee2020limsi_upv}. Our reasoning behind choosing English as the second language is the prominence of the usage of this language overall but more importantly among Persian speakers.\\
Since Persian is a low resource language, in order to perform the aforementioned task, we first needed to create a dataset of texts and label those texts with the correct sentiment scores. Thus, we begin by using the Twitter API and searching for a list of “Finglish” \footnote{English words that are written using the Persian/Farsi alphabet} words through the use of the API. After the data has been collected, 2 annotators labeled the 3640 tweets completely. A third annotator was then added to the project to label the tweets on which the previous two annotators did not agree on. \\
After the dataset collection and creation was completed, an ensemble model was created and used to detect the sentiment scores of the texts.\\
The rest of this paper is structured as follows: Section \ref{sec:related_work} provides a brief overview of related work. Next, we look at our dataset in detail in Section \ref{sec:data}. Our models, as well as our text cleaning and preparation steps are described in Section \ref{sec:methods}. We then report our results in Section \ref{sec:results} and the study is concluded in Section \ref{sec:conclusion}.
\section{Related Work}
\label{sec:related_work}
The prevalence of code-mixed textual data, due to the unstructured and uncontrolled nature of the web as well as social networks, have resulted in a focus on the topic throughout recent years, including multiple shared-tasks being defined on the subject \cite{patra2018sentiment, solorio2014overview, patwa2020sentimix, patwa2020semeval}. \\
One of the language pairs that has been the center of attention in code-mixed text analysis is Hindi-English \cite{srivastava2020iit, joshi2016towards, prabhu2016towards, sitaram2015sentiment, rudra2016understanding, jhanwar2018ensemble}. With the large population of multilingual individuals in India, such forms of texts have become quite common. Other language pairs (such as Bengali-English, Bambara-French, and Tamil-English), however, have been studied as well \cite{mandal2018preparing, konate2018sentiment, chakravarthi2020sentiment, chakravarthi2020corpus}.\\
Some studies attempt to solve the issue by hand engineering features which help in the task. In \cite{ghosh2017sentiment} various features (e.g. the number of code switches in the text) were introduced and employed in a multi-layer perceptron model. Number of sentiment and opinion lexicon, number of uppercase words, and POS tags are among some of the features which have been utilized. \\
Other studies try to find methods with less need for feature selection. For instance, it has been attempted to use cross-lingual word-embeddings \cite{singh2020sentiment} or subword embeddings \cite{prabhu2016towards} to accomplish the task. In the SemEval-2020 task on Spanglish and Hinglish \cite{patwa2020semeval} it is reported that BERT and ensemble methods are the most common and successful approaches. In \cite{singh2018automatic} an approach is introduced to help deal with different variations of the same word by substituting words with consideration to their context words. In \cite{aguilar2020lince} a benchmark for linguistic code-switching is presented, the aim of which is to enable evaluation of models. \\
To the best of our knowledge, the dataset annotated and presented as part of this study is the first dataset for the Code-mixed Persian-English sentiment analysis task. We also believe that there are no other studies on this specific subset of the topic available.
\section{Data}
\label{sec:data}
\setcode{utf8}
In this section, we describe the distributions and characteristics of our data. As described in Section \ref{sec:introduction}, our dataset consists of 3,640 tweets labeled with polarity values. Our dataset fields include the terms that were searched via the API that resulted in the tweets' retrieval, the text content of the tweets, the three labels assigned to each tweet, and the final label, which is calculated through majority voting.\\
We selected a list of 44 unique Finglish (English words transliterated to Persian) terms in order to collect this data. Some of the words in the list include: \<پرفکت> (perfect),
\<هپی> (happy), \<بورينگ> (boring), and \<سينگل> (single). \\
In our dataset, 69.2\% of the instances received unanimous labels by the first two annotators. A third annotator was then asked to label the rest of the data. In the resulting dataset, 15.7\% were labeled as positive and 59.7\% as negative (and the remaining 21.5\% were labeled as neutral). Table \ref{example-annotations} displays two examples of annotated data in our dataset. The majority of the data being negative can be explained by two facts: One is that due to the access restrictions of Twitter in Iran, only 9.24\% of Iranians use Twitter \cite{social_penetration:2020} and as a result the subset of users on the platform are mostly from a particular belief system \cite{khazraee2019mapping} which could result in the observed negative opinions. Another reason is that the data collection process was conducted in the last months of 2019 which included the beginning stages of the spread of the Coronavirus in Iran. Even though our keywords did not relate to the Coronavirus in any way the shift in spirit was observed in our data. We however believe that this issue reflects the current state of our society and thus the dataset can still be used for the task of code-mixed sentiment detection as the tweets do include the characteristics of code-mixed language and there are enough examples in the dataset to allow the model to learn attributes of polarities other than negative.\\
To preserve the privacy of the users, all user mentions in the texts have been replaced by @USERMENTION. However, since all tweets were public (at least at the time of collection) and were collected using the official Twitter API, Tweet IDs have been provided to allow use in future research should the need arise. This dataset is publicly available on GitHub. \footnote{\url{https://github.com/nazaninsbr/Persian-English-Code-mixed-Sentiment-Analysis}}
\begin{table}
\footnotesize
\centering
\caption{Example annotated tweets from our dataset}
\begin{tabular}{cc}
\hline
\textbf{Tweet} & \textbf{Label} \\ \hline
@USERMENTION \< زندگی خوشگل نيست، بورينگ و بی معنيه> & Negative\\
\<اوهوم هپی نيو آواتار و ازين حرفا بانو جان> & Positive\\
\hline
\end{tabular}
\label{example-annotations}
\end{table}
\section{Method}
\label{sec:methods}
In this section we will go over our text processing and feature extraction, and data representation methods as well as our model. \\
In the text processing step we aim to create a vectorized representation of the textual input in order to be able to fit the data into our machine learning model. To do so we take the following steps:
\begin{enumerate}
\item Finding the non-Persian words in the sentence: By the definition of our task we know that the text we are processing includes non-Persian words, however, we do not know which words in the sentence they are. The reason why this knowledge could be useful is that knowing which words they are would allow us to use methods such as translation to convert them to their original language. In order to find these words, we use a dataset of Persian words collected from Wikipedia \footnote{\url{https://github.com/behnam/persian-words-frequency/blob/master/persian-wikipedia.txt}}. The huge collection of articles available on Wikipedia ensures that the most frequent words of the language would be in the list. We then check the existence of every word in the list and if it is not in the list we add it to our non-Persian word candidates.
\item Translation: Next we translate the non-Persian words. First, we use an automatic tool, Yandex \footnote{\url{https://translate.yandex.com}}. This tool, however, faces difficulties when asked to translate Twitter specific slangs or expressions. Thus, we use common Twitter expression lists \footnote{\url{bit.ly/3h2qyNm}} and create a dictionary of our own.
\item Embedding creation: To create an embedding for our textual data we used the pretrained multilingual BERT \cite{devlin2018bert} model that Google has provided on their GitHub page \footnote{\url{https://github.com/google-research/bert}}. Since the model is multilingual, it would allow for creation of embeddings for words from more than one language which fits nicely with our problem since code-mixed data includes instances of more than one language. Further since the model uses the idea of subword tokenization and embedding, it could allow for a better understanding of slangs or other non-common words that could be made up of better known subwords.
\end{enumerate}
After our data passes these steps, it is fed into an ensemble model consisting of three Bidirectional Long Short-Term Memory (Bi-LSTM) networks.
\begin{itemize}
\item Our first network is a stacked Bi-LSTM network. The model is quite simple and aims to encode the general information available in the sentences.
\item Our second model adds the attention mechanism to the Bi-LSTM network which enables the model to pay more attention to the most important words in the sentence. The attention mechanism also helps with the encoding of long-distance dependencies and information.
\item Another method we use to make sure long-distance dependencies are accounted for and that information is not lost in our model is the use of pooling layers in our final model.
\end{itemize}
The final model takes the outputs of all three models and uses a weighted average to produce the output. To find the best weight assigned to the output of each model, we use the optimization algorithms offered by SciPy \cite{2020SciPy-NMeth}.
\section{Results}
\label{sec:results}
We used 10-fold cross-validation and averaged the metrics in order to present more reliable values. Additionally, to be able to compare our results, naive Bayes and random forest models were also used on the data. The results have been presented in Table~\ref{results-table}.
\begin{table}
\footnotesize
\centering
\caption{10-fold cross-validation model performance results: each reported metric is calculated by averaging the values across all 10 folds}
\begin{tabular}{ccccc}
\textbf{Model Name} & \textbf{Accuracy} & \textbf{Precision} & \textbf{Recall} & \textbf{F1}\\ \hline
Naive Bayes & 60.88 & 69.00 & 60.88 & 47.17\\
Random Forest & 62.12 & 60.37 & 62.12 & 57.67\\
Our Model & 66.17 & 64.16 & 65.99 & 63.66\\
\hline
\end{tabular}
\label{results-table}
\end{table}
We can see that our ensemble model outperforms the baseline models with regards to all metrics. Through our experiments, we find that the attention and pooling mechanisms both help with the performance of the models. We further find that the sum of all three models, offers better performance than each individual model, as each model appears to make up for the shortcomings of the other models in the ensemble.
\section{Conclusion}
\label{sec:conclusion}
In this study we presented a Persian-English code-mixed dataset. The dataset consisted of 3640 tweets collected through the use of the Twitter API. Each tweet was consequently labeled with its corresponding polarity score. We then used neural classification models to learn the polarity scores of these data. Our models employed Yandex and dictionary-based translation techniques to translate the code-mixed words in our texts. We further used pretrained BERT embeddings to represent our data. Our models reached an accuracy of 66.17\% and F1 of 63.66 on the data. \\
Future work could focus on other methods of dealing with the code-mixed words or ways in which we could find word-based polarity scores for our code-mixed words using sentence level scores.
\bibliographystyle{unsrt}
|
\section{Introduction}
\IEEEPARstart{T}{he} outbreak of novel coronavirus pneumonia (COVID-19) is affecting the entire world, which has caused a large number of deaths with an increase in the spread of COVID-19.
To control the spread of COVID-19 at the early stage, oropharyngeal-swab (OP-swab) sampling is commonly adopted with respect to sample collection and specimen sources for diagnosis \cite{wang2020comparison}.
\begin{figure}[htp!]
\centering{ \includegraphics[width=1.00\linewidth, height=0.84 \linewidth]{figure/Fig_all_new.jpg}}
\caption{ 9-degree-of-freedom (DOF) redundant rigid–flexible coupling robot }
\label{fig.9DOFs_RFC_Robot}
\end{figure}
However, protecting the safety of medical staff during the sampling process raises a new challenge because of susceptibility to infection from person to person through respiratory droplets and contact transmission in an unprotected environment \cite{xu2020open}.
A variety of related studies have reported that respiratory droplets, feces and urine are the routes of transmission \cite{hindson2020covid}.
To address these issues in OP-swab sampling, robotics could play a key role in disease prevention.
\begin{figure*}[htp!]
\centering{
\includegraphics[width=0.99\linewidth]{figure/block_diagram.png}}
\caption{Control framework of COVID-19 OP-swab sampling with a redundant rigid–flexible coupling robot}
\label{fig.block diagram}
\end{figure*}
Considering the high risk of infection of COVID-19 and negative tests of nasopharyngeal swabs (NP-swab) caused by irregular sampling, it is necessary to design an OP-swab sampling robot to assist healthcare staff through remote access.
Robot-assisted OP-swab sampling is a promising technique because it relieves the burden from medical staff, is convenient for large-scale deployment, is cost-effective, and offers sampling standardization.
As reported in \cite{yang2020combating}, an OP-swab robot could speed up the sampling process especially because there is a lack of qualified healthcare workers.
In \cite{li2020clinical} and \cite{liu2020},
a semi-automatic OP-swab robot (compared to the traditional human method) was developed with a teleoperation scheme that achieved good performance in clinical practice with a 95\% success rate.
The OP-swab robot system mainly includes a snake-like manipulator, two
haptic devices(Omega.3 and Omega.6) for teleoperation, an endoscope for visualization (assistance operating in the oral cavity) and a force-sensing system for safety protection.
All OP-swab sampling processes were controlled by experts' teleoperation, where one haptic device (Omega.3) control the tongue depressor and another haptic device (Omega.6) operate the manipulator.
Wang {\it{et al.}}~ \cite{wang2020design} have designed a low-cost robot for assistance in sampling of NP-swab; a swab gripper is attached to the rotation link with its extruded active 2-degree-of-freedom (DOF) end-effector for actuating the swab and a generic 6-DOF passive arm for global positioning.
In \cite{Danmark2020}, the team from the University of Southern Denmark and the Lifeline Robotics company developed the first automatic swab robot for COVID-19 sampling.
In \cite{Danmark2020}, the robot is integrated with a UR5 manipulator, a visual system and dexterous rotatable rigid connectors assembled with OP-swab which could perform the sampling task quickly using an automated scheme.
Of note, medical security remains vital to OP-swab sampling because human throat is fragile and easily injured.
Nonetheless, although many desirable results were obtained, the abovementioned end-effectors were designed on the basis of a rigid body structure, which may cause physical injuries of the oral cavity during improper operation or other medical malpractices.
On the basis of our previous work \cite{li2015novel}, a novel micro-pneumatic actuator (MPA) for throat swab sampling is developed to achieve flexible collection and integrate a force sensor, which provides safe, stable, and reliable sampling experience, as shown in Fig. \ref{fig.9DOFs_RFC_Robot}.
Furthermore, force-sensing actuation offers compliance, which is helpful
against shocks, particularly during interaction with oral cavity.
Compared with the rigid body-based sampling robots \cite{li2020clinical, wang2020design} and \cite{Danmark2020},
the developed MPA is made of soft material, which is safer and lighter.
Moreover, MPA is smaller and has a 7.5 mm cross-sectional diameter, which is convenient for working in human oral cavity.
To avoid contact with the OP-swab specimen and collision with the manipulator, it is essential to consider the constraints of the oral cavity space during motion planning.
Inspired by our previous work \cite{su2020improved} on remote center of motion (RCM) techniques in minimally invasive surgical robots, the second to the last link of the OP-swab sampling robot is constrained with the oral cavity center (OCC) constraint, which guarantees the absence of collision between the inserted end-effector with the oral cavity.
Although mechanical implementation is usually safer for an open minimally invasive surgery, programmable RCM is cost effective and convenient to be implemented
\cite{su2019improved}.
The traditional Cartesian adaptive control in \cite{su2019improved} was applied for surgical control with the RCM constraint but without considering the physical limits of the manipulator, such as the joint angle and velocity.
For practical applications, we propose an optimization strategy for the multi-constraint problem, where the joint angle and velocity, and the OCC constraint-based kinematic model are derived as a quadratic programming problem.
In this study, a comparison between the varying-parameter zeroing neural network (VP-ZNN) and the traditional gradient descent method \cite{mirrazavi2018unified} is proposed for the optimization problem that has achieved superior performance, which can converge at the finite time owing to the optimal solution changing over time.
To test the developed OP-swab sampling robot, which is shown in Fig.~\ref{fig.9DOFs_RFC_Robot}, the experiments are simultaneously conducted in both the human oral cavity phantom and volunteers, regarding the sampling time, operation complexity for medical staff, safety and effectiveness. The framework of COVID-19 OP-swab sampling using the RFC robot is shown in Fig.~\ref{fig.block diagram}.
The contributions of this study can be summarized as follows.
\begin{enumerate}
\item A rigid–flexible coupling robot is designed to assist COVID-19 OP-swab sampling, where the MPA for OP-swab sampling is developed to achieve flexible sampling and thereby provides compliant, safe, stable, and reliable sampling.
\item An improved motion planning method is proposed for the 9-DOF highly redundant robot-assisted OP-swab sampling system, which guarantees efficiency and accuracy, where the multi-constraint kinematics model is derived as an optimization problem; then, the VP-ZNN is applied to solve the optimization problem.
\item Substantive practical experiment testing of the developed OP-swab sampling robot system is demonstrated in the human oral cavity phantom and volunteers.
\end{enumerate}
\section{Design of the Rigid–Flexible Coupling Robot}
\label{sec.design}
\subsection{ Redundant Rigid–Flexible Coupling Robot}
The rigid–flexible coupling robot for OP-swab sampling consists of a UR5 manipulator and a self-developed rigid–flexible coupling manipulator. The overall structure of the 9-DOF redundant rigid–flexible coupling robot is shown in Fig.~\ref{fig.9DOFs_RFC_Robot}. The self-developed manipulator has 3 DOFs, including a linear motor (prismatic joint), a servo–$\alpha$ motor (revolute joint), and an MPA that can change the offset distance of throat swab from the center.
\begin{figure}[htp!]
\centering{
\includegraphics[width=1.0\linewidth, height=0.99 \linewidth]{figure/Bending_of_fingers.jpg}}
\caption{(a) Cross section of the micro-pneumatic actuator(MPA). (b) bending process of MPA. and (c) 2D modeling of MPA.}
\label{fig.Bending_of_fingers}
\end{figure}
\begin{figure
\flushleft
\subfigure[]{
\includegraphics[width=2.8cm,height=4cm]{figure/bengding.png}
}
\quad
\hspace{-8mm}
\subfigure[]{
\includegraphics[width=5.4cm,height=4cm]{figure/bengding_ange.png}
}
\caption{
Experimental results of MPA. (a) Deformation of MPA with a 50-shore hardness under different air pressures. (b) Relationship between bending angle $\gamma$ and air pressure}
\label{fig.cali}
\end{figure}
\subsection{Design, Modeling, and Fabrication of MPA}
The design of MPA refers to the wrinkle shape of the elastomer robot \cite{wang2017prestressed}. Figure \ref{fig.Bending_of_fingers} shows that MPA consists of $ n $ effective air chambers. The sectional view of MPA is shown in Fig.~\ref{fig.Bending_of_fingers}(a), which contains a chamber part and a cover part. These two parts are bonded together to form a single-DOF MPA. $ b $ is the wall thickness of the air chamber; $ c $ is the distance between the air chambers, and $ l_f $ is the total length of MPA. When MPA is inflated, the air chambers will expand and repel each other, which causes MPA to bend. The bending process of MPA is shown in Fig.~\ref{fig.Bending_of_fingers}(b).
Assuming that all air chambers have the same bending under the same pressure, it is easy to obtain the geometric relation:
\begin{equation}
\label{equ-gama}
\theta {\rm{ = }}\beta /n {\rm{ = 2}}\gamma /n
\end{equation}
Owing to the small value of $\theta$ after bending, the chord length is approximately equal to the arc length at the non-tensile layer. The distance between the center line and the non-tensile layer is $ m $, and the arc length $ c' $ is also approximately equal to the chord length; thus, $ c' $ is
\begin{equation}
\label{equ1}
c' = c + 2m\sin (\theta /2)
\end{equation}
The overall motion model is shown in Fig.~\ref{fig.Bending_of_fingers} (c), where the chord length $ d $ is
\begin{equation}
\label{equ2}
d = \frac{{c'\sin ({n}\theta /2)}}{{\sin (\theta /2)}}
\end{equation}
The forward kinematics of MPA is modeled as
\begin{equation}
\left\{ {\begin{array}{*{20}{c}}
{{x} = d\sin \left( {n\frac{\theta }{2}} \right){\rm{ + }}h\sin \left( \beta \right)}\\
~\\{{y} = d\cos \left( {n\frac{\theta }{2}} \right){\rm{ + }}h\cos \left( \beta \right)}
\end{array}} \right.
\label{equ3}
\end{equation}
According to the abovementioned information, we design MPA with 12 air chambers.
However, the number of effective air chambers, i.e., the number of effective joints is $ n=11 $ because only half of the head chamber and tail chamber will bend. The wall thickness $ b $ of each chamber is 1.5 mm , the width of each air chamber $ c $ is 4.5 mm and the total length $ l_f $ is 80 mm.
The chamber part and the cover part of MPA are made by 3D printing with a 50-shore hardness.
The air pressure is controlled by the proportional valve ITV1050-312L. Then, a deformation experiment of MPA was conducted, and the experimental results are shown in Fig.~\ref{fig.cali}(a), where the gravity is directly perpendicularly to the plane of the paper.
The relationship between the bending angle $\gamma$ and air pressure is expressed in Fig.~\ref{fig.cali}(b).
It is observed that the bending angle $\gamma$ is approximately proportional to the air pressure.
After data fitting, the linear equation for the angle $\gamma$ ($^{\circ}$) and pressure $P$ (kPa) is written as
\begin{equation}
\label{equ5}
\gamma {\rm{ = 0}}.160325P{\rm{ - 0}}.731885
\end{equation}
The correlation coefficient $ R $ is 0.99990 and the standard deviation $ S $ is 0.6464, which confirms that $\gamma$ and $P$ are linearly dependent. By combining equations (\ref{equ-gama}) and (\ref{equ5}), we obtain
\begin{equation}
\label{equ6}
\theta {\rm{ = 0}}.02915P{\rm{ - 0}}.13307
\end{equation}
Finally, the end-point position of MPA can be calculated by combining equations (\ref{equ1}), (\ref{equ2}), (\ref{equ3}), and (\ref{equ6}).
\begin{figure}[htp!]
\centering{
\includegraphics[width=0.99\linewidth]{figure/The_relationship_between_DeltaU_and_F.png}}
\caption{Relationship between $ F $ and $ \Delta U $ }
\label{fig.relationshipUF}
\end{figure}
\subsection{Force Sensing}
To detect the contact force $ F $ during OP-swab sampling, a strain gauge sensor (BF350-3AA23T0) attached to MPA is adopted. It is necessary to explore the mapping relationship between the force and voltage for force measurement calibration.
Actually, the current output voltage $ U_{now} $ is related to two factors, i.e., voltage $ \Delta U $ generated by the force, and voltage $ {U_{F{\rm{ = 0}}}} $ caused by the air pressure without a load.
Therefore, we have $\Delta U{\rm{ = }}{U_{{\rm{now}}}} - {U_{F{\rm{ = 0}}}}$. After 20 group experiments, the relationship between no-load voltage and air pressure is obtained by linear fitting: $ {U_{F = 0}}{\rm{ = }} - 10.693P + 2770.571 $ (mV). Force calibration was performed using the electronic balance (CX-668). The relationship between $ \Delta U $ and $ F $ is approximately linear; thus, we know that \begin{equation}
F = {\rm{0}}{\rm{.04378}}\Delta U + {\rm{0}}{\rm{.88828}}
\end{equation}
The details are shown in Fig.~\ref{fig.relationshipUF}.
The correlation coefficient $ R $ is 0.99461 and the standard deviation $ S $ is 2.65733.
\section{Optimization Control for the 9-DOF Redundant Manipulator}
\label{sec.control}
\subsection{Manipulator Kinematics Model with OCC Constraints}
The coordinate system of the 9-DOF redundant RFC manipulator is shown in Fig.~\ref{fig.coordinatesystem}. The forward kinematics model \cite{li2017asymmetric, li2018adaptive} of the 9-DOF RFC robot can be defined as follows:
\begin{equation}
{J}\dot q = {\dot l_{d}}
\label{eq.Jacobian}
\end{equation}
where $J \in \mathbb{R}^{m \times n}$ denotes the Jacobian matrix.
The joint angle and velocity constraints are defined as follows:
\begin{alignat}{2}
{q_i ^ - } \le q_i \le {q_i ^ + } \label{eq.limit0}\\
{{\dot q_i }^ - } \le \dot q_i \le {{\dot q_i }^ + }
\label{eq.limit1}
\end{alignat}
where $q_i^-$ and $q_i^ +$ present the lower and upper bounds of $q_i$, respectively; and $\dot q_i ^ -$ and $\dot q_i ^ +$ denote the lower and upper bounds of the joint velocity $\dot q$, respectively.
\begin{figure}[htp!]
\centering{
\includegraphics[width=0.99\linewidth]{figure/DH.png}}
\caption{Coordinate system of the 9-DOF redundant RFC manipulator.}
\label{fig.coordinatesystem}
\end{figure}
Actually, the constraints in (\ref{eq.limit0}) can be converted as
\begin{equation}\label{eq.limit2}
\sigma \left( {{q_i ^ - } - q_i } \right) \le \dot q_i \le \sigma \left( {{q_i ^ + } - q_i } \right)
\end{equation}
where $\sigma$ is the positive constant.
Therefore, according to (\ref{eq.limit1}) and (\ref{eq.limit2}), the joint angle and velocity constraints can be rewritten as a new constraint in the velocity level:
\begin{alignat}{2}
&{\rho_i ^ - } \le \dot q_i \le {\rho_i ^ + } \label{eq.limit3}\\
&\rho _i^ - = \max \left\{ {\dot q _i^ - ,\sigma \left( {q _i^ - - {q _i}} \right)} \right\} \nonumber \\
&{\rho_i ^ + } = \min \left\{ {\dot q _i^ + ,\sigma \left( {q _i^ + - {q _i}} \right)} \right\} \nonumber
\end{alignat}
Because the RFC robot is a 9-DOF highly redundant manipulator, there are infinite solutions of $\dot q$ in (\ref{eq.Jacobian}) by the inverse kinematics.
However, the convergence rate, accuracy, and computational complexity of the pseudoinverse-type solution in inverse kinematics cannot satisfy the requirements.
Consequently, we need to identify an optimization solution under multiple constraints; thus, the inverse kinematics problem can be expressed as a new optimization problem.
The first-priority optimization problem associated with OP-swab sampling is expressed as follows:
\begin{alignat}{2}
\text{min}~~~&\frac{1}{2}{\dot{q} ^T}W\dot{q} \label{eq.qpmin1} \\
\text{s.t.}~~&{J\left(q\right)}\dot{q} = {\dot{l}_{d}}\label{eq.qpst1} \\
&{\rho ^ - }\le {\dot{q }} \le &{\rho ^ + }\label{eq.qpstq1}
\end{alignat}
where $\dot{l}_{d}$ represents the reference velocity associated with OP-swab sampling tasks. The weight matrix $M$ is set as the identity matrix.
Considering the OCC constraint, the link-$L_{n-1}$ of the RFC manipulator needs to pass through the OCC, and $L_n$ performs the sampling tasks, which is different from the traditional RCM constraint in the last link. The geometric relationship is shown in Fig.~\ref{fig.constraintRCM}.
For the $n$-DOF OP-swab sampling robot, the forward kinematics mapping function of Cartesian position $ l_{n-2} \in \mathbb{R}^m$ and $l_{n-1} \in \mathbb{R}^m$ can be defined as follows:
\begin{equation}
\begin{aligned}
&{{l}_{n-2}=f_{n-2}( q)} \\
&{{l}_{n-1}=f_{n-1}( q)}
\end{aligned}
\label {eq2}
\end{equation}
\begin{figure}[htp!]
\centering{
\includegraphics[width=0.9\linewidth]{figure/RCM.png}}
\caption{Constraint with the oral cavity center.}
\label{fig.constraintRCM}
\end{figure}
Unlike common RCM constraints,
$l_{occ}$ should be always on the straight line straight line between $l_{n-1}$ and $l_{n-2}$ (second to the last link), where $l_{n-1}$ is the end position of the link $L_{n-1}$ and $l_{n-2}$ is the end position of the link $L_{n-2}$. During the actual OP-swab sampling, we want to keep the error of the OCC constraint $E_{occ}$ as small as possible. Lines 1 and 2 are constructed as follows: ${\overrightarrow{l_{n-2}r_{n-1}}}=l_{n-1}-l_{n-2}$, ${\overrightarrow{l_{n-2} l_{occ}}}=l_{occ}-l_{n-2}$, respectively.
Utilizing the relationship between $E_{occ}$ and the vector projection, $E_{occ}$ can be further written as follows:
\begin{alignat}{2}
{E_{occ}} &= \frac{{\overrightarrow {{l_{n-2}}{l_{occ}}} \times \overrightarrow {{l_{n-2}}{l_{n-1}}} }}{L}
\label{eq.OCCerror}
\end{alignat}
where $L=\left\| {{l_{n-2}} - {l_{n-1}}} \right\|$ is the length of the second to the last link.
The derivative of $E_{occ}$ in (\ref{eq.OCCerror}) with respect to time is reformulated as follows:
\begin{equation}\label{eq.OCC_constrant}
J_{occ}\dot q= \dot{E}_{occ}
\end{equation}
where $J_{occ} \in \mathbb{R}^{m \times n}$ denotes the Jacobian matrix of the OCC constraint.
Regarding the OCC constraint task, we want to keep the OCC error $E_{occ}$ at the minimum value.
In this section, it is necessary to find a feasible solution for multiple constraints and guarantee that the tracking error and OCC error always remain in a small range. For COVID-19 sampling tasks, we aim to reformulate OCC (\ref{eq.OCC_constrant}) and joint physical limits (\ref{eq.qpstq1}) in an optimization scheme and design a method to solve the optimization problem. Consequently, by simultaneously taking OCC, end-effector task, and joint physical constraints into account, the new multi-task optimization problem can be formulated as
\begin{alignat}{2}
\text{min} ~~~&\frac{1}{2}{{\dot q }^T}W\dot q \label{eq.multi-opt1}\\
\text{s.t.} ~~~&J\dot q = v_d\nonumber \\
&{J_{occ}}\dot q = v_{occ} \nonumber \\
&{\rho ^ - } \le \dot q \le {\rho ^ + } \nonumber
\end{alignat}
where $v_d={{\dot l}_{nd}}$ and $v_{occ}=0$.
Without the explicit expression of $l_n$ and $e_{occ}$, the actual trajectory will drift and the position error cannot converge to zero from a random initial position.
To overcome this issue in (\ref{eq.multi-opt1}), a feedback item associated with position signals is integrated into $v_d$ and $v_{occ}$:
\begin{alignat}{2}
&v_d = - {k_1}\left( {{f_n}\left( q \right) - {l_{nd}}} \right) + {{\dot l}_{nd}}\\
&v_{occ} = - {k_2}\left( {{l_{occ}}} \right)
\end{alignat}
Of note, we should manage the priority strategy for the multiple-task optimization problem by different weights. Therefore, the objective function is defined as
\begin{equation}
F(\dot q)= \frac{{{\varepsilon_0}}}{2}{{\dot q }^T}\dot q + \frac{{{\varepsilon_1}}}{2}{\left\| {J\dot q - v_d} \right\|^2} + \frac{{{\varepsilon_2}}}{2}{\left\| {{J_{occ}}\dot q-v_{occ} } \right\|^2} \label{eq.multi-opt2}
\end{equation}
The optimization problem in (\ref{eq.multi-opt2}) can be rewritten as follows:
\begin{alignat}{2}
\text{min} ~~~&F(\dot q) \\
\text{s.t.} ~~~&J\dot q = v_d \nonumber \\
&{J_{occ}}\dot q = v_{occ} \nonumber\\
&{\rho ^ - } \le \dot q \le {\rho ^ + }\nonumber
\end{alignat}
where $\varepsilon_0>0$, $\varepsilon_1>0$ and $\varepsilon_2>0$ are the constants employed to prioritize different tasks.
\subsection{Neural Network Design}\label{sec.VP-ZNN}
In this section, we describe the design of the VP-ZNN, which is used to solve the optimization problem in (\ref{eq.multi-opt2}). First, the multiple-task optimization problem in (\ref{eq.multi-opt2}) is converted into an equivalent problem; thus, the VP-ZNN is employed to solve it.
The Lagrange function of (\ref{eq.multi-opt2}) constraints is formulated as follows:
\begin{alignat}{2}
&\mathcal{L}\left( {\dot q ,{\xi _1},{\xi _2}} \right) = \frac{{{\varepsilon_1}}}{2}{\left\| {J\dot q - v_d} \right\|^2} + \frac{{{\varepsilon_2}}}{2}{\left\| {{J_{occ}}\dot q-v_{occ} } \right\|^2} \nonumber \\
& ~~~~~\frac{{{\varepsilon_0}}}{2}{{\dot q }^T}\dot q+ \xi _1^T\left( {{v_d} - J\dot q } \right) + \xi _2^T\left(v_{occ} { - {J_{occ}}\dot q } \right)
\label{eq.Lagrange}
\end{alignat}
where $\xi_1 \in \mathbb{R}^{m}$ and $\xi_2 \in \mathbb{R}^{m}$; $\nabla \mathcal{L} = {\left[ {\frac{{\partial \mathcal{L}}}{{\partial \dot q }},\frac{{\partial \mathcal{L}}}{{\partial {\xi _1}}},\frac{{\partial \mathcal{L}}}{{\partial {\xi_2}}}} \right]^T}$ indicates the gradient of (\ref{eq.Lagrange}).
As the KKT condition defined in \cite{ facchinei2014solving}, if $\nabla\mathcal{ L} $ is continuous, the optimization solution satisfies the following condition:
\begin{equation}\label{eq.Lagrangecondition}
\nabla \mathcal{L} = 0
\end{equation}
The state decision variable $l\left( t \right) = {\left[ {{{\dot q}},{{{\xi }}_1},{{{\xi }}_2}} \right]^T} \in R^{n+2m}$.
The problem in (\ref{eq.Lagrangecondition}) is equivalent to the following:
\begin{alignat}{2}
B(t)l(t)=P(t)
\end{alignat}
where $B\in \mathbb{R}^{(n+2m)\times (n+2m)}$ and $P(t)\in \mathbb{R}^{(n+2m)}$.
The bounds of state variable $\dot q^b$ are expressed as
\[\dot q^b= \left\{ \begin{array}{l}
{\rho ^ + },~~\dot q > {\rho ^ + }\\
\rho ,~~~{\rho ^ - } \le \dot q \le {\rho ^ + }\\
{\rho ^ - },~~\dot q < {\rho ^ - }
\end{array} \right.\]
Therefore, the bound of state variable $l(t)$ are defined as
\[{l^b}\left( t \right) = \left[ {\begin{array}{*{20}{c}}
{{{\dot q}^b}}&{\xi _1^b}&{\xi _2^b}
\end{array}} \right],~\xi _1^b,\xi _2^b \in \mathbb{R}\]
The error model of the novel VP-ZNN is expressed as
\begin{alignat}{2}
e\left( t \right) = B\left( t \right)l\left( t \right) - P\left( t \right)
\label{eq.novel VP-ZNNerror}
\end{alignat}
To ensure the model error convergence to zero, we define the following formulation,
\begin{alignat}{2}
\label{eq.errorconvergence}
&\frac{{de\left( t \right)}}{{dt}} = - \mu \exp \left( t \right)\Psi \left( {e\left( t \right)} \right)\\
&\Psi \left( {e\left( t \right)} \right) = \frac{{\left( {1 + \exp ( - \delta )} \right)\left( {1 - \exp ( - \delta {e_i}(t))} \right)}}{{\left( {1 - \exp ( - \delta )} \right)\left( {1 + \exp ( - \delta {e_i}(t))} \right)}} \nonumber
\end{alignat}
where $\mu>0$ is the constant that can adjust the convergence rate; ${{\Psi }\left( {{e_i}(t)} \right)}$ denotes the activation function, and $\xi \ge 2$, which makes $0 \le \left| {{e_i}(t)} \right| \le 1$1. Clearly, the error in (\ref{eq.errorconvergence}) converges to zero with exponential convergence.
The function in (\ref{eq.errorconvergence}) is expanded as
\begin{alignat}{2}
&B\left( t \right)\dot l \left( t \right) = - \dot B\left( t \right)l\left( t \right) + \dot P(t)\nonumber\\
&- \mu \exp \left( t \right) {\Psi }\left(l(t)-l^b(t)+{B(t)l(t) - P(t)} \right)
\label{eq.novel VP-ZNN1}
\end{alignat}
We further modify the VP-ZNN in (\ref{eq.novel VP-ZNN1}) as
\begin{alignat}{2}
\dot l\left( t \right) &= \left( {I - B\left( t \right)} \right)\dot l\left( t \right)-\dot{B}(t)l(t) + \dot P\left( t \right)\nonumber \\
&- \mu \exp \left( t \right) {\Psi }\left( l(t)-l^b(t)+{B\left( t \right)l\left( t \right) - P\left( t \right)} \right)
\end{alignat}
For comparison, the traditional gradient descent-based recurrent neural network is denoted as
\begin{alignat}{2}
\dot l\left( t \right) = \mu \left( { - l(t) + {P_\Omega }\left[ {l(t) - \left( {B(t)l(t) - P(t)} \right)} \right]} \right)
\end{alignat}
\begin{figure}[htp!]
\centering{
\includegraphics[width=0.99\linewidth]{figure/phantom.png}}
\caption{Demonstration with an oral cavity phantom. In our experiments, the 1:1 human oral cavity is tested.}
\label{fig.phantom}
\end{figure}
\begin{figure*}[htp!]
\centering{
\includegraphics[width=0.7\linewidth, height=0.4 \linewidth]{figure/Fig10-one.jpg}}
\caption{
Sampling process from vision detection to sampling tasks. In our experiments, many volunteers were tested, and the recognition rate,
control precision, sampling time, and sampling contact force were recorded.}
\label{fig.experiment}
\end{figure*}
\begin{figure*}
\centering
\subfigure[Motion trajectories.]{\label{fig.motiontraj}
\includegraphics[width=0.3\linewidth]{figure/Trajectory-01.png}}
\subfigure[Joint trajectories.]{\label{fig.Joint}
\includegraphics[width=0.35 \linewidth]{figure/joint-01.png}}
\subfigure[Joint angle velocity.]{\label{fig.Jointvel}
\includegraphics[width=0.35 \linewidth]{figure/jointvel-01.png}}
\subfigure[Tracking errors.]{\label{fig.trackingerror}
\includegraphics[width=0.35 \linewidth]{figure/error-01.png}}
\caption{
Tracking results: the manipulator track of the desired trajectories (sampling tasks) in the Cartesian space. (a) The 3D trajectories of the manipulator; (b) the joint angle trajectories; (c) the joint velocity trajectories (located in bounds); and (d) comparison experiments with \cite{su2020improved}.}
\label{fig.trackingresults}
\end{figure*}
\section{Experiments}
\label{sec.experiments}
In this section, we present the tests with an oral cavity phantom and volunteers using the OP-swab robot system, approval from the Institutional Review Board of The Chinese University of Hong Kong, Shenzhen was obtained(IRB number CUHKSZ-D-20210002). Then, the advantages and disadvantages are summarized and discussed. The parameters of the VP-ZNN are set as: $\varepsilon_0=0.1$, $\varepsilon_1=10$, $\varepsilon_2=10$, $\mu=0.01$, $k_1=10$, and $k_2=10$.
\subsection{Demonstration with Phantom Experiments}
\label{subsec.phantom}
For safety reasons, first, we conduct the experiments with an oral cavity phantom, which will help to confirm the safety of the OP-swab robot system. The oral cavity phantom has a 1:1 size of human oral cavity, which enables natural simulation of workflow across OP-swab sampling and allows us to work extremely close to conditions in practice. The procedures of experiments are as follows:
\begin{itemize}
\item The phantom oral cavity is detected and segmented with the RealSense Camera, which is configured on the RFC manipulator. Because the sampling areas are the left and right tonsils and the palate area, we recognize and locate the target position by Mask R-CNN.
\item The desired sampling trajectories are obtained by online motion planning with an oral cavity constraint, in which we locate the target position in the first step and then generate the trajectories. The desired trajectories are a piecewise straight line from the palate area to the left tonsils to the right tonsils in the Cartesian space, where the relationship between each axis and time is the minimum jerk curve.
\item The robot is driven to perform the sampling tasks. In this phase, the desired joint trajectories are obtained from optimization control methods. Moreover, multiple protection mechanisms with force sensing are activated during robot operation.
\end{itemize}
The experimental scenarios of the oral cavity phantom are shown in Fig.~\ref{fig.phantom}.
We collect the dataset for the oral cavity phantom which is trained by Mask R-CNN \cite{he2017mask}.
To obtain the category and location of the oral sampling areas, an oral visual detector is trained and obtained. First, an oral cavity image dataset is created by collecting from RealSense D435i. Second, the Mask R-CNN with a Inception v4 module as its backbone is trained on the oral cavity image dataset and an oral cavity detector is obtained. Finally, oral cavity coordinates are mapped to depth images to obtain the depth values of the oral cavity. We achieved a 95\% recognition success rate on the model, and the sampling time is less than 20 s.
\subsection{Demonstration with Volunteers Experiments}
On the basis of many trials of the OP-swab RFC robot system with the oral cavity phantom, the robustness and safety were significantly improved. Then, we conducted a number of volunteer experiments and achieved milestone significance for the future clinical application of OP-swab sampling. The procedures of experiments are the same as those described in Sec.~\ref{subsec.phantom}. The dataset of the human oral cavity is collected by 29 volunteers and trained using Mask R-CNN.
The experimental scenarios of different subjects are shown in Fig.~\ref{fig.experiment}. The desired trajectories are the same as in Sec.~\ref{subsec.phantom}. We design the piecewise straight-line trajectories from the palate to the right tonsils to the left tonsils in the Cartesian space, and the relationship between each axis and time is the minimum jerk curve. Figure~\ref{fig.trackingresults} shows the tracking results associated with one of the trials of planning trajectories. The motion trajectories in the Cartesian space are shown in Fig.~\ref{fig.motiontraj}. Figures~\ref{fig.Joint} and \ref{fig.Jointvel} show that the joint trajectories located in the physical limits and joint velocity are smooth within the velocity limits, respectively. Figure~\ref{fig.trackingerror} shows the tracking error in the Cartesian space, where all errors rapidly converge to a small value (0.02 mm). In addition, we added a comparison experiment shown in Fig.~\ref{fig.trackingerror}, where the proposed method has a faster convergence rate and smaller errors than \cite{su2020improved}. The average recognition rate, control precision, sampling time, and sampling contact force are recorded and shown in Table.~\ref{tab.samplingparameters}. The comparative experiments were conducted to examine the effectiveness of robotic OP-swab sampling. The swab quality was verified according to the threshold cycle (Ct) value of the selected reference gene (RNase P) by the RT-PCR test ~\cite{radonic2004guideline}. Figure ~\ref{fig.PRC_comparison} shows the RT-PCR test results compared with those for the manual scheme. OP-swab with Ct values of $\leq 37$ and $> 37$ are considered as qualified and unqualified samples, respectively, and the details are shown in Table ~\ref{tab.PCRparameters}. The test results show that the samples are qualified.
\begin{table}[ht]
\centering
\setlength{\tabcolsep}{3.5mm}
\caption{Average sampling parameters.}
\begin{tabular}{|l|c|c|}
\hline
\diagbox{Metric }{Types} & Phantom & Volunteers\\
\hline
Recognition rate & 95\% & 93\% \\
\hline
Control precision & $\le $0.2 mm & $\le$0.2 mm\\
\hline
Sampling time & 18 $\pm$ 2 s & 20 $\pm$ 4 s\\
\hline
Sampling force & $\approx$ 150 mN & $\approx$ 150 mN\\
\hline
\end{tabular}
\label{tab.samplingparameters}
\end{table}
\begin{figure}[htp!]
\centering{
\includegraphics[width=0.85\linewidth]{figure/ct.jpg}}
\caption{ RT-PCR test results.}
\label{fig.PRC_comparison}
\end{figure}
\begin{table}[ht]
\centering
\setlength{\tabcolsep}{1.5mm}
\caption{Ct distribution.}
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
Ct & (24–27)& (27–30)& (30–33)& (33–37)&\textgreater 37 & Qualified rate\\
\hline
Manual & 11 & 16& 3 & 0 &0 & 100\% \\
\hline
Robotic & 5 & 21& 3 &0 & 1 & 96.67\%\\
\hline
\end{tabular}
\label{tab.PCRparameters}
\end{table}
\section{Conclusion and Future Work} \label{sec.conclusion}
In this study, we designed a 9-DOF redundant RFC robot to assist the COVID-19 OP-swab sampling. Moreover, we formulate the sampling tasks, physical limits, and OCC constraints as a novel optimization problem. Then, a VP-ZNN is proposed to solve the multi-constraint optimization problem online. The experimental results of the oral cavity phantom and volunteer experiments demonstrate the effectiveness of the designed robot system and proposed control methods. The average sampling time on phantoms and volunteers are 18 s and 20 s respectively.
To maintain the sterility of the arm and effector between patients, the MPA is designed as a disposable device with a quick connector that is easily to be replaced, where the disposable protective film is attached to the MPA.
In the future, we will focus on improving the robustness of the system and move forward to clinical testing.
\appendices
\section*{Acknowledgment}
The authors would like to thank the Prof. Xi Zhu, Mr. Yi Liang, Ms. Kerui Yi and Mr. Xiaochuan Lin for their suggestions and engineering contributions to this work. This work was supported by the Shenzhen Fundamental Research grant (JCYJ20180508162406177, JCYJ20190806142818365) and the National Natural Science Foundation of China (U1613216, 62006204, 61903100) from The Chinese University of Hong Kong, Shenzhen. This work was also partially supported by the Shenzhen Institute of Artificial Intelligence and Robotics for Society.
\input{references.bbl}
\bibliographystyle{IEEEtran}
|
\section{Model of measurement and feedback}\label{app:model}
The displacement of the oscillator ($\delta x$) responds to a sum of thermal, back-action, and feedback forces:
\begin{equation}\label{eq:x0}
\chi_0^{-1} \delta x = \delta F_\t{th} + \delta F_\t{ba} + F_\t{fb}.
\end{equation}
Here, the susceptibility of the oscillator $\chi_0^{-1}[\Omega] = m(-\Omega^2 +\Omega_0^2 +i \Omega
\Gamma_0[\Omega])$ is well approximated by the test
mass pendulum mode at frequency $\Omega_0\approx 2\pi \cdot \SI{\fzero}{\Hz}$,
which is structurally damped, so that its damping rate is frequency dependent:
$\Gamma_0[\Omega]=(\Omega_0/Q_0)(\Omega_0/\Omega)$, with a quality factor $Q_0
\approx \num{e8}$.
The thermal force ($\delta F_\t{th}$) is characterized by its spectral density,
\begin{equation}
S_F^\t{th}[\Omega] = 4\hbar \left( n_\t{th}[\Omega] +\tfrac{1}{2}\right)\t{Im}\, \chi_0^{-1}[\Omega],
\end{equation}
where $n_\t{th}[\Omega] \approx k_B T/(\hbar \Omega) \approx 9\cdot 10^{12}
(\Omega_0/\Omega)$ is the average thermal phonon occupation. The back-action
force ($\delta F_\t{ba}$), arising from radiation pressure quantum
fluctuations, is characterized by
\begin{equation}
S_F^\t{ba}[\Omega] = \frac{16\hbar\mathcal{F}}{\lambda c}P_\t{cav} \mathrm{e}^{2r_\text{asqz}},
\end{equation}
where $P_\t{cav}\approx \SI{\Parm}{\kW}$ is the mean arm cavity power at
wavelength $\lambda = \SI{1064}{\nm}$, $\mathcal{F} \approx \num{45}$ is the effective finesse of the signal-recycled arm cavities, and $r_\text{asqz}$ quantifies the increase in
quantum fluctuations of the intracavity optical amplitude due to antisqueezing due to
the phase-squeezed vacuum injected at the interferometer's dark port; here
$10\log_{10}{e^{2r_\text{asqz}}} = \SI{8\pm1}{dB}$ \cite{hao20}.
The back-action force
can be quantified in terms of an average phonon occupation $n_\t{ba}$ via
$S_F^\t{ba} \equiv 4\hbar n_\t{ba}\, \t{Im}\, \chi_0^{-1}[\Omega_0]$, which
gives $n_\t{ba}\approx \num{1.0e12}$.
The feedback force $F_\t{fb}$ is based on a linear estimate of the oscillator's position,
\begin{equation}\label{eq:xe}
\delta x_\t{est} \equiv G (\delta x + \delta x_\t{imp});
\end{equation}
here, $\delta x_\t{imp}$ is the displacement imprecision (due to sensing
noise), and $G$ is the sensing function of the interferometer. Such an
estimate is obtained only when the interferometer is stabilized at its linear
operating point, achieved by a feedback loop that forces the test mass (modeled
by the actuation function $A$) based on a filtered (by $D_0$) record of the
error signal $x_\t{est}$.
We create an additional feedback path consisting of a digital filter $D_\t{T}$
in series with $D_0$ to produce the trap, and a parallel path consisting of the
digital filter $D_\t{C}$ to cold-damp the trapped oscillator. The combined
feedback force thus exerted is
\begin{equation}\label{eq:Ffb}
\begin{split}
F_\t{fb} &= A(D_0D_\t{T}+D_\t{C})\delta x_\t{est} + \delta F_\t{fb} \\
&\equiv -\chi_\t{fb}^{-1}(\delta x + \delta x_\t{imp}) + \delta F_\t{fb};
\end{split}
\end{equation}
here, $\delta F_\t{fb}$ models extraneous force fluctuations due to the
actuator. Solving for $\delta x$ between \cref{eq:x0,eq:Ffb} gives the
physical displacement fluctuations,
\begin{equation}\label{eq:xeff}
\delta x = \chi_\t{eff}\left( \delta F_\t{tot} -\chi_\t{fb}^{-1} \delta x_\t{imp} \right);
\end{equation}
here, $\chi_\t{eff} \equiv (\chi_0^{-1} + \chi_\t{fb}^{-1})^{-1}$ is the effective susceptibility of the oscillator, and
$\delta F_\t{tot} \equiv \delta F_\t{th}+\delta F_\t{ba} +\delta F_\t{fb}$ is the total force noise.
The oscillator can be trapped and cooled by synthesizing an effective susceptibility of the form,
\begin{equation}
\chi_\t{eff}^{-1} = m(-\Omega^2 +\Omega_\t{eff}^2 +i \Omega \Gamma_\t{eff}).
\end{equation}
We are able to do this by careful design of the effective loop filter $\chi_\t{fb}^{-1}$, which is switched on
in a sequence that both traps the oscillator, and keeps the interferometer unconditionally stable.
\begin{figure}[t!]
\centering
\includegraphics[width=0.75\columnwidth]{figure_4_feedback_schematic}
\caption{\label{fig:feedback}
Schematic of the physical system consisting of the intrinsic mechanical response $\chi_0$, the interferometer's
sensing function $G$, the digital filters $D_\t{0,T,C}$, and the actuation $A$; $G^{-1}$ denotes the
digital filter used to reconstruct the apparent displacement.
The feedback is subject to fluctuations arising from actuator force noise $\delta F_\t{fb}$, thermal noise $\delta F_\t{th}$, back-action noise $\delta F_\t{ba}$ and imprecision noise $\delta x_\t{imp}$.
\label{fig:schematic}}
\end{figure}
What we observe is the apparent displacement $\delta x_\t{obs}$ inferred
using an estimate for the inverse sensing function $G^{-1}$ (that forms part of
LIGO's calibration pipeline \cite{ligo_cal_17}). That is, $\delta x_\t{obs}
\equiv G^{-1} \delta x_\t{est} \approx \delta x + \delta x_\t{imp}$; amplitude
uncertainty in this estimate is at the $2\%$ level \cite{Sun20}. Using the
known expression for the physical displacement in \cref{eq:xeff}, the apparent
displacement is,
\begin{equation}
\delta x_\t{obs} = \chi_\t{eff}\left(\delta F_\t{tot} +\chi_0^{-1} \delta x_\t{imp}\right).
\end{equation}
This model produces the spectrum of the observed signal,
\begin{equation}
\begin{split}
S_x^\t{obs}[\Omega] = &
\frac{S_F^\t{tot}[\Omega]/m^2}{(\Omega_\t{eff}^2-\Omega^2)^2 + (\Omega \Gamma_\t{eff})^2} \\
&+ \frac{(\Omega_0^2-\Omega^2)^2 + (\Omega \Gamma_0[\Omega])^2}
{(\Omega_\t{eff}^2-\Omega^2)^2 + (\Omega \Gamma_\t{eff})^2}
S_x^\t{imp}[\Omega],
\end{split}
\end{equation}
that is used to fit the data in \cref{fig2}B in the main text. However the
apparent motion\,---\,since it contains correlations impressed by the feedback
of imprecision\,---\,cannot be directly compared to the spectrum of a physical
oscillator that is damped.
The spectrum of the physical motion of the oscillator ($\delta x$ in \cref{eq:xeff}),
\begin{equation}
\begin{split}
S_x[\Omega] &= \frac{S_F^\t{tot}[\Omega]/m^2}{(\Omega_\t{eff}^2-\Omega^2)^2 + (\Omega \Gamma_\t{eff})^2} \\
&+ \frac{\Omega_\t{fb}^4 + (\Omega \Gamma_\t{fb})^2}
{(\Omega_\t{eff}^2-\Omega^2)^2 + (\Omega \Gamma_\t{eff})^2}
S_x^\t{imp}[\Omega],
\end{split}
\end{equation}
can be directly compared against that of an oscillator trapped at frequency
$\Omega_\t{eff}$, and featuring a damped linewidth $\Gamma_\t{eff}$. Writing
$x=x_\t{zp}(b+b^\dagger)$ for the position of such an oscillator, with
zero-point motion $x_\t{zp}=\sqrt{\hbar/2m\Omega_\t{eff}}$ and creation
operator $b$, presumed to exist in a thermal state, we use the identities,
\begin{align}
\t{Var}[x] &= 2x_\t{zp}^2 \left( \avg{b^\dagger b}+\tfrac{1}{2} \right) \\
\t{Var}[x] &= \int S_x[\Omega]\, \frac{\mathrm{d}\Omega}{2\pi},
\end{align}
to assign an effective phonon number $n_\t{eff}\equiv \avg{b^\dagger b}$:
\begin{equation}
n_\t{eff}+\tfrac{1}{2} = \int \frac{S_x[\Omega]}{2x_\t{zp}^2}\, \frac{\mathrm{d}\Omega}{2\pi}.
\end{equation}
Here, the integral is understood to be evaluated in the frequency interval
where the oscillator susceptibility is realized.
When the imprecision noise is white (i.e. $S_x^\t{imp}[\Omega]\approx
S_x^\t{imp}[\Omega_\t{eff}]$), and the frequency-dependence of the
structurally-damped thermal phonon number can be neglected (i.e.
$n_\t{th}[\Omega] \approx n_\t{th}[\Omega_\t{eff}]$)\;---\;both inapplicable to
the current experiment, but useful to develop intuition\;---\;the phonon
occupation can be explicitly evaluated as,
\begin{equation}
\begin{split}
n_\t{eff}+\frac{1}{2} \approx& \left(n_\t{tot}[\Omega_\t{eff}]
+ \left(\frac{\Omega_\t{eff}}{\Gamma_0[\Omega_\t{eff}]}\right)^2 n_\t{imp}+ \frac{1}{2}\right)
\frac{\Gamma_0[\Omega_\t{eff}]}{\Gamma_\t{eff}[\Omega_\t{eff}]}\\
&\quad + n_\t{imp} \frac{\Gamma_\t{eff}[\Omega_\t{eff}]}{\Gamma_0[\Omega_\t{eff}]},
\end{split}
\end{equation}
where the factor in the parentheses in the first line is the total initial
occupation, consisting of the sum of thermal and back-action quanta
($n_\t{tot}=n_\t{th}+n_\t{ba}$), and an additional contribution
$(\Omega_\t{eff}/\Gamma_0)^2 n_\t{imp}$ due to fluctuations in the trap from
feedback of imprecision noise due to the active spring. Here, $n_\t{imp} \equiv
S_x^\t{imp}[\Omega_\t{eff}]/2S_x^\t{zp}[\Omega_\t{eff}]$, is the
phonon-equivalent imprecision ($S_x^\t{zp}[\Omega_\t{eff}] =
8x_\t{zp}^2/\Gamma_0[\Omega_\t{eff}]$ is the peak zero-point spectrum of the
trapped oscillator).
\subsection{Effect of actuator force noise}\label{app:Ffb}
It has been documented that the electrostatic drive (ESD) that is used to actuate the test masses produces excess
force noise that arises from a combination of charging effects and driver voltage noise \cite{Buik19}.
In the context of feedback cooling the test mass, this force (termed $\delta F_\t{fb}$ in \cref{eq:xeff}) acts
as an excess thermal force that heats the trapped oscillator, resulting in additional phonons ($n_\t{fb}$) that
add to the thermal occupation.
From Ref. \cite{Buik19}, it can be inferred that
\begin{equation*}
\sqrt{S_\t{F}^\t{fb}}\approx (4\cdot 10^{-18}\, \t{N/\sqrt{Hz}})\, \left(\frac{10\, \t{Hz}}{f} \right).
\end{equation*}
(By comparison, the typical actuation strength used to keep the interferometer locked is
$\sim 10^{-6}\, \t{N/\sqrt{Hz}}$ at 10 Hz.)
An extraneous phonon occupation $n_\t{fb,ex}$ can be associated with this force noise via,
$S_F^\t{fb} \approx 4\hbar m \Gamma_0 \Omega_\t{eff} n_\t{fb,ex}$. Assuming
the oscillator is trapped at $\Omega_\t{eff}\approx 2\pi\cdot \feff$ Hz, the equivalent phonon occupation
from excess ESD noise is, $n_\t{fb,ex} \lesssim 10^{-3}$.
\subsection{Effect of filter delay}
If the feedback is implemented with an overall delay $\tau$\;---\;for example arising from delays in the computation
of the digital filter\;---\;the trapping and cooling filter is modified to
\begin{equation*}
\begin{split}
\chi_\t{fb}^{-1} e^{i\Omega \tau} &= m(\Omega_\t{fb}^2 + i\Omega \Gamma_\t{fb})e^{i\Omega \tau} \\
&\approx m [\Omega_\t{fb}^2(1-\tau\Gamma_\t{fb}) + i\Omega(\Gamma_\t{fb}+\tau \Omega_\t{fb}^2)],
\end{split}
\end{equation*}
where in going to the second line, we assume
that the delay is small compared to the characteristic frequency at which it occurs, i.e. $\Omega \tau \ll 1$.
Thus, even in the absence of active damping (i.e. $\Gamma_\t{fb}=0$), delay in the loop manifests as
damping $\tau \Omega_\t{fb}^2$. This serves to stabilize the trapped oscillator.
Delay in other parts of the loop manifest as an overall phase factor in the closed-loop gain,
$\chi_\t{eff} e^{i\Omega \tau}$. Fits to the phase response of the closed-loop gain in the main text resolve
this overall phase shift at the level of $\approx \SI{\tdelay}{\milli\second}$, consistent with expected delays in the loop.
\section{Data analysis}
\subsection{Uncertainty in transfer function fits}\label{app:unc}
The standard deviation in the transfer function estimate $\hat{G}$ derived from signals with coherence $C$ over $N$ averages is given by~\cite{bendat2011random}
\begin{equation}
\sigma_{\hat{G}} = \sqrt{\frac{(1-C)}{2CN}} \hat{G}.
\end{equation}
Similarly, the coherence estimate $\hat{C}$ has standard deviation
\begin{equation}
\sigma_{\hat{C}} = \sqrt{\frac{2C}{N}} (1-C).
\end{equation}
Even assuming a worst-case true coherence $C$ within this range, most of the uncertainty in our data arises from the \SI{1}{\second} average duration and corresponding \SI{1}{\hertz} bin width.
We fit to a time-delayed resonator model,
\begin{equation}
\chi_\text{eff} \sim \frac{1}{\Omega_0^2 - \Omega^2 + \mathrm{i} \Omega_0 \Omega/Q} \mathrm{e}^{\mathrm{i}(\phi - \Omega t)},
\end{equation}
and propagate these uncertainties using orthogonal distance regression~\cite{boggs1988odr}.
\end{document}
|
\section{Introduction}
In the hyper-connected world today, collecting consumer statistics has been a common practice of companies to understand consumers' insights to improve services and products \cite{erlingsson2014rappor,ding2017collecting,adp2017learning}.
The statistical data collection is highly demonstrated in applications of Internet of Things (IoTs) such as Smart Homes, where energy consumption statistics from homeowners can be collected to optimize energy utilization, and Smart Cities, where road transport statistics from motor vehicles can be collected to improve the urban mobility \cite{miorandi2012internet}.
However, the collection of consumer data requires companies to protect consumer or user privacy to comply with enacted privacy laws and regulation \cite{regulation2016regulation}.
Local differential privacy (LDP) \cite{kasiviswanathan2011can,duchi2013local} is a promising privacy-preserving model for statistical aggregation of user data that prevents user privacy leakage from the data aggregator.
LDP achieves the privacy guarantee by introducing random noise into user data before transmitting them to the data aggregator while maintaining the users' statistics to be accurate.
Since the data aggregator cannot confidently know the raw user data, the users have plausible deniability and, therefore, their privacy is remained protected to some degree.
Given a mechanism $\mathcal{M}$ as a function of a value $v$ that perturbs $v$ to a value $s$ and returns $s$ as a noisy representation of $v$.
The formal definition of Local differential privacy is given below.
\begin{definition}[$\epsilon$-Local differential privacy \cite{erlingsson2014rappor}]
\label{def_LPD}
A randomized mechanism $\mathcal{M}$ satisfies $\epsilon$-Local differential privacy if for all pairs of values $v_1$ and $v_2$ given by users, and all set $S$ of the possible outputs of $\mathcal{M}$,
$$Pr[\mathcal{M}(v_1) \in S] \leq exp(\epsilon) \times Pr[\mathcal{M}(v_2) \in S]$$
where
the probability space is over the randomness of the mechanism $\mathcal{M}$.
\end{definition}
The parameter $\epsilon$, called \textit{privacy lost} or \textit{privacy budget}, takes a real positive value specified by the data aggregator to control the strength of the privacy protection.
The smaller the value of $\epsilon$, the higher the probability that the noisy representations of user values are the same or, in other words, the less likely the data aggregator realizes the true user values. Therefore, a smaller value of $\epsilon$ gives stronger protection of privacy.
However, the smaller the value of $\epsilon$, the worse the accuracy of the aggregated statistics because the noise eliminates useful statistical information about true user data.
Thus, the privacy budget $\epsilon$ needs to be carefully chosen to balance privacy protection and statistical accuracy \cite{hsu2014differential}.
In practice, $\epsilon$ is generally set to a value from 0.01 to 1 for a high privacy regime and from 1 to 10 for a low privacy regime \cite{erlingsson2014rappor}.
Discrete distribution estimation with LDP is the task of estimating the distribution of discrete user values from privatized data given by LDP mechanisms without assessing the true data \cite{kairouz2016discrete}.
In this paper, we consider the following scenario of discrete distribution estimation.
There is a set of $n$ users $U=\{ u_1, u_2, .., u_n \} $ and
each user $u_i$ has one value taken from a set of $d$ values $V=\{v_1, v_2, .., v_d\}$.
Let $c(v_j)$ being the number of users holding value $v_j$.
The (frequency) distribution of the user values is a vector $f=(f_{v_1}, f_{v_2}, .., f_{v_d})$ with $f_{v_j}$ being $c(v_j)/n$.
Discrete distribution estimation algorithms with LDP perturb user values at user locations and send the privatized data to the data aggregator for distribution aggregation.
Figure \ref{fig:discrete_distribution_estimation} illustrates the steps of discrete distribution estimation algorithms with LDP.
\begin{figure*}[h!]
\centering
\includegraphics[scale=0.49]{LDP_distribution_estimation.png}
\caption{Discrete distribution estimation with Local differential privacy}
\label{fig:discrete_distribution_estimation}
\end{figure*}
Algorithms for discrete distribution estimation with LDP have been proposed in the literature \cite{kairouz2016discrete,wang2016mutual,erlingsson2014rappor,adp2017learning,acharya2019hadamard}. However, there is a lack of comparison of these algorithms regarding the estimation accuracy and, thus, it is difficult to select a suitable LDP distribution estimation algorithm for a particular dataset.
Some of the existing work \cite{wang2016mutual,acharya2019hadamard} includes a comparison of discrete distribution algorithms but the comparison is solely made based on synthetic datasets that are dissimilar to real-world datasets, such as those are used in our evaluation.
In this paper, we review and present a comparative analysis on the performance of the existing discrete distribution estimation algorithms \cite{kairouz2016discrete,wang2016mutual,erlingsson2014rappor,adp2017learning,acharya2019hadamard} in terms of their accuracy on benchmarks of real-world and synthetic datasets.
The real-world datasets include the Statlog (Australian Credit Approval), the Adult, and the USCensusData1990 datasets that are publicly available at the UCI Machine Learning Repository \cite{asuncion2007uci}.
These datasets contain categorical individual values with dataset size from hundreds to millions of individuals and domain size up to a few hundred values.
The synthetic datasets are generated to be similar to the real-world benchmark datasets regarding dataset size and domain size.
\section{Algorithms}
\label{frequency_estimation}
\subsection{The Randomized response technique \cite{warner1965randomized}}
The Randomized response (RR) technique is perhaps the earliest proposed algorithm for distribution estimation that guarantees LDP.
This algorithm is originally designed for estimating the distribution of binary values, for example, being "yes" or "no" only.
The two user values can be represented as a single bit with value 1 indicating a user value and value 0 indicating the negation of the user value.
The RR algorithm perturbs a user value $v$ by returning this value with probability $p$ and returning the negation of the value with probability $(1-p)$, for $p \ne 1/2$.
The perturbation mechanism of the RR algorithm satisfies LDP with the privacy budget $$\epsilon=ln\frac{p}{1-p}.$$
The original distribution of user values is aggregated as follows.
Let $f'_v$ be the proportion of the perturbed values being $v$, the estimated proportion of value $v$ in the true user data is \cite{warner1965randomized}
$$\hat{f}_v=\frac{f'_v+p-1}{2p-1}.$$
\subsection{The generalized Randomized response algorithm \cite{kairouz2014extremal,kairouz2016discrete}}
The generalized Randomized response algorithm, named k-RR, is a generalization of the RR algorithm to the case that users have more than two values.
Given the privacy budget $\epsilon$, the k-RR algorithm privatizes a user value by sending this value with probability $\frac{e^\epsilon}{e^\epsilon + d -1}$
and sending one of the remaining user values with probability $\frac{1}{e^\epsilon + d -1}$ to the data aggregator.
When $d=2$, the k-RR algorithm is identical to the RR algorithm with the probability $p=\frac{e^\epsilon}{e^\epsilon+1}$.
The k-RR algorithm aggregates the distribution of user values using the maximum likelihood estimator \cite{kairouz2016discrete}.
\subsection{The k-subset algorithm \cite{wang2016mutual} }
The k-subset algorithm privatizes a user value by sending a set of values sampled from the user value domain to the data curator.
The number of the sampled values is defined by parameter $k$.
For a user value $v$, the privatized set $S_v$ has a probability of $\frac{d e^\epsilon}{ke^\epsilon + d - k}/(^d_k)$ to include $v$ into itself and a probability of $\frac{d }{ke^\epsilon + d - k}/(^d_k)$ to not include $v$.
When $k=1$, the k-subset algorithm is equivalent to the k-RR algorithm.
Given all the privatized sets of user values, the k-subset algorithm aggregates the true distribution of user values as follows.
Let $c'(v)$ be the number of times the value $v$ occurred in a privatized set.
Given $g_k = \frac{ke^\epsilon}{ke^\epsilon+d-k}$ and
$h_k = \frac{ke^\epsilon}{ke^\epsilon+d-k}\frac{k-1}{d-1}+\frac{d-k}{ke^\epsilon+d-k}\frac{k}{d-1}$,
the estimated proportion of the value $v$ in the true user data is \cite{wang2016mutual}
$$\hat{f}_v = \frac{c'(v)-h_k n}{(g_k - h_k)n}.$$
\subsection{The Basic RAPPOR algorithm \cite{erlingsson2014rappor}}
The Basic RAPPOR algorithm is a modified version of the RAPPOR algorithm \cite{erlingsson2014rappor} that maps a user value to a single bit in a bit sequence, or a one-hot vector.
A one-hot vector representing a value $v$ is a bit sequence with only the bit at the position $v$ set to 1 and the other bits set to 0.
To privatize a user value, the Basic RAPPOR algorithm applies a mechanism based on the Randomized response technique \cite{warner1965randomized}, simply called randomized response, to perturb every single bit of the represented one-hot vector.
The perturbed bit vector initially has all the bits set to 0.
For each of the output bits, the bit is set to 1 with probability $q$ if the corresponding bit in the input bit vector is 1 and with probability $p$ if the corresponding bit in the input bit vector is 0.
Given all the privatized bit vectors, the Basic RAPPOR algorithm aggregates the true distribution of user values as follows.
Let $c'(v)$ be the number of times the bit at the position $v$ in a privatized bit vector is set to 1.
The proportion of the value $v$ in the true user data is \cite{erlingsson2014rappor}
$$\hat{f}_v = \frac{c'(v) - pn}{(q-p)n}$$
\subsection{The Count Mean Sketch algorithm \cite{adp2017learning}}
The Count Mean Sketch (CMS) algorithm, as similar to the Basic RAPPOR algorithm,
represents a user value as a one-hot vector and applies randomized response to every bit of the one-hot vector to privatize the user value.
However, the CMS algorithm independently flips each bit of the one-hot vector with probability $\frac{1}{e^{\epsilon/2 + 1}}$ and keep each bit unchanged with the compliment probability.
The CMS algorithm aggregates the true distribution of user values from perturbed bit vectors as follows.
Let $c'(v)$ be the number of times the bit at the position $v$ in a perturbed bit vector is set to 1.
Given $c = \frac{e^\epsilon+1}{e^\epsilon-1}$,
the proportion of the value $v$ in the true user data is \cite{adp2017learning}
$$\hat{f}_v = c'(v)\frac{c+1}{2} + \bigg(c'(v)-n\bigg)\frac{c-1}{2}$$
\subsection{The Hadamard response algorithm \cite{acharya2019hadamard}}
The Hadamard response (HR) algorithm privatizes a user value $v$ in a domain of size $d$ by returning a value $v'$ in the domain of size $d'$ with $d \le d' \le 4d$.
To choose $v'$, suppose $d'$ is a power of two, the HR algorithm first creates a Hadamard matrix $H_{d'} = \{1,-1\}_{d' \times d'}$ as follows:
$H_1 = [+1]$.
$ H_m =
\begin{bmatrix}
H_{m/2} & H_{m/2} \\
H_{m/2} & -H_{m/2}
\end{bmatrix}
$
with $m = 2^j$ for $1\leq j\leq log(d')$.
The HR algorithm then creates a set of values $S_v$ of size $s$ with $s \le d'$.
The set $S_v$ includes all the elements in the row $(v+1)^{th}$ of the Hadamard matrix with row index starting from 0 and at the columns in the matrix with a '+1'.
To privatize $v$, the HR algorithm returns an element of $S_v$ with probability $\frac{e^\epsilon}{se^\epsilon + d' - s}$
and returns an element in the domain of size $d'$ but not in $S_v$ with probability $\frac{1}{se^\epsilon+d'-s}$.
When $d'=d$, $s=1$ and $S_v=\{v\}$, the HR algorithm is equivalent to the k-RR algorithm.
The original distribution of user values is aggregated using the following estimation.
Let $c'(v)$ be the number of the privatized values of $v$ that are in $S_v$. The estimated proportion of $v$ in the true user data is
$$\hat{f}(v) = \frac{2(e^\epsilon+1)}{e^\epsilon-1}\big(\frac{c'(v)}{n}-\frac{1}{2}\big).$$
\section{Experimental results}
In this section, we compare the performance of k-RR, Basic RAPPOR (bRAPPOR), CMS, k-subset, and HR in terms of accuracy on benchmark datasets.
The benchmark datasets should include discrete values specifying user attributes.
The compared algorithms first privatize these values and then estimate the discrete distribution of the original values based on the privatized data, without accessing the original data.
\subsection{Experiment setting}
Previous work having an evaluation of discrete distribution algorithms with LDP test the algorithms on synthetic data only
\cite{kairouz2016discrete,wang2016mutual,acharya2019hadamard} or on real-world datasets but the dataset are not made public \cite{erlingsson2014rappor,adp2017learning}.
We use publicly available real-world datasets as well as the synthetic dataset generated to be similar to these real-world datasets, in terms of dataset size and domain size.
For real-world datasets, we use three datasets of categorical individual values publicly available at the UC Irvine Machine Learning Repository \cite{asuncion2007uci}. These datasets include the Statlog (Australian Credit Approval), the Adult, and the USCensusData1990 datasets.
For each dataset, we select only the attributes that have a categorical data type for estimating the distribution of the categorical values in each attribute.
The number of samples in each dataset, the selected attributes, and the value domain size of each attribute are listed in Table \ref{tab:datasets}.
The attributes of each dataset are listed in the increasing order of the value domain sizes.
For detailed descriptions of these attributes, readers are referred to \cite{asuncion2007uci}.
\begin{table}[h!]
\centering
\caption{Real-world datasets with the number of individuals (n), the selected attributes and the value domain size (d).}
\label{tab:datasets}
\begin{tabular}{|c|r|c|r|}
\hline
Dataset& n &Attribute &d\\ \hline
\multirow{3}{*}{Statlog} & & A4 & 3\\ \cline{3-4}
& 690 & A6 & 8 \\ \cline{3-4}
&& A5 & 14\\ \hline
\multirow{3}{*}{Adult}& &Race & 5 \\ \cline{3-4}
& 32,561 & Occ & 15 \\ \cline{3-4}
&&Country & 42 \\ \hline
\multirow{3}{*}{USCensus1990}& &Military & 5 \\ \cline{3-4}
& 2,458,285 & Rvetserv & 12 \\ \cline{3-4}
&&Race & 63 \\ \cline{3-4}
&& PoB & 283 \\ \hline
\end{tabular}
\end{table}
We generate synthetic datasets of categorical values following the Geometric distribution.
For the synthetic datasets with the number of samples up to 20,000 (small datasets),
the generated values are taken from the two categorical domains of 20 and 100 values.
For the synthetic datasets with the number of samples from 20,000 to 400,000 (large datasets),
the generated values are taken from the two categorical domains of 100 and 500 values.
\begin{table*}[t!]
\centering
\caption{Performance (MAE) of the compared algorithms on the real world datasets for $\epsilon = 0.5$.}
\label{tab:high_privacy}
\begin{tabular}{|c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}}
\hline
\multirow{2}{*}{Dataset}& \multirow{2}{*}{Attribute}& \multicolumn{2}{c?{0.4mm}}{\textbf{k-RR}} & \multicolumn{2}{c?{0.4mm}}{\textbf{bRAPPOR}} & \multicolumn{2}{c?{0.4mm}}{\textbf{CMS}} & \multicolumn{2}{c?{0.4mm}}{\textbf{k-subset}} & \multicolumn{2}{c?{0.4mm}}{\textbf{HR}}\\ \cline{3-12}
& & Mean & Std.& Mean & Std. & Mean & Std. & Mean & Std. & Mean & Std.\\ \hline
\multirow{2}{*}{Statlog} & A4 &0.067 & 0.036 & \textbf{0.05} & 0.024 & 0.109 & 0.054 & 0.063 & 0.037 & 0.078 & 0.048 \\ \cline{2-12}
& A6 &0.083 & 0.03 & \textbf{0.043} & 0.014 & 0.089 & 0.03 & 0.082 & 0.022 & 0.071 & 0.023 \\ \cline{2-12}
& A5 &0.083 & 0.013 & \textbf{0.05} & 0.01 & 0.086 & 0.021 & 0.085 & 0.014 & 0.061 & 0.011 \\ \hline
\multirow{2}{*}{Adult}&Race &0.012 & 0.005 & \textbf{0.008} & 0.002 & 0.015 & 0.006 & 0.013 & 0.004 & 0.014 & 0.005 \\ \cline{2-12}
& Occ &0.024 & 0.005 & \textbf{0.009} & 0.002 & 0.016 & 0.003 & 0.015 & 0.003 & 0.015 & 0.003 \\ \cline{2-12}
&Country &0.008 & 0.002 & \textbf{0.005} & 0.001 & 0.01 & 0.002 & 0.01 & 0.001 & 0.013 & 0.002 \\ \hline
\multirow{2}{*}{USCensus1990}&Military &0.002 & 0.001 & \textbf{0.001} & 0.0 & 0.002 & 0.001 & 0.002 & 0.001 & 0.002 & 0.001 \\ \cline{2-12}
& Rvetserv &0.003 & 0.001 & \textbf{0.001} & 0.0 & 0.002 & 0.0 & 0.002 & 0.0 & 0.002 & 0.001 \\ \cline{2-12}
&Race &\textbf{0.001} & 0.0 & 0.016 & 0.0 & 0.023 & 0.0 & \textbf{0.001} & 0.0 & 0.002 & 0.0 \\ \cline{2-12}
& PoB &0.004 & 0.0 & 0.004 & 0.0 & 0.005 & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 \\ \hline
\end{tabular}
\end{table*}
\begin{table*}[t!]
\centering
\caption{Performance (MAE) of the compared algorithms on the real world datasets for $\epsilon = 1$.}
\label{tab:medium_privacy}
\begin{tabular}{|c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}}
\hline
\multirow{2}{*}{Dataset}& \multirow{2}{*}{Attribute}& \multicolumn{2}{c?{0.4mm}}{\textbf{k-RR}} & \multicolumn{2}{c?{0.4mm}}{\textbf{bRAPPOR}} & \multicolumn{2}{c?{0.4mm}}{\textbf{CMS}} & \multicolumn{2}{c?{0.4mm}}{\textbf{k-subset}} & \multicolumn{2}{c?{0.4mm}}{\textbf{HR}}\\ \cline{3-12}
& & Mean & Std.& Mean & Std. & Mean & Std. & Mean & Std. & Mean & Std.\\ \hline
\multirow{2}{*}{Statlog} & A4 &\textbf{0.033} & 0.02 & 0.05 & 0.023 & 0.044 & 0.025 & 0.034 & 0.02 & 0.106 & 0.04 \\ \cline{2-12}
& A6 &\textbf{0.041} & 0.013 & 0.044 & 0.013 & 0.043 & 0.014 & 0.044 & 0.013 & 0.048 & 0.015 \\ \cline{2-12}
& A5 &0.054 & 0.011 & 0.049 & 0.011 & 0.049 & 0.01 & 0.051 & 0.009 & \textbf{0.046} & 0.008 \\ \hline
\multirow{2}{*}{Adult}&Race & \textbf{0.006} & 0.002 & 0.008 & 0.003 & 0.008 & 0.003 & 0.007 & 0.002 & 0.008 & 0.003 \\ \cline{2-12}
& Occ &0.01 & 0.002 & \textbf{0.008} & 0.002 & \textbf{0.008} & 0.002 & \textbf{0.008} & 0.002 & 0.016 & 0.003 \\ \cline{2-12}
&Country & \textbf{0.005} & 0.001 & \textbf{0.005} & 0.001 & \textbf{0.005} & 0.001 & 0.006 & 0.001 & 0.008 & 0.001 \\ \hline
\multirow{2}{*}{USCensus1990}&Military &\textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 \\ \cline{2-12}
& Rvetserv &\textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 \\ \cline{2-12}
&Race &\textbf{0.001} & 0.0 & 0.016 & 0.0 & 0.021 & 0.0 & \textbf{0.001} & 0.0 & 0.01 & 0.0 \\ \cline{2-12}
& PoB & 0.002 & 0.0 & 0.004 & 0.0 & 0.005 & 0.0 & \textbf{0.001} & 0.0 & \textbf{0.001} & 0.0 \\ \hline
\end{tabular}
\end{table*}
\begin{table*}[t!]
\centering
\caption{Performance (MAE) of the compared algorithms on the real world datasets for $\epsilon = 2$.}
\label{tab:low_privacy}
\begin{tabular}{|c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}c|c?{0.4mm}}
\hline
\multirow{2}{*}{Dataset}& \multirow{2}{*}{Attribute}& \multicolumn{2}{c?{0.4mm}}{\textbf{k-RR}} & \multicolumn{2}{c?{0.4mm}}{\textbf{bRAPPOR}} & \multicolumn{2}{c?{0.4mm}}{\textbf{CMS}} & \multicolumn{2}{c?{0.4mm}}{\textbf{k-subset}} & \multicolumn{2}{c?{0.4mm}}{\textbf{HR}}\\ \cline{3-12}
& & Mean & Std.& Mean & Std. & Mean & Std. & Mean & Std. & Mean & Std.\\ \hline
\multirow{2}{*}{Statlog} & A4 &\textbf{0.013} & 0.008 & 0.047 & 0.024 & 0.024 & 0.013 & \textbf{0.013} & 0.008 & 0.031 & 0.013 \\ \cline{2-12}
& A6 & \textbf{0.018} & 0.005 & 0.047 & 0.014 & 0.025 & 0.006 & 0.027 & 0.008 & 0.029 & 0.008 \\ \cline{2-12}
& A5 & \textbf{0.022} & 0.004 & 0.051 & 0.01 & 0.027 & 0.006 & 0.031 & 0.006 & 0.032 & 0.007 \\ \hline
\multirow{2}{*}{Adult}&Race & \textbf{0.003} & 0.001 & 0.008 & 0.003 & 0.004 & 0.002 & 0.004 & 0.001 & 0.055 & 0.002 \\ \cline{2-12}
& Occ & \textbf{0.003} & 0.001 & 0.009 & 0.002 & 0.004 & 0.001 & 0.005 & 0.001 & 0.006 & 0.001 \\ \cline{2-12}
&Country &\textbf{0.003} & 0.0 & 0.005 & 0.001 & \textbf{0.003} & 0.0 & 0.004 & 0.0 & 0.007 & 0.001 \\ \hline
\multirow{2}{*}{USCensus1990}&Military & \textbf{0.0} & 0.0 & 0.001 & 0.0 & 0.001 & 0.0 & 0.001 & 0.0 & 0.039 & 0.0 \\ \cline{2-12}
& Rvetserv &\textbf{0.0} & 0.0 & 0.001 & 0.0 & 0.001 & 0.0 & 0.001 & 0.0 & 0.0 & 0.0 \\ \cline{2-12}
&Race & \textbf{0.0} & 0.0 & 0.016 & 0.0 & 0.018 & 0.0 & 0.0 & 0.0 & 0.003 & 0.0 \\ \cline{2-12}
& PoB & 0.001 & 0.0 & 0.004 & 0.0 & 0.004 & 0.0 & \textbf{0.0} & 0.0 & 0.001 & 0.0 \\ \hline
\end{tabular}
\end{table*}
We implement the k\_RR, the Basic RAPPOR, the CMS, and the k-subset algorithms following \cite{kairouz2016discrete,wang2016mutual,erlingsson2014rappor,adp2017learning}. The implementation of the Hadamard algorithm is publicly provided as listed in \cite{acharya2019hadamard}.
The algorithms are executed with default values of parameters.
The performance of the algorithms are compared in three privacy regimes with $\epsilon$ = 0.5 for high privacy regime, $\epsilon$ = 1 for medium privacy regime, and $\epsilon$ = 2 for low privacy regime.
The accuracy of the distribution estimation is measured by Mean Absolute Error (MAE) which is defined as
$$MAE(\hat{f},f) =\frac{1}{d} \sum^d_{i=1}|\hat{f}_{v_i} - f_{v_i}|$$
where
$\hat{f}_{v_i}$ and $f_{v_i}$ are the estimated and the true proportions of user value $v_i$ respectively.
The MAE value is the average of the absolute differences between all the elements of the estimated distribution and the true distribution of user values.
The smaller the MAE value, the more accurate the distribution estimation.
\subsection{Results}
Table \ref{tab:high_privacy}, Table \ref{tab:medium_privacy} and Table \ref{tab:low_privacy} show the accuracy of the evaluated algorithms on the real world datasets for the three privacy regimes.
The result for the Statlog and the Adult datasets is the average of the results from a hundred runs.
The result of the USCensus dataset is the average of the results from 10 runs.
The column Mean lists the average of the MAE values obtained from all the runs of the algorithm and the column Std. lists the standard deviation of the MAE values.
The higher the value of Std., the more fluctuated the estimation accuracy.
The best performing results (or the lowest MAE values) are highlighted in bold.
As can be seen from the tables, the Basic RAPPOR algorithm performs best for most of the datasets in the high privacy regime ($\epsilon= 0.5$).
In the medium privacy regime ($\epsilon= 1$), k-RR performs better than the other algorithms for the datasets with a small number of samples
while on the datasets with a large number of samples, k-subset and HR are fairly competitive with k-RR.
In the low privacy regime ($\epsilon= 2$), the k-RR algorithm generally performs better than the other algorithms.
\begin{figure*}[t!]
\subfloat[][$\epsilon$=0.5]{
\includegraphics[scale=0.42]{MAE_Geo_ep05d20n1000.png}
}
\subfloat[][$\epsilon$=1]{
\includegraphics[scale=0.42]{MAE_Geo_ep10d20n1000.png}
}
\subfloat[][$\epsilon$=2]{
\includegraphics[scale=0.42]{MAE_Geo_ep20d20n1000.png}
}
\caption{Performance (MAE) of the compared algorithms on the small synthetic datasets ($n \leq 20,000$) with d=20.}
\label{fig:synthetic_smalld20}
\end{figure*}
\begin{figure*}[t!]
\subfloat[][$\epsilon$=0.5]{
\includegraphics[scale=0.42]{MAE_Geo_ep05d100n1000.png}
}
\subfloat[][$\epsilon$=1]{
\includegraphics[scale=0.42]{MAE_Geo_ep10d100n1000.png}
}
\subfloat[][$\epsilon$=2]{
\includegraphics[scale=0.42]{MAE_Geo_ep20d100n1000.png}
}
\caption{Performance (MAE) of the compared algorithms on the small synthetic datasets ($n \leq 20,000$) with d=100.}
\label{fig:synthetic_smalld100}
\end{figure*}
\begin{figure*}[t!]
\subfloat[][$\epsilon$=0.5]{
\includegraphics[scale=0.42]{MAE_Geo_ep05d100n20000.png}
}
\subfloat[][$\epsilon$=1]{
\includegraphics[scale=0.42]{MAE_Geo_ep10d100n20000.png}
}
\subfloat[][$\epsilon$=2]{
\includegraphics[scale=0.42]{MAE_Geo_ep20d100n20000.png}
}
\caption{Performance (MAE) of the compared algorithms on the large synthetic datasets ($20,000 < n \leq 400,000$) with d=100.}
\label{fig:synthetic_larged100}
\end{figure*}
\begin{figure*}[t!]
\subfloat[][$\epsilon$=0.5]{
\includegraphics[scale=0.42]{MAE_Geo_ep05d500n20000.png}
}
\subfloat[][$\epsilon$=1]{
\includegraphics[scale=0.42]{MAE_Geo_ep10d500n20000.png}
}
\subfloat[][$\epsilon$=2]{
\includegraphics[scale=0.42]{MAE_Geo_ep20d500n20000.png}
}
\caption{Performance (MAE) of the compared algorithms on the large synthetic datasets ($20,000 < n \leq 400,000$) with d=500.}
\label{fig:synthetic_larged500}
\end{figure*}
Figure \ref{fig:synthetic_smalld20}, Figure \ref{fig:synthetic_smalld100},
Figure \ref{fig:synthetic_larged100} and Figure \ref{fig:synthetic_larged500}
illustrate the performance of the compared algorithms on the synthetic datasets.
According the figures, in the high privacy regime ($\epsilon=0.5$),
Basic RAPPOR generally outperforms the other algorithms for the small datasets with small value domain size (Figure \ref{fig:synthetic_smalld20}a)
and for the large datasets with large value domain size (Figure \ref{fig:synthetic_larged500}a).
HR is better than the other algorithms for the small datasets with large value domain size (Figure \ref{fig:synthetic_smalld100}a)
while k-subset is the best performing algorithm for the large datasets with small value domain size (Figure \ref{fig:synthetic_larged100}a).
In the medium privacy regime ($\epsilon=1$),
k-RR, k-subset, and HR are competitive with each other and generally give the best estimation.
In the low privacy regime ($\epsilon=2$),
k-RR generally has the highest accuracy but the performance of k-subset and HR are often comparable with the performance of k-RR.
The experimental results indicate that the relative accuracy of the compared algorithms are dependent on the characteristics of the evaluation datasets, such as dataset size and value domain size, as well as the privacy protection level.
The dependency of the algorithm performance on dataset characteristics has also been previously observed in \cite{acharya2019hadamard} for a set of synthetic datasets.
However, our results further confirm the finding for real-world datasets and a set of synthetic datasets with similar characteristics to these real-world data.
The performance dependency of the compared algorithms on dataset characteristics is likely due to their approach to privatizing user values.
While k-RR and k-subset privatize user values by directly perturbing the domain values only,
HR perturbs user values using the domain values and a pre-defined number of the values outside the user value domain.
Basic RAPPOR and CMS privatize user values using one-hot vector representations representing a large number of the values outside the user value domain.
The perturbation steps of the algorithms differently introduce noise into their privatized data depending on value domain size, dataset size, and privacy protection level.
The estimation accuracy of the algorithms are, therefore, differently sensitive to these parameters.
\section{Conclusions}
We have presented a comparative analysis on the performance of discrete distribution estimation algorithms with Local differential privacy on real-world and synthetic datasets.
The real-world datasets include three datasets of categorical individual values publicly available at the UC Irvine Machine Learning Repository \cite{asuncion2007uci}: the Statlog (Australian Credit Approval), the Adult, and the USCensusData1990 datasets.
The synthetic datasets are generated to be similar to the real-world datasets regarding dataset size and domain size.
We conclude that
the Basic RAPPOR algorithm generally performs best, in terms of estimation accuracy, for the evaluation datasets in the high privacy regime.
In the medium privacy regime, k-RR, k-subset, and HR give fairly comparable results but and generally are better than Basic RAPPOR and CMS.
In the low privacy regime, the k-RR algorithm often gives the best estimation.
Our empirical evaluation is based on the execution of the algorithms with default values of parameters.
To further understand the relative performance of the algorithms, it would be necessary to compare the algorithms for different parameter settings.
For example, the accuracy of the Basic RAPPOR algorithm would be improved by adopting the optimal choice of parameter values as discussed in \cite{wang2017locally}.
In the limited scope of this paper, we have only focused on the estimation accuracy of the algorithms.
Further work will be required to evaluate the algorithms in other aspects such as computational complexity and communication cost.
\section*{Acknowledgment}
The work has been supported by the Cyber Security Research Centre Limited whose activities are partially funded by the Australian Government’s Cooperative Research Centres Programme.
We would like to thank anonymous reviewers for their constructive
and helpful comments on an earlier version of the manuscript.
\bibliographystyle{plain}
|
\section{Case Study: President Biden's 2021 Inauguration}
To provide a contemporary case study using the dynamic keyword selection method, we started the dynamic monitor on January 11, 2021 (at 10:40:44 Pacific Standard Time), with a single keyword, ``inauguration.'' We used the Twitter data collection architecture developed by \cite{cao2020reliable} to stream data. We stopped collecting data using this dynamic monitor on January 22, 2021 at 15:21:06 Pacific Standard Time.
We show in Figure~\ref{fig:casestudy1} the amount of data collected by this dynamic monitor (by hour in the top panel, and by day in the bottom panel). The bottom panel of Figure~\ref{fig:casestudy1} provides what a static monitor would have collected, daily, during this period. During the first full day of data collection, the dynamic monitor pulled 499,111 tweets; rising by January 14, 2021 to 2,470,410 on that day. The number of tweets collected by the dynamic monitor peaked on January 20, 2021 (the day of the inauguration), pulling 3,434,650 tweets. In total, the dynamic monitor collected 19,723,508 tweets.
\begin{figure}[htb!]
\centering
\begin{tabular}{cc}
\includegraphics[width = 0.45\textwidth, trim={0.65cm, 0.7cm, 0.2cm, 0.2cm}, clip]{figs/hourly_count.pdf} \\
\end{tabular}
\caption{Hourly Number of Tweets containing \#inauguration from Jan 06, 2021 to Jan 22, 2021.}
\label{fig:casestudy1}
\end{figure}
\subsection{Data Collection and Update Procedure}
We pulled discussions related to \#inauguration from Twitter's historical and streaming APIs. In order to adapt the keyword set to account for dynamically changing conversations, we utilized our data visualization platform in a human-assisted update procedure. Specifically, on each day of the experiment at $~5pm$ PST, we used the closet neighbor table and tsne plots on our webpage (see \ref{fig:interface}, \ref{fig:interface1}) to collaboratively determine whether to add or subtract keywords for data collection.
\subsection{Comparison with the Static Monitor}
\begin{figure}[hbt!]
\setlength{\tabcolsep}{0pt}
\centering
\begin{tabular}{cc}
\includegraphics[height=5cm, scale = 0.40, trim={2cm, 0.1cm, 1.5cm, 0.5cm}]{figs/static_day9.png} &
\includegraphics[height=5cm, scale = 0.40, trim={2cm, 0.1cm, 1.5cm, 0.5cm}]{figs/dynamic_day9.png}
\end{tabular}
\vspace{-20pt}
\caption{Word cloud from the static monitor (left) and the dynamic monitor (right) on day 9 of the experiment showing the most frequently used hashtags in the statically and dynamically obtained Twitter datasets. The dynamic monitor covers a wider range of topics and has a more uniform distribution in terms of the frequency of different keywords. }
\label{fig:compStatic}
\end{figure}
As shown in Figure \ref{fig:compStatic}, the semi-automated dynamic approach is visibly better than a static data collection procedure, producing a more uniform distribution of popular keywords. The dynamic approach keeps old frequently-discussed hashtags (\#inaugurationday) while detecting new viral hashtags (\#trump, \#biden), providing more effective general topic coverage than the static monitor.
\subsection{Observation and Discussion}
We have the following observations in the data analysis and keyword update for this case study.
\paragraph{Word Ambiguity} General terms are ambiguous and may induce irrelevant information. In day 2, we saw a surge in the keyword frequency of ``safety" and ``security" in our filtered corpus. We then decided to include them in our keywords set and expect to capture more text about data security. However, with the knowledge that ``safety" and ``security" are general and ambiguous terms, we take extra caution in the following days. When taking a closer look at the corpus that contains the keywords ``security", we dropped it and include more specific terms like ``datasecurity" and ``databreach" instead.
\paragraph{Convergence} The dynamic discussion would
``converge" with no new keywords added in the end. In our case study, we did not find new keywords to add to the keyword set in the last three days. Since we did not set a upper limit in the keyword set, it is possible that we captured all the topics after 9 days. In practice, we believe the convergence may or may not happen depending on whether there is a constraint in terms of the number of keywords in total or number of new keywords each round.
\paragraph{Forecast Information} One difficulty in our case study is the usage of the forecast information. Due to the short time horizon we are working on, there is not enough data for us to train reliable predictive models on the keywords, which impose difficulty in fully automatic update. Fortunately, our method is flexible enough to also account for human intervention. In the future, we aim to further test various type of update rules.
\section{Conclusion and Discussion}
In this paper, we design and implement a novel dynamic keyword search method for tracking and monitoring fast-evolving online discussions. This closes the gap between the traditional static keyword search method and the highly dynamic data sources in social media. We use word embedding models for finding relevant keyword automatically and use indicators from predictive time series models to help the decision making in keyword updates. We allow for both semi-automatic and fully-automatic data collection. The whole system is build using modern data collection, storage and visualization tools.
To test the current deployment, we simulate on data collected from the \#MeToo movement and also analyze a practical use case that covers the recent President Inauguration. Our simulation and case study reflects the effectiveness of our framework.
The remaining work on our process will focus on the time series predictive modeling. While we have deployed our process with a preliminary time series predictive model (which we demonstrate in the \#inauguration case study), we plan to more rigorously test effectiveness of different time series approaches, including traditional linear time series models like ARIMA and deep learning forecasting models like LSTM.
\section{Evaluation}
\begin{figure}[hbt]
\begin{tabular}{l}
\includegraphics[height=5cm, scale = 0.30, trim={2cm, 0.09cm, 0.9cm, 0.5cm}]{figs/MeToo_2017_percent.png} \\
\\
\includegraphics[height=5cm, scale = 0.30, trim={2cm, 0.09cm, 0.9cm, 0.5cm}]{figs/2yesallwomen_notallmen.png} \\
\\
\end{tabular}
\caption{{\bf Upper}: {\bf Upper}: Percentage of monthly tweets containing \#MeToo, with an unexpected surge in October 2017. {\bf Bottom}: Percentage of tweets with "\#yesallwomen" (pink) or "\#notallmen"(blue) throughout 2017.}
\label{fig:totalcounts}
\end{figure}
\subsection{Simulation Designed using \#Metoo Data}
\paragraph{\#MeToo Data:} Building on a series of women's rights marches and protests, the \#MeToo movement went viral on Twitter in October 2017 after media outlets widely publicized sexual assault allegations against Harvey Weinstein. Women and men across the globe adopted the \#MeToo hashtag to stand in solidarity against sexual harassment and bring to light gender inequality issues across many spheres, from Hollywood to the tech industry. To study this evolving set of issues, our research team obtained a large collection of Twitter data about the \#MeToo movement; this data was obtained directly from Twitter through Boolean filtering and is publicly available information. Demonstrating the scale of women's rights discussions and the shift to \#MeToo, we show the total number of tweets in our data, as well as the percentage of these tweets that contain ``\#MeToo'' in Figure \ref{fig:totalcounts}.
\paragraph{Evolution of Topics:} As shown in Figure \ref{fig:janclouds}(a), January 2017 sees an upsurge of the hashtag ``\#womensmarch'', which can be attributed to the Women's March on Washington, the U.S.'s largest single-day public demonstration that took place the day after President Trump's inauguration on the 20th to promote gender equality and civil rights. The controversial 2016 U.S. presidential campaign inspired many partisan discussions in January, with politically contentious hashtags like ``\#theresistance'', ``\#notmypresident'', ``\#trumpleaks'', ``\#alternativefacts,'' and ``\#fakenews''. The months February through September witness similar lines of discussion about politics, gender inequality, and feminism, albeit with evolving top terms. For instance, come June 2017, hashtags like ``\#imwithher'', ``\#everydaysexism'', and ``\#feminismiscancer'' rise to the forefront of online conversation. As this discourse gains momentum, we see the emergence of movements and their corresponding anti-movements: take for instance ``\#yesallwomen'' (expressing prevalence of sexual harassment faced by women) vs. ``\#notallmen'' (a hashtag to defend males), both which are prominent in August 2017, as shown in Figure \ref{fig:janclouds} (b). Come October 2017, high-profile sexual assault allegations including those against Harvey Weinstein catalyzes the ``\#MeToo'' movement and spurs others to come forward with allegations against Larry Nassar, Kevin Spacey, Roy Price, and more over the next few months.
\paragraph{Simulation of a Dynamic Monitor:} The shifts in the language used to describe a general set of issues and the advent of \#MeToo represent an ideal scenario to utilize a dynamic keyword algorithm. While theoretically we could have run our dynamic keyword algorithm in real-time, in order to better test our method, we simulate it on historical data from January 2017 to December 2017 to evaluate how well our algorithm tracks the evolving \#MeToo movement. The aim of this simulation is to capture the most frequent keywords for each month starting in January 2017 and ending in December 2017. We simulate our dynamic keyword search process as follows: starting in January, we filter the overall corpus of historical January data with our seed keywords. Note that by filtering down the large set of historical data, we simulate gathering data from the Twitter API. We then analyze this filtered set with our dynamic keyword approach to determine a new round of keywords for the subsequent month. In this simulation, the top hashtags each month in the full \textbf{unfiltered} historical data represents the ground truth target set for evaluation.
\begin{figure}[hbt!]
\setlength{\tabcolsep}{0pt}
\centering
\begin{tabular}{cc}
\includegraphics[scale=0.45, trim={1.2cm, 1cm, 1cm, 1cm}, clip]{figs/WH_Jan17.png}&
\includegraphics[scale=0.45, trim={1.2cm, 1cm, 1cm, 1cm}, clip]{figs/WH_Aug17.png} \\
(a) & (b)
\end{tabular}
\caption{(a) Word cloud containing 40 most frequently used hashtags in January 2017; (b)
Word cloud of 40 most frequently used hashtags in August 2017. The size of the words indicate the frequency. There is a significant topic shift on Twitter between these two months.}
\label{fig:janclouds}
\end{figure}
\subsection{Evaluation Baselines and Metrics}
We compare our method with two baselines: Static and Last-Top.
\begin{enumerate}
\item Static: uses the top $n$ keywords from January 2017 for each of the following months throughout the simulation.
\item Last-Top: assumes previously trending hashtags can expose trending conversations in the current month and uses the keywords set from the last month.
\end{enumerate}
To make comparison between the methods more tractable, we set $n=15$ keywords at any time $t$, although in practice it is certainly possible to use more.
Since we have access to the top trending hashtags for each month in the whole dataset, we set these to be the ground truth, and evaluate the most frequently used hashtags pulled by various monitors against this. We use the Jaccard similarity index and F1-score for the evaluation.
To calculate F1-score, we regard the proportion of correctly retrieved top trending hashtags in the retrieved hashtags ($R$) and in the ground truth hashtags ($G$) as precision and recall.
For the Jaccard similarity, we look at the union and the intersection of the retrieved hashtags ($R$) and the ground truth hashtags ($G$). Importantly, all monitors begin with the same set of keywords--the top $15$ most frequently used hashtags in our January 2017 \#MeToo data.
\begin{align}
F_1 = 2 \cdot \frac{\text{Precision}\cdot \text{recall}}{\text{Precision}+ \text{recall}}
\end{align}
\begin{align}
J = \frac{R \cap G}{R \cup G}
\end{align}
\subsection{Result Analysis}
\begin{figure}[h]
\centering
\begin{tabular}{cc}
\includegraphics[height=5.5cm, scale = 0.40, trim={1cm, 0.1cm, 1.5cm, 0.5cm}]{figs/jaccard_sim_metoo.png} \\
\\
\includegraphics[height=5.5cm, scale = 0.40, trim={1cm, 0.1cm, 1.5cm, 0.5cm}]{figs/f1_score_metoo.png}
\end{tabular}
\caption{The similarity between the set of top 20 hashtags in the subset of data recovered by each algorithm (dynamic, last-top, and static) and the target set of top 20 hashtags in the full month of data in terms of Jaccard index (top) and F1 score (bottom). Dynamic algorithm outperforms the baselines at most timesteps. In particular, the dynamic monitor covers 75\% more top trending keywords than the static monitor on average.}
\label{fig:results}
\end{figure}
\begin{table}[hbt!]
\caption{Global Performance Comparison}
\label{tab:global}
\centering
\begin{tabular}{|l|l|l|l|}
\hline
& {\small Jaccard } & {\small Avg. F1} & {\small Avg. F1} \\
& & {\small Weighted} & {\small Unweighted} \\
\hline
Dynamic & $ .5406$ & $.6976$ & $.7083$ \\
Last-Top & $.508$ & $.6665$ & $.6041$ \\
Static & $.4594$ & $.6199$ & $.5166$\\
\hline
\end{tabular}
\label{fig:F1Table}
\end{table}
Figure \ref{fig:results} shows the performance of our algorithms and the baselines. Table \ref{tab:global} shows the weighted average of Jaccard similarity index and F1-score of each algorithm as percentages with respect to the to target set (top hashtags for the full-month data), where the weight is the proportion of the size of monthly data to the entire corpus size. We also compare with an unweighted average F1-score. In all of the metrics, our method outperforms the Last-Top baseline and the conventional Static monitor typically used in social science research. This indicates that our method can better capture the trending topics in the \#MeToo data without losing track of the newly emerging keywords and hashtags. \newline
For a more granular analysis, the fact that all monitors start with the same keywords explains identical performance in the first month. The Dynamic and Last-Top monitors perform most poorly in October 2017: this is due to the fact that neither picks up the sudden emergence of \#MeToo. Through preliminary analysis, we find that this dramatic surge in usage of \#MeToo occurs within a few hours, and that even powerful deep learning models perform poorly in prediction tasks with this dataset-specific anomaly.
\paragraph{Keyword Evolution in \#MeToo Simulation:} Figure \ref{fig:MeTooKeywords} demonstrates the keywords retrieved by our algorithm, which reflects the evolution of the topics in the early stage of the \#MeToo movement.
\section{Implementation}
\subsection{Data Visualization Platform}
To enable researchers to collaboratively adapt their social media data collection process to dynamically changing discussions, we are building a data visualization platform which integrates a browser-based user interface on the frontend with scripts for data streaming and AI-driven predictive modelling on the backend. Below, we detail the features and capabilities of this tool, before demonstrating its use on real-time monitoring of 2021 Twitter discussions about \#inauguration.
\paragraph{Frontend}
The frontend of our platform is built with JavaScript, HTML, and CSS. We use the JavaScript library echarts to render figures and tables which are dynamically updated by the backend to reflect real-time data. Our frontend code uses the echarts API to make our figures interactive: users can ``zoom in" on specific numerical values. We host our frontend using Github pages and plan to release our code to allow any research group to host their own personalized version of the data visualization platform. Figures \ref{fig:interface1}, \ref{fig:interface2}, and \ref{fig:interface3} show the interface visualizations for \#insurrection, a tracked keyword in our 2021 \#inauguration case study. We provide a demo of our frontend UI: https://mayasrikanth.github.io/social-media-trends/index.html.
\begin{figure}[H]
\begin{tabular}{cc}
\includegraphics[width = 0.47\textwidth, trim={5cm, 1cm, 0.2cm, 5cm}, clip]{figs/insurrection_tsne.png}
\end{tabular}
\caption{Interactive tsne (t-distributed stochastic neighbor embedding) plot of the closest 30 neighbors to the tracked keyword \#insurrection (shown in blue). Size of bubble corresponds to corpus frequency of keyword. }
\label{fig:interface1}
\end{figure}
\begin{figure}[H]
\begin{tabular}{cc}
\includegraphics[width = 0.47\textwidth, trim={0.3cm, 0cm, 0.2cm, 0.3cm}, clip]{figs/insurrection_forecast.jpg}
\end{tabular}
\caption{Forecast plot generated after training ARIMA on log-transformed \#insurrection frequency data, which predicts log frequencies 15 timesteps into future and provides a 95\% confidence interval. ARIMA $p,d,q$ values were naively grid-searched. MSE for forecasts is $0.259$. X-axis shows last 50 hours in case study. }
\label{fig:interface2}
\end{figure}
\begin{figure}[hbt!]
\begin{tabular}{cc}
\includegraphics[height=5.0cm, scale = 0.60, trim={0.5cm, 0.1cm, 1.5cm, 0.5cm}]{figs/insurrection_table.png}
\end{tabular}
\caption{Table of 30 closest neighbors to \#insurrection, sorted by a linear combination of keyword cosine distance to \#insurrection and corpus frequency.}
\label{fig:interface3}
\end{figure}
\paragraph{Backend}
Our code streams data using Twitter API and stores it in a cloud compute service (Oracle Cloud or GCP). On the cloud, several scripts preprocess data, train GloVe and other predictive time series models, and update the frontend interface. Then on the frontend, the user can select new keywords or drop old ones to customize their data collection process. Time intervals for updating all frontend visualizations are customizable, with a lower bound of $\approx 15$ minutes.
\subsection{Predictive Model Guided Decision Making}
Our dynamic monitor design lends itself to semi-automated and fully -automated data collection processes.
\paragraph{Semi-automated}
We consider a semi-automated data collection process to mean a human-assisted one. That is, a group of researchers interested in tracking a particular set of topics on social media can utilize the AI-driven keyword recommendations on the frontend to alter their keyword set throughout data collection. The time series forecasting and word embeddings can uncover new discussion topics and indicate whether existing hashtags are increasing or decreasing in frequency: given this real-time information, researchers can adjust their data streaming.
\paragraph{Fully-automated}
Our implementation leaves room for a fully-automated approach which sorts candidate keywords using a linear combination of predictive factors, such as:
\begin{equation}
s_{i} = \alpha \cdot m_{i} + \beta \cdot d_{i} + \gamma \cdot f_{i} + \delta \cdot v_{i}
\end{equation}
where $s_{i}$ is the ``virality" score for a given keyword, $m_{i}$ is the slope of the projected frequency trend-line, $d_{i}$ is the average cosine distance from the current set of keywords, $f_{i}$ is the current corpus frequency of the keyword, and $v_{i}$ is the variance of the keyword's frequency trend-line. Keywords can be sorted according to this metric, and the first $3-5$ keywords can be automatically added to the set. Further, removal criterion can be imposed--e.g. we can drop keywords that are relatively old and have low usage in the corpus. The scaling factors for these variables can be customized to fit the research objective: for instance, if a researcher wants to track niche topics with low predicted popularity, they can reduce the weight of $f_{i}$.
We take the semi-automated approach in our studies, as it better fits our research objectives of testing a human-in-the-loop dynamic data collection method. In future iterations, we will include framework for the fully-automated keyword selection in our code.
\section{Introduction}
Conversations on social media platforms like Twitter are often dynamic \citep{bruns}. As new events occur, the language in tweets changes --- words rise and fall in popularity and evolve with time. Influencers, celebrities, and political leaders often change the topics discussed online to sell their products, brands, and ideas. Furthermore, extremist groups, state-sponsored organizations targeting dissidents, and those intent on harassing and intimidating others online often permute the syntax of existing keywords or create new hashtags to avoid detection by social media platforms (and others who might be monitoring them) \citep{Badawy_etal_2018}. To understand the natural evolution of online discussions and detect abusive conversations in real-time, it is important to develop data collection methods which track shifts in online discourse.
Many researchers who use social media data from Twitter collect data using a set of static (unchanging) keywords and hashtags, e.g., \citep{cao2020reliable}. But as previous research shows, static data collection methods fall short when social media conversations change, either because the language used to discuss some topic alters or the hashtags are syntactically modified \citep{king_lam_roberts_2017, liu_etal_2019}. Thus, there is a need for building dynamic social media monitors that can adapt to changes in social media conversations.
Developing a dynamic social media data collection monitor that can update keywords and hashtags is a challenging task. Prior research has proposed methods that require human intervention, or are semi-automated \citep{king_lam_roberts_2017, liu_etal_2019}. Other researchers may prefer fully-automated methods. Either way, a dynamic monitor requires the integration of a number of different methods: it needs to start with a collection of social media posts on a certain topic, which can then be analyzed by natural-language processing tools to determine if there are new keywords or hashtags emerging in the data over time. The dynamic monitor then needs a predictive modeling step, where it forecasts the likelihood that the new language on the topic will continue to grow. Finally, based on the predictive model, the dynamic monitor then needs to adjust the keywords or hashtags it collects information on, and needs to continue to analyze whether new keywords or hashtags should continue to be included in the monitor.
\begin{figure*}[t]
\centering
\begin{tabular}{c}
\includegraphics[width = 0.7\textwidth]{figs/Dynamic_Keyword_Backend.pdf} \\
\end{tabular}
\caption{Workflow of our Data Collection, Storage, Analysis, and Visualization Platform. Here we use Twitter APIs as an example for data collection and we use Google Cloud Platform (GCP) as an example for data streaming and storage. GCP Monitor works in a sequential way while the backend analyses can be parallel.}
\label{fig:flowchart}
\end{figure*}
We design and implement a dynamic monitor for collecting data on fast-evolving online discussions. We allow for both semi-automatic and fully-automatic data collection. Our final dynamic monitor design uses word embeddings, corpus frequencies, and predictive time series modelling to visualize trends in a real-time social media discussion, recommend new keywords for data streaming, and facilitate social media data collection. We provide the code so that other researchers can use these tools \footnote{https://github.com/mayasrikanth/DynamicMonitor}. Figure \ref{fig:flowchart} demonstrates the four components of our framework, which include data collection and storage, data analysis, and visualization.
\begin{figure*}[htbp]
\centering
\begin{tabular}{c}
\includegraphics[width = 0.9\textwidth]{figs/metoo_keyword_timelines.pdf}
\end{tabular}
\caption{Keywords update (add and remove) by the dynamic monitor for data collection in historical \#MeToo simulation. The evolving keyword set aligns with real-world events: throughout early to mid 2017, the monitor tracks important movements (e.g. `\#womensmarchoonwashington in Jan 2017 and `\#sciencemarch' in Apr 2017) and gender-equality hashtags (`\#womenwhowork', `\#womeninpolitics). In late 2017, the monitor picks up the viral \#MeToo movement with `\#metoo' in Oct, as well as related figures like `\#harveyweinstein' and `\#alfranken'. The monitor also picks up traditionally anti-\#MeToo hashtags throughout the simulation, with `\#sjw', `\#feminismishate', `\#mensrights', `\#fakecases'. }
\label{fig:MeTooKeywords}
\end{figure*}
\begin{figure*}[htbp]
\centering
\begin{tabular}{c}
\includegraphics[width = 0.9\textwidth, trim={0cm, 4cm, 0cm, 0cm}, clip]{figs/inauguration_keywords_timeline.pdf}
\end{tabular}
\caption{Keywords update (add and remove) by the dynamic monitor for data collection in real-time case study of 2021 \#inauguration discussions on Twitter. Emergence of `\#nationalguard' on Jan 12 likely corresponds with the 2021 capitol protests (as does `\#capitolriot' and `\#capitolbuilding' on Jan 16), and emergence of `databreach' on Jan 13 is likely linked to the hacking of Parler's data. As time goes on, we see that the monitor begins to focus on politicians, detecting various forms of `\#biden' (`\#joebiden', `\#bidenharris' `\#bideninauguration', `\#bidenharrisinauguration'), `\#trump' (`\#donaldtrump'), and `\#kamalaharris'. After inauguration on Jan 20th, all \#inauguration discussions naturally lost traction (see Figure \ref{fig:casestudy1} for declining frequencies after the fact).}
\label{fig:InaugurationKeywords}
\end{figure*}
Our work makes the following contributions:
\begin{enumerate}
\item By combining word embeddings with predictive time series modeling, our methodology allows for fully-automated, semi-automated, or completely by-hand updating of keywords used to pull social media data over extended periods.
\item By providing an online interface, we give analysts the means to visualize the various components of the dynamic monitor: using our code, researchers can oversee the operations of a dynamic monitor and alter the course of their data collection.
\item By conducting simulations and case studies with both historical data from the prominent 2017 \#MeToo movement and the tumultuous 2021 Presidential Inauguration, we put the deployed components of our process into use on real-world, fast-evolving online discussions to demonstrate their efficacy.
\end{enumerate}
More specifically, Figure \ref{fig:MeTooKeywords} summarizes our dynamic monitor updates of keywords in the simulation, where we only use GloVe embeddings and keyword frequency information to model the keywords relations and apply human judgement for keywords updates. Our method achieves 37.1\% higher F-1 score on average than the traditional static monitor in tracking the top trending keywords for each month in 2017 (Table \ref{fig:F1Table}). Figure \ref{fig:InaugurationKeywords} shows dynamic monitor updates of keywords used to collect real-time \#inauguration data in 2021.
\section{Method}
At present, we have built and deployed the word embedding component of our dynamic keyword and hashtag monitoring process (focusing at this point on hashtags). The predictive modeling component of our process is still under development, as we discuss later, we have implemented some time series predictive models and will demonstrate their utility in future research.
\subsection{Word Embeddings to Model Keyword Relations}
At time $t$, we train the GloVe model \cite{pennington2014glove} using $K_t$ to produce 50-dimensional word embedding representations $G_t$ of all tokens in our filtered corpus $K_t$.
GloVe can represent linear substructures in data. It is a log-bilinear model with a weighted least-squares objective, and aims to learn word vectors such that their dot product equals the logarithm of the words’ probability of co-occurrence. In the resulting word vector space, cosine similarity indicates linguistic or semantic similarity between two words, while vector differences capture analogies between pairs of words. These embeddings allow us to project each token into euclidean space, where we can use distance metrics to measure "closest" neighbors to each of our keywords $s \in s_t$.
\subsection{Predictive Method to Forecast Keyword Frequency Trend }
Our latest implementation supports time series forecasting with ARIMA (auto-regressive integrated moving average) for univariate frequency data. Specifically, we visualize keyword frequencies and predict their trajectory within a confidence interval to determine whether a discussion topic is increasing or decreasing.
Prior to any forecasting, we apply log transform to all corpus counts in order to stabilize the variance \citep{Lutkepohl2012} and induce stationarity in all series by differencing lags. For each keyword, we grid search ARIMA model parameters and select those which maximize performance (minimize mean-squared error) on the validation set. The best model is then used to forecast frequencies $10-15$ time-steps into the future.
While linear models like ARIMA do not outperform larger deep learning models when data is abundant, they can be informative in earlier stages of data streaming when textual data is sparse. For applications with more abundant data, we plan to include options for training more robust deep learning models in the future.
In order to collect enough data to allow for meaningful time series prediction during the data collection process, our dynamic implementation design pulls $7$ days worth of data from Twitter's REST API using the starting set of keywords as the query.
\subsection{Proposed Algorithm}
With $G_t$ and $P_t$ defined above, our algorithm proceeds as follows.
For each $s \in s_t$ at time $t$, from embedding space $G_t$ we find $C_{30}$, the set of 30 closest neighbors to keyword $s$, defining closest with the cosine similarity metric. From this set of thirty words, we choose the most relevant hashtags or mentions in the domain or event we are tracking. We define these neighbors as $n_s \in C_s \subseteq C_{30}$.
For each $n_s \in C_s$, we use our time series model $P_t$ to predict future frequencies for the hashtag or mention. If we predict $n_s$ is declining in future time periods, we drop these values from the set $C_s$. We define $C_{s'} \subseteq C_s$ as the set of neighbors \emph{without} declining time series predictions.
Finally, define $s_{t+1}$, the set of keywords we track in the next time period, as $C_{s'}$.
\begin{algorithm}[hbt!]
\DontPrintSemicolon
\KwInput{$s_t$: keyword set at $t$, $K_{t}$: filtered corpus at $t$}
\KwOutput{\newline $s_{t+1} = \{\}$}
\KwData{\newline $G_{t} \leftarrow$: obtain 50-dimension GloVe embeddings trained on $K_{t}$. $P_{t} \leftarrow$: update time series models with latest frequency data from corpus}
\For{$s \in s_{t}$} {
{1. $C_{30}$: 30 closest neighbors to $s$ in embedding space $G_{t}$.}\\
{2. $C_{s}$: choose relevant hashtags or mentions from $C_{30}$ } \\
{3. $C_{s'}$: Discard hashtags from $C_{s}$ with declining trend lines in time series prediction or low corpus counts.}\\
{4. $s_{t+1} \longleftarrow C_{s'}$}
}
\textbf{Return} ${s_{t+1}}$.
\caption{Dynamic Algorithm}
\label{humaninput}
\end{algorithm}
\subsection{Platform Features vs. Case Study Methods}
The case studies we conduct on historical \#MeToo data and real-time 2021 \#inauguration discussions use a simplified version of the algorithm above. In both studies, the human-assisted decision process to drop or add keywords is informed \textbf{only by corpus frequencies and GloVe embeddings}: we do not use time series models to
predict the trajectory of keywords.
However, we observed from independent analysis that even linear time series models can produce reasonable short-term estimates of keyword frequency trajectories. Thus, our data visualization platform code includes interactive charts with time series forecasts (see Figure \ref{fig:interface2}), as well as code for training ARIMA on real-time Twitter data.
\section{Task Formulation}
\paragraph{Notation} We define $s_t$ as the set of keywords we are interested in tracking at timestep $t$. Using $s_t$, we can filter a corpus $K_t$ using APIs from different social media platforms. For example, Twitter provides various APIs that filter tweets containing specific keywords and hashtags. We define $G_t$ as the semantic representation of words included in the filtered corpus. In prevalent word embedding models, we can use a vector to represent each word's semantic relation with other words. We use $P_t$ to represent the future trends for each word. $P_t$ can be directions (increases or decreases) or specific frequencies. Our goal then is to build a system for dynamic data collection such that maximizes the information coverage for evolving discussions around certain topics. As such, given keywords $s_t$ and the corresponding corpus $K_t$, we aim to update the keyword set $s_{t+1}$ according to the patterns in $G_t$ and $P_t$.
\paragraph{Data Collection and Storage} We assume data collection using the APIs from social media platforms can be conducted efficiently. In practice, there exist additional difficulties in conducting this data filtering in a large scale and storing the data reliably \cite{cao2020reliable}. However, in this paper, we focus on the process of decision making for updating the keywords and the visualization for facilitating the decision making when human intervention is needed.
\paragraph{Decision Making on Updating Keywords} Based on $K_0,...,K_t$, we first need to generate the semantic representation $G_t$ and the future trend $P_t$. Given $G_t$ and $P_t$, there are different many ways to determine the updated keywords, and in many contexts it is desirable that this updating process remain flexible. When there is a smooth trend in topic shifting or there are emerging events that slowly change the direction or sentiment of the discussion, fully-automatic updating using simple rules is sufficient. However, in some contexts human intervention guided by information of $G_t$ and $P_t$ is needed. For example, swift topic shifting is hard to forecast using historical data. In these situations, human intervention may be helpful.
\paragraph{User Interface} Demonstrating $G_t$ and $P_t$ for decision making is important for closing the loop for future data collection. The interface should consist of user-friendly elements that enable clear illustration of the semantic relation between keywords and the forecasts of future trends.
\section{Related Work}
\paragraph{Information Retrieval}
Existing information retrieval methods like Okapi BM25 \citep{Sparck_Jones_2000} which take a probabilistic approach to ranking documents use a complex and rigid weighting scheme with various parameters that require refinement. Deep learning information retrieval methods like BERT \citep{deshmukh2020irbert} are non-transparent and require abundant data to produce sensible rankings. These approaches focus on optimizing information retrieval in largely static databases and returning results that are most similar or relevant to user's query. Our task is different: social media data is extremely dynamic and prediction intervals occur on smaller scales. Therefore, rankings must be produced efficiently and robustly on relatively small datasets which are streamed real-time. Further, the search process for new keywords begins with a known keyword, but could end in new discoveries which alter the course of data collection. Our target user is uncertain about what they are looking for, and their ``information retrieval'' process is highly informed by real-time predictions and newly exposed keywords. Finally, our application requires flexibility in ranking documents (or keywords): in one use case, corpus frequency may be the strongest indicator of relevance, while in another use case, semantic similarity to an existing set of keywords may be more important.
\paragraph{Word Embedding Models} Previous work has shown that word embedding models serve as an efficient information retrieval mechanism on vast corpora \citep{mci/Galke2017}. In previous work, we show the ability of word embedding models to uncover conversational threads and emergent hashtags in online conversation \citep{liu_etal_2019}. As our application requires real-time processing of streamed social media data, we iteratively train GloVe word embeddings \citep{pennington2014glove} on incoming data to efficiently create a vector representation of the corpus and retrieve information about closest neighbors based on metrics like cosine similarity. To provide greater flexibility and transparency to the ranking scheme, our final implementation sorts keywords based on a linear combination of keyword frequency information and embedding information.
\paragraph{Keyword Selection}
Prior research has focused on the problem of keyword selection \citep{Wang_Chen_Liu_Emery_2016}. This is different from our work in that we assume that the analyst begins with subject-matter expertise that provides keywords and hashtags to seed the dynamic monitor. Other related work has used similar static keyword search query approaches \cite[e.g.][]{oconner:2010,conover_2012,barbera_etal_2015}, which fall short in the study of dynamic debates with rapidly evolving conversation. Another approach uses semi-automated approaches for keyword selection and search by crafting a semi-supervised dynamic keyword methodology \citep{zhengetal2017}. Their approach differs from ours, as we use relatively easy-to-estimate word embedding models and straightforward time-series predictive models, making our approach more intuitive and likely faster computationally. Further, related deep learning methods \cite{Wang_Chen_Liu_Emery_2016} often lack transparency and require vast data to perform well, which precludes their usage in applications with sparse data. Finally, related work has proposed semi-automated keyword selection combining computer and human input \cite{king_lam_roberts_2017}: these are based on complex rules are more effort-intensive, as researchers must themselves apply these rules in their decision process.
We offer an AI-driven dynamic keyword searching methodology that is fast and efficient (like fully-automated methods), yet provides more transparency and intuition than these methods. We also offers unprecedented visualizations and a interface for user decision making.
|
\section{Introduction}
The deterministic logarithmic Schr\"odinger equation
has wide applications in quantum mechanics, quantum optics, nuclear physics, transport and diffusion phenomena, open quantum system, Bose--Einstein condensations and so on (see e.g. \cite{AZ03,BM76,Hef85,Lau08,Yas78,Zlo10}).
It takes the form of
\begin{align*}
\partial_t u(t,x)&=\mathbf i \Delta u(t,x)+\mathbf i \lambda u(t,x)\log(|u(t,x)|^2)+\mathbf i V(t,x,|u|^2)u(t,x),\; x\in \mathbb R^d,\; t>0,\\\nonumber
u(0,x)&=u_0(x),\; \quad x\in \mathbb R^d,
\end{align*}
where $\Delta$ is the Laplacian operator on $\mathcal{O} \subset \mathbb{R}^{d}$ with $\mathcal{O}$ being either $\mathbb{R}^{d}$ or a bounded domain with homogeneous Dirichlet or periodic boundary condition,
$t$ is time, $x$ is spatial coordinate, $\lambda\in \mathbb R/\{0\}$ characterizes the force of nonlinear interaction, and $V$ is a real-valued function. While retaining many of the known features of the linear Schr\"odinger equation, Bialynicki--Birula and Mycielski show that only such a logarithmic nonlinearity satisfies the condition of
separability of noninteracting systems (see \cite{BM76}). The logarithmic nonlinearity makes the logarithmic Schr\"odinger equation unique among nonlinear wave equations.
For instance, the longtime dynamics of the logarithmic Schr\"odinger equation is essentially different from the Schr\"odinger equation. There is a faster dispersive phenomenon when $\lambda<0$ and the convergence of the modulus of the solution to a universal Gaussian profile (see \cite{CG18}), and no dispersive phenomenon when $\lambda>0$ (see \cite{Caz83}).
In this paper, we are mainly focus on the well-posedness of the following stochastic logarithmic Schr\"odinger (SlogS) equation,
\begin{align}\label{SlogS}
d u(t)&=\mathbf i \Delta u(t)dt+\mathbf i \lambda u(t)\log(|u(t)|^2)dt+ \widetilde g(u)\star dW(t),\; t>0\\\nonumber
u(0)&=u_0,
\end{align}
where $W(t)=\sum_{k\in \mathbb N^+}Q^{\frac 12}e_k\beta_k(t)$, $\{e_k\}_{k\in\mathbb N^+}$ is an orthonormal basis of $L^2(\mathcal O;\mathbb C)$ with $\{\beta_k\}_{k\in\mathbb N^+}$ being a sequence of independent Brownian motions on a probability space $(\Omega,$
$\mathcal F, (\mathcal F_t)_{t\ge 0},\mathbb P).$
\iffalse
Here, $\widetilde g$ is a continuous function, and $\circ$ denotes the stochastic integral in Stratonovich sense.
The parameter $\lambda\in \mathbb R$ and $W(t)=\sum_{k\in \mathbb N^+}Q^{\frac 12}e_k\beta_k(t)$, $\{e_k\}_{k\in\mathbb N^+}$ is an orthonormal basis of $L^2(\mathcal O)$ with $\{\beta_k\}_{k\in\mathbb N^+}$ being a sequence of independent Brownian motions on a probability space $(\Omega,$ $\mathcal F, (\mathcal F_t)_{t\ge 0},\mathbb P).$ The spatial domain $\mathcal O\subset \mathbb R^d, d\in \mathbb N^+,$ is either $\mathbb R^d$ or a bounded domain equipped with homogenous Dirichlet or periodic boundary condition.
\fi
Here $\widetilde g$ is a continuous function and $\widetilde g(u)\star dW(t)$ is defined by
\begin{align*}
&\widetilde g(u)\star dW(t) =-\frac 12\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u|^2)|^2u\Big)dt\\\nonumber
&\quad-\mathbf i \sum_{k\in \mathbb N^+}g(|u|^2)g'(|u|^2) |u|^2 u
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_k dt+ \mathbf i g(|u|^2)u dW(t)
\end{align*}
if $\widetilde g(x)=\mathbf i g(|x|^2)x$, and by
$$\widetilde g(u)\star dW(t)= dW(t)$$ if $\widetilde g=1$.
We would like to remark that when $W$ is $L^2(\mathcal O;\mathbb R)$-valued and $\widetilde g(x)=\mathbf i g(|x|^2)x$, $\widetilde g(u)\star dW(t)$ is just the classical Stratonovich integral.
The SlogS equation \eqref{SlogS} could be derived from the deterministic model by using Nelson's mechanics {\cite{Nel66}}. Applying the Madlung transformation $u(t,x)=\sqrt{\rho(t,x)}e^{\mathbf i {S(t,x)} }$, \cite{Nas86} obtains a fluid expression of the solution as follows,
\begin{align*}
\partial_tS(t,x)&=-|\nabla S(t,x)|^2- \frac {1}4 \frac {\delta I}{\delta \rho}(\rho(t,x)){+\lambda}\log(\rho){+V(t,x,\rho(t,x))},\; \\\nonumber
\partial_t\rho(t,x)&=-2div(\rho(t,x) \nabla S(t,x)),\; S(0,x_0)=S_0(x_0)\;, \rho(0)=\rho_0,
\end{align*}
where $I(\rho)=\int_{\mathcal O}|\nabla \log(\rho)|^2$$\rho dx$ is the Fisher information.
If $V$ is random and fluctuates rapidly, the term $\mathbf iV u$ can be approximated by some multiplicative Gaussian noise $\widetilde g(u)\dot W,$ which plays an important role in the theory of measurements continuous in time in open quantum systems (see e.g. \cite{BG09}). Then we could use the inverse of Madlung transformation and formally obtain the stochastic logarithmic Schr\"odinger equation
\begin{align*}
\partial_t u(t,x)&=\mathbf i \Delta u(t,x)+\mathbf i \lambda u(t,x)\log(|u(t,x)|^2)+\widetilde g(u(t,x))\dot W(t,x),\; x\in \mathcal O,\; t>0\\
u(0,x)&=u_0(x), \quad x\in \overline{\mathcal O}.
\end{align*}
The main assumption on $W$ and $\widetilde g$ is stated as follows.
\begin{ap}\label{main-as}
The diffusion operator is the Nemystkii operator of $\widetilde g$. Wiener process $W$ and $\widetilde g$ satisfies one of the following condition,
\begin{enumerate}
\item[{\rm Case 1.}] $\{W(t)\}_{t\ge 0}$ is $L^2(\mathcal O;\mathbb C)$-valued and $\widetilde g=1;$
\item[{\rm Case 2.}] $\{W(t)\}_{t\ge 0}$ is $L^2(\mathcal O;\mathbb C)$-valued,
$\widetilde g(x)=\mathbf ig(|x|^2)x,$ $g \in \mathcal C^2_b(\mathbb R)$ and satisfies the growth condition
\begin{align*}
\sup_{x\in [0,\infty)}|g(x)|+\sup_{x\in [0,\infty)}|g'(x)x|+\sup_{x\in [0,\infty)}|g''(x)x^2|\le C_g;
\end{align*}
\item[{\rm Case 3.}] $ {\{W(t)\}_{t\ge 0}}$ is $L^2(\mathcal O;\mathbb R)$-valued, $\widetilde g(x)=\mathbf ig(|x|^2)x,$ $g \in \mathcal C^1_b(\mathbb R)$ and satisfies the growth condition
\begin{align*}
\sup_{x\in [0,\infty)}|g(x)|+\sup_{x\in [0,\infty)}|g'(x)x|\le C_g.
\end{align*}
\end{enumerate}
\end{ap}
\begin{ap}\label{ass-strong}
Assume that $g$ satisfies
\begin{align}\label{con-g}
(x+y)(g(|x|^2)-g(|y|^2))\le C_g|x-y|, \quad x,y\in [0,\infty).
\end{align}
When $W(t)$ is $L^2(\mathcal O;\mathbb C)$-valued, we in addition assume that $g$ satisfies following one-side Lipschitz continuity
\begin{align}\label{con-g1}
|(\bar y-\bar x)(g'(|x|^2)g(|x|^2)|x|^2x-g'(|y|^2)g(|y|^2)|y|^2y)| \le C_g|x-y|^2, \quad x,y\in \mathbb C.
\end{align}
\end{ap}
A typical example is $\widetilde g(u)=u$, and then Eq. \eqref{SlogS} becomes the SlogS equation driven by linear multiplicative noise in \cite{BRZ17}.
There are two main difficulties in proving the well-posedness of the SlogS equation.
On one hand, the random perturbation in SlogS equation destroys a lot of physical conservation laws, like the mass and energy conservation laws in Case 1 and Case 2, and the energy conservation law in Case 3.
Similar phenomenon has been observed in stochastic nonlinear Schr\"odinger equation with polynomial nonlinearity (see \cite{BD03}).
On the other hand, the logarithmic nonlinearity in SlogS equation is not locally Lipschitz continuous.
The contraction mapping arguments via Strichartz estimates (see e.g. \cite{BRZ16,BD03,Hor18}) for stochastic nonlinear Schr\"odinger equation with smooth nonlinearity are not applicable here.
We only realize that in Case 2, when the driving noise is a linear multiplicative noise ($\widetilde g(u)=\mathbf i u$), \cite{BRZ17} uses a rescaling technique, together with maximal monotone operator theory to obtain a unique global mild solution in some Orlicz space.
As far as we know, there has no results concerning the well-posedness of the SlogS equation driven by additive noise or general multiplicative noise.
To show the well-posedness of the considered model, we introduce an energy regularized problem inspired by \cite{BCST19} where the authors use the regularized problem to study error estimates of numerical methods for deterministic logarithmic Schr\"odinger equation.
The main idea is firstly constructing a proper approximation of $\log(|x|^2)$ denoted by $f_{\epsilon}(|x|^2).$ Then it induces the regularized entropy
$F_{\epsilon}$ which is an approximation of the entropy $F(\rho)=\int_{\mathcal O}(\rho\log(\rho)-\rho)dx,$ where $\rho=|u|^2$.
The RSlogS equation is defined by
\begin{align}\label{Reg-SlogS}
du^{\epsilon}=\mathbf i \Delta u^{\epsilon}dt+\mathbf i u^{\epsilon} f_{\epsilon}(|u^{\epsilon}|^2)dt+ \widetilde g(u^{\epsilon})\star dW(t),
\end{align}
whose regularized energy is $\frac 12\|\nabla u\|^2+\frac {\lambda}2 F_{\epsilon}(u)$.
Denoting $\mathbb H:=L^2=L^2(\mathcal O;\mathbb C)$ with the product $\<u,v\>:=\int_{\mathbb R^d} Re(u\bar v)dx,$ for $u,v\in \mathbb H,$
we obtain the existence and uniqueness of the solution of regularized SlogS equation by proving $\epsilon$-independent estimate in $\mathbb H^1:=W^{1,2}$ and the weighted $L^2$-space $L^2_{\alpha}:=\{v\in L^2 | \; x \mapsto (1+|x|^2)^{\frac \alpha 2}v(x)\in L^2\}$ with the norm $\|v\|_{L^2_\alpha(\mathbb R^d)}:=\|(1+|x|^2)^{\frac \alpha 2}v(x)\|_{L^2(\mathbb R^d)}$.
Then we are able to prove that the limit of $\{u^{\epsilon}\}_{\epsilon>0}$ is convergent to a unique stochastic process $u$ which is shown to be the unique mild solution of \eqref{SlogS}. Meanwhile, the sharp convergence rate of $\{u^{\epsilon}\}_{\epsilon>0}$ is given when $\mathcal O=\mathbb R^d,$ or $\mathcal O$ is a bounded domain in $\mathbb R^d$ equipped with homogenous Dirichlet or periodic boundary condition.
Our main result is formulated as follows.
\begin{tm}\label{mild-general}
Let $T>0$, Assumptions \ref{main-as} and \ref{ass-strong} hold, $u_0\in \mathbb H^1\cap L^2_{\alpha},$ $\alpha\in(0,1],$ be $\mathcal F_0$ measurable and has any finite $p$th moment. Assume that $\sum_{i\in\mathbb N^+} \|Q^{\frac 12}e_i\|_{L^2_{\alpha}}^2+\|Q^{\frac 12}e_i\|_{\mathbb H^1}^2 <\infty$ when $\widetilde g=1$ and that $\sum_{i\in\mathbb N^+} \|Q^{\frac 12}e_i\|_{\mathbb H^1}^2+\|Q^{\frac 12}e_i\|_{W^{1,\infty}}^2<\infty$ when $\widetilde g(x)= \mathbf i g(|x|^2)x$.
Then there exists a unique mild solution $u\in C([0,T];\mathbb H)$ of Eq. \eqref{SlogS}.
Moreover, for $p\ge 2$, there exists $C(Q,T,\lambda,p,u_0)>0$ such that
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u(t)\|_{\mathbb H^1}^p\Big]+\mathbb E\Big[\sup_{t\in [0,T]}\|u(t)\|_{L^2_{\alpha}}^p\Big]\le C(Q,T,\lambda, p,u_0).
\end{align*}
\end{tm}
When $W(t)$ is $L^2(\mathcal O;\mathbb R)$-valued, the well-posedness of SlogS equation with a super-linearly growing diffusion coefficient is also proven (see Theorem \ref{mild-d1}).
The reminder of this article is organized as follows. In section 2, we introduce the RSlogS equation and show the local well-posedness of RSlogS equation driven by both additive and multiplicative noise. Section 3 is devoted to $\epsilon$-independent estimate of the mild solution in $\mathbb H^1$ and $L^2_{\alpha}$ of the RSlogS equation. In section 4, we prove the main result by passing the limit of the sequence of the regularized mild solutions and providing the sharp strong convergence rate. Several technique details are postponed to the Appendix. Throughout this article, $C$ denotes
various constants which may change from line to line.
\section{Regularized SLogS equation}\label{sec-loc}
In this section, we show the well-posedness of the solution for Eq. \eqref{Reg-SlogS} (see Appendix for the definition of the solution). We would like to remark that there are several choices of the regularization function $f_{\epsilon}(|x|^2)$. For instance, one may take $f_{\epsilon}(|x|^2)=\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$ (see Lemma \ref{sec-reg} in Appendix for the necessary properties) or $f_{\epsilon}(|x|^2)=\log(|x|^2+\epsilon)$ (see e.g. \cite{BCST19} and references therein for more choices of regularization functions). If the regularization function $f_{\epsilon}$ enjoys the same properties of $\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$, then one can follow our approach to obtain the well-posedness of Eq. \eqref{Reg-SlogS}. In the following, we first present the local well-posedness of Eq. \eqref{Reg-SlogS}, and then derive global existence and uniform estimate of its solution. For simplicity, we always assume that $0<\epsilon \ll 1$.
\subsection{Local well-posedness of Regularized SLogS equation}
In this part, we give the detailed estimates to get the local well-posedness in $\mathbb H^2$ of Eq. \eqref{Reg-SlogS} if $d\le 3$ via the regularization function $f_{\epsilon}(|x|^2)=\log(|x|^2+\epsilon).$
In the case of $d\ge 3$, one could use the regularization function like $f_{\epsilon}(|x|^2)=\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$ to get the local well-posedness in $\mathbb H^1$. Assume that $u_0\in \mathbb H^2$ when using the regularization $\log(|x|^2+\epsilon)$ and that $u_0\in \mathbb H^1$ when applying the regularization $\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2}).$
\iffalse
Recall that the regularized SlogS equation
\begin{align}\label{mul-reg}
du^{\epsilon}=\mathbf i \Delta u^{\epsilon} dt+\mathbf i \lambda u^{\epsilon}f_{\epsilon}(|u^{\epsilon}|^2)dt+ \mathbf i g(u^{\epsilon}) \circ dW(t).
\end{align}
\fi
\iffalse
{\color{red}
Under Assumption \ref{main-as}, when $\widetilde g(x)=\mathbf i |x|^2x,$
\begin{align*}
du^{\epsilon}&=\mathbf i \Delta u^{\epsilon} dt+\mathbf i \lambda u^{\epsilon}f_{\epsilon}(|u^{\epsilon}|^2)dt
-\frac 12\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon}|^2)|^2u^{\epsilon}\Big)dt\\
&\quad-\mathbf i \sum_{k\in \mathbb N^+}g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2) |u^{\epsilon}|^2 u^{\epsilon}
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_k dt+ \mathbf i g(|u^{\epsilon}|^2)u^{\epsilon} dW(t).
\end{align*}
}
When $W(t)$ becomes $L^2(\mathcal O;\mathbb R)$-valued, the term $$\mathbf i\sum_{k\in \mathbb N^+}g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2) |u^{\epsilon}|^2 u^{\epsilon}
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_k =0.$$
\fi
Denote by $\mathbb M_{\mathcal F}^p(\Omega;C([0,T];\mathbb H^2))$ with $p\in [1,\infty)$ the space of process $v: [0,T]\times \Omega \to \mathbb H^2$ with continuous paths in $\mathbb H^2$ which is $\mathcal F_t$-adapted and satisfies
\begin{align*}
\|v\|^p_{\mathbb M_{\mathcal F}^p(\Omega;C([0,T];\mathbb H^2))}
:=\mathbb E\Big[\sup_{t\in[0,T]}\|u(t)\|_{\mathbb H^2}^p\Big]<\infty.
\end{align*}
Let $\tau\le T$ be an $\mathcal F_t$-stopping time.
And we call $v\in \mathbb M_{\mathcal F}^p(\Omega;C([0,\tau);\mathbb H^2))$, if there exists $\{\tau_n\}_{n\in \mathbb N^+}$ with $\tau_n\nearrow \tau$ as $n\to \infty$ a.s., such that $v\in \mathbb M_{\mathcal F}^p(\Omega;C([0,\tau_n];\mathbb H^2))$ for $n\in \mathbb N^+.$
Next we show the existence and uniqueness of the local mild solution (see Definition \ref{global} in Appendix).
For the sake of simplicity, let us ignore the dependence on $\epsilon$ and write $u_R:=u^{\epsilon}_R,$ where $u_R$ is the solution of the truncated equation
\begin{align}\label{trun-reg}
du_R=&\mathbf i \Delta u_R dt+\mathbf i \lambda \Theta_R(u_R,t) u_Rf_{\epsilon}(|u_R|^2)dt\\\nonumber
&-\frac 12 \Theta_R(u_R,t) \sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u_R|^2)|^2u_R\Big)dt\\\nonumber
& -\mathbf i \Theta_R(u_R,t)\sum_{k\in \mathbb N^+}g(|u_R|^2)g'(|u_R|^2) |u_R|^2 u_R
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_kdt\\\nonumber
&+ \Theta_R(u_R,t) \mathbf i g(|u_R|^2)u_R dW(t).
\end{align}
Here, $\Theta_R(u,t):=\theta_R(\|u\|_{\mathcal C([0,t];\mathbb H^2)}), R>0,$ with
a cut-off function $\theta_R$, that is, a positive $C^{\infty}$ function on $\mathbb R^+$ which has a compact support, and
\begin{equation*}
\theta_R(x)=
\left\{
\begin{aligned}
&0, \quad \text{for}\quad x\ge 2R,\\
&1,\quad \text{for}\quad x\in [0,1].
\end{aligned}
\right.
\end{equation*}
\iffalse
One may take other type truncated function, such as a bounded Lipschitz function,
$$\theta_R(x)=1, \; x\in [0,R], \; \theta_R(x)=0, \; x\in [2R,\infty)\; \text{and}\; \theta_R(x)=2-\frac x{R}, \; x\in (R,2R].$$
\fi
\begin{lm}\label{trun}
Let Assumption \ref{main-as} hold, $d\le 3$, and $f_{\epsilon}(|x|^2)=\log(|x|^2+\epsilon)$. Assume in addition that $g \in \mathcal C^3_b(\mathbb R)$ when $W(t)$ is $L^2(\mathcal O;\mathbb C)$-valued and that $g \in \mathcal C^2_b(\mathbb R)$ when $W(t)$ is $L^2(\mathcal O;\mathbb R)$-valued.
Suppose that the $Q$-Wiener process $W(t)$ satisfies $\sum_{i\in\mathbb N^+}\|Q^{\frac 12}e_i\|_{\mathbb H^2}^2<\infty$, and $u_0\in \mathbb H^2$ is $\mathcal F_0$-measurable and has any finite $p$th moment. Then there exists a unique global solution to \eqref{trun-reg} with continuous $\mathbb H^2$-valued path.
\end{lm}
{\noindent \bf{Proof}}
Let $S(t)=\exp(\mathbf i\Delta t)$ be the $C_0$-group generated by $\mathbf i\Delta.$
For fixed $R>0$, we use the following notations, for $t\in [0,T],$
{\small
\begin{align*}
\Gamma_{det}^R u(t):&=\mathbf i\int_0^t S(t-s) \Big( \Theta_R(u,s)\lambda f_{\epsilon}(|u(s)|^2)u(s)\Big)ds,\\
\Gamma_{mod}^R u(t):&=-\frac 12\int_0^t S(t-s)\Big(\Theta_R(u,s)\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u(s)|^2)|^2u(s)
\Big)\Big)ds\\
-\mathbf i \int_0^tS(t-s)&\Big(
\Theta_R(u,s)\sum_{k\in \mathbb N^+}g(|u(s)|^2)g'(|u(s)|^2) |u(s)|^2 u(s)
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_k \Big)ds,\\
\Gamma_{Sto}^R u(t):&=\mathbf i\int_0^t S(t-s) \Big(\Theta_R(u,s)g(|u(s)|^2)u(s)\Big)dW(s).
\end{align*}
W}e look for a fixed point of the following operator given by
\begin{align*}
\Gamma^R u(t):=S(t)u_0+\Gamma_{det}^R u(t)+\Gamma_{mod}^R u(t)
+\Gamma_{Sto}^R u(t), u\in \mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2)),
\end{align*}
where $r$ will be chosen later.
The unitary property of $S(\cdot)$ yields that
$$\|S(\cdot)u_0\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}\le \|u_0\|_{\mathbb H^2}.$$
Now, we define a stopping time $\tau=\inf\{t\in [0,T]: \|u\|_{C([0,t];\mathbb H^2)}\ge 2R\}\wedge r.$
By using the properties of $\log(\cdot)$ and the Sobolev embedding $\mathbb H^2\hookrightarrow L^\infty(\mathcal O)$, we have
\begin{align*}
\|\Gamma_{det}^R u\|_{C([0,r];\mathbb H^2)}&\le Cr|\lambda| \max(|\log(\epsilon)|,\log(\epsilon+4R^2))(\sup_{t\in [0,r]}\|u(t)\|_{\mathbb H^2})\\
&\quad+Cr|\lambda|(1+\epsilon^{-1})\Big(\sup_{t\in [0,r]}\|u(t)\|_{\mathbb H^2}+\sup_{t\in [0,r]}\|u(t)\|_{\mathbb H^2}^{2}\Big)\\
&\le C(\epsilon,\lambda)r(1+2R+R^2),
\end{align*}
\iffalse
\begin{align*}
&\|\Gamma_{mod}^R u\|_{C([0,r];L^2)}\\
&\le C(g,\lambda) r \sum_{k} |Q^{\frac 12}e_k|_{L^{\infty}}^2\sup_{s\in[0,r]}\Big(\|u(s)\|+\|u(s)\|_{L^6}^3\Big),\\
&\|\Gamma_{mod}^R u\|_{C([0,r];\mathbb H^1)}\\
&+C(g,\lambda) \Big\|\int_0^r\sum_{k} \nabla Q^{\frac 12}e_k Q^{\frac 12}e_k
\Big(|g(|u(s)|^2)|^2u(s)+g(|u(s)|^2)g'(|u(s)|^2)2|u(s)|^2u(s)\Big)ds\Big\|\\
& +C(g,\lambda) \Big\|\int_0^r \sum_{k} |Q^{\frac 12}e_k|^2
\Big(|g(|u(s)|^2)g'(|u(s)|^2)4Re(\bar u\nabla u)u+ |g(|u(s)|^2)|^2\nabla u \\
&\quad+g(|u(s)|^2)g'(|u(s)|^2)2|u(s)|^2\nabla u(s)+g(|u(s)|^2)g'(|u(s)|^2)4Re(\bar u\nabla u) u(s)\\
&\quad +|g'(|u(s)|^2)|^2 4 Re(\bar u\nabla u)|u(s)|^2 u(s)+g(|u(s)|^2)g''(|u(s)|^2) 4 Re(\bar u\nabla u) |u(s)|^2 u(s)\Big)ds\Big\|,\\
&\|\Gamma_{mod}^R u\|_{C([0,r];\mathbb H^2)}\\
&+C(g,\lambda) \Big\|\int_0^r\sum_{k} \nabla Q^{\frac 12}e_k \nabla Q^{\frac 12}e_k
\Big(|g(|u(s)|^2)|^2u(s)+g(|u(s)|^2)g'(|u(s)|^2)2|u(s)|^2u(s)\Big)ds\Big\|\\
&+C(g,\lambda) \Big\|\int_0^r\sum_{k} \Delta Q^{\frac 12}e_k Q^{\frac 12}e_k
\Big(|g(|u(s)|^2)|^2u(s)+g(|u(s)|^2)g'(|u(s)|^2)2|u(s)|^2u(s)\Big)ds\Big\|,\\
&\quad +C(g,\lambda) \Big\|\int_0^r\sum_{k} \nabla Q^{\frac 12}e_k Q^{\frac 12}e_k
\Big(4g(|u(s)|^2)g'(|u|^2)Re(\bar u\nabla u)u(s)
+|g(|u(s)|^2)|^2\nabla u\\
&\quad+|g'(|u(s)|^2)|^2 4 Re(\bar u\nabla u)|u(s)|^2u(s)
+g(|u(s)|^2)g''(|u(s)|^2)4 Re(\bar u\nabla u) |u(s)|^2u(s)\\
&\quad+g(|u(s)|^2)g'(|u(s)|^2)2|u(s)|^2\nabla u(s)
+g(|u(s)|^2)g'(|u(s)|^2)4 Re(\bar u\nabla u) u(s)\Big)ds\Big\|\\
& +C(g,\lambda) \Big\|\int_0^r \sum_{k} \nabla Q^{\frac 12}e_k Q^{\frac 12}e_k
\Big(|g(|u(s)|^2)g'(|u(s)|^2)(8Re(\bar u\nabla u)u +2|u|^2u)\Big)ds\Big\|
\\
&+C(g,\lambda) \Big\|\int_0^r \sum_{k} |Q^{\frac 12}e_k|^2
\Big( g'(|u(s)|^2)g'(|u(s)|^2)(16 (Re(\bar u\nabla u))^2u+4Re(\bar u\nabla u)|u|^2u)\\
&\quad
+g(|u(s)|^2)g''(|u(s)|^2)(16 (Re(\bar u\nabla u))^2u+4Re(\bar u\nabla u)|u|^2u)\\
&\quad+g(|u(s)|^2)g'(|u(s)|^2) (8|\nabla u|^2u+8Re(\bar u \Delta u)u+12Re(\bar u \nabla u)\nabla u)+2|u|^2\Delta u\Big)ds\Big\|\\
&+C(g,\lambda) \Big\|\int_0^r \sum_{k} |Q^{\frac 12}e_k|^2
\Big(g(|u(s)|^2)g'(|u(s)|^2)2Re(\bar u\nabla u)\nabla u \Big)ds\Big\|\\
&+C(g,\lambda) \Big\|\int_0^r \sum_{k} |Q^{\frac 12}e_k|^2
\Big(|g'(|u(s)|^2)|^2 4 Re(\bar u\nabla u)|u(s)|^2 u(s)
+g'(|u(s)|^2)g''(|u(s)|^2)16 (Re(\bar u\nabla u))^2 |u|^2 u\\
&\quad+|g'(|u(s)|^2)|^2 (8Re(\bar u\Delta u)|u(s)|^2 u(s)+4|u|^2|\nabla u|^2u+4Re(\bar u\nabla u)|u(s)|^2 \nabla u(s))
\\
&\quad +C(g,\lambda) \Big\|\int_0^r \sum_{k} |Q^{\frac 12}e_k|^2\Big(
(g'(|u(s)|^2)g''(|u(s)|^2)+g(|u(s)|^2)g'''(|u(s)|^2)) 16 Re(\bar u\nabla u)^2 |u(s)|^2 u(s)
\\
&+g(|u(s)|^2)g''(|u(s)|^2)(8Re(\bar u\Delta u)|u(s)|^2 u(s)+4|u|^2|\nabla u|^2u+4Re(\bar u\nabla u)|u(s)|^2 \nabla u(s))
\Big)ds\Big\|.
\end{align*}
\fi
and
\begin{align*}
&\|\Gamma_{mod}^R u\|_{C([0,r];\mathbb H^2)}
\le C(\lambda,g)r\sum_{k\in \mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2\Big(R+R^3+R^5\Big).
\end{align*}
Integrating over $\Omega$ yields that
\begin{align*}
\|\Gamma_{det}^R u\|_{\mathbb M_{\mathcal F}^p(\Omega;C([0,r];\mathbb H^2))}+\|\Gamma_{mod}^R u\|_{\mathbb M_{\mathcal F}^p(\Omega;C([0,r];\mathbb H^2))}
\le& C(\lambda,g)r\sum_{k\in \mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2\Big(R+R^3+R^5\Big).
\end{align*}
The Burkerholder inequality yields that for $p\ge 2$,
\begin{align*}
\|\Gamma_{sto}^R u\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}
\le& C(g) \Big(\mathbb E\Big[\Big(\int_0^r\sum_{k\in\mathbb N^+} \|g(|u(s)|^2)u(s)Q^{\frac 12}{e_k}\|_{\mathbb H^2}^2ds\Big)^{\frac p2}\Big]\Big)^{\frac 1p}\\
\le& C(g)r^{\frac 12}\left(\sum_{k\in \mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2\right)^{\frac 12}\Big(R+R^3\Big).
\end{align*}
Therefore, $\Gamma^R$ is well-defined on $\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2)).$
Now we turn to show the contractivity of $\Gamma^R.$
Let $u_1, u_2\in \mathbb M_{\mathcal F}^p(\Omega;C([0,r];\mathbb H^2)),$ and define the stopping times
$\tau_j=\inf\{t\in [0,T]: \|u_j\|_{C([0,r];\mathbb H^2)}\ge2R\}\wedge r, j=1,2.$
For a fixed $\omega,$ let us assume that $\tau_1\le \tau_2$ without the loss of generality.
Then direct calculation leads to
\begin{align*}
&\|\Gamma_{det}^R u_1-\Gamma_{det}^R u_2\|_{C([0,r];\mathbb H^2)}\\
\le &
C(\lambda)r \Big\|\Big(\Theta_R(u_2,\cdot)-\Theta_R(u_1,\cdot)\Big) f_{\epsilon}(|u_2|^2)u_2\Big\|_{C([0,r];\mathbb H^2)} \\
&+C(\lambda)r \Big\|\Theta_R(u_1,\cdot) \Big( f_{\epsilon}(|u_1|^2)u_1- f_{\epsilon}(|u_2|^2)u_2\Big)\Big\|_{C([0,r];\mathbb H^2)}\\
\le& C(\lambda)r\|u_1-u_2\|_{C([0,r];\mathbb H^2)} \|f_{\epsilon}(|u_2|^2)u_2\|_{C([0,\tau_2];\mathbb H^2)}\\
&\quad+C(\lambda)r\|f_{\epsilon}(|u_2|^2)u_2-f_{\epsilon}(|u_1|^2)u_1\|_{C([0,\tau_1];\mathbb H^2)}\\
\le& C(\lambda,\epsilon)r\|u_1-u_2\|_{C([0,r];\mathbb H^2)}(1+R^2),
\end{align*}
and
\begin{align*}
\|\Gamma_{mod}^R u_1-\Gamma_{mod}^R u_2\|_{C([0,r];\mathbb H^2)}
\le C(\lambda,\epsilon) \sum_{k\in\mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2 r\|u_1-u_2\|_{C([0,r];\mathbb H^2)}(1+R^6).
\end{align*}
By applying the Burkerholder inequality, we obtain
\begin{align*}
&\|\Gamma_{Sto}^R u_1-\Gamma_{Sto}^R u_2\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}\\
\le& Cr \Big\|\Theta_R(u_1,\cdot)(g(|u_1|^2)u_1-g(|u_2|^2)u_2) \Big\|_{L^p(\Omega; L^2([0,r];\mathbb L_2^{2}))}\\
&+ Cr \Big\|(\Theta_R(u_1,\cdot)-\Theta_R(u_2,\cdot))g(|u_2|^2)u_2 \Big\|_{L^p(\Omega; L^2([0,r];\mathbb L_2^{2}))}\\
\le& C(\lambda,\epsilon) (\sum_{k\in\mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2)^{\frac 12} r^{\frac 12}\|u_1-u_2\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}(1+R^3).
\end{align*}
{ where $\mathbb L_2^{2}$ is the space of Hilbert--Schmidt operators form $U_0=Q^\frac 12(\mathbb L^2(\mathcal O))$ to $\mathbb H^2.$}
Combining all the above estimates, we have
\begin{align*}
&\quad\|\Gamma^R u_1-\Gamma^R u_2\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}\\
&\le C(\lambda,\epsilon)\left(1+\sum_{k\in\mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2\right) (r^{\frac 12}+r)\|u_1-u_2\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}(1+R^6),
\end{align*}
which implies that there exists a small $r>0$ depending on $Q,R,\lambda,\epsilon$ such that
$\Gamma^R$ is a strict contraction in $\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))$ and has a fixed point $u^{R,1}$ satisfying $\Gamma^R(u^{R,1})=u^{R,1}.$
Assume that we have found the fixed point on each interval $[(l-1)r,l r],$ $l\le k$ for some $k\ge 1$.
Define
\begin{align*}
u^{R,k}=S(\cdot )u_0+ \Gamma^{R}_{det}u^{R,k}+\Gamma^{R}_{mod}u^{R,k}+\Gamma^{sto}_{det}u^{R,k},
\end{align*}
on $[0,kr].$
In order to extend $u^{R,k}$ to $[kr,(k+1)r]$, we repeat the previous arguments to show that on the interval $[kr,(k+1)r],$ there exists a fixed point of the map $\Gamma^R$ defined by
\begin{align*}
\Gamma^R u(t):=S(t)u^{R,k}({kr})+\Gamma_{det}^{R,k} u(t)+\Gamma_{mod}^{R,k} u(t)
+\Gamma_{Sto}^{R,k} u(t), u\in \mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2)).
\end{align*}
Here we use the following notations,
{\small
\begin{align*}
\Gamma_{det}^{R,k} u(t):&=\mathbf i\int_0^t S(t-s) \Big( \widetilde \Theta_R(u,k,s)\lambda f_{\epsilon}(|u(s)|^2)u(s)\Big)ds,\\
\Gamma_{mod}^{R,k} u(t):&=-\frac 12\int_0^t S(t-s)\Big(\widetilde \Theta_R(u,k,s)\sum_{j\in\mathbb N^+}|Q^{\frac 12}e_j|^2\Big(|g(|u(s)|^2)|^2u(s)\Big)\Big)ds \\
-\mathbf i \int_0^tS(t-s)&\Big(\widetilde
\Theta_R(u,k,s)\sum_{k\in \mathbb N^+}g(|u(s)|^2)|g'(|u(s)|^2) |u(s)|^2 u(s)
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_k \Big)ds,\\
\Gamma_{Sto}^{R,k} u(t):&=\mathbf i\int_0^t S(t-s) \Big(\widetilde \Theta_R(u,k,s)g(|u(s)|^2)u(s)\Big)dW^k(s),
\end{align*}
w}here $t\in [0,r]$, $W^k(s)=W(s+kr)-W(kr),$ $u\in \mathbb M_{\mathcal F_{kr}}^p(\Omega;C([0,r];\mathbb H^2))$
and
$$\widetilde \Theta_R(u,k,s)=\theta_R(\|u^{R,k}\|_{C([0,kr];\mathbb H^2)}+\|u\|_{C([0,r];\mathbb H^2)}).$$
For any different $v_1,v_2\in \mathbb M_{\mathcal F_{kr}}^p(\Omega; C([0,r];\mathbb H^2))$, we define the stopping times $\tau_i=\inf\{t\in [0,T-kr]: \|u^{R,k}\|_{C([0,kr];\mathbb H^2)}+\|v_i\|_{C([0,t];\mathbb H^2)}\ge 2R\}\wedge r,$ $ i=1,2$ and assume that $\tau_1\le \tau_2$ for the convenience.
Then the same procedures yield that
\iffalse
\begin{align*}
&\|\Gamma_{det}^R v_1-\Gamma_{det}^R v_2\|_{C([0,r];\mathbb H^2)}\\
&\le C(\lambda)r\|v_1-v_2\|_{C([0,r];\mathbb H^2)} \|f_{\epsilon}(|v_2(s)|^2)v_2(s)\|_{C([0,\tau_2];\mathbb H^2)}\\
&\quad+C(\lambda)r\|f_{\epsilon}(|v_2(s)|^2)v_2(s)-f_{\epsilon}(|v_1(s)|^2)v_1(s)\|_{C([0,\tau_1];\mathbb H^2)}\\
&\le C(\lambda,\epsilon)r\|v_1-v_2\|_{C([0,r];\mathbb H^2)}(1+R^6)
\end{align*}
and
\begin{align*}
&\|\Gamma_{mod}^R v_1-\Gamma_{mod}^R v_2\|_{C([0,r];\mathbb H^2)}\\
&\le C(\lambda,\epsilon) \sum_{k\in\mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2 r\|v_1-v_2\|_{C([0,r];\mathbb H^2)}(1+R^6).
\end{align*}
By applying the Burkerholder inequality, we obtain that
\begin{align*}
&\|\Gamma_{Sto}^R v_1-\Gamma_{Sto}^R v_2\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}\\
&\le C(\lambda,\epsilon) (\sum_{k\in\mathbb N^+}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2)^{\frac 12} r^{\frac 12}\|v_1-v_2\|_{\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^2))}(1+R^6).
\end{align*}
\fi
this map is a strict contraction and has a fixed point $u^{R,k+1}$ for a small $r>0.$ Now, we define a process $u^R$ as
$u^{R}(t):=u^{R,k}(t)$ for $t\in [0,kr]$ and $u^{R}(t):=u^{R,k+1}(t)$ for $t\in [kr,kr+1].$
It can be checked that $u^R$ satisfies \eqref{trun-reg} by the induction assumption and the definition of $\Gamma^R$.
Meanwhile, the uniqueness of the mild solution can be obtained by repeating the previous arguments.
\iffalse
\begin{equation}
\left\{
\begin{array}{lr}
\theta_R(x)=1, \; x\in [0,R] \\
\theta_R(x)=0, \; x\in [2R,\infty)\\
\end{array}
\right.
\end{equation}
\fi
\hfill$\square$
\begin{prop}\label{mul}
Let the condition of Lemma \ref{trun} hold.
There exists a unique local mild solution to \eqref{Reg-SlogS} with continuous $\mathbb H^2$-valued path. And the solution is defined
on a random interval $[0,\tau^*_{\epsilon}(u_0,\epsilon))$, where $\tau^*_{\epsilon}(u_0,\epsilon,\omega)$ is a stopping time such that $\tau^*_{\epsilon}(u_0,\epsilon,\omega)=+\infty$ or $\lim\limits_{t\to\tau^*_{\epsilon}}\|u^{\epsilon}(t)\|_{\mathbb H^2}=+\infty.$
\end{prop}
{\bf{Proof}}
When $\widetilde g=1$, one can follow the same steps in the proof of \cite[Theorem 3.1]{BD03} to complete the proof.
It suffices to consider the multiplicative noise case.
Let $\{u^R\}_{R\in \mathbb N^+}\in \mathbb M_{\mathcal F}(\Omega; C([0,T];\mathbb H^2))$ be a sequence of solution constructed in Lemma \ref{trun}. And define a stopping time sequence $\tau_{R}:=\inf\{t\in[0,T]: \|u^R\|_{C([0,t];\mathbb H^2)}\ge R\}\wedge T.$
Then $\tau_R>0$ is well-defined since $\|u^R\|_{C([0,t];\mathbb H^2)}$ is an increasing, continuous and $\mathcal F_t$-adapted process.
We claim that if $R_1\le R_2$, then $\tau_{R_1}\le \tau_{R_2}$ and $u^{R_1}=u^{R_2}, a.s.$ on $[0,\tau_{R_1}].$
Let $\tau_{R_2,R_1}:=\inf\{t\in[0,T]: \|u^{R_2}\|_{C([0,t];\mathbb H^2)}\ge R_1\}\wedge T.$
Then it holds that $\tau_{R_2,R_1}\le \tau_{R_2}$ and $\Theta_{R_2}(u^{R_2},t)=\Theta_{R_1}(u^{R_2},t)$ on $t\in [0,\tau_{R_2,R_1}].$ This implies that $\{(u^{R_2},\tau_{R_2,R_1})\}$ is a solution of \eqref{trun-reg} and that $u^{R_1}=u^{R_2}, a.s.$ on $\{t\le \tau_{R_2,R_1}\}.$ Thus we conclude that $\tau_{R_1}=\tau_{R_2,R_1}, a.s.$ and that $u^{R_1}=u^{R_2}$ for $\{t\le \tau_{R_1}\}$.
Now consider the triple $\{u,(\tau_{R})_{R\in\mathbb N^+},\tau_{\infty}\}$ defined by $u(t):=u^R(t)$ for $t\in [0,\tau^R]$ and $\tau_{\infty}=\sup_{R\in\mathbb N^+}\tau_{R}.$ From Lemma \ref{trun}, we know that
$u\in \mathbb M_{\mathcal F}^p(\Omega, C([0,\tau];\mathbb H^2))$ satisfies \eqref{Reg-SlogS} for $t\le \tau_R$.
The uniqueness of the local solution also holds.
If we assume that $(u,\tau)$ and $(v,\sigma)$ are local mild solutions of \eqref{Reg-SlogS}, then $u(t)=v(t), a.s.$ on $\{t<\sigma \wedge \tau\}.$
Let $R_1,R_2\in \mathbb N^+.$
Set $\tau_{R_1,R_2}:=\inf\{t\in [0,T]: \max(\|u\|_{C([0,t];\mathbb H^2)}, \|v\|_{C([0,t];\mathbb H^2)})\ge R_1\}\wedge \sigma_{R_2} \wedge \sigma_{R_2}.$
Then we have on $\{t\le\tau_{R_1,R_2}\},$ $(u,\tau_{R_1,R_2}),$ $(v,\tau_{R_1,R_2})$ are local mild solutions of \eqref{Reg-SlogS}.
The uniqueness
in Lemma \ref{trun} leads to $u=v$ on $\{t\le \tau_{R_1,R_2}\}.$ Letting $R_1,R_2\to \infty$, we complete the proof.\hfill$\square$
If we assume that $g\in \mathcal C^{\mathbf s}_p(\mathbb R),$ $u_0\in \mathbb H^{\mathbf s}$, $\mathbf s\ge 2$, following the same procedures, we can also obtain the local existence of the solution $u_{\epsilon}$ in $\mathcal C([0,\tau^*_{\epsilon }]; \mathbb H^{\mathbf s})$ when $d\le 3$.
When $d\ge 3$, one needs to use another regularization function $\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$ and additional assumption on $g$.
In this case, we can get the local well-posedness in $\mathbb M_{\mathcal F}^p(\Omega; C([0,r];\mathbb H^1))$ based on Lemma \ref{sec-reg} in Appendix and previous arguments. Since its proof is similar to that in Proposition \ref{mul}, we omit these details here and leave them to readers.
\begin{prop}\label{weak-loc}
Let Assumption \ref{main-as} hold, and $f_{\epsilon}(|x|^2)$ $=\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$.
Suppose that $d\in \mathbb N^+$, $u_0\in \mathbb H^2$ is $\mathcal F_0$-measurable and has any finite $p$th moment for $p\geq 1$.
Assume in addition that $Q$-Wiener process $W$ satisfies $\sum_{k\in \mathbb N^+}\|Q^{\frac 12}e_k\|_{L^2_\alpha}^2+\|Q^{\frac 12}e_k\|_{\mathbb H^1}^2<\infty$ when $\widetilde g=1$, and $\sum_{k\in \mathbb N^+}\|Q^{\frac 12}e_k\|_{W^{1,\infty}}^2+\|Q^{\frac 12}e_k\|_{\mathbb H^1}^2<\infty $ when $\widetilde g(x)=\mathbf i g(|x|^2)x$.
Then there exists a unique local mild solution to \eqref{Reg-SlogS} with continuous $\mathbb H^1$-valued path.
And the solution is defined
on a random interval $[0,\tau_{\epsilon}^*(u_0,\epsilon,\omega))$, where $\tau_{\epsilon}^*(u_0,\epsilon,\omega)$ is a stopping time such that $\tau^*(u_0,\epsilon,\omega)=+\infty$ or $\lim_{t\to\tau^*_{\epsilon}}\|u^{\epsilon}(t)\|_{\mathbb H^1}=+\infty.$
\end{prop}
\iffalse
{\color{blue}
To explicitly get the bound of $u^{\epsilon},$ we consider $f_{\epsilon}(\rho)=\log(\rho+\epsilon).$ Then it is not difficult to obtain $$F_{\epsilon}:=\int_0^{\rho}f_{\epsilon}(s)ds=(\epsilon+\rho)\log(\epsilon+\rho)-\rho-\epsilon\log\epsilon.$$
Then $F_{\epsilon}$ can be viewed as $\mathscr O(\epsilon)$-perturbation of $\rho\log \rho$ if $\rho>0$.
The regularized energy for the deterministic logarithmic equation is defined by
$$H_{\epsilon}(u)=\frac 12\|\nabla u\|^2-\frac \lambda{2}F_{\epsilon}(|u|^2).$$
Now we are going to show the local well-posedness of \eqref{add-reg}.
\begin{prop}\label{add}
Assume that $Q^{\frac 12}\in \mathbb L_2^{0,1}$, $u_0\in \mathbb H^2$ is $\mathcal F_0$ measurable. Then there exists a solution to \eqref{add-reg} with continuous $\mathbb H^2$ valued path. And the solution is defined
on a random interval $[0,\tau^*(u_0,\epsilon))$, where $\tau^*(u_0,\epsilon,\omega)$ is a stopping time such that $\tau^*(u_0,\epsilon,\omega)=+\infty$ or $\lim_{t\to\tau^*}\|u^{\epsilon}\|_{\mathbb H^2}=+\infty.$
\end{prop}
\textbf{Proof}
Consider the mild form of $\eqref{add-reg},$
\begin{align*}
u^{\epsilon}(t)=S(t)u_0+\mathbf i\lambda\int_0^tS(t-s)f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon}ds+\mathbf i\int_0^tS(t-s)dW(s).
\end{align*}
Denote the stochastic convolution $Z(t)=\mathbf i\int_0^tS(t-s)dW(s)$ which is well-defined and belongs to $C([0,\infty);\mathbb H^2).$ Set $v=u^{\epsilon}-Z(t).$
Then it sufficient to prove that
\begin{align*}
\Gamma(v)(t):=S(t)u_0+\mathbf i\lambda\int_0^tS(t-s)f_{\epsilon}(|v(s)+Z(s)|^2)(v+Z(s))ds
\end{align*}
has a fixed point in a ball of $C([0,T];\mathbb H^2)$ for sufficiently small $T(u_0,\epsilon).$
Then by using Sobolev embedding theorem $L^{\infty}\hookrightarrow \mathbb H^2$, it is not hard to prove that $\Gamma$ is well-defined map on $C([0,T];\mathbb H^2)$, that is,
\begin{align*}
\|\Gamma(v)\|_{\mathbb H^2}&\le \|S(t)u_0\|_{\mathbb H^2}+|\lambda|\int_0^t\|f_{\epsilon}(|v(s)+Z(s)|^2)(v+Z(s))\|_{\mathbb H^2}ds\\
&\le \|u_0\|_{\mathbb H^2}+C|\lambda| \int_0^t\|f_{\epsilon}(|v(s)+Z(s)|^2)(v(s)+Z(s))\|ds\\
&+ C|\lambda| \int_0^t\Big\|f'_{\epsilon}(|v(s)+Z(s)|^2)Re\Big((\bar v(s)+\bar Z(s))(\nabla v(s)+\nabla Z(s))\Big)(v(s)+Z(s))\Big\|ds\\
&+ C|\lambda| \int_0^t\Big\|f_{\epsilon}(\epsilon+|v(s)+Z(s)|^2)(\nabla v(s)+\nabla Z(s))\Big\|ds\\
&+ C|\lambda| \int_0^t\Big\|f''_{\epsilon}(|v(s)+Z(s)|^2)\Big(Re\Big((\bar v(s)+\bar Z(s))(\nabla v(s)+\nabla Z(s))\Big)\Big)^2(v(s)+Z(s))\Big\|ds\\
&+C|\lambda| \int_0^t\Big\|f'_{\epsilon}(|v(s)+Z(s)|^2)Re\Big((\bar v(s)+\bar Z(s))(\Delta v(s)+\Delta Z(s))\Big)(v(s)+Z(s))\Big\|ds\\
&+C|\lambda| \int_0^t\Big\|f'_{\epsilon}(|v(s)+Z(s)|^2) |\nabla (v(s)+ Z(s))|^2(v(s)+Z(s))\Big\|ds\\
&+C|\lambda| \int_0^t\Big\|f'_{\epsilon}(|v(s)+Z(s)|^2)\Big(Re\Big((\bar v(s)+\bar Z(s))(\nabla v(s)+\nabla Z(s))\Big)\nabla (v(s)+Z(s))\Big\|ds\\
&+C|\lambda| \int_0^t\Big\|f_{\epsilon}(|v(s)+Z(s)|^2) \Delta (v(s)+Z(s))\Big\|ds.\\
&\le \|u_0\|_{\mathbb H^2}+C|\log(\epsilon)|T|\lambda|(1+\sup_{t\in [0,T]}\|v\|_{\mathbb H^2}+\sup_{t\in [0,T]}\|Z\|_{\mathbb H^2})\\
&+CT|\lambda|(1+\epsilon^{-\frac 12})\Big(1+\sup_{t\in [0,T]}\|v\|_{\mathbb H^2}+\sup_{t\in [0,T]}\|Z\|_{\mathbb H^2}+\sup_{t\in [0,T]}\|\Delta v\|^{2}+\sup_{t\in [0,T]}\|\Delta Z\|^{2}\Big)\\
&\le \|u_0\|_{\mathbb H^2}+C(\lambda,\epsilon)T\Big(1+\sup_{t\in [0,T]}\| v\|_{\mathbb H^2}+\sup_{t\in [0,T]}\|Z\|_{\mathbb H^2}+\sup_{t\in [0,T]}\| v\|_{\mathbb H^2}^2+\sup_{t\in [0,T]}\|Z\|^2_{\mathbb H^2}\Big)
\end{align*}
Fixed $\omega\in \Omega$. We are going to prove that $\Gamma$ is also local Lipschitz map on the ball $B_{R_1}^{\tau_1}:=\{v\in \mathbb H^2: \|v\|_{C([0,\tau_1];\mathbb H^2)}\le R_1\}$ in $C([0,T];\mathbb H^2),$ where $R$ and $\tau_1$ are random variables to be chosen later.
Now, assume that $v, v_1\in B_{R_1}^{\tau_1}$. Then we have that
\begin{align}\label{wel-fix}
\|\Gamma(v)\|_{C([0,\tau_1];\mathbb H^2)}\le \|u_0\|_{\mathbb H^2}+C(\lambda,\epsilon)\tau_1(1+\sup_{t\in [0,\tau_1]}\|Z\|_{\mathbb H^2}+\sup_{t\in [0,\tau_1]}\|Z\|_{\mathbb H^2}^{2}))\\\nonumber+C(\lambda,\epsilon)\tau_1(1+R_1)R_1,
\end{align}
and that
\begin{align}\label{loc-lip}
\|\Gamma(v)-\Gamma(v_1)\|_{C([0,\tau_1];\mathbb H^2)}
&\le C'_{\epsilon,\tau_1,\lambda}\tau_1\|v-v_1\|_{C([0,\tau_1];\mathbb H^2)},
\end{align}
where $C'_{\epsilon,\tau_1,\lambda}:=C_{\lambda}(1+\epsilon^{-3})(1+\sup_{t\in [0,\tau_1]}\|Z\|_{\mathbb H^2}^6+R_1^6).$
Then we define a real-valued continuous $\mathcal F_t$-adapted process $D(t):=C'_{\epsilon,t,\lambda}+C(\lambda,\epsilon)t (1+R_1)$ and define two stopping time, $\tau_1:=\inf\{t\in[0,T]: tD_t\ge \frac 12 \}\wedge T$ and $\widetilde \tau_1:=\inf\{t\in [0,T]: \|u_0\|_{\mathbb H^2}+C(\lambda,\epsilon)t(\sup_{s\in [0,t]}\|Z\|_{\mathbb H^2}+\sup_{s\in [0,t]}\|Z\|_{\mathbb H^2}^{2}))\ge \frac {R_1} 2\}\wedge T$.
Define $\widehat \tau_1=\tau_1\wedge \widetilde \tau_1.$
Then the above estimates \eqref{wel-fix} and \eqref{loc-lip} imply that $\Gamma$ is a contraction map on $B_{R_1}^{\tau_1\wedge \tau_2}$ and that there exists a unique $\mathcal F_t$-adapted solution $u_1:=u(t\wedge \widehat \tau_1)$ of \eqref{add-reg} in $C([0,\widehat \tau_1];\mathbb H^2)$, a.s. Applying the iterative procedures, we can construct a sequence $\{u_m,\widehat \tau_m\}_{m\ge 1}$ such that $\widehat \tau_m$ are increasing stopping time and that $u_m(t)=u_m(t\wedge\widehat \tau_m)$ solves \eqref{add-reg}, a.s.
Given the pair $\{u_m,\widehat \tau_m\}$ at the $m$-step. On
$B_{R_{m+1}}^{\sigma_m}$, we have that for $v\in C([0,T];\mathbb H^2),$
\begin{align*}
\Gamma_{m}(v)(t)&:=S(t)u_{m}(\widehat \tau_m)\\
&\quad+\mathbf i \lambda \int_{0}^{t} S(t-s)f_{\epsilon}(|v(s)+Z( \widehat \tau_m+s)|^2)(v(s)+Z( \widehat \tau_m+s))ds.
\end{align*}
By using the estimates \eqref{wel-fix} and \eqref{loc-lip},
we have that for $v\in B_{R_{m+1}}^{\sigma_m},$
\begin{align*}
\|\Gamma_m(v)\|_{C([0,\sigma_m];\mathbb H^2)}&\le \|u_{m}(\widehat \tau_m)\|_{\mathbb H^2}\\
&+C(\lambda,\epsilon)\sigma_m(1+\sup_{t\in [0,\sigma_m]}\|Z(\widehat \tau_m+t)\|_{\mathbb H^2}+\sup_{t\in [0,\tau_1]}\|Z(\widehat \tau_m+t)\|_{\mathbb H^2}^{2})
\\\nonumber
&+C(\lambda,\epsilon)\sigma_m(1+R_{m+1})R_{m+1},
\end{align*}
and that
\begin{align*}
\|\Gamma_m(v)-\Gamma_m(\widetilde v)\|_{C([0,\sigma_m];\mathbb H^2)}
&\le C'_{\epsilon,\widehat \tau_m+\sigma_m,\lambda}\sigma_m\|v-\widetilde v\|_{C([0,\sigma_m];\mathbb H^2)},
\end{align*}
where $C'_{\epsilon,\widehat \tau_m+\sigma_m,\lambda}:=C_{\lambda}(1+\epsilon^{-3})(1+\sup_{t\in [0,\widehat \tau_m+\sigma_m]}\|Z\|_{\mathbb H^2}^6+R_{m+1}^6).$
Then repeating the argument at first step, define $D_t^m:=C(\lambda,\epsilon)(1+R_{m+1})+C'_{\epsilon,\widehat \tau_m+\sigma_m,\lambda}$ and
\begin{align*}
\sigma_m&=\inf\{t\in[0,T-\widehat \tau_m]: t D_t^{m}\ge \frac 12\} \wedge (T-\widehat \tau_m)\\
&\wedge \inf\{t\in [0,T-\widehat \tau_m]: \|u_{\widehat \tau_m}\|_{\mathbb H^2}+C(\lambda,\epsilon)t(\sup_{s\in [0,\widehat \tau_m+t]}\|Z\|_{\mathbb H^2}\\
&\quad +\sup_{s\in [0,\widehat \tau_m+t]}\|Z\|_{\mathbb H^2}^{2})\ge \frac {R_{m+1}} 2\}
\end{align*}
It follows that $\Gamma_m$ has a fixed point $v_{m+1}$ in $\mathcal C([0,\sigma_m]; \mathbb H^2).$ Thus, define $\widehat \tau_{m+1}:=\widehat \tau_m+\sigma_m$,
$u_{m+1}(t)=u_{m}(t)$ for $t\in [0,\widehat \tau_m]$ and $u_{m+1}(t)=v_{m+1}((t-\widehat \tau_m)\wedge (\sigma_m))$, for $t\in [\widehat \tau_m,\sigma_m].$ Thus by iterative procedures, we obtain the sequence $\{(u_m,\widehat \tau_m)\}_{m\ge 1}.$ Define $\tau^*(u_0,\epsilon)=\lim\limits_{m\to \infty} \widehat \tau_m$, and $u_{\epsilon}:=u_m 1_{[0,\tau^*(u_0,\epsilon))}.$
Then it follows that $\tau(u_0,\epsilon)$ is a stopping time, $u_{\epsilon}$ is the solution of \eqref{add-reg} on $[0,\tau^*(u_0,\epsilon).$
The blow-up alternative is obtained immediately.
\qed
}
\fi
\subsection{Global existence and uniform estimate of regularized SlogS equation}
Due to the blow-up alternative results in section \ref{sec-loc}, it suffices to prove that $\sup_{t\in[0,\tau_{\epsilon}^*)}\|u^{\epsilon}(t)\|_{\mathbb H^2}<\infty,$ or $\sup_{t\in[0,\tau_{\epsilon}^*)}\|u^{\epsilon}(t)\|_{\mathbb H^1}<\infty, a.s.$ under corresponding assumptions.
In the following, we present several a priori estimates in strong sense to achieve our goal.
To simplify the presentation, we omit some procedures
like mollifying the unbounded operator $\Delta$ and taking the limit on the regularization parameter.
More precisely, the mollifier $\widetilde \Theta_m$, $m\in \mathbb N^+$ may be defined by the Fourier transformation (see e.g. \cite{BD03})
\begin{align*}
\mathbb F(\widetilde \Theta_m v)(\xi)=\widetilde \theta(\frac {|\xi|}{m})\widehat v(\xi),\; \xi\in \mathbb R^d,
\end{align*}
where $\widetilde \theta$ is a positive $\mathcal C^{\infty}$ function on $\mathbb R^+$, has a compact support satisfying
$\theta(x)=0,$ for $x\ge 2$ and $\theta(x)=1,$ for $0\le x\le 1.$
Another choice of mollifier is via Yosida approximation $\Theta_{m}:=m(m-\Delta)^{-1}$ for $m\in \mathbb N^+$ (see e.g. \cite{Hor18}).
This kind of procedure is introduced to make that the It\^o formula can be applied rigorously to deducing several a priori estimates.
If $\mathcal O$ becomes a bounded domain equipped with periodic or homogenous Dirichlet boundary condition, the mollifier can be chosen as the Galerkin projection, and the approximated equation becomes the Galerkin approximation (see e.g. \cite{CH17,CHL16b,CHLZ17,CHS19}).
In this section, we assume that $u_0$ has finite $p$-moment for all $p\ge 1$ for simplicity.
We will also use Assumption \ref{main-as} to obtain the global existence of the mild solution of the regularized SlogS equation.
When $d\le 3$ and $f_{\epsilon}(x)=\log(x+\epsilon),$ we assume that $\sum_{k}\|Q^{\frac 12}e_k\|_{\mathbb H^2}^2<\infty.$ When $d\in \mathbb N^+$ and $f_{\epsilon}(x)=\log(\frac {x+\epsilon}{1+\epsilon x}),$ we assume that $Q^{\frac 12}\in \mathbb L^1_2$ for the additive noise and $\sum_{k}\|Q^{\frac 12}e_k\|^2_{W^{1,\infty}}<\infty$ for the multiplicative noise.
\subsection{A priori estimates in $L^2$ and $\mathbb H^1$}
\begin{lm}\label{mass-pri}
Let $T>0.$
Under the condition of Proposition \ref{mul} or Proposition \ref{weak-loc}, assume that $(u^{\epsilon},\tau_{\epsilon}^*)$ be a local mild solution in $\mathbb H^1$.
Then for any $p\ge2,$ there exists a positive constant $C(Q,T,\lambda,p,u_0)>0$ such that
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau^*_{\epsilon}\wedge T)}\|u^{\epsilon}(t)\|^p\Big]\le
C(Q,T,\lambda,p,u_0).
\end{align*}
\end{lm}
{\noindent\bf{Proof}}
Take any stopping time $\tau<\tau^*_{\epsilon}\wedge T, a.s.$ Using the It\^{o} formula to
$M^k(u^{\epsilon}(t)),$ where $M(u^{\epsilon}(t)):=\|u^{\epsilon}(t)\|^2$ and $k\in \mathbb N^+$ or $k\ge 2$, we obtain that for $t\in [0,\tau]$ and the case $\widetilde g=1,$
\begin{align*}
&M^k(u^{\epsilon}(t))\\
=&M^k(u_0^{\epsilon})
+2k(k-1)\int_0^tM^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+} \<u^{\epsilon}(s),Q^{\frac 12}e_i\>^2ds
\\
&+k\int_{0}^tM^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+} \|Q^{\frac 12}e_i\|^2 ds
+2k\int_{0}^tM^{k-1}(u^{\epsilon}(s))\<u^\epsilon(s),dW(s)\>,
\end{align*}
and for $t\in [0,\tau]$ and the case $\widetilde g(x)=\mathbf i g(|x|^2)x$,
\begin{align*}
&\quad M^k(u^{\epsilon}(t))\\
&=M^k(u_0^{\epsilon})+ {2k(k-1)}\int_0^tM^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+} \<u^{\epsilon}(s),\mathbf i g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)Q^{\frac 12}e_i\>^2ds
\\
&\quad
+2k\int_{0}^tM^{k-1}(u^{\epsilon}(s))\<u^\epsilon(s),\mathbf i g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s) dW(s)\>+2k\int_0^tM^{k-1}(u^{\epsilon}(s))\\
&\quad\times\Big\<u^{\epsilon}(s), -\mathbf i \sum_{k\in \mathbb N^+}Im(Q^{\frac 12}e_k) Q^{\frac 12}e_k g'(|u^{\epsilon}(s)|^2)g(|u^{\epsilon}(s)|^2)|u^{\epsilon}(s)|^2u^{\epsilon}(s)\Big\>ds.
\end{align*}
In particular, if $W(t,x)$ is real-valued and $\widetilde g(x)=\mathbf i g(|x|^2)x$, we have $M^k(u^{\epsilon}(t))=M^k(u^{\epsilon}_0),$ for $t\in[0,T]$ a.s.
By Assumption \ref{main-as} and conditions in Propositions \ref{mul} and \ref{weak-loc}, using the martingale inequality, the H\"older inequality and the Young inequality and Gronwall's inequality, we achieve that for all $k\ge1,$
\begin{align*}
\sup_{t\in [0,\tau]}\mathbb E\Big[M^k(u^{\epsilon}(t))\Big]\le C(T,k,u_0,{Q}).
\end{align*}
Next, taking the supreme over $t$ and repeating the above procedures, we have that
\begin{align*}
&\mathbb E\Big[\sup_{t\in [0,\tau]}M^k(u^{\epsilon}(t))\Big]\\
\le& \mathbb E\Big[M^k(u_0^{\epsilon})\Big]+C(k)\mathbb E\Big[\int_0^\tau M^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+} (1+\|u^{\epsilon}(s)\|^2)\|Q^{\frac 12}e_i\|_{U}^2ds\Big]
\\
&\quad+C(k)\mathbb E\Big[\Big(\int_{0}^{\tau} M^{2k-2}(u^{\epsilon}(s))(1+\|u^{\epsilon}(s)\|^4) \sum_{i\in\mathbb N^+}\|Q^{\frac 12}e_i\|_{U}^2ds\Big)^{\frac 12}\Big],
\end{align*}
where $U=L^2$ for additive noise case and $U=L^{\infty}$ for multiplicative noise case.
Applying the estimate of $\sup\limits_{t\in [0,\tau]}\mathbb E\Big[M^{k}(u^{\epsilon}(t))\Big]$, we complete the proof by taking $p=2k$.
\hfill$\square$
\begin{lm}\label{h1-pri}
Let $T>0$. Under the condition of Proposition \ref{mul} or Proposition \ref{weak-loc}, assume that $(u^{\epsilon},\tau_{\epsilon}^*)$ is a local mild solution in $\mathbb H^1$.
Then for any $p\ge2,$ there exists $C(Q,T,\lambda,p,u_0)>0$ such that
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau^*_{\epsilon}\wedge T)}\|u^{\epsilon}\|_{\mathbb H^1}^p\Big]\le
C(Q,T,\lambda,p,u_0).
\end{align*}
\end{lm}
{\noindent\bf{Proof}}
Take any stopping time $\tau<\tau^*_{\epsilon}\wedge T, a.s.$
Applying the It\^{o} formula to
the kinetic energy $K(u^{\epsilon}(t)):=\frac 12\|\nabla u^{\epsilon}(t)\|^2,$ for $k\in \mathbb N^+$ and using integration by parts,
we obtain that for $t\in [0,\tau],$
\iffalse
&=K^k(u^{\epsilon}_0)+
k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<-\Delta u^{\epsilon}(s),\mathbf i \Delta u^{\epsilon}(s)+\mathbf i \lambda f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)\>ds\\
&\quad+
\frac 12 k(k-1)\int_0^{t}K^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\<\nabla u^{\epsilon}, \nabla Q^{\frac 12}e_i\>^2ds\\
&\quad
+k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\<\nabla Q^{\frac 12}e_i,\nabla Q^{\frac 12}e_i\>ds\\
&\quad
+k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u,\nabla dW(s)\>\\
\fi
\begin{align*}
&\quad K^k(u^{\epsilon}(t))\\
&=K^k(u^{\epsilon}_0)
+k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^{\epsilon}(s),\mathbf i 2\lambda f'_{\epsilon}(|u^{\epsilon}(s)|^2)Re(\bar u^{\epsilon}(s)\nabla u^{\epsilon}(s)) u^{\epsilon}(s)\>ds\\
&\quad+
\frac 12k(k-1)\int_0^{t}K^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\<\nabla u^{\epsilon}(s), \nabla Q^{\frac 12}e_i\>^2ds\\
&\quad
+\frac k 2\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\<\nabla Q^{\frac 12}e_i,\nabla Q^{\frac 12}e_i\>ds\\
&\quad
+k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^\epsilon(s),\nabla dW(s)\>
\end{align*}
for the additive noise case,
and {\small
\begin{align*}
&\quad K^k(u^{\epsilon}(t))\\
&=K^k(u^{\epsilon}_0)+
k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^{\epsilon}(s),\mathbf i 2\lambda f'_{\epsilon}(|u^{\epsilon}(s)|^2)Re(\bar u^{\epsilon}(s)\nabla u^{\epsilon}(s)) u^{\epsilon}(s)\>ds\\
&\quad+
\frac 12 k(k-1)\int_0^{t}K^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\<\nabla u^{\epsilon}(s), \mathbf i \nabla(g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s) Q^{\frac 12}e_i)\>^2ds\\
&\quad
+\frac k2\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\<\nabla(g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s) Q^{\frac 12}e_i),\nabla(g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s) Q^{\frac 12}e_i)\>ds\\
&\quad
+k \int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^{\epsilon}(s),\mathbf i \nabla g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s) dW(s)\>\\
&\quad+\frac k2 \int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\Delta u^{\epsilon}(s),\sum_{i\in \mathbb N^+}|Q^{\frac 12}e_i|^2(g(|u^{\epsilon}(s)|^2))^2u^{\epsilon}(s)\>ds\\
&\quad-k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\Delta u^{\epsilon}(s),\mathbf i \sum_{i\in \mathbb N^+}g(|u^{\epsilon}(s)|^2) g'(|u^{\epsilon}(s)|^2) |u^{\epsilon}(s)|^2 u^{\epsilon}(s)
Im(Q^{\frac 12}e_i)Q^{\frac 12}e_i\>ds
\end{align*}
f}or the multiplicative noise case.
Applying integration by parts, in the multiplicative noise case, we further obtain
\begin{align*}
& K^k(u^{\epsilon}(t))\\
\le& K^k(u^{\epsilon}_0)+
k \int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^{\epsilon},\bi2 \lambda f'_{\epsilon}(|u^{\epsilon}(s)|^2)Re(\bar u^{\epsilon} \nabla u^{\epsilon})u^{\epsilon}(s)\>ds\\
&+
C_k \int_0^{t}K^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\<\nabla u^{\epsilon}, g(|u^{\epsilon}|^2)u^{\epsilon} \nabla Q^{\frac 12}e_i)\>^2+\<\nabla u^{\epsilon}, g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i)\>^2\Big)ds\\
&
+C_k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\|g(|u^{\epsilon}|^2) u^{\epsilon} \nabla Q^{\frac 12}e_i\|^2+
\\
&\quad+\|g(|u^{\epsilon}|^2)\nabla u^{\epsilon} Q^{\frac 12}e_i\|^2+\|g'(|u^{\epsilon}|^2) Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i\|^2\Big)ds\\
&
+k \int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^\epsilon,\mathbf i \nabla g(|u^{\epsilon}|^2)u^{\epsilon} dW(s)\>\\
&+C_k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(|\<\nabla u^{\epsilon},(\nabla Im(Q^{\frac 12}e_i) Q^{\frac 12}e_i+\nabla Q^{\frac 12}e_i Im(Q^{\frac 12}e_i)) g'(|u^{\epsilon}|^2) g(|u^{\epsilon}|^2) |u^{\epsilon}|^2u^{\epsilon}\>|\\
&\quad+|\<\nabla u^{\epsilon}, Im(Q^{\frac 12}e_i)Q^{\frac 12}e_i g''(|u^{\epsilon}|^2) g(|u^{\epsilon}|^2) |u^{\epsilon}|^2u^{\epsilon} Re(\bar u^{\epsilon}\nabla u^{\epsilon})\>| \\
&\quad +|\<\nabla u^{\epsilon}, Im(Q^{\frac 12}e_i)Q^{\frac 12}e_i|g'(|u^{\epsilon}|^2)|^2 |u^{\epsilon}|^2u^{\epsilon} Re(\bar u^{\epsilon}\nabla u^{\epsilon})\>| \\
&\quad +|\<\nabla u^{\epsilon}, Im(Q^{\frac 12}e_i)Q^{\frac 12}e_ig'(|u^{\epsilon}|^2) g(|u^{\epsilon}|^2) (2Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon}+|u^{\epsilon}|^2\nabla u^{\epsilon})\>|\Big)ds.
\end{align*}
By using the property of $g$ in Assumption \ref{main-as} and conditions on $Q$, and applying H\"older's, Young's and Burkholder's inequalities, we achieve that for small $\epsilon_1>0,$
\iffalse
Direct calculation leads to
\begin{align*}
&\<\nabla(g(|u^{\epsilon}|^2)u^{\epsilon} Q^{\frac 12}e_i),\nabla(g(|u^{\epsilon}|^2)u^{\epsilon} Q^{\frac 12}e_i)\>\\
&=\|g(|u^{\epsilon}|^2)\nabla u^{\epsilon} Q^{\frac 12}e_i\|^2
+\|g(|u^{\epsilon}|^2) u^{\epsilon} \nabla Q^{\frac 12}e_i\|^2+4\|g'(|u^{\epsilon}|^2) Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i\|^2\\
&\quad+2\<g(|u^{\epsilon}|^2)\nabla u^{\epsilon} Q^{\frac 12}e_i,g(|u^{\epsilon}|^2) u^{\epsilon} \nabla Q^{\frac 12}e_i\>\\
&\quad+4\<g(|u^{\epsilon}|^2)\nabla u^{\epsilon} Q^{\frac 12}e_i,g'(|u^{\epsilon}|^2) Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i\>\\
&\quad+4\<g(|u^{\epsilon}|^2) u^{\epsilon} \nabla Q^{\frac 12}e_i,g'(|u^{\epsilon}|^2) Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i\>,\\
&\<\Delta u^{\epsilon},|Q^{\frac 12}e_k|^2(g(|u^{\epsilon}|^2))^2u^{\epsilon}\>\\
&=-2\<\nabla u^{\epsilon}, Re(\overline{ Q^{\frac 12}e_i}\nabla Q^{\frac 12}e_i)(g(|u^{\epsilon}|^2))^2u^{\epsilon}\>\\
&\quad-\<\nabla u^{\epsilon},|Q^{\frac 12}e_i|^2(g(|u^{\epsilon}|^2))^2\nabla u^{\epsilon}\>\\
&\quad-4\<\nabla u^{\epsilon},|Q^{\frac 12}e_i|^2g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon}) u^{\epsilon}\>.
\end{align*}
\fi
\begin{align*}
&\quad\mathbb E[\sup_{r\in[0,t]}K^k(u^{\epsilon}(r))]\\
&\le \mathbb E[K^k(u^{\epsilon}_0)]+
C_k |\lambda| \mathbb E\Big[ \int_{0}^{t}\sup_{r\in[0,s]} K^{k-1}(u^{\epsilon}(r)) \|\nabla u^{\epsilon}(r)\|^2 ds\Big]
\\
&\quad+C_kT\mathbb E\Big[\sup_{r\in[0,t]}K^{k-2}(u^{\epsilon}(r))\sum_{i\in \mathbb N^+}\|\nabla u^{\epsilon}(r)\|^2(1+\|\nabla u^{\epsilon}(r)\|^2+\|u^{\epsilon}(r)\|^2) \|Q^{\frac 12}e_i\|_{U}^2\Big]\\
&\quad+ C_k\mathbb E\Big[\Big(\int_0^{t}K^{2k-2}(u^{\epsilon}(s))\sum_{i\in\mathbb N^+}\|\nabla u^{\epsilon}(r)\|^2(1+\|\nabla u^{\epsilon}(r)\|^2+\|u^{\epsilon}(r)\|^2 \|\nabla Q^{\frac 12}e_i\|_U^2ds\Big)^{\frac 12}\Big]\\
&\le \mathbb E[K^k(u^{\epsilon}_0)]+C(T,\lambda,k,\epsilon_1,Q)
+\epsilon_1 \mathbb E[\sup_{r\in[0,t]}K^k(u^{\epsilon}(r))]\\
&\quad+C_k \mathbb E\Big[ \int_{0}^{t}\sup_{r\in[0,s]} K^{k}(u^{\epsilon}(r)) dr\Big],
\end{align*}
where $\|Q^{\frac 12}e_i\|_U^2=\|Q^{\frac 12}e_i\|^2_{\mathbb H^1}$ for the additive noise case and $\|Q^{\frac 12}e_i\|_U^2=\|Q^{\frac 12}e_i\|^2_{\mathbb H^1}+\|Q^{\frac 12}e_i\|^2_{W^{1,\infty}}$ for the multiplicative noise case.
Applying Gronwall's inequality, we complete the proof by taking $p=2k$.
\qed
From the above proofs of Lemmas \ref{mass-pri} and \ref{h1-pri},
it is not hard to see that to obtain $\epsilon$-independent estimates, the boundedness restriction $\sup_{x\ge 0}|g(x)|<\infty$ may be not necessary in the case that $W(t,x)$ is real-valued.
We present such result in the following which is the key of the global well-posedness of an SlogS equation with super-linear growth diffusion in next section.
\begin{lm}\label{lm-con-h1}
Let $T>0$ and $(u^{\epsilon},\tau^*_{\epsilon})$ be a local mild solution in $\mathbb H^{\mathbf s}$, $\mathbf s\ge 1$ for any $p\ge 1.$ Assume that $u_0\in \mathbb H^1\cap L^2_{\alpha}$, for some $\alpha\in (0,1]$, is $\mathcal F_0$-measurable and has any finite $p$th moment, and $W(t,x)$ is real-valued with $\sum_{i} \|Q^{\frac 12}e_i\|_{\mathbb H^1}^2+\|Q^{\frac 12}e_i\|_{W^{1,\infty}}^2<\infty.$ Let $\widetilde g(x)=\mathbf ig(|x|^2)x,$ $g \in \mathcal C^1_b(\mathbb R)\cap \mathcal C(\mathbb R)$ satisfy the growth condition and the embedding condition,
\begin{align*}
&\sup_{x\in [0,\infty)}|g'(x)x|\le C_g,\\
&\|vg(|v|^2)\|_{L^q}\le C_d (1+\|v\|_{\mathbb H^1}+\|v\|_{L^2_{\alpha}}),
\end{align*}
for some $q\ge 2$, where $C_g>0$ depends on $g$, and $C_d>0$ depends on $\mathcal O$, $d$, $\|v\|$.
Then it holds that $M(u^{\epsilon}_t)=M(u_0)$ for $t\in [0,\tau_{\epsilon}).$
Furthermore, there exists a positive constant $C(Q,T,\lambda,p,u_0)$ such that
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau^*_{\epsilon}\wedge T)}\|u^{\epsilon}(t)\|_{\mathbb H^1}^p\Big]\le
C(Q,T,\lambda,p,u_0).
\end{align*}
\end{lm}
\textbf{Proof}
The proof is similar to those of Lemmas \ref{mass-pri} and \ref{h1-pri}. We only need to modify the estimation involved with $g.$
The mass conservation is not hard to be obtained since the calculations in Lemmas \ref{mass-pri} only use the assumptions that $g\in \mathcal C(\mathbb R)$ and $W(t,x)$ is real-valued. Therefore, we focus on the estimate in $\mathbb H^1.$
We only show estimation about $\mathbb E\Big[K^k(u^{\epsilon}(t))\Big]$ since the proof on $\mathbb E\Big[\sup_{t\in[0,\tau^*_{\epsilon}\wedge T)}K^k(u^{\epsilon}(t))\Big]$ is similar.
Then following the same steps in Lemma \ref{h1-pri}, we get that for $\frac 1q+\frac 1{q'}=\frac 12$,
\begin{align*}
&\mathbb E\Big[K^k(u^{\epsilon}(t))\Big]\\
\le& \mathbb E\Big[K^k(u^{\epsilon}_0)\Big]+
k \mathbb E\Big[\int_0^{t}K^{k-1}(u^{\epsilon}(s))\<\nabla u^{\epsilon},\bi2 \lambda f'_{\epsilon}(|u^{\epsilon}(s)|^2)Re(\bar u^{\epsilon} \nabla u^{\epsilon})u^{\epsilon}(s)\>ds\Big]\\
&+
\mathbb E\Big[C_k\int_0^{t}K^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\<\nabla u^{\epsilon}, g(|u^{\epsilon}|^2)u^{\epsilon} \nabla Q^{\frac 12}e_i)\>^2+\<\nabla u^{\epsilon}, g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i)\>^2\Big)ds\Big]\\
&
+C_k\mathbb E\Big[\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\|g(|u^{\epsilon}|^2) u^{\epsilon} \nabla Q^{\frac 12}e_i\|^2+\|g'(|u^{\epsilon}|^2) Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} Q^{\frac 12}e_i\|^2\Big)ds\Big]\\
\le& \mathbb E\Big[K^k(u^{\epsilon}_0)\Big]+
C_k \mathbb E\Big[\int_0^{t}K^{k}(u^{\epsilon}(s))ds\Big]\\
&+
\mathbb E\Big[C_k\int_0^{t}K^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\|g(|u^{\epsilon}|^2)u^{\epsilon}\|_{L^q}^2 \|\nabla Q^{\frac 12}e_i\|_{L^{q'}}^2+ \|\nabla u^{\epsilon}\|^2\|Q^{\frac 12}e_i\|^2\Big)ds\Big].
\end{align*}
Applying the embedding condition on $g$, mass conservation law and the procedures in the proof of Proposition \ref{wei-hol-1}, we complete the proof by taking $p=2k$ and Gronwall's inequality.
\qed
The embedding condition is depending on the assumption on $Q$ and $d$. One example which satisfies the embedding condition on $g$ and is not bounded is $g(x)=\log(c+x),$ for $x\ge 0$ and $c>0.$ Let us verify this example on $\mathcal O=\mathbb R^d.$
If the domain $\mathcal O$ is bounded, one can obtain the similar estimate.
We apply the Gagliardo--Nirenberg interpolation inequality and get that for $q>2$ and small enough $\eta>0,$
\begin{align*}
\|g(|v|^2)v\|_{L^q}^q&\le \int_{c+|v|^2\ge 1}|\log(c+|v|^2)|^q|v|^qdx+ \int_{c+|v|^2\le 1}|\log(c+|v|^2)|^q|v|^qdx\\
&\le C\Big(\|v\|_{L^q}^q+\|v\|_{L^{q+q\eta}}^{q+q\eta}+\|v\|_{L^{q-q\eta}}^{q-q\eta}\Big)\\
&\le C\Big(\|v\|^{1-\alpha_0}\|\nabla v\|^{\alpha_0}+(\|v\|^{1-\alpha_1}\|\nabla v\|^{\alpha_1})^{1+\eta}+(\|v\|^{1-\alpha_2}\|\nabla v\|^{\alpha_2})^{1-\eta}\Big)^q\\
&\le C \Big({\|\nabla v\|^{q\alpha_0}}+\|\nabla v\|^{(q+q\eta)\alpha_1}+\|\nabla v\|^{(q-q\eta)\alpha_2}\Big),
\end{align*}
where $\alpha_0=\frac {d(q-2)}{2q}, \alpha_1=\frac {d(q(1+\eta)-2)}{2q(1+\eta)}$ and $\alpha_2=\frac {d(q(1-\eta)-2)}{2q(1-\eta)}$ satisfies $\alpha_i\in (0,1), i=0,1,2.$
When $q=2,$ similar calculations, together with the interpolation inequality in Lemma \ref{wei-hol}, yield that
\begin{align*}
\|g(|v|^2)v\|^
&\le C\Big(\|v\|^2+\|v\|_{L^{2+2\eta}}^{2+2\eta}+\|v\|_{L^{2-2\eta}}^{2-2\eta}\Big)\\
&\le C \Big(\|\nabla v\|^{2}+\|\nabla v\|^{(2+2\eta)\alpha_1}+\|v\|_{L^2_{\alpha_2}}^{\frac {d\eta}{\alpha_2}}\Big),
\end{align*}
where $\alpha_1=\frac {d(2(1+\eta)-2)}{4(1+\eta)}\in(0,1)$ and $\alpha_2\in(\frac {d\eta}{2-2\eta},1).$
\subsection{$L^2_{\alpha}$-estimate and modified energy}
Beyond the $L^2$ and $\mathbb H^1$ estimates, we also need the uniform boundedness in $L_{\alpha}^2,$ $\alpha>0$ to show the strong convergence of $\{u^{\epsilon}\}_{\epsilon>0}$ when $\mathcal O=\mathbb R^d.$
We would like to mention that when $\mathcal O$ is a bounded domain, such estimate in $L_{\alpha}^2,$ $\alpha>0$ is not necessary. To this end, the following useful weighted interpolation inequality is introduced.
\begin{lm}\label{wei-hol}
Let $d\in \mathbb N^+$ and $\eta\in (0,1)$.
Then for $\alpha>\frac {d\eta}{2-2\eta}$, it holds that for some $C=C(d)>0$,
\begin{align*}
\|v\|_{L^{2-2\eta}}\le C\|v\|^{1-\frac {d\eta}{2\alpha(1-\eta)}}\|v\|_{L^2_{\alpha}}^{\frac {d\eta}{2\alpha(1-\eta)}},\; v\in L^2\cap L^2_{\alpha}.
\end{align*}
\end{lm}
\textbf{Proof}
Using the Cauchy-Schwarz inequality and $\alpha>\frac {d\eta}{2-2\eta}$, we have that for any $r>0,$
\begin{align*}
\|v\|_{L^{2-2\eta}}^{2-2\eta}&\le \int_{|x|\le r}|v(x)|^{2-2\eta}dx+\int_{|x|\ge r}\frac {|x|^{ \alpha({2-2\eta})}|v(x)|^{2-2\eta}}{|x|^{\alpha({2-2\eta})}}dx\\
&\le Cr^{ d \eta} \|v\|^{ {2-2\eta}}+C\|v\|_{L^2_{\alpha}}^{{2-2\eta}}(\int_{|x|\ge r}\frac 1{|x|^{\frac {\alpha(2-2\eta)}{\eta}}}dx)^{\eta}\\
&\le Cr^{ d \eta} \|v\|^{ {2-2\eta}}+C r^{- \alpha(2-2\eta)+d\eta} \|v\|_{L^2_{\alpha}}^{{2-2\eta}}.
\end{align*}
Let $r=(\frac {\|v\|_{L^2_{\alpha}}}{\|v\|})^{\frac {1} {\alpha}}$, we complete the proof.
\qed
\begin{prop}\label{wei-hol-1}
Let $T>0$, $\mathcal O=\mathbb R^d$, $d\in \mathbb N^+$.
Under the condition of Proposition \ref{mul} or Proposition \ref{weak-loc}, let $(u^{\epsilon},\tau_{\epsilon}^*)$ be a local mild solution in $\mathbb H^{\mathbf s}$, $\mathbf s\ge 1$ for any $p\ge 1.$
Let $u_0\in L^2_{\alpha}\cap \mathbb H^1$, for some $\alpha\in (0,1]$, have any finite $p$th moment.
Then the solution $u^{\epsilon}$ of regularized problem satisfies for $\alpha\in (0,1],$
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau_{\epsilon}^*\wedge T)}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2p}\Big]&\le C(Q,T,\lambda,p,u_0).
\end{align*}
\end{prop}
{\noindent\bf{Proof}}
We first introduce the stopping time $$\tau_R=\inf\Big\{t\in [0,T]: \sup_{s\in[0,t]}\|u^{\epsilon}(s)\|_{L^2_{\alpha}}\ge R\Big\}\wedge \tau_{\epsilon}^*,$$ then show that $\mathbb E\Big[\sup\limits_{t\in[0,\tau_{R}]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2p}\Big]\le C(T,u_0,Q,p)$ independent of $R$. After taking $R\to \infty$, we get $\tau_R=\tau_{\epsilon}^*, a.s.$ For simplicity, we only prove uniform upper bound when {\color{red}$p=1$.}
Taking $0<t\le t_1\le \tau_R,$
and applying the It\^o formula to $\|u^{\epsilon}\|_{L^2_{\alpha}}^{2}=\int_{\mathbb R^d}(1+|x|^2)^{\alpha}|u^{\epsilon}|^2dx$,
we get
\begin{align*}
\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}
&=\|u^{\epsilon}_0\|_{L^2_{\alpha}}^{2}
+\int_0^t 2 \<(1+|x|^2)^{\alpha}u^{\epsilon}(s),\mathbf i \Delta u^{\epsilon}(s)\>ds+\int_0^t 2 \sum_{i\in \mathbb N^+}\<(1+|x|^2)^{\alpha}Q^{\frac 12}e_i, Q^{\frac 12}e_i\>ds\\
&\quad+\int_0^t 2\<(1+|x|^2)^{\alpha}u^{\epsilon}(s), \mathbf i f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)\>ds
+\int_0^t 2\<(1+|x|^2)^{\alpha}u^{\epsilon}(s),dW(s)\>
\end{align*}
for additive noise case,
and
\begin{align*}
\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}
&=\|u^{\epsilon}_0\|_{L^2_{\alpha}}^{2}
+2\int_0^t \<(1+|x|^2)^{\alpha}u^{\epsilon}(s),\mathbf i \Delta u^{\epsilon}(s)+\mathbf i f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)\>ds\\
&\quad-\int_0^t \<(1+|x|^2)^{\alpha}u^{\epsilon}(s)(s), (g(|u^{\epsilon}(s)|^2))^2u^{\epsilon}(s)\sum_{i}|Q^{\frac 12}e_i|^2\>ds\\
&\quad+\int_0^t \sum_{i}\<(1+|x|^2)^{\alpha} g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)Q^{\frac 12}e_i, g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)Q^{\frac 12}e_i\>ds\\
&\quad -2\int_0^t \<(1+|x|^2)^{\alpha}u^{\epsilon}(s), \mathbf i \sum_{k\in \mathbb N^+}g(|u^{\epsilon}(s)|^2) g'(|u^{\epsilon}(s)|^2) |u^{\epsilon}(s)|^2 u^{\epsilon}(s)
Im(Q^{\frac 12}e_k)Q^{\frac 12}e_k \>ds\\
&\quad +2\int_0^t\<(1+|x|^2)^{\alpha}u^{\epsilon}(s),\mathbf i g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)dW(s)\>
\end{align*}
for multiplicative noise case.
Using integration by parts, then taking supreme over $t\in [0,t_1]$ and applying the Burkerholder inequality, we deduce
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,t_1]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}\Big]
&\le \mathbb E\Big[\|u^{\epsilon}_0\|_{L^2_{\alpha}}^{2}\Big]
+C_{\alpha} \mathbb E \Big[\int_0^T\Big| \<(1+|x|^2)^{\alpha-1} x u^{\epsilon}(s),\nabla u^{\epsilon}(s)\> \Big|ds\Big]\\
&\quad+C\Big(\mathbb E \Big[\int_0^{t_1}\sum_{i\in\mathbb N^+}\|(1+|x|^2)^{\frac \alpha 2}{{u^{\epsilon}(s)}}\|^2 \|(1+|x|^2)^{\frac \alpha 2} Q^{\frac 12}e_i\|^2 ds\Big]\Big)^{\frac 12}
\end{align*}
for additive noise case,
and
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,t_1]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}\Big]
&\le \mathbb E\Big[\|u^{\epsilon}_0\|_{L^2_{\alpha}}^{2}\Big]
+C_{\alpha} \mathbb E \Big[\int_0^T\Big| \<(1+|x|^2)^{\alpha-1} x u^{\epsilon}(s),\nabla u^{\epsilon}(s)\> \Big|ds\Big]\\
&\quad+C\Big(\mathbb E \Big[\int_0^{t_1}\sum_{i\in\mathbb N^+}\|(1+|x|^2)^{\frac \alpha 2}u^{\epsilon}(s)\|^4 \|Q^{\frac 12}e_i\|_{L^{\infty}}^2 ds\Big]\Big)^{\frac 12}
\end{align*}
for multiplicative noise case.
Then Young's and Gronwall's inequalities, together with a priori estimate of $u^{\epsilon}$ in $\mathbb H^1$, lead to the desired result.
\qed
\begin{cor}\label{cor-con-l2a}
Under the condition of Lemma \ref{lm-con-h1}, the solution $u^{\epsilon}$ of regularized problem satisfies for $\alpha\in (0,1],$
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau_{\epsilon}^*\wedge T)}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2p}\Big]&\le C(Q,T,\lambda,p,u_0).
\end{align*}
\end{cor}
It is not possible to obtain the uniform bound of the exact solution in $L^2_{\alpha}$ for $\alpha\in (1,2]$ like the deterministic case. The main reason is that the rough driving noise leads to low H\"older regularity in time and loss of uniform estimate in $\mathbb H^2$ for the mild solution. We can not expect that the mild solution of \eqref{Reg-SlogS} enjoys $\epsilon$-independent estimate in $\mathbb H^2.$ More precisely, we prove that applying the regularization $f_{\epsilon}(|x|^2)=\log(|x|^2+\epsilon)$ in Proposition \ref{mul}, one can only expect $\epsilon$-dependent estimate in $\mathbb H^2$. We omit the tedious calculation and procedures, and present a sketch of the proof for Lemma \ref{h2-pri} and Propostion \ref{wei-hol-2} in Appendix.
\begin{lm}\label{h2-pri}
Let $T>0$ and $d=1$. Under the condition of Proposition \ref{mul}, assume that $(u^{\epsilon},\tau_{\epsilon}^*)$ is the local mild solution in $\mathbb H^1$. In addition assume that $\sum_{i\in \mathbb N^+}\|Q^{\frac 12}e_i\|_{W^{2,\infty}}^2<\infty$ for the multiplicative noise. Then for any $p\ge2,$ there exists a positive $C(Q,T,\lambda,p,u_0)$ such that
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau_{\epsilon}^*\wedge T)}\|u^{\epsilon}(t)\|_{\mathbb H^2}^{2p}\Big]\le
C(Q,T,\lambda,p,u_0)(1+\epsilon^{-2p}).
\end{align*}
\end{lm}
\begin{prop}\label{wei-hol-2}
Assume that $\mathcal O=\mathbb R^d$. Let $T>0$ and $d=1$. Under the condition of Proposition \ref{mul}, assume that $(u^{\epsilon},\tau_{\epsilon}^*)$ is the local mild solution in $\mathbb H^1$.
Let $u_0\in L^2_{\alpha}$, for some $\alpha\in (1,2]$. In addition assume that $\sum_{i}\|Q^{\frac 12}e_i\|_{L^2_{\alpha}}^2<\infty$ in the additive noise case and that $\sum_{i\in \mathbb N^+}\|Q^{\frac 12}e_i\|_{W^{2,\infty}}^2<\infty$ for the multiplicative noise. Then the solution $u^{\epsilon}(t)$ of regularized problem satisfies for $\alpha\in (1,2],$
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau_{\epsilon}^*\wedge T)}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2p}\Big]&\le C(Q,T,\lambda,p,u_0)(1+\epsilon^{-2p}).
\end{align*}
\end{prop}
The above results indicate that both spatial and temporal regularity for SLogS equation are rougher than deterministic LogS equation.
In the following, we present the behavior of the regularized energy for the RSlogS equation.
When applying $f_{\epsilon}(|x|^2)=\log(|x|^2+\epsilon),$ the modified energy of \eqref{Reg-SlogS} becomes $$H
_{\epsilon}(u^{\epsilon}(t)):=K(u^{\epsilon})-\frac \lambda{2}F_{\epsilon}(|u^{\epsilon}|^2)$$ with $F_{\epsilon}(|u^{\epsilon}|^2)=\int_{\mathcal O}\Big((\epsilon+|u^{\epsilon}|^2)\log(\epsilon+|u^{\epsilon}|^2)-|u^{\epsilon}|^2-\epsilon\log\epsilon\Big) dx.$
When using another regularization function $f_{\epsilon}(|u|^2)=\log(\frac {\epsilon+|u|^2}{1+\epsilon|u|^2}),$ its regularized entropy in the modified energy $H_{\epsilon}$ becomes $$ F_{\epsilon}(|u^{\epsilon}|^2)=\int_{\mathcal O}\Big(|u^{\epsilon}|^2\log(\frac {|u^{\epsilon}|^2+\epsilon}{1+|u^{\epsilon}|^2\epsilon})+\epsilon \log(|u^{\epsilon}|^2+\epsilon)-\frac1{\epsilon}\log(\epsilon |u^{\epsilon}|^2+1)-\epsilon\log(\epsilon)\Big) dx.$$
In general, the modified energy is defined by the regularized entropy $\widetilde F(\rho)=\int_\mathcal O\int_0^{\rho}f_{\epsilon}(s)dsdx,$ where $f_{\epsilon}(\cdot)$ is a suitable approximation of $\log(\cdot).$
We remark that the regularized energy is well-defined when $\mathcal O$ is a bounded domain. The additional constant term $\epsilon\log(\epsilon)$ ensures that the regularized energy is still well-defined when $\mathcal O=\mathbb R^d.$ We leave the proof of Proposition \ref{eng-pri} in Appendix.
\begin{prop}\label{eng-pri}
Let $T>0$. Under the condition of Proposition \ref{mul} or Proposition \ref{weak-loc}, assume that $(u^{\epsilon},\tau_{\epsilon}^*)$ is the local mild solution in $\mathbb H^1$.
Then for any $p\ge2,$ there exists a positive constant $C(Q,T,\lambda,p,u_0)$ such that
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau_{\epsilon}\wedge T)}|H_{\epsilon}(u^{\epsilon}(t))|^p\Big]\le
C(Q,T,\lambda,p,u_0).
\end{align*}
\end{prop}
Below, we present the global existence of the unique mild solution for Eq. \eqref{Reg-SlogS} in $\mathbb H^2$ based on Proposition \ref{mul}, Lemma \ref{h2-pri}, Proposition \ref{weak-loc} and Lemma \ref{h1-pri}, as well as a standard argument in \cite{BD03}.
\begin{prop}\label{prop-well-reg}
Let Assumption \ref{main-as} hold and $(u^{\epsilon},\tau_{\epsilon}^*)$ be the local mild solution in Proposition \ref{mul} or Proposition \ref{weak-loc}.
Then the mild solution $u^{\epsilon}$ in $\mathbb H^1$ is global, i.e., $\tau^{*}_{\epsilon}=+\infty, a.s.$
In addition assume that $d=1$ in Proposition \ref{mul}, the mild solution $u^{\epsilon}$ in $\mathbb H^2$ is global.
\end{prop}
\section{Well-posedness for SLogS equation}
Based on the a priori estimates of the regularized problem, we are going to prove the strong convergence of any sequence of the solutions of the regularized problem.
This immediately implies that the existence and uniqueness of the mild solution in $L^2$
for SLogS equation.
\subsection{Well-posedness for SLogS equation via strong convergence approximation}
In this part, we not only show the strong convergence of a sequence of solutidregularized problems, but also give the explicit strong convergence rate.
The strong convergence rate of the regularized SLogS equation will make a great contribution to the numerical analysis of numerical schemes for the SLogS equation. And this topic will be studied in a companion paper.
For the strong convergence result, we only present the mean square convergence rate result since the proof of the strong convergence rate in $L^q(\Omega), q\ge 2$ is similar.
In this section, the properties of regularization function $f_{\epsilon}$ in Lemmas \ref{frist-loc} and \ref{sec-reg} will be frequently used.
\iffalse
\begin{cor}\label{add-wei-l2a}
Assume that $\mathcal O=\mathbb R^d$, $d\le 3$.
Let $u_0\in \mathbb H^2\cap L^2_{\alpha}$, for some $\alpha\in (1,2)$, $\sum_{i}\|Q^{\frac 12}e_i\|_{L^2_{\alpha}}^2<\infty$. Then the solution $u^{\epsilon}$ of regularized problem satisfies
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}\|_{L^2_{\alpha}}^{2p}\Big]&\le C(T,u_0,Q,p)(1+\epsilon^{-\frac 1{2-d}p}).
\end{align*}
\end{cor}
\textbf{Proof}
Due to the proof of Proposition \ref{wei-est}, we have that
$$\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}\|_{L^2_{1}}^{2p}\Big]\le C(T,u_0,Q,p).$$
Integration by parts and H\"older's inequality yield that for some small $\eta>0,$
\begin{align*}
&\|(1+|x|^2)^{\frac \alpha 2-\frac 12}\nabla u^{\epsilon}\|^2\\
&=\<(1+|x|^2)^{\alpha -1} \nabla u^{\epsilon}, \nabla u^{\epsilon}\>\\
&=-\<(1+|x|^2)^{\alpha -1}u^{\epsilon}, \Delta u^{\epsilon}\>
-2(\alpha-1)\<(1+|x|^2)^{\alpha -2} x \nabla u^{\epsilon}, u^{\epsilon}\>\\
&\le C\|u^{\epsilon}\|_{L^2_{\max(2\alpha-2,0)}}\|\Delta u^{\epsilon}\|+ C(\eta)|\alpha-1|\|u^{\epsilon}\|^2_{L^2_{\alpha}}+\eta |\alpha-1|\|\nabla u^{\epsilon}\|^2_{L^2_{\max(\alpha-3,0)}}.
\end{align*}
Then if $\alpha\le \frac 32$, we use Proposition \ref{wei-est} and get the desired result.
If $\alpha\in (\frac 32, 2)$, we use the result of $\alpha\le \frac 32$ and complete the proof.
\qed
\subsection{Additive noise case}
\begin{lm}\label{add-con}
Suppose that $d\le 3$, $\sum_{i}\|Q^{\frac 12}e_i\|_{\mathbb H^2}^2<\infty,$ $u_0\in \mathbb H^2\cap L^2_{\alpha}$ for some $\alpha\in (\frac d2,2]$. Let $(\epsilon_n)_{n\le N^+}$ be a decreasing sequence satisfying $\lim_{n\to \infty}\epsilon_n=0$.
Then $u^{\epsilon_n}$ is strongly convergent in $C([0,T];\mathbb H)$ and satisfies for $n\ge m$
\begin{align*}
\mathbb E \Big[\sup_{t\in[0,T]}\|u^{\epsilon_n}-u^{\epsilon_m}\|^{2p}\Big]
\le C(u_0,T,Q)\epsilon_{m}^{2p}.
\end{align*}
\end{lm}
\textbf{Proof}
Subtracting the equation of $u^{\epsilon_n}$ from the equation of $u^{\epsilon_m}$, we have that
\begin{align*}
d(u^{\epsilon_m}-u^{\epsilon_n})=\mathbf i \Delta (u^{\epsilon_m}-u^{\epsilon_n})
+\mathbf i \lambda (f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n})dt.
\end{align*}
Then multiplying the above equation by the conjugate of $u^{\epsilon_m}-u^{\epsilon_n}$, using Lemma \ref{loc} and mean-value theorem, we arrive that
\begin{align*}
&\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\
&=\int_0^t 2\<u^{\epsilon_m}-u^{\epsilon_n},\mathbf i \lambda f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}\>ds\\
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
\\
&\quad+4|\lambda| \int_0^t |Im \<u^{\epsilon_m}(s)-u^{\epsilon_n}(s),(f_{\epsilon_m}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}\>| ds\\
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds\\
&\quad+4|\lambda| \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}
\Big\|\frac {(\epsilon_m-\epsilon_n)|u^{\epsilon_n}|^2}{\epsilon_m+|u^{\epsilon_n}|^2}\Big\|_{L^{\infty}} ds\\
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_m \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds.
\end{align*}
Now, if the domain $\mathcal O$ is bounded, the convergence rate is immediately obtained. Indeed, by using the H\"older inequality $\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\le |\mathcal O|^{\frac 12}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|$ on the last equality and using the Gronwall's inequality, we get
\begin{align*}
\sup_{t\in[0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\le e^{6|\lambda| T} 2|\lambda| |\mathcal O|\epsilon_m^2.
\end{align*}
Taking the $p$ moment, we complete the proof for the bounded domain.
Next, we deal with the case that $\mathcal O=\mathbb R^d.$
Lemma \ref{wei-hol} and Young's inequality yield that
\begin{align*}
\epsilon_m\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^1}
&\le \epsilon_m C_d \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^{1-\frac d{2\alpha}}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac d{2\alpha}}\\
&\le C_d\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2
+C_d \epsilon_m^{\frac 2{1+\frac d{2\alpha}}} \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac d{2\alpha}\frac 2{1+\frac d{2\alpha}}}.
\end{align*}
Applying Gronwall's inequality, we obtain
\begin{align*}
\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2
&\le C_de^{Ct} \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac {2d}{2\alpha +d}} \epsilon_m^{\frac {4\alpha}{2\alpha+d}}.
\end{align*}
Similarly, taking $p$ moment on the above inequality and using Proposition \ref{wei-est}, we get that
\begin{align}\label{est-rd}
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^{2p}\Big]
&\le C_de^{CT} \sup_{t\in [0,T]}\mathbb E\Big[\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac {2d}{2\alpha +d}p}\Big] \epsilon_m^{\frac {4\alpha}{2\alpha+d}p}\\\nonumber
&\le C_{p,d}e^{CT}(\epsilon_m^{\frac {4\alpha-d}{2\alpha+d}p}+\epsilon_m^{\frac {4\alpha}{2\alpha+d}p}\epsilon_n^{-\frac d{2\alpha+d}p}).
\end{align}
Based on the above estimate, we claim that there exists a sequence $\{\epsilon_n\}$ such that $u^{\epsilon_n}$ is strongly convergent in $L^2$.
More precisely, let us choose $\epsilon_n=n^{-k},$ for some $k\ge1.$
Then consider the summation $u^{\epsilon_j}+\sum_{j=1}^N u^{\epsilon_{j+1}}-u^{\epsilon_j},$ we only need to show that the limit of the summation sequence is well-defined in the norm $M_{\mathcal F}^{p}(\Omega;\mathcal C([0,T];L^2)).$
To this end, we show that $\sum_{j=1}^{\infty} \|u^{\epsilon_{j+1}}-u^{\epsilon_j}\|_{M_{\mathcal F}^{p}(\Omega;\mathcal C([0,T];L^2))}$ is summable. Using \eqref{est-rd}, we have that
\begin{align*}
\sum_{j=1}^{\infty} \|u^{\epsilon_{j+1}}-u^{\epsilon_j}\|_{M_{\mathcal F}^{p}(\Omega;\mathcal C([0,T];L^2))}
&\le
\sum_{j=1}^{\infty}
C_{p,d}e^{CT}(\epsilon_j^{\frac {4\alpha-d}{4\alpha+2d}}+\epsilon_j^{\frac {4\alpha}{4\alpha+2d}}\epsilon_{j+1}^{-\frac d{4\alpha+2d}})\\
&\le \sum_{j=1}^{\infty}
C_{p,d}e^{CT}j^{-\frac {4\alpha-d}{4\alpha+2d}k}.
\end{align*}
By choosing $k$ such that $\frac {4\alpha-d}{4\alpha+2d}k>1,$ we conclude that $u^{\epsilon_n}$ is a Cauchy sequence in $M_{\mathcal F}^{p}(\Omega;\mathcal C([0,T];L^2)).$
\qed
\fi
In the multiplicative noise case, Assumption \ref{ass-strong} is needed to obtain the strong convergence rate of the solution of Eq. \eqref{Reg-SlogS}.
We remark that the assumption can be weaken if one only wants to obtain the strong convergence instead of deriving a convergence rate.
Some sufficient condition for \eqref{con-g1} in Assumption \ref{ass-strong} is
$$\Big|(g'(|x|^2)g(|x|^2)|x|^2-g'(|y|^2)g(|y|^2)|y|^2)(|x|^2-|y|^2)\Big|\le C_g|x-y|^2, x,y\in \mathbb C$$
or
$$
\Big|(g'(|x|^2)g(|x|^2)|x|^2-g'(|y|^2)g(|y|^2)|y|^2)(|x|+|y|)\Big|\le C_g|x-y|, x,y\in \mathbb C.
$$
Functions like $1, \frac 1{c+x},\frac x{c+x}, \frac x{c+x^2}, \log(\frac {c+|x|^2}{1+c|x|^2})$ with $c>0$, etc., will satisfy Assumption \ref{ass-strong}.
\iffalse
\begin{prop}\label{wei-est-mul}
Assume that $\mathcal O=\mathbb R^d$, $d\le 3$.
Let $u_0\in \mathbb H^2\cap L^2_{\alpha}$, for some $\alpha\in (0,2]$, $\sum_{i}\|Q^{\frac 12}e_i\|_{\mathbb H^1}^2<\infty$. Then the solution $u^{\epsilon}$ of regularized problem satisfies for $\alpha\in (0,1],$
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}\|_{L^2_{\alpha}}^{2p}\Big]&\le C(T,u_0,Q,p),
\end{align*}
and
for $\alpha\in (1,2],$
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}\|_{L^2_{\alpha}}^{2p}\Big]&\le C(T,u_0,Q,p)(1+\epsilon^{-\frac 2{2-d}p}),
\end{align*}
\end{prop}
\fi
\iffalse
\begin{lm}\label{mul-con}
Suppose that $d\le 3$, $\sum_{i}\|Q^{\frac 12}e_i\|_{\mathbb H^2}^2+\|Q^{\frac 12}e_i\|_{L^2_{\alpha}}^2<\infty,$ $u_0\in \mathbb H^2\cap L^2_{\alpha}$ for some $\alpha\in (\frac d2,2]$. If $d=3,$ we in addition assume that $\sum_{i}\|\nabla Q^{\frac 12}e_i\|_{L^{\infty}}^2<\infty.$ Let $(\epsilon_n)_{n\le N^+}$ be a decreasing sequence satisfying $\lim_{n\to \infty}\epsilon_n=0$.
Then $u^{\epsilon_n}$ is strongly convergent in $C([0,T];\mathbb H)$ and satisfies for $n\ge m$
\begin{align*}
\sup_{t\in[0,T]}\mathbb E \Big[\|u^{\epsilon_n}-u^{\epsilon_m}\|^2\Big]
\le C(u_0,T,Q)\epsilon_{m}^2.
\end{align*}
\end{lm}
\textbf{Proof}
Subtracting the equation of $u^{\epsilon_n}$ from the equation of $u^{\epsilon_m}$, we have that
\begin{align*}
d(u^{\epsilon_m}-u^{\epsilon_n})&=\mathbf i \Delta (u^{\epsilon_m}-u^{\epsilon_n})
+\mathbf i \lambda (f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n})dt\\
&\quad -\frac 12\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)dt\\
&\quad+\mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(t)
\end{align*}
Then using the It\^o formula, Lemma \ref{loc} and mean-value theorem, we arrive that
\begin{align*}
&\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\
&=\int_0^t 2\<u^{\epsilon_m}-u^{\epsilon_n},\mathbf i \lambda f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}\>ds\\
&\quad-\int_0^t \<u^{\epsilon_m}-u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)\>ds\\
&\quad+2\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\\
&\quad+\int_0^t \<g(|u^{\epsilon_m}|^2)u^{\epsilon_m}- g(|u^{\epsilon_n}|^2)u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)ds\\
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_m \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds\\
&\quad+2\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\\
&\quad+\int_0^t \<g(|u^{\epsilon_m}|^2)u^{\epsilon_m}- g(|u^{\epsilon_n}|^2)u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)\>ds.
\end{align*}
By taking expectation and using the property \eqref{con-g}, we have that
\begin{align*}
&\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_m \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds\\
&\quad-\mathbb E\Big[\int_0^t \<u^{\epsilon_m}-u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)\>ds\Big]\\
&\quad+\mathbb E\Big[\int_0^t \<g(|u^{\epsilon_m}|^2)u^{\epsilon_m}- g(|u^{\epsilon_n}|^2)u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)\>ds\Big]\\
&=\int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_m \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds\\
&\quad-\mathbb E\Big[\int_0^t \sum_{k\in\mathbb N^+}\<|Q^{\frac 12}e_k|^2
g(|u^{\epsilon_m}|^2)u^{\epsilon_m},g(|u^{\epsilon_n}|^2)u^{\epsilon_n}-g(|u^{\epsilon_m}|^2)u^{\epsilon_n}\>ds\Big]\\
&\quad-\mathbb E\Big[\int_0^t \sum_{k\in\mathbb N^+}\<|Q^{\frac 12}e_k|^2
g(|u^{\epsilon_n}|^2)u^{\epsilon_n},g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_m}\>ds\Big]\\
&= \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_m \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds\\
&\quad+\mathbb E\Big[\int_0^t \sum_{k\in\mathbb N^+}\<|Q^{\frac 12}e_k|^2
(g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))u^{\epsilon_m},(g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))u^{\epsilon_n}\>ds\Big]\\
&\le \int_0^t (4|\lambda|+C(g,Q)) \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_m \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds.
\end{align*}
If the domain $\mathcal O$ is bounded, the convergence rate is immediately obtained by using the H\"older inequality $\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\le |\mathcal O|^{\frac 12}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|$ on the last equality and using the Gronwall's inequality, we get
\begin{align}\label{bond-est}
\mathbb E\Big[\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le e^{(6|\lambda| +C(g,Q))T} 2|\lambda| |\mathcal O|\epsilon_m^2.
\end{align}
Similarly, taking supreme over $t$ and then taking expectation, together with the Burkholder and Young inequalities, we can get for a small $\eta>0$,
\begin{align*}
&\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
&\le e^{(6|\lambda| +C(g,Q))T} C(\lambda,Q,T)\epsilon_m^2\\
&\quad+C\mathbb E\Big[\sup_{t\in [0,T]}\Big|\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\Big|\Big]\\
&\le e^{(6|\lambda| +C(g,Q))T} C(\lambda,Q,T)\epsilon_m^2\\
&\quad+C\mathbb E\Big[\Big(\int_0^T\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\|u^{\epsilon_m}-u^{\epsilon_n}\|^4ds\Big)^{\frac 12}\Big]\\
&\le e^{(6|\lambda| +C(g,Q))T} C(\lambda,Q,T)\epsilon_m^2\\
&\quad+C\mathbb E\Big[\sup_{s\in [0,T]}\|u^{\epsilon_m}-u^{\epsilon_n}\|^2\Big(\int_0^T\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\|u^{\epsilon_m}-u^{\epsilon_n}\|^2ds\Big)^{\frac 12}\Big]\\
&\le e^{(6|\lambda| +C(g,Q))T} C(\lambda,Q,T)\epsilon_m^2+\eta \mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
&\quad+C(\eta)\mathbb E\Big[\int_0^T\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\|u^{\epsilon_m}-u^{\epsilon_n}\|^2ds\Big].
\end{align*}
Taking $\eta<\frac 12$, and using the \eqref{bond-est}, we have that
$$\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le C(\lambda,g,T,Q)\epsilon_m^2.$$
Next, we deal with the case that $\mathcal O=\mathbb R^d.$
Lemma \ref{wei-hol} and Young's inequality yield that
\begin{align*}
\epsilon_m\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^1}
&\le \epsilon_m C_d \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^{1-\frac d{2\alpha}}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac d{2\alpha}}\\
&\le C_d\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2
+C_d \epsilon_m^{\frac 2{1+\frac d{2\alpha}}} \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac d{2\alpha}\frac 2{1+\frac d{2\alpha}}}.
\end{align*}
Similarly, taking expectation on the above inequality and using Proposition \ref{wei-est-mul}, we get that
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^{2}\Big]
&\le C_de^{CT} \sup_{t\in [0,T]}\mathbb E\Big[\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^{\frac {2d}{2\alpha +d}}\Big] \epsilon_m^{\frac {4\alpha}{2\alpha+d}}\\\nonumber
&\le C_{d}e^{CT}(\epsilon_m^{\frac {4\alpha-d}{2\alpha+d}}+\epsilon_m^{\frac {4\alpha}{2\alpha+d}}\epsilon_n^{-\frac d{2\alpha+d}}).
\end{align*}
Based on the above estimate, we claim that there exists a sequence $\{\epsilon_n\}$ such that $u^{\epsilon_n}$ is strongly convergent in $L^2$.
More precisely, let us choose $\epsilon_n=n^{-k},$ for some $k\ge1.$
Then consider the summation $u^{\epsilon_j}+\sum_{j=1}^N u^{\epsilon_{j+1}}-u^{\epsilon_j},$ we only need to show that the limit of the summation sequence is well-defined in the norm $M_{\mathcal F}^{2}(\Omega;\mathcal C([0,T];L^2)).$
To this end, we show that $\sum_{j=1}^{\infty} \|u^{\epsilon_{j+1}}-u^{\epsilon_j}\|_{M_{\mathcal F}^{2}(\Omega;\mathcal C([0,T];L^2))}$ is summable. Using \eqref{est-rd}, we have that
\begin{align*}
\sum_{j=1}^{\infty} \|u^{\epsilon_{j+1}}-u^{\epsilon_j}\|_{M_{\mathcal F}^{2}(\Omega;\mathcal C([0,T];L^2))}
&\le
\sum_{j=1}^{\infty}
C_{d}e^{CT}(\epsilon_j^{\frac {4\alpha-d}{4\alpha+2d}}+\epsilon_j^{\frac {4\alpha}{4\alpha+2d}}\epsilon_{j+1}^{-\frac d{4\alpha+2d}})\\
&\le \sum_{j=1}^{\infty}
C_{d}e^{CT}j^{-\frac {4\alpha-d}{4\alpha+2d}k}.
\end{align*}
By choosing $k$ such that $\frac {4\alpha-d}{4\alpha+2d}k>1,$ we conclude that $u^{\epsilon_n}$ is a Cauchy sequence in $M_{\mathcal F}^{2}(\Omega;\mathcal C([0,T];L^2)).$
\qed
\fi
The main idea of the proof lies on showing that for a decreasing sequence $\{\epsilon_{n}\}_{n\in \mathbb N^+}$
satisfying $\lim\limits_{n\to\infty}\epsilon_n=0$, $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}$ must be a Cauchy sequence in $C([0,T];L^p(\Omega; \mathbb H)),$ $p\ge 2.$ As a consequence, we obtain that there exists a limit process $u$ in $C([0,T];L^p(\Omega; \mathbb H))$ which is shown to be independent of the sequence $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}$ and is the unique
mild solution of the mild form of \eqref{SlogS}.
\textbf{[Proof of Theorem \ref{mild-general}]}
Based on Proposition \ref{prop-well-reg}, we can construct a sequence of mild solutions $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}$ of Eq. \eqref{Reg-SlogS} with $f_{\epsilon^n}(|x|^2)=\log(\frac {\epsilon_n+|x|^2}{1+\epsilon_n |x|^2})$. Here the decreasing sequence $\{\epsilon_{n}\}_{n\in \mathbb N^+}$ satisfies $\lim\limits_{n\to\infty}\epsilon_n=0.$ We use the following steps to complete the proof. For simplicity, we only present the details for $p=2$ since the procedures for $p>2$ are similar.
Step 1: $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}$ is a Cauchy sequence in $L^2(\Omega; C([0,T];\mathbb H)).$
Fix different $n,m\in \mathbb N^+$ such that $n<m.$
Subtracting the equation of $u^{\epsilon_n}$ from the equation of $u^{\epsilon_m}$, we have that
\begin{align*}
d(u^{\epsilon_m}-u^{\epsilon_n})=\mathbf i \Delta (u^{\epsilon_m}-u^{\epsilon_n})dt
+\mathbf i \lambda (f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n})dt
\end{align*}
for additive noise case,
and
\begin{align*}
&\quad d(u^{\epsilon_m}-u^{\epsilon_n})\\
&=\mathbf i \Delta (u^{\epsilon_m}-u^{\epsilon_n})dt
+\mathbf i \lambda (f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n})dt\\
&\quad -\frac 12\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)dt\\
&\quad-\mathbf i \sum_{k\in \mathbb N^+} Im(Q^{\frac 12}e_k) Q^{\frac 12}e_k \Big(g'(|u^{\epsilon_m}|^2)g(|u^{\epsilon_m}|^2)|u^{\epsilon_m}|^2u^{\epsilon_m}- g'(|u^{\epsilon_n}|^2)g(|u^{\epsilon_n}|^2)|u^{\epsilon_n}|^2u^{\epsilon_n}\Big)dt\\
&\quad+\mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(t)
\end{align*}
for multiplicative noise case.
Then using the It\^o formula to $\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2$, the properties of $f_{\epsilon}$ in Lemma \ref{sec-reg}, the mean value theorem and the Gagliardo--Nirenberg interpolation inequality, we obtain that for $\eta'(2-d)\le 2,$
\begin{align}\label{main-add}
&\quad\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\\nonumber
&=\int_0^t 2\<u^{\epsilon_m}-u^{\epsilon_n},\mathbf i \lambda f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}\>ds\\\nonumber
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda| \int_0^t |Im \<u^{\epsilon_m}(s)-u^{\epsilon_n}(s),(f_{\epsilon_m}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}\>| ds\\\nonumber
&\le \int_0^t 6|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds+4|\lambda| \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}
\Big\|\frac {(\epsilon_m-\epsilon_n)|u^{\epsilon_n}|}{\epsilon_m+|u^{\epsilon_n}|^2}\Big\|_{L^{\infty}} ds\\\nonumber
&\quad +2|\lambda|\int_0^t \|\log(1+\frac {(\epsilon_n-\epsilon_m)|u^{\epsilon_n}|^2}{1+\epsilon_m|u^{\epsilon_n}|^2}) |u^{\epsilon_n}|\|^2ds \\\nonumber
&\le \int_0^t 6|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}d+2|\lambda| C\epsilon_n^{\eta'}\int_0^t\|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds
\end{align}
for additive noise case,
and
\begin{align*}
&\quad\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\
&=\int_0^t 2\<u^{\epsilon_m}-u^{\epsilon_n},\mathbf i \lambda f_{\epsilon_m}(|u^{\epsilon_m}|^2)u^{\epsilon_m}-f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}\>ds\\
&\quad-\int_0^t \<u^{\epsilon_m}-u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(|g(|u^{\epsilon_m}|^2)|^2u^{\epsilon_m}- |g(|u^{\epsilon_n}|^2)|^2u^{\epsilon_n}\Big)\>ds\\
&\quad-2\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n},\mathbf i\sum_{k\in \mathbb N^+}Im(Q^{\frac 12}e_k) Q^{\frac 12}e_k \Big(g'(|u^{\epsilon_m}|^2)g(|u^{\epsilon_m}|^2)|u^{\epsilon_m}|^2u^{\epsilon_m}- g'(|u^{\epsilon_n}|^2)g(|u^{\epsilon_n}|^2)|u^{\epsilon_n}|^2u^{\epsilon_n}\Big)\>ds\\
&\quad+2\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\\
&\quad+\int_0^t \<g(|u^{\epsilon_m}|^2)u^{\epsilon_m}- g(|u^{\epsilon_n}|^2)u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2\Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}- g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)\>ds\\
&\le \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds +2|\lambda| C\epsilon_n^{\eta'}\int_0^t \|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds\\
&\quad+2\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\\
&\quad+\int_0^t \<(g(|u^{\epsilon_m}|^2)-g(|u^{\epsilon_n}|^2))u^{\epsilon_n},\sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|^2(g(|u^{\epsilon_m}|^2)- g(|u^{\epsilon_n}|^2))u^{\epsilon_m}\>ds\\
&-2\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n},\mathbf i \sum_{k\in \mathbb N^+}Im(Q^{\frac 12}e_k) Q^{\frac 12}e_k\Big(g'(|u^{\epsilon_m}|^2)g(|u^{\epsilon_m}|^2)|u^{\epsilon_m}|^2u^{\epsilon_m}- g'(|u^{\epsilon_n}|^2)g(|u^{\epsilon_n}|^2)|u^{\epsilon_n}|^2u^{\epsilon_n}\Big)\>ds
\end{align*}
for multiplicative noise case.
By using \eqref{con-g} and \eqref{con-g1} in Assumption \ref{ass-strong} and the assumptions on $Q$, we have that
\begin{align}\label{main-mul}
&\quad\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\\nonumber
&\le \int_0^t (4|\lambda|+C(g,Q)) \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds\\\nonumber
&\quad+2|\lambda| C\epsilon_n^{\eta'}\int_0^t \|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds+\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>.
\end{align}
Next we show the strong convergence of the sequence $\{u^{\epsilon_n}\}_{n\in\mathbb N^+}$ in the following different cases.
Case 1: $\mathcal O$ is a bounded domain.
By using the H\"older inequality $\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\le |\mathcal O|^{\frac 12}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|$ on \eqref{main-add} and \eqref{main-mul}, and using the Gronwall's inequality, we get
\begin{align*}
\sup_{t\in[0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\le C(\lambda,T,|\mathcal O|) (\epsilon_n+\epsilon_n^{\eta'})(1+\sup_{t\in[0,T]}\|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'})
\end{align*}
for additive noise case.
In the multiplicative noise case, taking supreme over $t$ and then taking expectation on \eqref{main-mul}, together with the Burkholder and Young inequalities, we get that for a small $\kappa>0$,
\begin{align*}
&\quad\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
&\le C(\lambda,T,|\mathcal O|,Q) (\epsilon_n+\epsilon_n^{\eta'})+C\mathbb E\Big[\sup_{t\in [0,T]}\Big|\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\Big|\Big]\\
&\le C(\lambda,T,|\mathcal O|,Q) (\epsilon_n+\epsilon_n^{\eta'})+C\mathbb E\Big[\Big(\int_0^T\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\|u^{\epsilon_m}-u^{\epsilon_n}\|^4ds\Big)^{\frac 12}\Big]\\
&\le C(\lambda,T,|\mathcal O|,Q) (\epsilon_n+\epsilon_n^{\eta'})+C\mathbb E\Big[\sup_{s\in [0,T]}\|u^{\epsilon_m}-u^{\epsilon_n}\|\Big(\int_0^T\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\|u^{\epsilon_m}-u^{\epsilon_n}\|^2ds\Big)^{\frac 12}\Big]\\
&\le C(\lambda,T,|\mathcal O|,Q) (\epsilon_n+\epsilon_n^{\eta'})+\kappa\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
&\quad+C(\kappa)\mathbb E\Big[\int_0^T\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\|u^{\epsilon_m}-u^{\epsilon_n}\|^2ds\Big].
\end{align*}
Taking $\kappa<\frac 12$, we have that
$$\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le C(Q,T,\lambda,p,u_0,|\mathcal O|)(\epsilon_n+\epsilon_n^{\eta'}).$$
Case 2: $\mathcal O=\mathbb R^d$.
Since $u_0\in L_2^{\alpha}, \alpha\in (0,1],$
using the interpolation inequality in Lemma \ref{wei-hol} implies that for any $\eta\in [0,1)$ and $\alpha>\frac {\eta d}{2(1-\eta)}$ (i.e., $\eta\in (0,\frac {2\alpha} {2\alpha+d})$),
\begin{align*}
&\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\\nonumber
\le& \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\int_0^t \epsilon_{n}^{\frac \eta2}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|\|u^{\epsilon_n}\|_{L^{2-2\eta}}^{1-\eta}ds\\\nonumber
&+2|\lambda| C\epsilon_n^{\eta'}\int_0^t \|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds\\\nonumber
\le& \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^{2} ds
+2|\lambda|\int_0^t\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2+
2|\lambda| \epsilon_{m}^{\eta} \int_0^t\|u^{\epsilon_n}\|_{L^{2-2\eta}}^{2-2\eta}ds\\\nonumber
&+2|\lambda| C\epsilon_n^{\eta'}\int_0^t \|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds\\\nonumber
\le&\int_0^t 6|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+2|\lambda|C \epsilon_{n}^{\eta} \int_0^t\|u^{\epsilon_n}\|_{L^{2}_\alpha}^{\frac {d\eta}{\alpha}}\|u^{\epsilon_n}\|^{2-2\eta-\frac{d\eta}{\alpha} }ds\\\nonumber
&+2|\lambda| C\epsilon_n^{\eta'}\int_0^t \|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds
\end{align*}
for additive noise case,
and
\begin{align*}
&\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\\nonumber
\le&\int_0^t C\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+C \epsilon_{m}^{\eta} \int_0^t\|u^{\epsilon_n}\|_{L^{2}_\alpha}^{\frac {d\eta}{\alpha}}\|u^{\epsilon_n}\|^{2-2\eta-\frac{d\eta}{\alpha} }ds\\\nonumber
&+2|\lambda| C\epsilon_n^{\eta'}\int_0^t \|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}ds+\Big|\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>\Big|\nonumber
\end{align*}
for multiplicative noise case.
Then taking supreme over $t$, taking expectation, using \eqref{con-g}, Lemma \ref{mass-pri} and Proposition \ref{wei-hol-1}, and applying Gronwall's inequality, we have that for $\alpha\in (0,1],$ $\eta\in (0,\frac {2\alpha} {2\alpha+d})$ and $\eta'(d-2)\le 2$,
\begin{align}\label{main-rd-sec-wei}
&\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\\nonumber
\le& C(T,Q,u_0,g)\mathbb E\Big[\sup_{[0,T]}\Big(\|u^{\epsilon_n}\|_{L^{2}_\alpha}^{\frac {d\eta}{\alpha}}\|u^{\epsilon_n}\|^{2-2\eta-\frac{d\eta}{\alpha} }+\|u^{\epsilon_n}\|_{L^{2+2\eta'}}^{2+2\eta'}\Big)\Big](\epsilon_n^{\eta}+\epsilon_n^{\eta'})\\\nonumber
\le& C(T,Q,u_0,g,\alpha,\eta)\epsilon_n^{\min(\eta,\eta')}.
\end{align}
\iffalse
The restriction on $\alpha$ lies the fact that we do not have the uniform control on $\|u^{\epsilon_n}\|_{L^2_{\alpha}}$ for $\alpha>1.$
If in addition assume that $u_0\in L_2^{\alpha}, \alpha\in (1,2],$ and $d=1,$
similar arguments, together with Lemmas \ref{mass-pri} and \ref{h1-pri}, and the proof of Proposition \ref{wei-hol-2}, yield that for $\eta\in (0,\frac {2\alpha} {2\alpha+d}),$
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]&\le C(\lambda,T) \epsilon_n^{\min(\eta(1-\beta\frac {d}{\alpha}),\eta')},
\end{align*}
where $\beta=\frac 12$ for $\alpha\in (1,\frac 32]$, $\beta=\frac 34$ for $\alpha\in (\frac 32,2)$ and $\beta=1$ for $\alpha=2.$
\fi
Step 2: The limit process $u$ of $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}$ in $\mathbb M_{\mathcal F}^2(C([0,T];\mathbb H))$ satisfies \eqref{SlogS} in mild form.
We use the multiplicative noise case to present all the detailed procedures.
It suffices to prove that each term in the mild form of RSlogS equation \eqref{Reg-SlogS}
convergenes to the corresponding part in
\begin{align*}
&S(t)u_0+\mathbf i\lambda\int_0^tS(t-s)\log(|u|^2)uds-\frac 12\int_0^tS(t-s)(g(|u|^2))^2u
\sum_{k}|Q^{\frac 12}e_k|^2ds\\\nonumber
&-\mathbf i \int_0^tS(t-s)g'(|u|^2)g(|u|^2)|u|^2u
\sum_{k}Im(Q^{\frac 12}e_k)Q^{\frac 12}e_kds+\mathbf i\int_0^tS(t-s)g(|u|^2)udW(s)\\
&:=S(t)u_0+V_1+V_2+V_3+V_4.
\end{align*}
We first claim that all the terms $V_1$-$V_4$ make sense.
By Lemma \ref{h1-pri} and Proposition \ref{wei-hol-1}, we have that for $p\ge2,$
\begin{align*}
\sup_{n}\mathbb E\left[\sup_{t\in[0,T]}\|u^{\epsilon_n}(t)\|_{\mathbb H^1}^p\right]+\sup_{n}\mathbb E\left[\sup_{t\in[0,T]}\|u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^p\right]&\le C(u_0,T,Q).
\end{align*}
By applying the Fourier transform and
Parseval's theorem, using the Fatou theorem and strong convergence of $(u^{\epsilon_n})_{n\in \mathbb N^+}$ in $\mathbb M_{\mathcal F}^2(C([0,T];\mathbb H))$, we obtain
\begin{align*}
&\mathbb E\Big[\sup_{t\in[0,T]}\|u(t)\|_{\mathbb H^1}^2\Big]+\mathbb E\Big[\sup_{t\in[0,T]}\|u(t)\|_{L^2_{\alpha}}^2\Big]\\
\le& \sup_{n}\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon_n}(t)\|_{\mathbb H^1}^2\Big]
+\sup_{n}\mathbb E[\sup_{t\in[0,T]}\|u^{\epsilon_n}(t)\|_{L^2_{\alpha}}^2]\le C(u_0,T,Q).
\end{align*}
\iffalse
If the domain $\mathcal O=\mathbb R^d$, we get
\begin{align*}
&\mathbb E[\sup_{t\in[0,T]}\|u^{\epsilon_n}\|_{\mathbb H^1}^2]
+\mathbb E[\sup_{t\in[0,T]}\|u^{\epsilon_n}\|_{L^2_{\alpha}}^2]
\\
&=\mathbb E\Big[\sup_{t\in[0,T]}\int_{\mathbb R^d} |\widehat u^{\epsilon_n}(\xi)|^2(1+|\xi|^2)d\xi\Big]+\mathbb E\Big[\sup_{t\in[0,T]}\int_{\mathbb R^d} |u^{\epsilon_n}(x)|^2(1+|x|^2)^{\alpha}dx\Big].
\end{align*}
The strong convergence of $(u^{\epsilon_n})_{n\in \mathbb N^+}$ in $\mathbb M_{\mathcal F}^2(C([0,T];L^2))$ and the uniform
equicontinuity of $u^{\epsilon_n}$ implies that
\begin{align*}
&\sup_{t\in[0,T]}|\widehat u^{\epsilon_n}(\xi)-\widehat u(\xi)|^2\to0,\; d\xi \times d\mathbb P,\\
&\sup_{t\in[0,T]}|u^{\epsilon_n}(x)- u(x)|^2\to0,\; dx\times d\mathbb P.
\end{align*}
By using the Fatou theorem, we have that
\fi
Then
the Gagliardo--Nirenberg interpolation inequality yields that for small $\eta',\eta>0$,
\begin{align*}
\|\log(|u|^2)u\|^2&=
\int_{|u|^2\ge 1} (\log(|u|^2))^2|u|^2dx+\int_{|u|^2\le 1}(\log(|u|^2))^2|u|^2dx\\
&\le \int_{|u|^2\ge 1} |u|^{2+{2\eta'}}dx+\int_{|u|^2\le 1}|u|^{2-{2\eta}}dx\\
&\le C(\|u\|_{L^{2+2\eta}}^{2+2\eta'}+\|u\|_{L^{2-2\eta}}^{2-2\eta})\\
&\le C(\|u\|^{2-2\eta}_{L^{2-2\eta}}+\|\nabla u\|^{d\eta'}\|u\|^{2\eta'+2-d\eta'}).
\end{align*}
When $\mathcal O=\mathbb R^d$, we use the weighted version of the interpolation inequality in Lemma \ref{wei-hol} to deal with the term $\|u\|_{L^{2-2\eta}}^{2-2\eta}$,
and have that for small $\eta<\frac {2\alpha}{2\alpha+d}$,
$$\|\log(|u|^2)u\|^2 \le C(\|\nabla u\|^{d\eta'}\|u\|^{2\eta'+2-d\eta'}+\|u\|_{L^2_{\alpha}}^{\frac {2d\eta}{\alpha}}\|u\|^{2-2\eta-\frac {2d\eta}{\alpha}}).$$
This implies that $V_1$ makes sense in $\mathbb M_{\mathcal F}^2(C([0,T];\mathbb H))$ by Proposition \ref{wei-hol-1}, Lemmas \ref{mass-pri} and \ref{h1-pri}.
Meanwhile, we can show that $V_2$-$V_4 \in \mathbb M_{\mathcal F}^2(C([0,T];\mathbb H))$ by using the Minkowski and Burkerholder inequalities due to our assumption on $g$ and $Q$.
Next, we show that the mild form of $u^{\epsilon_n}$ converges to
$S(t)u_0+V_1+V_2+V_3+V_4.$ To prove that
\begin{align*}
&\lim_{n\to\infty}\mathbb E\Big[\sup_{t\in[0,T]}\Big\|\int_0^tS(t-s)f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}ds-V_1\Big\|^2\Big]=0,
\end{align*}
we use the following decomposition of $f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}-\log(|u|^2)u.$
When $|u|>|u^{\epsilon_n}|$,
\begin{align*}
&f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}-\log(|u|^2)u\\
=&(f_{\epsilon_n}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u|^2))u^{\epsilon_n}+f_{\epsilon_n}(|u|^2)(u^{\epsilon_n}-u)+(f_{\epsilon_n}(|u|^2)-\log(|u|^2))u,
\end{align*}
and when $|u|<|u^{\epsilon_n}|,$
\begin{align*}
&f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}-\log(|u|^2)u\\
=&(\log(|u^{\epsilon_n}|^2)-\log(|u|^2))u+\log(|u^{\epsilon_n}|^2)(u^{\epsilon_n}-u)+(f_{\epsilon_n}(|u^{\epsilon_n}|^2)-\log(|u^{\epsilon_n}|^2))u^{\epsilon_n}.
\end{align*}
For convenience, let us show the estimate for $|u|>|u^{\epsilon_n}|$, the other case will be estimated in a similar way.
By using the H\"older inequality and the mean-valued theorem, we have that for small $\gamma>0,$
\begin{align*}
&\Big|(f_{\epsilon_n}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u|^2))u^{\epsilon_n}\Big|\\
\le& \Big|(\frac {|u|^2-|u^{\epsilon_n}|^2}{\epsilon_n+|u^{\epsilon_n}|^2})^{\frac 12-\frac 12\gamma}(f_{\epsilon_n}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u|^2))^{\frac 12+\frac 12\gamma}u^{\epsilon_n}\Big|\\
\le& \Big|(|u|-|u^{\epsilon}_n|)^{\frac 12-\frac 12\gamma}(|u|+|u^{\epsilon}_n|)^{\frac 12-\frac 12\gamma} \frac {|u^{\epsilon_n}|}{(\epsilon_n+|u^{\epsilon_n}|^2)^{\frac 12-\frac 12\gamma}}
(f_{\epsilon_n}(|u^{\epsilon_n}|^2)+f_{\epsilon_n}(|u|^2))^{\frac 12+\frac 12\gamma}\Big|.
\end{align*}
Then it implies that for small enough $\eta>0,$
\begin{align*}
&\int_{|u|>|u^{\epsilon_n}|}\Big|(f_{\epsilon_n}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u|^2))u^{\epsilon_n}\Big|^2dx\\
\le& \int_{|u|>|u^{\epsilon_n}|} \frac {|u^{\epsilon_n}|^2}{(\epsilon_n+|u^{\epsilon_n}|^2)^{1-\gamma}} |u-u^{\epsilon_n}|^{1-\gamma}
(|u|+|u^{\epsilon}|)^{1-\gamma}(f_{\epsilon_n}(|u^{\epsilon_n}|^2)+f_{\epsilon_n}(|u|^2))^{1+\gamma} dx\\
\le& \int_{|u|>|u^{\epsilon_n}|, \epsilon_n+|u^{\epsilon_n}|\le 1} \frac {|u^{\epsilon_n}|^2}{(\epsilon_n+|u^{\epsilon_n}|^2)^{1-\gamma}} |u-u^{\epsilon_n}|^{1-\gamma} (|u|+|u^{\epsilon}|)^{1-\gamma}(f_{\epsilon_n}(|u^{\epsilon_n}|^2)+f_{\epsilon_n}(|u|^2))^{1+\gamma} dx\\
& +\int_{|u|>|u|^{\epsilon_n}, \epsilon_n+|u^{\epsilon_n}|\ge 1} \frac {|u^{\epsilon_n}|^2}{(\epsilon_n+|u^{\epsilon_n}|^2)^{1-\gamma}} |u-u^{\epsilon_n}|^{1-\gamma} (|u|+|u^{\epsilon}|)^{1-\gamma}(f_{\epsilon_n}(|u^{\epsilon_n}|^2)+f_{\epsilon_n}(|u|^2))^{1+\gamma} dx\\
\le& C\int_{|u|>|u^{\epsilon_n}|, \epsilon_n+|u^{\epsilon_n}|\le 1} \frac {|u^{\epsilon_n}|^2}{(\epsilon_n+|u^{\epsilon_n}|^2)^{1-\gamma}} |u-u^{\epsilon_n}|^{1-\gamma} (|u|+|u^{\epsilon}|)^{1-\gamma}
((\epsilon_n+|u^{\epsilon_n}|^2)^{-\eta}+(\epsilon_n+|u|^2)^{\eta})dx\\
& +C\int_{|u|>|u|^{\epsilon_n}, \epsilon_n+|u^{\epsilon_n}|\ge 1} \frac {|u^{\epsilon_n}|^2}{(\epsilon_n+|u^{\epsilon_n}|^2)^{1-\gamma}} (|u|+|u^{\epsilon}|)^{1-\gamma} |u+u^{\epsilon}|^{1-\gamma}(\epsilon_n+|u|^2)^{\eta} dx.
\end{align*}
Now choosing $2-2\gamma+2\eta \le 2$, using the H\"older inequality and the weighted interpolation inequality in Lemma \ref{wei-hol}, we have that for $\alpha>\frac {\eta d}{1+\gamma-2\eta}$,
\begin{align*}
&\quad\int_{|u|>|u^{\epsilon_n}|}\Big|(f_{\epsilon_n}(|u^{\epsilon_n}|^2)-f_{\epsilon_n}(|u|^2))u^{\epsilon_n}\Big|^2dx\\
&\le C\int_{\mathbb R^d}|u^{\epsilon_n}|^{2\gamma-2\eta} |u-u^{\epsilon_n}|^{1-\gamma} (|u|+|u^{\epsilon}|)^{1-\gamma} dx\\
&\quad+C\int_{\mathbb R^d}|u^{\epsilon_n}|^{2\gamma-2\eta} |u-u^{\epsilon_n}|^{1-\gamma} (|u|+|u^{\epsilon}|)^{1-\gamma}(\epsilon_n+|u|^2)^{\eta} dx\\
&\le C\|u-u^{\epsilon_n}\|^{1-\gamma}\Big\| (|u|^{2\gamma-2\eta}+|u|^{2\gamma})(|u|+|u^{\epsilon}|)^{1-\gamma}\Big\|_{L^{\frac 2{1+\gamma}}}\\
&\le C\|u-u^{\epsilon_n}\|^{1-\gamma}\Big(\|u\|^{1+\gamma}+\|u\|_{L^{2-\frac {4\eta}{1+\gamma}}}^{1+\gamma-2\eta}\Big)\\
&\le C\|u-u^{\epsilon_n}\|^{1-\gamma}\Big(\|u\|^{1+\gamma}+\|u\|_{L^{2}_{\alpha}}^{\frac {\eta d}{\alpha(1+r)}}\|u\|^{1+\gamma-2\eta-\frac {\eta d}{\alpha(1+\gamma)}}\Big).
\end{align*}
For the term $f_{\epsilon_n}(|u|^2)(u^{\epsilon_n}-u),$ we similarly have that for $\eta'<\frac {2}{\min(d-2,0)}$ and $\eta<\frac{2\alpha}{2\alpha+d},$
\begin{align*}
&\quad\int_{|u|>|u^{\epsilon}|}|f_{\epsilon_n}(|u|^2)(u^{\epsilon_n}-u)|^2dx\\
&\le \int_{|u|>|u^{\epsilon}|,\epsilon_n+|u|^2\le 1} (\epsilon_n+|u|^2)^{-\frac \eta 2}|u^{\epsilon_n}-u|^2dx+\int_{|u|>|u^{\epsilon}|, \epsilon_n+|u|^2\ge 1} (\epsilon_n+|u|^2)^{\frac \eta 2}|u^{\epsilon_n}-u|^2dx\\
&\le C\|u^{\epsilon_n}-u\| \Big(\|u^{\epsilon_n}-u\|+\|u\|_{L^{2+2\eta'}}^{1+\eta'} +\|u^{\epsilon_n}\|_{L^{2-2\eta}}^{1-\eta}+\|u\|_{L^{2-2\eta}}^{1-\eta}\Big)\\
&\le C\|u^{\epsilon_n}-u\| \Big(\|u^{\epsilon_n}-u\|+\|\nabla u\|^{\frac {\eta' d}{2}}\|u\|^{1+\eta'-\frac {\eta' d}{2}}+\|u\|_{L^2_{\alpha}}^{\frac {d\eta}{2\alpha}}\|u\|^{1-\eta-\frac {d\eta}{2\alpha}}+\|u^{\epsilon_n}\|_{L^2_{\alpha}}^{\frac {d\eta}{2\alpha}}\|u^{\epsilon_n}\|^{1-\eta-\frac {d\eta}{2\alpha}}\Big).
\end{align*}
For the term $(f_{\epsilon_n}(|u|^2)-\log(|u|^2))u,$ the mean-valued theorem, the property that $\log(1+|x|)\le |x|$ and the Gagliardo--Nirenberg interpolation inequality
\begin{align*}
\|u\|_{L^q(\mathbb R^d)}\le C_1\|u\|^{1-\alpha}\|\nabla u\|^{\alpha}, \quad q=\frac {2d}{d-2\alpha},\; \text{for} \;\alpha\in (0,1]
\end{align*}
with $q=2\eta'+2$, yield that for $\eta'(d-2)\le 2$ and $\eta\le \frac{\alpha}{2\alpha-d},$
\begin{align*}
&\int_{\mathcal O}|(f_{\epsilon_n}(|u|^2)-\log(|u|^2))u|^2 dx\le C\epsilon_n^{\eta'}\|u\|_{L^{2\eta'+2}}^{2\eta'+2}+C\epsilon_n^{\eta}\|u\|_{L^{2-2\eta}}^{2-2\eta}\\
\le&
C\epsilon_n^{\eta'}\|\nabla u\|^{\frac {\eta' d}{2}}\|u\|^{1+\eta'-\frac {\eta' d}{2}}+C\epsilon_n^{\eta}
\|u^{\epsilon_n}\|_{L^2_{\alpha}}^{\frac {d\eta}{\alpha}}\|u^{\epsilon_n}\|^{1-\eta-\frac {d\eta}{\alpha}}.
\end{align*}
Combining the above estimates, using the a priori estimate of $u^{\epsilon}$ and $u$ in Lemmas \ref{mass-pri} and \ref{h1-pri} and Proposition \ref{wei-hol-1}, and applying the strong convergence of $u^{\epsilon_n}$ \eqref{main-rd-sec-wei},we obtain that
\begin{align*}
\lim_{n\to\infty}\mathbb E\Big[\sup_{t\in[0,T]}\Big\|\int_0^tS(t-s)f_{\epsilon_n}(|u^{\epsilon_n}|^2)u^{\epsilon_n}-\log(|u|^2)uds\Big\|^2\Big]=0.
\end{align*}
The Minkowski inequality, \eqref{con-g} and \eqref{con-g1} yield that
\begin{align*}
&\quad\Big\|-\int_0^tS(t-s)\frac 12(g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}
\sum_{k}|Q^{\frac 12}e_k|^2ds-V_2\Big\|\\
&\le \sum_{k}\|Q^{\frac 12}e_k\|_{L^{\infty}}^2\int_0^T\|g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}-g(|u|^2))^2u\|ds\\
&\le C_gT\sum_{k}\|Q^{\frac 12}e_k\|_{L^{\infty}}^2\sup_{t\in[0,T]}\|u^{\epsilon_n}(t)-u(t)\|,
\end{align*}
and
\begin{align*}
&\quad\Big\|-\mathbf i \int_0^tS(t-s)(g'(|u^{\epsilon_n}|^2)g(|u^{\epsilon_n}|^2)|u^{\epsilon_n}|^2 u^{\epsilon_n}
\sum_{k}Im (Q^{\frac 12}e_k)Q^{\frac 12}e_kds-V_3\Big\|\\
&\le C_gT\sum_{k}\|Q^{\frac 12}e_k\|_{L^{\infty}}^2\sup_{t\in[0,T]}\|u^{\epsilon_n}(t)-u(t)\|.
\end{align*}
The Burkerholder inequality and the unitary property of $S(\cdot)$ yield that
\begin{align*}
&\quad\mathbb E\Big[\sup_{t\in[0,T]}\Big\|\mathbf i \int_0^tS(t-s)g(|u^{\epsilon_n}|^2)u^{\epsilon_n}dW(s)-V_4\Big\|^2\Big]\\
&\le C\mathbb E\Big[ \int_0^T\sum_{k}\|Q^{\frac 12}e_k\|_{L^{\infty}}^2\|g(|u^{\epsilon_n}|^2)u^{\epsilon_n}-g(|u|^2)u\|^2ds\Big]
\le C\mathbb E\Big[\sup_{t\in[0,T]} \|u^{\epsilon_n}(t)-u(t)\|^2\Big].
\end{align*}
Combining the above estimates and the strong convergence of $u^{\epsilon_n}$, we complete the proof of step 2.
Step 3: $u$ is independent of the choice of the sequence of $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}.$ Assume that $\widetilde u$ and $u$ are two different limit processes of two different sequences of $\{u^{\epsilon_n}\}_{n\in\mathbb N^+}$ and $\{u^{\epsilon_m}\}_{m\in\mathbb N^+}$, respectively.
Then by step 2, they both satisfies Eq. \eqref{SlogS}. By repeating the procedures in step 1, it is not hard to obtain that $\widetilde u=u.$
\iffalse
\|v\|_{L^{1+\epsilon}}\le C\|v\|^{1-\frac {(1-\epsilon)d}{2\alpha(1+\epsilon)}}\|v\|_{L^2_{\alpha}}^{\frac {(1-\epsilon)d}{2\alpha(1+\epsilon)}}.
\fi
\qed
The procedures in the above proof immediately yield the following convergence rate result for $u^{\epsilon}$ in the regularized problem \eqref{Reg-SlogS} and the H\"older regularity estimate of $u^{\epsilon}$ and $u^0.$
\begin{cor}\label{strong-con}
Let the condition of Theorem \ref{mild-general} hold. Assume that $u^{\epsilon}$ is the mild solution in Proposition \ref{prop-well-reg}, $\epsilon\in (0,1).$ For $p\ge 2$, there exists $C(Q,T,\lambda,p,u_0)>0$ such that for any $\eta'(d-2)\le 2,$
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u(t)-u^{\epsilon}(t)\|^p\Big] \le C(Q,T,\lambda, p,u_0)(\epsilon^{\frac p2}+\epsilon^{\frac {\eta'p} 2})
\end{align*}
when $\mathcal O$ is bounded domain, and
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u(t)-u^{\epsilon}(t)\|^{p}\Big]
&\le C(Q,T,\lambda, p,u_0,\alpha)(\epsilon^{\frac {\alpha p}{2\alpha+d}}+\epsilon^{\frac {\eta'p} 2})
\end{align*}
when $\mathcal O=\mathbb R^d$.
\end{cor}
\iffalse
We would like to remark that if one apply the regularization function $f_{\epsilon}(x)=\log(\epsilon+x),$ then the convergence rate in Corollary \ref{strong-con} is independent of the parameter $\eta'.$
\fi
\begin{cor}\label{stro-contin}
Let the condition of Theorem \ref{mild-general} hold. Assume that $u^{\epsilon}$ is the mild solution in Proposition \ref{prop-well-reg}, $\epsilon\in (0,1)$ and $u^{0}$ is the mild solution of Eq. \eqref{SlogS}. For $p\ge 2$, there exists $C(Q,T,\lambda,p,u_0)>0$ such that for $\epsilon\in [0,1],$
\begin{align*}
\mathbb E\Big[\|u^{\epsilon}(t)-u^{\epsilon}(s)\|^p\Big] \le C(Q,T,\lambda, p,u_0)|t-s|^{\frac p2}.
\end{align*}
\end{cor}
{\noindent\bf{Proof}}
By means of the mild form of $u^{\epsilon},$ $\epsilon\in [0,1)$, the priori estimates of $u^{\epsilon}$ in $\mathbb H^1\cap L_\alpha^2$ in Lemmas \ref{h1-pri} and \ref{wei-hol-1}, and in Step 2 of the proof of Theorem \ref{mild-general}, and the Burkholder inequalty, we obtain the desirable result.
\qed
\subsection{Well-posedness of SlogS equation with super-linearly
growing diffusion coefficients}
In this part, we extends the scope of $\widetilde g$, which allows the diffusion with super-linear growth, for the well-posedness of SlogS equation driven by conservative multiplicative noise. For instance, it includes the example $\widetilde g(x)=\mathbf i g(|x|^2)x=\mathbf i x\log(c+|x|^2),$ for $c>0$.
\begin{tm}\label{mild-d1}
Let $W(t)$ be $L^2(\mathcal O;\mathbb R)$-valued and $g \in \mathcal C^1_b(\mathbb R)\cap \mathcal C(\mathbb R)$ satisfy the growth condition and the embedding condition,
\begin{align*}
&\sup_{x\in [0,\infty)}|g'(x)x|\le C_g,\\
&\|vg(|v|^2)\|\le C_d (1+\|v\|_{\mathbb H^1}+\|v\|_{L^2_{\alpha}})
\end{align*}
for some $q\ge 2$, $ \alpha\in [0,1]$, where $C_g>0$ depends on $g$, \;$C_d>0$ depends on $\mathcal O$, $d$, $\|v\|$ and $v\in \mathbb H^1\cap L^2_{\alpha}$.
Assume that $d=1$, $u_0\in \mathbb H^1\cap L^2_{\alpha}, $ $\alpha\in(0,1],$ and $\sum_{i\in \mathbb N^+}\|Q^{\frac 12}e_i\|_{\mathbb H^1}^2+\|Q^{\frac 12}e_i\|_{W^{1,\infty}}^2<\infty$. Then
there exists a unique mild solution $u$ in $C([0,T];\mathbb H)$ for Eq. \eqref{SlogS} satisfying
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u(t)\|_{\mathbb H^1}^p\Big]+\mathbb E\Big[\sup_{t\in [0,T]}\|u(t)\|_{L^2_{\alpha}}^p\Big]\le C(Q,T,\lambda, p,u_0).
\end{align*}
\end{tm}
{\noindent\bf{Proof}}
By Proposition \ref{prop-well-reg} and Lemma \ref{h1-pri}, we can introduce the truncated sample space $$\Omega_R(t):=\left\{\omega: \sup_{s\in[0,t]}\|u^{\epsilon_m}(s)\|_{L^{\infty}}\le R, \sup_{s\in [0,t]}\|u^{\epsilon_n}(s)\|_{L^{\infty}}\le R\right\},$$
where $n\le m.$
The Gagliardo--Nirenberg interpolation inequality in $d=1$, the priori estimate in $\mathbb H^1$ and the continuity in $L^2$ of $u^{\epsilon_n}$
imply that $u^{\epsilon_n}$ are continuous in $L^{\infty}$ a.s.
Define a stopping time $$\tau_R:=\inf\{t\ge 0: \min(\sup_{s\in[0,t]}\|u^{\epsilon_m}(s)\|_{L^{\infty}},\sup_{s\in[0,t]}\|u^{\epsilon_n}(s)\|_{L^{\infty}})\ge R\}\wedge T.$$ Then on $\Omega_R(T),$ we have $\tau_R=T.$
Let us take $f_{\epsilon}(x)=\log(x+\epsilon), x>0$ for convenience.
It is obvious that $\Omega_R(t)\to \Omega$ as $R\to \infty$ and that for any $p\ge 1$,
\begin{align*}
&\quad\mathbb P\Big(\sup_{t\in[0,T]}\min(\|u^{\epsilon_m}(t)\|_{L^{\infty}},\|u^{\epsilon_n}(t)\|_{L^{\infty}})\ge R\Big)\\
\le& C\frac 1{R^p}\Big(\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon_m}(t)\|_{L^{\infty}}^p\Big]+\sup_{t\in[0,T]} \mathbb E\Big[\|u^{\epsilon_m}(t)\|_{L^{\infty}}^p\Big]\Big).
\end{align*}
Step 1: $\{u^{\epsilon_n}\}_{n\in \mathbb N^+}$ forms a Cauchy sequence in $ \mathbb M_{\mathcal F}^2(\Omega;C([0,T];\mathbb H^2))$.
Following the same steps like the proof of Theorem \ref{mild-general}, applying the It\^o formula on $\Omega_R(t)$ for $t\in (0,\tau_R)$ yields that
\begin{align*}
&\quad\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\\
\le& \int_0^t 4|\lambda| \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}ds\\
&+\int_0^t \sum_{k\in\mathbb N^+}\<|Q^{\frac 12}e_k|^2
(g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))u^{\epsilon_m},(g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))u^{\epsilon_n}\>ds\\
&+\int_0^t\<u^{\epsilon_m}-u^{\epsilon_n}, \mathbf i \Big(g(|u^{\epsilon_m}|^2)u^{\epsilon_m}-g(|u^{\epsilon_n}|^2)u^{\epsilon_n}\Big)dW(s)\>.
\end{align*}
Taking expectation on $\Omega_R(t)$ yields that
\begin{align*}
&\quad\mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
\le& \int_0^t 4|\lambda| \mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 \Big]ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \mathbb E\Big[\mathbb I_{\Omega_R(t)} \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\Big]ds\\
&\quad+\int_0^t \sum_{k\in\mathbb N^+}|Q^{\frac 12}e_k|_{L^{\infty}}^2 \mathbb E\Big[\mathbb I_{\Omega_R(t)}\<
(g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))u^{\epsilon_m},(g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))u^{\epsilon_n}\>\Big]ds.
\end{align*}
\iffalse
&\quad+C(Q)\int_0^t \mathbb E\Big[\mathbb I_{\Omega_R}\int_{\mathcal O}\frac {(|u^{\epsilon_m}|^2+|u^{\epsilon_n}|^2)|u^{\epsilon_n}||u^{\epsilon_m}|}{(1+\min(|u^{\epsilon_m}|^2,|u^{\epsilon_m}|^2))^2} |u^{\epsilon_m}(t)-u^{\epsilon_n}(t)|^2dx\Big]ds\\
&\quad+C(Q)\int_0^t (\mathbb P({\Omega_R}))^{\frac 1p} \Big(\mathbb E\Big[\|g(|u^{\epsilon_n}|^2)-g(|u^{\epsilon_m}|^2))^2u^{\epsilon_n}u^{\epsilon_m}\|_{L^1}^q\Big]\Big)^{\frac 1q}ds\\
&\le \int_0^t 4|\lambda| \mathbb E\Big[\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 \Big]ds
+4|\lambda|\epsilon_m^{\frac 12} \int_0^t \mathbb E\Big[\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\Big]ds\\
&\quad+C(Q)\int_0^t \mathbb E\Big[\mathbb I_{\Omega_R^c}\int_{\mathcal O}\Big(\max(|u^{\epsilon_m}|,|u^{\epsilon_m}|)+\max(|u^{\epsilon_m}|,|u^{\epsilon_m}|)^3\Big) |u^{\epsilon_m}(t)-u^{\epsilon_n}(t)|^2dx\Big]ds\\
&\quad+C(Q)\int_0^t (\mathbb P({\Omega_R}))^{\frac 1{p_1}} \Big(\mathbb E\Big[\Big(\|u^{\epsilon_n}\|\|u^{\epsilon_m}\|_{L^{2+2\eta}}^{(1+\eta)}+\|u^{\epsilon_m}\|\|u^{\epsilon_n}\|_{L^{2+2\eta}}^{(1+\eta)}\Big)^q\Big]\Big)^{\frac 1q}ds\\
&\le \int_0^t 4|\lambda| \mathbb E\Big[\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 \Big]ds
+4|\lambda|\epsilon_m^{\frac 12} \int_0^t \mathbb E\Big[\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\Big]ds\\
&\quad+C(Q)\int_0^t \mathbb E\Big[(1+R^3)\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]ds\\
&\quad+C(u_0,Q,T,p_1){R^{-\frac p{p_1}}}.
\fi
Making use of the assumptions on $g$, we get
\begin{align*}
&\quad\mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
&\le \int_0^t 4|\lambda| \mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 \Big]ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\Big]ds\\
&\quad+C(Q)\int_0^t \mathbb E\Big[\mathbb I_{\Omega_R(t)}\int_{\mathcal O} {(|u^{\epsilon_m}|^2+|u^{\epsilon_n}|^2)|u^{\epsilon_n}||u^{\epsilon_m}|} |u^{\epsilon_m}(t)-u^{\epsilon_n}(t)|^2dx\Big]ds\\
&\le\int_0^t 4|\lambda| \mathbb E\Big[\mathbb I_{\Omega_R(t)} \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 \Big]ds
+4|\lambda|\epsilon_n^{\frac 12} \int_0^t \mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|_{L^1}\Big]ds\\
&\quad+C(Q)\int_0^t \mathbb E\Big[(1+R^4)\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]ds.
\end{align*}
If $\mathcal O$ is bounded, then H\"older inequality and Gronwall's inequality yield that
\begin{align*}
\mathbb E\Big[\mathbb I_{\Omega_R(t)} \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]
&\le C(u_0,Q,T,\lambda)e^{(1+R^4 )T}\epsilon_n.
\end{align*}
On the other hand, the Chebyshev inequality and the a priori estimate lead to
\begin{align*}
&\mathbb E\Big[\mathbb I_{\Omega^c_R(t)} \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]
\le (\mathbb P(\Omega^c_R(t)))^{\frac 1{p_1}}\Big(\mathbb E\Big[\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^{2q_1}\Big]\Big)^{\frac 1q_1},
\end{align*}
where $\frac 1{p_1}+\frac 1{q_1}=1.$
From the above estimate, choosing $p\gg p_1$ and denote by $\kappa=\frac {p}{p_1}$, we conclude that
\begin{align*}
&\mathbb E\Big[ \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le C(u_0,Q,T,\lambda,p,p_1)\Big(e^{(1+R^4)T}\epsilon_n+R^{-\kappa}\Big)
\end{align*}
Then one may take $R=(\frac {c}{T} |\log(\epsilon)|)^{\frac 14}$ for $c\in (0,1)$ and get
\begin{align*}
&\mathbb E\Big[ \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le C(u_0,Q,T,\lambda,p,p_1)(\epsilon_n^{1-c}+(\frac {c}{T} |\log(\epsilon_n)|)^{-\frac \kappa 4}).
\end{align*}
By further applying the Burkerholder inequality to the stochastic integral, we achieve that for any $\kappa>0,$
\begin{align*}
\mathbb E\Big[ \sup_{t\in [0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]
&\le C(u_0,Q,T,\lambda,p,p_1)|\log(\epsilon_n)|^{-\frac \kappa 4}.
\end{align*}
When $\mathcal O=\mathbb R^d$, we just repeat the procedures in the proof of the case that $g$ is bounded and obtain that for $\eta\in (0,\frac{2\alpha}{2\alpha+d}]$ and $\alpha\in (0,1],$
\begin{align*}
&\quad\mathbb E\Big[\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\\
&\le\int_0^t 4|\lambda| \mathbb E\Big[\mathbb I_{\Omega_R(t)} \|u^{\epsilon_m}(s)-u^{\epsilon_n}(s)\|^2 \Big]ds
+C \epsilon_{m}^{\eta} \int_0^t\mathbb E\Big[\mathbb I_{\Omega_R(t)} \|u^{\epsilon_n}\|_{L^{2}_\alpha}^{\frac {d\eta}{\alpha}}\|u^{\epsilon_n}\|^{2-2\eta-\frac{d\eta}{\alpha} }\Big]ds\\
&\quad+C(Q)\int_0^t \mathbb E\Big[(1+R^4)\mathbb I_{\Omega_R(t)}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]ds.
\end{align*}
By using Gronwall's inequality and the estimate of $\mathbb P(\mathbb I_{\Omega_R(t)}^c)$, we immediately have that for $\eta\in (0,\frac {2\alpha}{2\alpha+d}), \alpha\in (0,1]$ and any $\kappa>0,$
\begin{align*}
&\mathbb E\Big[ \|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le C(u_0,Q,T,\lambda,p,p_1)\Big(e^{(1+R^4)T}\epsilon_n^{\eta}+R^{-\kappa}\Big)
\end{align*}
Taking $R=(\frac {\eta c}{T} |\log(\epsilon_n)|)^{\frac 14}$ for $c\eta \in (0,1)$ and using the Burkerholder inequality, we have for $\eta\in (0,\frac {2\alpha}{2\alpha+d}), \alpha\in (0,1]$ and any $\kappa>0,$
\begin{align*}
&\mathbb E\Big[ \sup_{t\in[0,T]}\|u^{\epsilon_m}(t)-u^{\epsilon_n}(t)\|^2\Big]\le C(u_0,Q,T,\lambda,p,p_1,\eta)|\log(\epsilon_n)|^{-\frac {\kappa}4}.
\end{align*}
Step 2. $u$ is the mild solution.
Let us use the same notations and procedures as in step 2 of the proof in Theorem \ref{mild-general}.
To show that the mild form $u^{\epsilon_n}$ converges to
$S(t)u_0+V_1+V_2+V_3+V_4.$ We only need to estimate $V_2$ and $V_4$ since $V_3=0$.
Define
$$\Omega_{R_1}(t):=\{\omega: \sup_{s\in[0,t]}\|u^{\epsilon_n}(s)\|_{L^{\infty}}\le R_1, \sup_{s\in [0,t]}\|u(s)\|_{L^{\infty}}\le R_1\},$$
and a stopping time $$\tau_{R_1}:=\inf\{t\ge 0: \min(\sup_{s\in[0,t]}\|u(s)\|_{L^{\infty}},\sup_{s\in[0,t]}\|u^{\epsilon_n}(s)\|_{L^{\infty}})\ge R_1\}\wedge T.$$ Then on $\Omega_{R_1}(T),$ we have $\tau_{R_1}=T$. The Minkowski inequality and the properties of $g$ yield that on $\Omega_{R_1}(t),$ for a small enough $\eta_1>0$,
\begin{align*}
&\Big\|-\frac 12\int_0^tS(t-s)(g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}
\sum_{k}|Q^{\frac 12}e_k|^2ds-V_2\Big\|\\
\le& \sum_{k}|Q^{\frac 12}e_k|_{L^{\infty}}^2\int_0^T\|g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}-g(|u|^2))^2u\|ds\\
\le& C_gT\sum_{k}|Q^{\frac 12}e_k|_{L^{\infty}}^2\Big(1+R_1^2\Big)\sup_{t\in[0,T]}\|u^{\epsilon_n}-u\|.
\end{align*}
On the other hand, for any $p_2>0$,
\begin{align*}
\mathbb E\Big[\mathbb I_{\Omega_{R_1}^c(t)}\Big\|-\frac 12\int_0^tS(t-s)(g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}
\sum_{k}|Q^{\frac 12}e_k|^2ds-V_2\Big\|^2\Big]\le C(u_0,Q,T,p_2)R_1^{-p_2}.
\end{align*}
Taking $R_1=\mathscr O(|\log(\epsilon_n)|^{\frac{ \kappa_1} {4(4+p_2)}}),\kappa_1<\kappa,$ we have that
\begin{align*}
\lim_{n\to \infty}\sup_{t\in[0,T]}\mathbb E\Big[\Big\|\int_0^t-\frac 12S(t-s)(g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}
\sum_{k}|Q^{\frac 12}e_k|^2ds-V_2\Big\|^2\Big]=0,
\end{align*}
which immediately implies that
\begin{align*}
\lim_{n\to \infty}\mathbb E\Big[\sup_{t\in[0,T]}\Big\|-\frac 12\int_0^tS(t-s)(g(|u^{\epsilon_n}|^2))^2u^{\epsilon_n}
\sum_{k}|Q^{\frac 12}e_k|^2ds-V_2\Big\|^2\Big]=0.
\end{align*}
The Burkerholder inequality and the unitary property of $S(\cdot)$ yield that
\begin{align*}
&\mathbb E\Big[\sup_{t\in[0,\tau_R]}\Big\|\int_0^t\mathbf i S(t-s)g(|u^{\epsilon_n}|^2)u^{\epsilon_n}dW(s)-V_4\Big\|^2\Big]\\
\le& C\mathbb E\Big[ \int_0^T\sum_{k}\|Q^{\frac 12}e_k\|_{L^{\infty}}^2\|g(|u^{\epsilon_n}|^2)u^{\epsilon_n}-g(|u|^2)u\|^2ds\Big]\\
\le& C(1+R_1^{2})\mathbb E\Big[\sup_{t\in[0,T]} \|u^{\epsilon_n}(t)-u(t)\|^2\Big].
\end{align*}
On the other hand, the Chebyshev inequality, together with the a priori estimate of $u^{\epsilon_n}$, implies that
\begin{align*}
&\mathbb E\Big[\sup_{t\ge \tau_{R_1}}\Big\|\int_0^t\mathbf i S(t-s)g(|u^{\epsilon_n}|^2)u^{\epsilon_n}dW(s)-V_4\Big\|^2\Big]
\\
=&\mathbb E\Big[ \sup_{t\in [0,T]}\mathbb I_{\Omega_{R_1}^c} \Big\|\int_0^tS(t-s)g(|u^{\epsilon_n}|^2)u^{\epsilon_n}dW(s)-V_4\Big\|^2\Big]\le C(u_0,Q,T,p)R_1^{-p_2}.
\end{align*}
Taking $R_1=\mathscr O(|\log(\epsilon_n)|^{\frac{ \kappa_1} {4(2+p_2)}}),\kappa_1<\kappa,$ we have that
\begin{align*}
\lim_{n\to \infty}\mathbb E\Big[\sup_{t\in[0,T]}\Big\|\int_0^t\mathbf i S(t-s)g(|u^{\epsilon_n}|^2)u^{\epsilon_n}dW(s)-V_4\Big\|^2\Big]
=0.
\end{align*}
Combining the above estimates and the strong convergence of $u^{\epsilon_n}$, we complete the proof.
\qed
\begin{rk}
One may extend the scope of $\widetilde g$ to an abstract framework by similar arguments. Here the assumption $d=1$ lies on the fact that in $\mathbb H^1$ is an algebra by Sololev embedding theorem. When considering the case $d\ge2$, one may use $\mathbb H^{\mathbf s},{\mathbf s>\frac d2}$ as the underlying space for the local well-posedness. However, as stated in Lemma \ref{h2-pri}, it seems impossible to get the uniform bound of $u^{\epsilon}$ in $\mathbb H^{\mathbf s}$ for $\mathbf s\ge 2.$
\end{rk}
\iffalse
\|v\|_{L^{1+\epsilon}}\le C\|v\|^{1-\frac {(1-\epsilon)d}{2\alpha(1+\epsilon)}}\|v\|_{L^2_{\alpha}}^{\frac {(1-\epsilon)d}{2\alpha(1+\epsilon)}}.
\fi
\section{Appendix}
The original problem and regularized problem can be rewritten into the equivalent evolution forms
\begin{align}\label{evo}
du&=A udt+F(u)dt+G(u)dW(t),\\\nonumber
u(0)&=u_0,
\end{align}
where $A=\mathbf i \Delta,$ $F$ is the Nemystkii operator of drift coefficient function, and $G$ are the Nemystkii operator of diffusion coefficient function. Then the mild solution of the above evolution is defined as follows.
\begin{df}\label{global}
A continuous $\mathbb H$-valued $\mathcal F_t$ adapted process $u$ is a solution to \eqref{evo} if it satisfies $\mathbb P$-a.s for all $t\in [0,T],$
\begin{align*}
u(t)=S(t)u_0+\int_0^t S(t-s) F(u(s))ds+\int_0^tG(u(s))dW(s),
\end{align*}
where $S(t)$ is the $C_0$-group generated by $A.$
\end{df}
\begin{df}\label{local}
A local mild solution of \eqref{evo} is $(u,\tau):=(u,\tau_n,\tau)$ satisfying $\tau_n\nearrow \tau, a.s.,$ as $n\to \infty$, $u \in \mathbb M_{\mathcal F}^p(\Omega;C([0,\tau);\mathbb H^{\mathbf s}), \mathbf s>0, p\ge 1$ and that
\begin{align*}
u(t)&=S(t)u_0+\int_0^t S(t-s)F(u(s))ds\\
&+\int_0^t S(t-s) G(u(s))dW(s), a.s.,
\end{align*}
for $t\le \tau_n$ in $\mathbb H^2$ for $n\in \mathbb N^+$.
Solutions of \eqref{evo} are called unique, if $$\mathbb P\Big(u_1(t)=u_2(t), \forall t\in [0,\sigma_1\wedge \sigma_2)\Big)=0.$$
for all local mild solution $(u_1,\sigma_1)$ and $(u_2,\sigma_2).$ The local solution $(u,\tau)$ is called a global mild solution if $\tau=T, a.s.$ and $u\in \mathbb M_{\mathcal F}^p(\Omega;C([0,T];\mathbb H^{\mathbf s}).$
\end{df}
\begin{lm}\label{frist-loc}
Let $\epsilon\in (0,1)$. Then $f_{\epsilon}(x)=\log({|x|^2+\epsilon}), x\in\mathbb C,$ satisfies
\begin{align*}
|Im[(f_{\epsilon}(x_1)-f_{\epsilon}(x_2))(\bar x_1-\bar x_2)]|\le 4|x_1-x_2|^2,
\end{align*}
\end{lm}
\textbf{Proof}
Without loss of generality, we assume that $0<|x_2|\le |x_1|.$
Notice that
\begin{align*}
Im[(f_{\epsilon}(x_1)-f_{\epsilon}(x_2))(\bar x_1-\bar x_2)]
=\frac 12(\log(\epsilon+|x_1|^2)-\log(\epsilon+|x_2|^2))Im(\bar x_1x_2-\bar x_2x_1).
\end{align*}
Direct calculation yields that
\begin{align*}
|Im(\bar x_1x_2-\bar x_2x_1)|\le 2|x_2||x_1-x_2|.
\end{align*}
Using the fact that
\begin{align*}
|\log(\epsilon+|x_1|^2)-\log(\epsilon+|x_2|^2)|&=2|\log((\epsilon+|x_1|^2)^{\frac 12})-\log((\epsilon+|x_2|^2)^{\frac 12})|,
\end{align*}
we obtain
\begin{align*}
&|Im[(f_{\epsilon}(x_1)-f_{\epsilon}(x_2))(\bar x_1-\bar x_2)]|\\
&\le 2|\log((\epsilon+|x_1|^2)^{\frac 12})-\log((\epsilon+|x_2|^2)^{\frac 12})||x_2||x_1-x_2|.
\end{align*}
The mean value theorem leads to the desired result.
\qed
\begin{lm}\label{sec-reg}
Let $\epsilon\in (0,1)$. Then $f_{\epsilon}(|x|^2)=\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$ satisfies the following properties,
\begin{align*}
|f_{\epsilon}(|x|^2)|&\le |\log(\epsilon)|,\\
|d_{|x|}f_{\epsilon}(|x|^2)|&\le \frac {2(1-\epsilon^2)|x|}{(\epsilon+|x|^2)(1+\epsilon|x|^2)},\\
|Im[(f_{\epsilon}(|x_1|^2)x_1-f_{\epsilon}(|x_2|^2)x_2)(\bar x_1-\bar x_2)]|&\le 4(1-\epsilon^2)|x_1-x_2|^2.
\end{align*}
\end{lm}
\textbf{Proof}
The proof of first and second estimates are derived by the property of $\log(\cdot).$ The last estimate is proven by similar arguments in the proof of Lemma \ref{frist-loc}.
\qed
\textbf{[Proof of Proposition \ref{eng-pri}]} Due to Lemma \ref{h1-pri}, it suffices to prove
\begin{align*}
\mathbb E\Big[\sup\limits_{t\in[0,T]} (F_{\epsilon}(|u^{\epsilon}(t)|^2))^{p}\Big]\le C(u_0,T,Q,p).
\end{align*}
Let us take $f_{\epsilon}(|x|^2)=\log(|x|^2+\epsilon)$ as an example to illustrate the procedures. The desirable estimate in case that
$f_{\epsilon}(|x|^2)=\log(\frac{|x|^2+\epsilon}{1+\epsilon |x|^2})$ can be obtained similarly.
Using the property of logarithmic function, we have that for small $\eta>0,$
\begin{align*}
|F_{\epsilon}(|u^{\epsilon}(t)|^2)|
&=\Big|\int_{\mathcal O}\Big((\epsilon+|u^{\epsilon}(t)|^2)\log(\epsilon+|u^{\epsilon}(t)|^2)-|u^{\epsilon}(t)|^2-\epsilon\log(\epsilon)\Big)dx\Big|\\
&\le \|u^{\epsilon}(t)\|^2+\int_{\mathcal O}|u^{\epsilon}(t)|^2\log(\epsilon+|u^{\epsilon}(t)|^2)dx+\Big|\int_{\mathcal O}\epsilon(\log(\epsilon+|u^{\epsilon}(t)|^2)-\log(\epsilon))dx\Big|\\
&\le 2\|u^{\epsilon}(t)\|^2+\| u^{\epsilon}(t)\|_{L^{2-2\eta}}^{2-2\eta}+\| u^{\epsilon}(t)\|_{L^{2+2\eta}}^{2+2\eta},
\end{align*}
where we use the following estimation,
\iffalse
Take any stopping time $\tau<\tau^*, a.s.$
Using the It\^{o} formula to
$H^k(u^{\epsilon}(t)),$ for $k\in \mathbb N^+$ or $k\ge 2$,
we obtain that for $t\in [0,\tau],$
\begin{align*}
&H^k(u^{\epsilon}(t))\\
&=H^k(u^{\epsilon}_0)+
k\int_0^{t}H^{k-1}(u^{\epsilon}(s))\<-\Delta u^{\epsilon}(s),\mathbf i \Delta u^{\epsilon}(s)+\mathbf i \lambda f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)\>ds\\
&\quad+
k\int_0^{t}H^{k-1}(u^{\epsilon}(s))\<-\lambda (f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}),\mathbf i \Delta u^{\epsilon}(s)+\mathbf i \lambda f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)\>ds\\
&\quad+
k(k-1)\int_0^{t}H^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\<\nabla u^{\epsilon}, \nabla Q^{\frac 12}e_i\>^2+\lambda^2\<f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon},Q^{\frac 12}e_i\>^2\Big)ds\\
&\quad
+k\int_0^{t}H^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\<\nabla Q^{\frac 12}e_i,\nabla Q^{\frac 12}e_i\>-\lambda\<f_{\epsilon}(|u^{\epsilon}(s)|^2)Q^{\frac 12}e_i,Q^{\frac 12}e_i\>\\
&\quad -2\lambda\<f'_{\epsilon}(|u^{\epsilon}(s)|^2) Re(u^{\epsilon}Q^{\frac 12}e_i)u^{\epsilon},Q^{\frac 12}e_i\> \Big)ds\\
&\quad
+k\int_0^{t}H^{k-1}(u^{\epsilon}(s))\Big(\<\nabla u,\nabla dW(s)\>-\lambda\<f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s),dW(s)\>\Big)\\
&=H^k(u^{\epsilon}_0)
+
k(k-1)\int_0^{t}H^{k-2}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\<\nabla u^{\epsilon}, \nabla Q^{\frac 12}e_i\>^2+\lambda^2\<f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon},Q^{\frac 12}e_i\>^2\Big)ds\\
&\quad
+k\int_0^{t}H^{k-1}(u^{\epsilon}(s))\sum_{i\in \mathbb N^+}\Big(\<\nabla Q^{\frac 12}e_i,\nabla Q^{\frac 12}e_i\>-\lambda\<f_{\epsilon}(|u^{\epsilon}(s)|^2)Q^{\frac 12}e_i,Q^{\frac 12}e_i\>\\
&\quad -2\lambda\<f'_{\epsilon}(|u^{\epsilon}(s)|^2) Re(u^{\epsilon}Q^{\frac 12}e_i)u^{\epsilon},Q^{\frac 12}e_i\> \Big)ds\\
&\quad
+k\int_0^{t}H^{k-1}(u^{\epsilon}(s))\Big(\<\nabla u,\nabla dW(s)\>-\lambda\<f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s),dW(s)\>\Big).
\end{align*}
Then taking the supreme over $t$, and using the martingale inequality, the H\"older inequality and the Young inequality, yield that for small $\epsilon>0,$
\begin{align*}
&\mathbb E[\sup_{t\in [0,\tau]}H^k(u^{\epsilon}(t))]\\
&=\mathbb E[H^k(u^{\epsilon}_0)]+
k(k-1)T\mathbb E\Big[\sup_{s\in [0,\tau]}H^{k-2}(u^{\epsilon}(s))\sup_{s\in[0,\tau]}\sum_{i\in \mathbb N^+}\Big(\|\nabla u^{\epsilon}\|^2 \|\nabla Q^{\frac 12}e_i\|^2\\
&\quad +\lambda^2\|f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon}\|^2\|Q^{\frac 12}e_i\|^2\Big)\Big]\\
&\quad
+kT\mathbb E\Big[\sup_{s\in [0,\tau]}H^{k-1}(u^{\epsilon}(s))\sup_{s\in[0,\tau]}\sum_{i\in \mathbb N^+}\Big(\|\nabla Q^{\frac 12}e_i\|^2+|\lambda| \|f_{\epsilon}(|u^{\epsilon}|^2)\|^2\|Q^{\frac 12}e_i\|_{L^4}^2\\
&\quad +2|\lambda| \|f'_{\epsilon}(|u^{\epsilon}(s)|^2)|u^{\epsilon}|^2\|^2 \|Q^{\frac 12}e_i\|_{L^4}^2 \Big)\Big]\\
&\quad
+ k\mathbb E\Big[\Big(\int_0^{\tau}H^{2k-2}(u^{\epsilon}(s))\sum_{i\in\mathbb N^+}\|\nabla u\|^2\|\nabla Q^{\frac 12}e_i\|^2ds\Big)^{\frac 12}\Big]\\
&\quad+ k|\lambda| \mathbb E\Big[\Big(\int_0^{\tau}H^{2k-2}(u^{\epsilon}(s)) \|f_{\epsilon}(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)\|^2\|Q^{\frac 12}e_i\|^2ds\>\Big)^{\frac 12}\Big].
\end{align*}
\fi
for any small enough $\eta>0$,
\begin{align*}
&\int_{\mathcal O}|u^{\epsilon}(t)|^2\log(\epsilon+|u^{\epsilon}(t)|^2)dx\\
&=
\int_{\epsilon +|u^{\epsilon}|^2\ge 1} f_{\epsilon}(|u^{\epsilon}|^2)|u^{\epsilon}|^2dx+\int_{\epsilon +|u^{\epsilon}|^2\le 1} f_{\epsilon}(|u^{\epsilon}|^2)|u^{\epsilon}|^2dx\\
&\le \int_{\epsilon +|u^{\epsilon}|^2\ge 1} (\epsilon +|u^{\epsilon}|^2)^{2\eta}|u^{\epsilon}|^2dx+\int_{\epsilon +|u^{\epsilon}|^2\le 1}(\epsilon+|u^{\epsilon}|^2)^{-2\eta}|u^{\epsilon}|^2dx\\
&\le \|u\|_{L^{2-2\eta}}^{2-2\eta}+\|u\|_{L^{2+2\eta}}^{2+2\eta}.
\end{align*}
Then by
the Gagliardo--Nirenberg interpolation inequality in a bounded domain $\mathcal O$, i.e.,
\begin{align*}
\|u\|_{L^q(\mathcal O)}\le C_1\|u\|^{1-\alpha}\|\nabla u\|^{\alpha}+C_2\|u\|, \quad q=\frac {2d}{d-2\alpha}, \; \text{for} \; \alpha\in (0,1],
\end{align*}
we have that
\begin{align*}
\int_{\mathcal O}F_{\epsilon}(|u^{\epsilon}(s)|^2)dx&\le |\int_{\mathcal O}\Big((\epsilon+|u^{\epsilon}|^2)\log(\epsilon+|u^{\epsilon}|^2)-|u^{\epsilon}|^2-\epsilon\log\epsilon\Big) dx|\\
&\le C(1+\|u^{\epsilon}\|_{L^{2+2\eta}}^{2+2\eta})\le C(\|u^{\epsilon}\|+\|\nabla u^{\epsilon}\|^{\frac {d\eta}{2\eta+2}}\|u^{\epsilon}\|^{1-\frac {d\eta}{2\eta+2}}).
\end{align*}
Taking $p$th moment and applying Lemma \ref{h1-pri}, we complete the proof for the case that $\mathcal O$ is a bounded domain.
When $\mathcal O=\mathbb R^d,$ we need to control $\|u^{\epsilon}\|_{L^{2-2\eta}}.$ By using the weighted interpolation inequality in Lemma \ref{wei-hol}
with $\alpha>\frac {d\eta}{2-2\eta}, \alpha\in (0,1]$,
and applying the Gagliardo--Nirenberg interpolation inequality,
\begin{align*}
\|u\|_{L^q(\mathbb R^d)}\le C_1\|u\|^{1-\alpha}\|\nabla u\|^{\alpha}, \quad q=\frac {2d}{d-2\alpha},\; \text{for} \;\alpha\in (0,1],
\end{align*}
where $q=2\eta+2.$
Based on Lemma \ref{h1-pri} and Lemma \ref{wei-hol-1}, we complete the proof by using the Young inequality and taking $p$th moment.
\qed
\iffalse
&\le C(\|u^{\epsilon}\|+\|u^{\epsilon}\|_{L^{2+2\eta}}^{2+2\eta}+\|u^{\epsilon}\|_{L^{2-2\eta}}^{2-2\eta})\le C(\|u^{\epsilon}\|+\|\nabla u^{\epsilon}\|^{\frac {d\eta}{2\eta+2}}\|u^{\epsilon}\|^{1-\frac {d\eta}{2\eta+2}})\\
\|f_{\epsilon}(|u^{\epsilon}|^2)\|^2&\le \int_{\epsilon +|u^{\epsilon}|^2\ge 1} (\epsilon +|u^{\epsilon}|^2)^{\eta}dx+\int_{\epsilon +|u^{\epsilon}|^2\le 1}|\log(\epsilon+|u^{\epsilon}|^2)|^2dx\\
&\le C(\epsilon+\|u^{\epsilon}\|_{L^{2\eta}}^{2\eta}+|\log(\epsilon)|^2)\le C(1+\|u^{\epsilon}\|_{L^{2}}^{2}+|\log(\epsilon)|^2)\\
\|f_{\epsilon}'(|u^{\epsilon}|^2)|u^{\epsilon}|^2\|&= \int_{\mathcal O} (f'_{\epsilon}(|u^{\epsilon}|^2))^2 |u^{\epsilon}|^4dx\le C.
\fi
\textbf{[Sketch Proof of Lemma \ref{h2-pri}]}
Due to the loss of the regularity of the solution in time, we can not establish the bound in $\mathbb H^2$ through $\frac {\partial u^{\epsilon}} {\partial t}$ like in the deterministic case.
\iffalse
It is not difficult to obtain that
\begin{align*}
d \Delta u^{\epsilon}&=\mathbf i \Delta^2 u^{\epsilon}dt+\mathbf i\lambda f(|u^{\epsilon}|^2)\Delta u^{\epsilon}dt+\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}dt\\
&\quad+\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}dt\\
&\quad +\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}dt\\
&\quad+\mathbf i 2\lambda f'(|u^{\epsilon}|^2)|\nabla u^{\epsilon}|^2u^{\epsilon}dt+\Delta dW.
\end{align*}
applying the It\^o formula to $\|\Delta u^{\epsilon}\|^{2p}$ and using integration by parts, we have that
\begin{align*}
\|\Delta u^{\epsilon}(t)\|^{2p}
&= \|\Delta u^{\epsilon}(0)\|^{2p}
+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u,\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}\>ds\\
&\quad+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u,\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}\>ds\\
&\quad+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u,\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}\>ds\\
&\quad+\int_{0}^t2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u^{\epsilon}, \Delta dW(t)\>\\
&\quad+\int_{0}^t p(2p-2) \|\Delta u^{\epsilon}(s)\|^{2p-2}\sum_{i\in\mathbb N}\<\Delta u^{\epsilon}, \Delta Q^{\frac 12}e_k\>^2ds\\
&\quad+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\sum_{i\in\mathbb N}\< \Delta Q^{\frac 12}e_k, \Delta Q^{\frac 12}e_k\>ds.
\end{align*}
Taking supreme over $t$, taking expectation and using the Burkerholder inequality yield that for $t\le t_1\le \tau,$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2p}\Big]
+2p \mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\<\Delta u,\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}\>dr\Big]\\
&\quad+2p\mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\<\Delta u,\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}\>dr\Big]\\
&\quad+2p\mathbb E\Big[\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{2p-2}\<\Delta u,\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}\>dr\Big]\\
&\quad+2p\mathbb E\Big[\Big(\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{4p-4}
\sum_{i\in\mathbb N}\<\Delta u^{\epsilon}, \Delta Q^{\frac 12}e_i\>^2dr\Big)^{\frac 12}\Big]\\
&\quad+p(2p-2) \mathbb E\Big[\int_{0}^t \sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\sum_{i\in\mathbb N}\<\Delta u^{\epsilon}, \Delta Q^{\frac 12}e_k\>^2dr\Big]\\
&\quad+2p\int_{0}^{t_1}\sup_{r\in[0,t]} \mathbb E\Big[ \|\Delta u^{\epsilon}(r)\|^{2p-2}\sum_{i\in\mathbb N}\< \Delta Q^{\frac 12}e_k, \Delta Q^{\frac 12}e_k\>dr\Big].
\end{align*}
The properties of $f_{\epsilon}$,
$xf'_{\epsilon}(|x|^2)\le \frac 1{2\sqrt{\epsilon}}$, $x^3f''_{\epsilon}(|x|^2)\le \frac 1{2\sqrt{\epsilon}}$ and
$x^2f'_{\epsilon}(|x|^2)\le 1,$
for $x\ge 0,$
together with the Gagliardo--Nirenberg interpolation inequality, the H\"older and Young inequalities, yield that for some small $\epsilon_1>0,$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2p}\Big]
+8|\lambda|p \mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\|\Delta u\| \|f'(|u^{\epsilon}|^2)| u^{\epsilon}| |\nabla u^{\epsilon}|^2\|dr\Big]\\
&\quad+8|\lambda|p\mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\|\Delta u\| \|f''(|u^{\epsilon}|^2)|u^{\epsilon}|^3 |\nabla u|^2\|dr\Big]\\
&\quad+4|\lambda|p\mathbb E\Big[\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{2p-2}
\|\Delta u\| \|f'(|u^{\epsilon}|^2)|u^{\epsilon}|^2|\Delta u^{\epsilon}|\|dr\Big]\\
&\quad+2p\mathbb E\Big[\Big(\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{4p-2}
\sum_{i\in\mathbb N}\|\Delta Q^{\frac 12}e_i\|^2dr\Big)^{\frac 12}\Big]\\
&\quad+p(2p-2) \mathbb E\Big[\int_{0}^t \sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p}\sum_{i\in\mathbb N} \|\Delta Q^{\frac 12}e_k\|^2dr\Big]\\
&\quad+2p\mathbb E\Big[ \int_{0}^{t_1}\sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\sum_{i\in\mathbb N}\| \Delta Q^{\frac 12}e_k\|^2dr\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2p}\Big]
+C(\lambda,p) \epsilon^{-\frac 12}\mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-1}\|\Delta u\|^{\frac {d}2} \|\nabla u^{\epsilon}\|^{\frac {4-d}2}dr\Big]\\
&\quad+C(\lambda,p) \mathbb E\Big[\int_{0}^t \sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p}\sum_{i\in\mathbb N} (1+\|\Delta Q^{\frac 12}e_k\|^2)dr\Big]\\
&\quad+C(\lambda,p) \mathbb E\Big[ \int_{0}^{t_1}\sup_{r\in[0,t]} \Big(\epsilon_1 \|\Delta u^{\epsilon}(r)\|^{2p}+C(\epsilon_1)(\sum_{i\in\mathbb N}\| \Delta Q^{\frac 12}e_k\|^2)^{2p}\Big)dr\Big].
\end{align*}
Applying the H\"older and Young inequalities, and using the Gronwall's equality, we obtain that for $d=1,$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]
\le C(u_0,T,\lambda,p)(1+\epsilon^{- 2p}).
\end{align*}
\fi
\iffalse
{\color{red}
The case that $\mathcal O=\mathbb R^d.$
Denote $p'$ as the conjugate pair of $p\ge 1$ such that $\frac 1p +\frac 1{p'}=1.$
Deterministic Strichartz estimate:
Let $r_i,q_i\in [2,\infty],$ $\frac 2{q_i}+\frac d{r_i}=\frac d2, (q_i,r_i,d)\neq (2,\infty,2)$ with $i=1,2.$
Let $v\in L^2(\mathcal O),$ $J\subset \mathbb R$ an interval with $0\in J,$
and $w\in L^{q_2'}(J,L^{r_2'}(\mathbb R^d)).$ Then there is a constant independent of $J, w$ and $v$ such that
\begin{align}\label{stri-est}
\|S(\cdot)v\|_{L^{q_1}(J,L^{r_1})}&\le C\|v\|_{L^{2}},\\\nonumber
\|\int_0^{\cdot}S(\cdot-s)w(s)ds\|_{L^{q_1}(J,L^{r_1})}&\le C\|f\|_{L^{q_2'}(J,L^{r_2'})}.
\end{align}
In particular, if $S(\cdot)v$ and $\int_0^{\cdot} U(\cdot-s)w(s)ds$ belong to $\mathcal C_b(J,L^2),$ we have
\begin{align}\label{stri-est1}
\|S(\cdot)v\|_{\mathcal C_b(J,L^2)}&\le C\|v\|_{L^{2}},\\\nonumber
\|\int_0^{\cdot}S(\cdot-s)w(s)ds\|_{\mathcal C_b(J,L^2)}&\le C\|f\|_{L^{q_2'}(J,L^{r_2'})}.
\end{align}
Now, starting from the mild form of $\Delta u, $ we obtain that
\begin{align*}
\|\Delta u^{\epsilon}(t)\|&\le C\|\Delta u^{\epsilon}_0\|
+2|\lambda| \Big\|\int_0^t S(t-s) f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}ds\Big\|\\
&\quad+4|\lambda| \Big\|\int_0^tS(t-s) f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}ds\Big\|\\
&\quad +2|\lambda| \Big\|\int_0^t S(t-s)f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}ds\Big\|\\
&\quad +\Big\|\int_0^{t} S(t-s)\Delta dW(s)\Big\|.
\end{align*}
The properties of $f_{\epsilon}$,
$xf'_{\epsilon}(|x|^2)\le \frac 1{2\sqrt{\epsilon}}$, $x^3f''_{\epsilon}(|x|^2)\le \frac 1{2\sqrt{\epsilon}}$ and
$x^2f'_{\epsilon}(|x|^2)\le 1,$
for $x\ge 0,$
together with the Strichartz estimate, the H\"older and Young inequalities, yield that
\begin{align*}
\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]
&\le C_p\mathbb E\Big[\|\Delta u^{\epsilon}_0\|^{2p}+\sum_{i\in\mathbb N} \|\Delta Q^{\frac 12}e_i\|^2\Big]\\
&\quad+ C_p\int_{0}^{t_1}\mathbb E\Big[\|\nabla u^{\epsilon}\|^{2p}+\|\Delta u^{\epsilon}\|^{2p}\Big]dr\\
&\quad + \epsilon^{-p}\mathbb E\Big[\||\nabla u|^2\|_{L^{q'}([0,t_1];L^{r'})}^{2p}\Big].
\end{align*}
By choosing that large enough $r,$ $q=\frac 2{d(\frac 12-\frac 1r)},$ and using the Gagliardo--Nirenberg interpolation inequality, we have that for $\alpha=\frac d 4-\frac d {4\gamma'}$ close to $0$
\begin{align*}
\||\nabla u|^2\|_{L^{q'}([0,t_1];L^{r'})}^{2p}
&\le \Big(\int_{0}^{t_1}\|\nabla u\|_{L^{2r'}}^{2q'}dr\Big)^{\frac {2p}{q'}}\\
&\le \Big(\int_{0}^{t_1}\|\Delta u\|_{L^{2}}^{\alpha 2q'}\|\nabla u\|_{L^{2}}^{2q'(1-\alpha)}dr\Big)^{\frac {2p}{q'}}\\
&\le \sup_{t\in [0,t_1]}\|\Delta u\|_{L^{2}}^{ 4p\alpha} \sup_{t\in[0,t_1]}\|\nabla u\|_{L^{2}}^{4p(1-\alpha)}.
\end{align*}
The above estimates implies that for a small $\epsilon_1>0$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\\
&\le C_p\mathbb E\Big[\|\Delta u^{\epsilon}_0\|^{2p}+\sum_{i\in\mathbb N} \|\Delta Q^{\frac 12}e_i\|^{2p}\Big]+ C_p\int_{0}^{t_1}\mathbb E\Big[\|\nabla u^{\epsilon}\|^{2p}+\|\Delta u^{\epsilon}\|^{2p}\Big]dr\\
&\quad +\epsilon_1\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big] + C(\epsilon_1)\epsilon^{-p\frac 1{1-2p\alpha}}\mathbb E\Big[ \sup_{t\in[0,t_1]}\|\nabla u\|_{L^{2}}^{\frac {4p(1-\alpha)}{1-{2p\alpha}}}\Big].
\end{align*}
Applying Lemma \ref{h1-pri} and Gronwall's inequality, we obtain for a sufficient small $\epsilon>1$,
\begin{align*}
\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\le C(1+\epsilon^{-(\frac 12+\epsilon_1)2p})
\end{align*}
and complete the proof.
}
\fi
\iffalse
Let us first consider the behavior of $\|\nabla u\|_{L^4}^{4p}, p\ge 1.$
For convenience, here we only give the procedures for $\|\nabla u\|_{L^4}^4.$ . By the It\^o formula, we have
\begin{align*}
\|\nabla u^{\epsilon}(t)\|_{L^4}^4&=\|\nabla u^{\epsilon}_0\|_{L^4}^4
+4\int_{0}^t\<|\nabla u^{\epsilon}|^2\nabla u^{\epsilon}, \mathbf i \nabla \Delta u^{\epsilon} \>ds
\\
&\quad+4\int_{0}^t\<|\nabla u^{\epsilon}|^2\nabla u^{\epsilon},\mathbf i\lambda f_{\epsilon}(|u^{\epsilon}|^2) \nabla u^{\epsilon}\>ds\\
&\quad +8\int_{0}^t\<|\nabla u^{\epsilon}|^2\nabla u^{\epsilon}, \mathbf i \lambda f_{\epsilon}'(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon}\>ds\\
&\quad +4\int_{0}^t\<|\nabla u|^2\nabla u,\nabla dW(s)\>\\
&\quad +2\int_{0}^t\sum_{i\in\mathbb N^+}\<|\nabla u^{\epsilon}|^2\nabla Q^{\frac 12}e_i,\nabla Q^{\frac 12}e_i\>ds\\
&\quad +4\int_{0}^t\sum_{i\in\mathbb N^+}\<Re(\nabla \bar u^{\epsilon} \nabla Q^{\frac 12}e_i) u^{\epsilon},\nabla Q^{\frac 12}e_i\>ds.
\end{align*}
Taking expectation yields that
\begin{align*}
\mathbb E\Big[\|\nabla u^{\epsilon}(t)\|_{L^4}^4\Big]&\le \mathbb E\Big[\|\nabla u^{\epsilon}_0\|_{L^4}^4\Big]+8C\mathbb E\Big[\int_{0}^t\|\nabla u^{\epsilon}\|^4_{L^4}ds\Big]\\
&\quad+6C\mathbb E\Big[\int_{0}^t\sum_{i\in\mathbb N^+}\|\nabla u^{\epsilon }\|_{L^4}\|u^{\epsilon}\|_{L^4} \|\nabla Q^{\frac 12}e_i\|_{L^4}^2ds\Big].
\end{align*}
Then the H\"older and Young inequalities and Sobolev embedding yield that $$\mathbb E\Big[\|\nabla u^{\epsilon}(t)\|_{L^4}^4\Big]\le C(T,u_0).$$
Similarly, we have $$\mathbb E\Big[\|\nabla u^{\epsilon}(t)\|_{L^4}^{4p}\Big]\le C(T,u_0,p).$$
\fi
According to Lemma \ref{h1-pri}, it suffices to bound $\|\Delta u^{\epsilon}\|^2.$ We present the procedures of the estimation of $\mathbb E[\|\Delta u\|^2]$ for the conservative multiplicative noise case. One can easily follow the procedures to obtain the estimate of $\mathbb E[\sup_{t\in[0,\tau]}\|\Delta u(t)\|^2]$ for both additive and multiplicative noises.
By using the It\^o formula to $\|\Delta u^{\epsilon}\|^2$ we obtain that
\begin{align*}
\|\Delta u^{\epsilon}(t)\|^2&= \|\Delta u^{\epsilon}_0\|^2
+2\int_0^t \<\Delta u^{\epsilon}(s),II_{det} \>ds\\
&\quad+ 2\int_0^t \<\Delta u^{\epsilon}(s),II_{mod} \>ds+2II_{Sto},
\end{align*}
where
\begin{align*}
II_{Sto}&:=\int_0^t\<\Delta u^{\epsilon},\mathbf i g(|u^{\epsilon}|^2)\nabla u^{\epsilon}\nabla dW(s)\>+\int_0^t\<\Delta u^{\epsilon},\mathbf i g(|u^{\epsilon}|^2)u^{\epsilon}\Delta dW(s)\>\\
&+\int_0^t\<\Delta u^{\epsilon},\mathbf i2g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon} \nabla u^{\epsilon})\nabla u^{\epsilon} dW(s)+\int_0^t\<\Delta u^{\epsilon}, \mathbf i2g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon} \nabla u^{\epsilon})u^{\epsilon}\nabla dW(s)\>\\
&+\int_0^t\<\Delta u^{\epsilon},\mathbf i4g''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon} \nabla u^{\epsilon}))^2u^{\epsilon} dW(s)+\int_0^t\<\Delta u^{\epsilon},\mathbf i 2g'(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon} \Delta u^{\epsilon}))u^{\epsilon} dW(s)\>\\
&+\int_0^t\<\Delta u^{\epsilon},2g'(|u^{\epsilon}|^2) |\nabla u|^2 u^{\epsilon} dW(s)\>,\\
II_{det} &:= \mathbf i \Delta^2 u^{\epsilon}+\mathbf i\lambda f(|u^{\epsilon}|^2)\Delta u^{\epsilon}dt+\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}\\
&\quad+\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}\\
&\quad +\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon},
\end{align*}
and $II_{mod}$ is the summation of all terms involving the second derivative of the It\^o modified term produced by the Stratonovich integral. Here for simplicity, we omit the presentation of the explicit form for $II_{Stra}.$
\iffalse
Taking supreme over $t\in[0,\tau]$, taking expectation and using the Gagliardo--Nirenberg interpolation inequality $\|\nabla v\|_{L^4}\le C\|\Delta v\|^{\frac d4}\|\nabla v\|^{1-\frac d4}$, we get for $d=1,$
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,t_1]}\|\Delta u^{\epsilon}(t)\|^{2p}\Big]&\le \Big(\mathbb E\Big[\|\Delta u^{\epsilon}_0\|^{2p}\Big]+C(T,p)\Big(\sum_{i\in \mathbb N^+}\|Q^{\frac 12}e_i\|_{\mathbb H^2}^2\Big)^p\Big)\\
&\quad
+C(\lambda,p)\epsilon^{-\frac {2}{2-d}p}\int_0^{\tau} \mathbb E\Big[\|\nabla u\|^{\frac {4}{2-d}p}\Big]ds\\
&\quad +C(\lambda,p) \int_0^{\tau}\sup_{r\in[0,s]} \|u^{\epsilon}\|_{\mathbb H^2}^{2p}ds.
\end{align*}
Then the Gronwall inequality yields
\begin{align*}
\mathbb E\Big[ \sup_{r\in [0,\tau]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\le C(1+\epsilon^{-\frac {2}{2-d}p}).
\end{align*}
It is not difficult to obtain that
\begin{align*}
d \Delta u^{\epsilon}&=\mathbf i \Delta^2 u^{\epsilon}dt+\mathbf i\lambda f(|u^{\epsilon}|^2)\Delta u^{\epsilon}dt+\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}dt\\
&\quad+\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}dt\\
&\quad +\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}dt\\
&\quad+\mathbf i 2\lambda f'(|u^{\epsilon}|^2)|\nabla u^{\epsilon}|^2u^{\epsilon}dt+\Delta dW.
\end{align*}
Now, applying the It\^o formula to $\|\Delta u^{\epsilon}\|^{2p}$ and using integration by parts, we have that
\begin{align*}
\|\Delta u^{\epsilon}(t)\|^{2p}
&= \|\Delta u^{\epsilon}(0)\|^{2p}
+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u,\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}\>ds\\
&\quad+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u,\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}\>ds\\
&\quad+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u,\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}\>ds\\
&\quad+\int_{0}^t2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\<\Delta u^{\epsilon}, \Delta dW(t)\>\\
&\quad+\int_{0}^t p(2p-2) \|\Delta u^{\epsilon}(s)\|^{2p-2}\sum_{i\in\mathbb N}\<\Delta u^{\epsilon}, \Delta Q^{\frac 12}e_k\>^2ds\\
&\quad+\int_{0}^t 2p \|\Delta u^{\epsilon}(s)\|^{2p-2}\sum_{i\in\mathbb N}\< \Delta Q^{\frac 12}e_k, \Delta Q^{\frac 12}e_k\>ds.
\end{align*}
Taking supreme over $t$, taking expectation and using the Burkerholder inequality yield that for $t\le t_1\le \tau,$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2p}\Big]
+2p \mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\<\Delta u,\bi4\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})\nabla u^{\epsilon}\>dr\Big]\\
&\quad+2p\mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\<\Delta u,\mathbf i 4\lambda f''(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon}))^2u^{\epsilon}\>dr\Big]\\
&\quad+2p\mathbb E\Big[\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{2p-2}\<\Delta u,\mathbf i 2\lambda f'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\Delta u^{\epsilon} )u^{\epsilon}\>dr\Big]\\
&\quad+2p\mathbb E\Big[\Big(\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{4p-4}
\sum_{i\in\mathbb N}\<\Delta u^{\epsilon}, \Delta Q^{\frac 12}e_i\>^2dr\Big)^{\frac 12}\Big]\\
&\quad+p(2p-2) \mathbb E\Big[\int_{0}^t \sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\sum_{i\in\mathbb N}\<\Delta u^{\epsilon}, \Delta Q^{\frac 12}e_k\>^2dr\Big]\\
&\quad+2p\int_{0}^{t_1}\sup_{r\in[0,t]} \mathbb E\Big[ \|\Delta u^{\epsilon}(r)\|^{2p-2}\sum_{i\in\mathbb N}\< \Delta Q^{\frac 12}e_k, \Delta Q^{\frac 12}e_k\>dr\Big].
\end{align*}
The properties of $f_{\epsilon}$,
$xf'_{\epsilon}(|x|^2)\le \frac 1{2\sqrt{\epsilon}}$, $x^3f''_{\epsilon}(|x|^2)\le \frac 1{2\sqrt{\epsilon}}$ and
$x^2f'_{\epsilon}(|x|^2)\le 1,$
for $x\ge 0,$
together with the Gagliardo--Nirenberg interpolation inequality, the H\"older and Young inequalities, yield that for some small $\epsilon_1>0,$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2p}\Big]
+8|\lambda|p \mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\|\Delta u\| \|f'(|u^{\epsilon}|^2)| u^{\epsilon}| |\nabla u^{\epsilon}|^2\|dr\Big]\\
&\quad+8|\lambda|p\mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\|\Delta u\| \|f''(|u^{\epsilon}|^2)|u^{\epsilon}|^3 |\nabla u|^2\|dr\Big]\\
&\quad+4|\lambda|p\mathbb E\Big[\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{2p-2}
\|\Delta u\| \|f'(|u^{\epsilon}|^2)|u^{\epsilon}|^2|\Delta u^{\epsilon}|\|dr\Big]\\
&\quad+2p\mathbb E\Big[\Big(\int_{0}^{t_1} \|\Delta u^{\epsilon}(r)\|^{4p-2}
\sum_{i\in\mathbb N}\|\Delta Q^{\frac 12}e_i\|^2dr\Big)^{\frac 12}\Big]\\
&\quad+p(2p-2) \mathbb E\Big[\int_{0}^t \sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p}\sum_{i\in\mathbb N} \|\Delta Q^{\frac 12}e_k\|^2dr\Big]\\
&\quad+2p\mathbb E\Big[ \int_{0}^{t_1}\sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p-2}\sum_{i\in\mathbb N}\| \Delta Q^{\frac 12}e_k\|^2dr\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2p}\Big]
+C(\lambda,p) \epsilon^{-\frac 12}\mathbb E\Big[\int_{0}^{t_1} \sup_{r\in[0,s]} \|\Delta u^{\epsilon}(r)\|^{2p-1}\|\Delta u\|^{\frac {d}2} \|\nabla u^{\epsilon}\|^{\frac {4-d}2}dr\Big]\\
&\quad+C(\lambda,p) \mathbb E\Big[\int_{0}^t \sup_{r\in[0,t]} \|\Delta u^{\epsilon}(r)\|^{2p}\sum_{i\in\mathbb N} (1+\|\Delta Q^{\frac 12}e_k\|^2)dr\Big]\\
&\quad+C(\lambda,p) \mathbb E\Big[ \int_{0}^{t_1}\sup_{r\in[0,t]} \Big(\epsilon_1 \|\Delta u^{\epsilon}(r)\|^{2p}+C(\epsilon_1)(\sum_{i\in\mathbb N}\| \Delta Q^{\frac 12}e_k\|^2)^{2p}\Big)dr\Big].
\end{align*}
Applying the H\"older and Young inequalities, and using the Gronwall's equality, we obtain that for $d=1,$
\begin{align*}
&\mathbb E\Big[ \sup_{r\in [0,t_1]}\|\Delta u^{\epsilon}(r)\|^{2p}\Big]
\le C(u_0,T,\lambda,p)(1+\epsilon^{- 2p}).
\end{align*}
First consider the behavior of $\|\nabla u\|_{L^4}^{4p}, p\ge 1.$
For convenience, here we only give the procedures for $\|\nabla u\|_{L^4}^4.$ By the It\^o formula, we have
\begin{align*}
\|\nabla u^{\epsilon}(t)\|_{L^4}^4&=\|\nabla u^{\epsilon}_0\|_{L^4}^4
+4\int_{0}^t\<|\nabla u^{\epsilon}|^2\nabla u^{\epsilon}, \mathbf i \nabla \Delta u^{\epsilon} \>ds
\\
&\quad+4\int_{0}^t\<|\nabla u^{\epsilon}|^2\nabla u^{\epsilon},\mathbf i\lambda f(|u^{\epsilon}|^2) \nabla u^{\epsilon}\>ds\\
&\quad +8\int_{0}^t\<|\nabla u^{\epsilon}|^2\nabla u^{\epsilon}, \mathbf i \lambda f'(|u^{\epsilon}|^2)(Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon}\>ds\\
&\quad +I_{Stra}+I_{sto}+I_{sec},
\end{align*}
where
\begin{align*}
I_{Sto}&=4\int_0^t\<|\nabla u|^2\nabla u,\mathbf i g(|u^{\epsilon}|^2)u^{\epsilon}\nabla dW(s)\>+4\int_0^t\<|\nabla u|^2\nabla u, \mathbf i g(|u^{\epsilon}|^2)\nabla u^{\epsilon} dW(s)\>\\
&+8\int_0^t\<|\nabla u|^2\nabla u, \mathbf i g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon} \nabla u^{\epsilon})u^{\epsilon}\nabla dW(s)\>,
\end{align*}
\begin{align*}
I_{Stra}&=4\int_0^t\<|\nabla u|^2\nabla u,- \sum_{i\in\mathbb N^+}Re(\overline{Q^{\frac 12}e_k }\nabla Q^{\frac 12}e_k)(g(|u^{\epsilon})|^2)^2u^{\epsilon} \>ds\\
&+4\int_0^t\<|\nabla u|^2\nabla u,-2 \sum_{i\in\mathbb N^+}|Q^{\frac 12}e_k|^2g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2) Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon} \>ds\\
&+4\int_0^t\<|\nabla u|^2\nabla u,-\frac 12 \sum_{i\in\mathbb N^+}|Q^{\frac 12}e_k|^2(g(|u^{\epsilon})|^2)^2\nabla u^{\epsilon} \>ds\\
&+4\int_0^t\<|\nabla u|^2\nabla u,- \sum_{i\in\mathbb N^+}Im(Q^{\frac 12}e_k) Im(\nabla Q^{\frac 12}e_k) g'(|u^{\epsilon}|^2)|u^{\epsilon}|^2u^{\epsilon}\>ds\\
&+4\int_0^t\<|\nabla u|^2\nabla u,- \sum_{i\in\mathbb N^+}|Im(Q^{\frac 12}e_k)|^2g''(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})|u^{\epsilon}|^2u^{\epsilon}\>ds\\
&+4\int_0^t\<|\nabla u|^2\nabla u,- \sum_{i\in\mathbb N^+}|Im(Q^{\frac 12}e_k)|^2g'(|u^{\epsilon}|^2)Re(\bar u^{\epsilon}\nabla u^{\epsilon})u^{\epsilon}\>ds\\
&+4\int_0^t\<|\nabla u|^2\nabla u,-\frac 12 \sum_{i\in\mathbb N^+}|Im(Q^{\frac 12}e_k)|^2g'(|u^{\epsilon}|^2)|u^{\epsilon}|^2\nabla u^{\epsilon}\>ds.
\end{align*}
and
\begin{align*}
I_{sec}&=2\int_{0}^t\sum_{i\in\mathbb N^+}\<|\nabla u^{\epsilon}|^2\nabla (g(|u^{\epsilon}|^2)u^{\epsilon}Q^{\frac 12}e_i), \nabla(g(|u^{\epsilon}|^2)u^{\epsilon} Q^{\frac 12}e_i)\>ds\\
&\quad +4\int_{0}^t\sum_{i\in\mathbb N^+}\<Re(\nabla \bar u^{\epsilon}\mathbf i \nabla(g(|u^{\epsilon}|^2)u^{\epsilon} Q^{\frac 12}e_i) ) \nabla u^{\epsilon},\mathbf i \nabla(g(|u^{\epsilon}|^2)u^{\epsilon} Q^{\frac 12}e_i)\>ds.\\
\end{align*}
Taking expectation and applying H\"older and Young inequalities, together with the Sobolev embedding theorem, lead that
\begin{align*}
&\mathbb E\Big[\|\nabla u^{\epsilon}(t)\|_{L^4}^4\Big]\\
&\le \mathbb E\Big[\|\nabla u^{\epsilon}_0\|_{L^4}^4\Big]+C(1+\sum_{i\in \mathbb N^+}(\| Q^{\frac 12}e_i\|^2_{L^{\infty}}+\|\nabla Q^{\frac 12}e_i\|_{L^{\infty}}^2)\mathbb E\Big[\int_{0}^t\|\nabla u^{\epsilon}\|^4_{L^4}ds\Big]\\
&\quad+C(1+\sum_{i\in \mathbb N^+}(\| Q^{\frac 12}e_i\|^2_{L^{\infty}}+\|\nabla Q^{\frac 12}e_i\|_{L^{\infty}}^2)\mathbb E\Big[\int_{0}^t\|u^{\epsilon}\|_{\mathbb H^1}^4 ds\Big].
\end{align*}
Then Lemma \ref{mul-h1} and Gronwall's inequality yield that $$\mathbb E\Big[\|\nabla u^{\epsilon}(t)\|_{L^4}^4\Big]\le C(T,u_0).$$
Similarly, we have $\mathbb E\Big[\|\nabla u^{\epsilon}(t)\|_{L^4}^{4p}\Big]\le C(T,u_0,p).$
{\color{red}
By repeating the same procedures in the proof Lemma \ref{add-h2} and , we obtain that for any $p\ge 2$ or $p\in \mathbb N^+,$
\begin{align*}
&\mathbb E\Big[\|\Delta u^{\epsilon}(t)\|^{2}\Big]\\
&\le C\Big(\mathbb E\Big[\|\Delta u^{\epsilon}_0\|^{2}\Big]+\sum_{i\in \mathbb N^+}\|Q^{\frac 12}e_i\|_{\mathbb H^2}^2\Big)
+C(\lambda,p,T)\epsilon^{-\frac {2}{2-d}}\int_0^{\tau} \mathbb E\Big[\|\nabla u\|^{\frac {4}{2-d}}\Big]ds\\
&\quad +C(T,\lambda,p) \int_0^{t}\Big(1+\mathbb E\Big[\|u^{\epsilon}\|_{\mathbb H^2}^{2}\Big]+\mathbb E\Big[\|\nabla u\|^{\frac {4}{2-d}}\Big]\Big)ds,
\end{align*}
which implies $\mathbb E\Big[\|\Delta u^{\epsilon}(t)\|^{2}\Big]\le C(u_0,T,\lambda,p)(1+\epsilon^{-1})$ by using Gronwall's inequality.
By taking $L^2$ norm on the mild form of $\Delta u$, taking supreme over $t$, and applying the Burkerholder inequality to $I_{sto},$ it is not hard to get
$\mathbb E\Big[\sup_{t\in [0,\tau]}\|\Delta u^{\epsilon}(t)\|^{2}\Big]\le C(u_0,T,\lambda)(1+\epsilon^{-1}).$
}
\fi
Taking expectation and using the Gagliardo--Nirenberg interpolation inequality $\|\nabla v\|_{L^4}\le C\|\Delta v\|^{\frac 14}\|\nabla v\|^{\frac 34}$ in $d=1$, we obtain that
\begin{align*}
&\mathbb E\Big[ \|\Delta u^{\epsilon}(t)\|^{2}\Big]\\
&\le \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2}\Big]
+C(\lambda,p) \epsilon^{-\frac 12}\mathbb E\Big[\int_{0}^{t} \|\Delta u^{\epsilon}(r)\|(1+ \|\nabla u^{\epsilon}\|_{L^4}^{2})dr\Big]\\
&+C(\lambda,p) \mathbb E\Big[\int_{0}^{t} \|\Delta u^{\epsilon}(r)\|\Big(\| \sum_{i} \Delta Q^{\frac 12}e_i Q^{\frac 12}e_i (g(|u^{\epsilon}|^2))^2 u^{\epsilon}\|+\| \sum_{i} |\nabla Q^{\frac 12}e_i|^2 (g(|u^{\epsilon}|^2))^2 u^{\epsilon}\|
\\
&+ \|\sum_{i}\nabla Q^{\frac 12}e_iQ^{\frac 12}e_ig(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2)|u^{\epsilon}|^{2}\nabla u^{\epsilon} \| +\|\sum_{i}|Q^{\frac 12}e_i|^2g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2)|\nabla u^{\epsilon}|^2u^{\epsilon} \| \\
&+\|\sum_{i}|Q^{\frac 12}e_i|^2(g(|u^{\epsilon}|^2)g''(|u^{\epsilon}|^2)+(g'(|u^{\epsilon}|^2))^2)|\nabla u^{\epsilon}|^2|u^{\epsilon}|^3 \|
+\| \sum_{i} \nabla Q^{\frac 12}e_i Q^{\frac 12}e_i (g(|u^{\epsilon}|^2))^2 \nabla u^{\epsilon}\| \Big)dr\Big]\\
&+C(\lambda,p) \mathbb E\Big[\int_{0}^{t}\sum_{i\in\mathbb N}\Big(\|g(|u^{\epsilon}|^2)\nabla u^{\epsilon}\nabla Q^{\frac 12}e_i\|^2+\|g(|u^{\epsilon}|^2)u^{\epsilon}\Delta Q^{\frac 12}e_i\|^2\\
&+\|g'(|u^{\epsilon}|^2)|\nabla u^{\epsilon}|^2u^{\epsilon}Q^{\frac 12}e_i\|^2
+\|g'(|u^{\epsilon}|^2)\nabla u^{\epsilon}|u^{\epsilon}|^2Q^{\frac 12}e_i\|^2\\
&+\|g''(|u^{\epsilon}|^2)|\nabla u^{\epsilon}|^2|u^{\epsilon}|^3Q^{\frac 12}e_i\|^2+
\|g'(|u^{\epsilon}|^2)|\nabla u^{\epsilon}|^2u^{\epsilon}Q^{\frac 12}e_i\|^2
\Big)dr\Big]\\
&=: \mathbb E\Big[\|\Delta u^{\epsilon}(0)\|^{2}\Big]
+C(\lambda,p) \epsilon^{-\frac 12}\mathbb E\Big[\int_{0}^{t} \|\Delta u^{\epsilon}(r)\|(1+ \|\nabla u^{\epsilon}\|_{L^4}^{2})dr\Big]\\
&+C(\lambda,p) \mathbb E\Big[\int_{0}^{t} \|\Delta u^{\epsilon}(r)\| A(r) dr\Big]
+C(\lambda,p) \mathbb E\Big[\int_{0}^{t} B(r) dr\Big].
\end{align*}
Now applying the H\"older inequality, using the properties of $g$, using the Gagliardo--Nirenberg interpolation inequality, we obtain that for a small $\eta>0,$
\begin{align*}
&A(r)\le
\sum_{i} \|\Delta Q^{\frac 12}e_i\|\| Q^{\frac 12}e_i\|_{L^{\infty}} \|(g(|u^{\epsilon}|^2))^2 u^{\epsilon}\|_{L^{\infty}}+\sum_{i} \|\nabla Q^{\frac 12}e_i\|_{L^4}^2 \|(g(|u^{\epsilon}|^2))^2 u^{\epsilon}\|_{L^{\infty}}
\\
&+ \sum_{i}\|\nabla Q^{\frac 12}e_i\|_{L^4}\|Q^{\frac 12}e_i\|_{L^{\infty}}\|g(|u^{\epsilon}|^2)\|_{L^{\infty}}\|\nabla u^{\epsilon} \|_{L^4} +\sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2\|\nabla u^{\epsilon}\|_{L^4}^2\|g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2)u^{\epsilon}\|_{L^{\infty}} \\
&+ \sum_{i}\|Q^{\frac 12}e_i\|_{L^{\infty}}^2\|\nabla u^{\epsilon}\|_{L^4}^2
\|(g(|u^{\epsilon}|^2)g''(|u^{\epsilon}|^2)+(g'(|u^{\epsilon}|^2))^2)|u^{\epsilon}|^3 \|_{L^{\infty}}\\
&
+ \sum_{i} \|\nabla Q^{\frac 12}e_i\|_{L^4} \|Q^{\frac 12}e_i\|_{\infty} \|g(|u^{\epsilon}|^2)\|_{L^{\infty}}^2 \|\nabla u^{\epsilon}\|_{L^4}\\
&\le
\sum_{i}\Big(\|\nabla Q^{\frac 12}e_i\|_{L^4}^2+\|\nabla Q^{\frac 12}e_i\|_{L^{\infty}}^2+\|\Delta Q^{\frac 12}e_i\|^2+\|Q^{\frac 12}e_i\|_{L^{\infty}}^2\Big)
\Big(1+\|\nabla u\|_{L^4}+\|\nabla u\|_{L^4}^2\Big)\\
&\quad \times \Big(\|(g(|u^{\epsilon}|^2))^2 u^{\epsilon}\|_{L^{\infty}}+\|g(|u^{\epsilon}|^2)\|_{L^{\infty}} +\|g(|u^{\epsilon}|^2)g'(|u^{\epsilon}|^2)u^{\epsilon}\|_{L^{\infty}}\\
&\quad + \|(g(|u^{\epsilon}|^2)g''(|u^{\epsilon}|^2)+(g'(|u^{\epsilon}|^2))^2)|u^{\epsilon}|^3 \|_{L^{\infty}}+\|g(|u^{\epsilon}|^2)\|_{L^{\infty}}^2 \Big)\\
&\le C\sum_{i}\Big(\|\nabla Q^{\frac 12}e_i\|_{L^4}^2+\|\nabla Q^{\frac 12}e_i\|_{L^{\infty}}^2+\|\Delta Q^{\frac 12}e_i\|^2+\|Q^{\frac 12}e_i\|_{L^{\infty}}^2\Big)
\Big(1+\|\nabla u^{\epsilon}\|_{L^4}+\|\nabla u^{\epsilon}\|_{L^4}^2\Big).
\end{align*}
Similarly, we have that for a small $\eta>0$,
\begin{align*}
B(r)&\le \sum_{i}\Big(\|g(|u^{\epsilon}|^2)\|^2_{L^{\infty}}\|\nabla u^{\epsilon}\|_{L^4}^2\|\nabla Q^{\frac 12}e_i\|_{L^4}^2+\|g(|u^{\epsilon}|^2)u^{\epsilon}\|^2_{L^{\infty}}\|\Delta Q^{\frac 12}e_i\|^2\\
&+\|g'(|u^{\epsilon}|^2) u^{\epsilon}\|_{L^{\infty}}^2\|\nabla u^{\epsilon}\|_{L^4}^4\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
+\|g'(|u^{\epsilon}|^2)|u^{\epsilon}|^2\|_{L^{\infty}}^2\|\nabla u^{\epsilon}\|_{L^4}^2\|Q^{\frac 12}e_i\|_{L^{4}}^2\\
&+\|g''(|u^{\epsilon}|^2)|u^{\epsilon}|^3\|_{L^{\infty}}^2\|\nabla u^{\epsilon}\|_{L^4}^4\|Q^{\frac 12}e_i\|_{L^{\infty}}^2+
\|g'(|u^{\epsilon}|^2)u^{\epsilon}\|_{L^{\infty}}^2\|\nabla u^{\epsilon}\|_{L^4}^4\|Q^{\frac 12}e_i\|_{L^{\infty}}^2
\Big)\\
&\le C\sum_{i}\Big(\|\nabla Q^{\frac 12}e_i\|_{L^4}^2+\|\nabla Q^{\frac 12}e_i\|_{L^{\infty}}^2+\|\Delta Q^{\frac 12}e_i\|^2+\|Q^{\frac 12}e_i\|_{L^{\infty}}^2\Big)
\Big(1+\|\nabla u^{\epsilon}\|_{L^4}^4\Big).
\end{align*}
Combining the above estimates, and using the Young inequality and Gronwall inequality imply that
\begin{align*}
\mathbb E\Big[\|\Delta u^{\epsilon}(t)\|^{2}\Big]\le C(u_0,T,Q,p,\eta)(1+\epsilon^{- 2}).
\end{align*}
Now, taking supreme over $t$, then taking expectation, and applying the Burkerholder inequality to the $III_{sto}dW(t),$ we achieve that
for sufficient small $\eta>0,$
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,\tau]}\|\Delta u^{\epsilon}(t)\|^{2}\Big]\le C(u_0,T,Q,\eta)(1+\epsilon^{-2}).
\end{align*}
\iffalse
\textbf{Proof}[Proof of Proposition \ref{prop-well-reg}]
Based on Proposition \ref{mul} and Lemma \ref{h2-pri}, Proposition \ref{weak-loc} and Lemma \ref{h1-pri}, it is not hard to see that $u^{\epsilon}$ is a global solution. Next we focus on the H\"older continuity estimate. For convenience, assume that $0\le s\le t$. Then the mild formulation of $u^{\epsilon}$ yield that
\begin{align*}
&\|u^{\epsilon}(t)-u^{\epsilon}(s)\|\\
&\le \|(S(t-s)-I)u^{\epsilon}(s)\|+\|\int_{s}^tS(t-r)f_{\epsilon}(|u^{\epsilon}(r)|^2)u^{\epsilon}(r)dr\|\\
&\quad+\|\int_s^tS(t-r)dW(r)\|.
\end{align*}
The Gagliardo--Nirenberg interpolation inequality yield that for a small $\eta>0$,
\begin{align*}
\|f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon}\|^2&=
\int_{\epsilon +|u^{\epsilon}|^2\ge 1} f_{\epsilon}^2(|u^{\epsilon}|^2)|u^{\epsilon}|^2dx+\int_{\epsilon +|u^{\epsilon}|^2\le 1} f_{\epsilon}^2(|u^{\epsilon}|^2)|u^{\epsilon}|^2dx\\
&\le \int_{\epsilon +|u^{\epsilon}|^2\ge 1} (\epsilon +|u^{\epsilon}|^2)^{\eta}|u^{\epsilon}|^2dx+\int_{\epsilon +|u^{\epsilon}|^2\le 1}(\epsilon+|u^{\epsilon}|^2)^{-\eta}|u^{\epsilon}|^2dx\\
&\le C(\|u^{\epsilon}\|^2+\|u^{\epsilon}\|_{L^{2+2\eta}}^{2+2\eta}+\|u^{\epsilon}\|_{L^{2-2\eta}}^{2-2\eta}).
\end{align*}
If $\mathcal O$ is bounded, we immediately have
$$\|f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon}\|^2 \le C(\|u^{\epsilon}\|^2+\|\nabla u^{\epsilon}\|^{d\eta}\|u^{\epsilon}\|^{2\eta+2-d\eta}).$$
When $\mathcal O=\mathbb R^d$, we need use the weighted version of the interpolation inequality to deal with the term $\|u^{\epsilon}\|_{L^{2-2\eta}}^{2-2\eta}$.
Using the Cauchy-Schwarz inequality and $\alpha\in (0,1]$, we have that for $r>0,$ small $\eta>0$ and $\alpha>\frac {d\eta}{2(1-\eta)},$
\begin{align*}
\|v\|_{L^{2-2\eta}}^{2-2\eta}&\le \int_{|x|\le r}|v(x)|^{2-2\eta}dx+\int_{|x|\ge r}\frac {|x|^{(1-\eta)2\alpha}|v(x)|^{2-2\eta}}{|x|^{(1-\eta)2\alpha}}dx\\
&\le Cr^{d\eta} \|v\|^{2-2\eta}+C\|v\|_{L^2_{\alpha}}^{2-2\eta}(\int_{|x|\ge r}\frac 1{|x|^{(1-\eta)2\alpha\frac 1\eta}}dx)^{\eta}\\
&\le Cr^{d\eta} \|v\|^{2-2\eta}+C r^{d\eta-2\alpha+2\alpha \eta } \|v\|_{L^2_{\alpha}}^{2-2\eta}.
\end{align*}
Let $r=(\frac {\|v\|_{L^2_{\alpha}}}{\|v\|})^{\frac 1 {\alpha}}$, we get
\begin{align}\label{wei-sob}
\|v\|_{L^{2-2\eta}}\le C \|v\|_{L^2_{\alpha}}^{\frac {d\eta}{2\alpha(1-\eta)}}\|v\|^{1-\frac {d\eta}{2\alpha(1-\eta)}}.
\end{align}
Thus we conclude that for small $\eta>0$,
$$\|f_{\epsilon}(|u^{\epsilon}|^2)u^{\epsilon}\|^2 \le C(\|u^{\epsilon}\|^2+\|\nabla u^{\epsilon}\|^{d\eta}\|u^{\epsilon}\|^{2\eta+2-d\eta}+\|v\|_{L^2_{\alpha}}^{\frac {d\eta}{\alpha}}\|v\|^{2-2\eta-\frac {d\eta}{\alpha}}).$$
Now using Lemma \ref{h1-pri}, Proposition \ref{wei-est} and the continuity of $S(\cdot)$, we have
\begin{align*}
&\mathbb E\Big[\|u^{\epsilon}(t)-u^{\epsilon}(s)\|^{2p}\Big]\\
&\le \mathbb E\Big[\|(S(t-s)-I)u^{\epsilon}(s)\|^{2p}\Big]+
\mathbb E\Big[\|\int_{s}^tS(t-r)f_{\epsilon}(|u^{\epsilon}(r)|^2)u^{\epsilon}(r)dr\|^{2p}\Big]\\
&\quad+\mathbb E\Big[\|\int_s^tS(t-r)dW(r)\|^{2p}\Big]\\
&\le C_p(t-s)^{p}.
\end{align*}
By the Kolmogorov continuous modification theorem, for a.s. every $\omega$, $u^{\epsilon}(t)$ is $\frac 12-\epsilon$-H\"older continuous in time and thus is uniformly equicontinuous in $L^2$ for all $\epsilon.$
\qed
\fi
\qed
\textbf{[Proof of Proposition \ref{wei-hol-2}]}
We follow the steps in the proof of Proposition \ref{wei-hol-1} to present the proof in the case of $p=2$. For convenience, we present the proof for the multiplicative noise case.
Applying the It\^o formula to $\|u^{\epsilon}\|_{L^2_{\alpha}}^{2}=\int_{\mathbb R^d}(1+|x|^2)^{\alpha}|u^{\epsilon}|^2dx$,
using integration by parts, then taking supreme over $t$, and applying Burkerholder inequality, we deduce that
\begin{align*}
\mathbb E\Big[\sup_{t\in [0,T]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}\Big]
&\le \mathbb E\Big[\|u_0\|_{L^2_{\alpha}}^{2}\Big]
+2\alpha \mathbb E \Big[\int_0^T \Big| \<(1+|x|^2)^{\alpha-1} x u^{\epsilon}(s),\nabla u^{\epsilon}\> \Big|ds\Big]\\
&+\mathbb E \Big[\sup_{t\in [0,T]}\Big|\int_0^t\<(1+|x|^2)^{\alpha}u^{\epsilon}(s),\mathbf i g(|u^{\epsilon}(s)|^2)u^{\epsilon}(s)dW(s)\>\Big|\Big]\\
&\le \mathbb E\Big[\|u_0\|_{L^2_{\alpha}}^{2}\Big]
+C_{\alpha} \mathbb E \Big[\int_0^T\Big| \<(1+|x|^2)^{\alpha-1} x u^{\epsilon}(s),\nabla u^{\epsilon}\> \Big|ds\Big]\\
&+C \mathbb E \Big[\int_0^T\sum_{i\in\mathbb N^+}\Big|\<(1+|x|^2)^{\alpha}u^{\epsilon}(s), \mathbf ig(|u^{\epsilon}(s)|^2)u^{\epsilon}(s) Q^{\frac 12}e_i\>\Big|^2\Big|ds\Big].
\end{align*}
By H\"older's inequality, for $\alpha\in (1,2]$, we have that
\begin{align*}
\Big| \<(1+|x|^2)^{\alpha-1} x u^{\epsilon},\nabla u^{\epsilon}\> \Big|
&\le C\|u^{\epsilon}\|_{L^{2}_{\alpha}}\|(1+|x|^2)^{\frac \alpha 2-\frac 12}\nabla u^{\epsilon}\|.
\end{align*}
Integration by parts and H\"older's inequality yield that for some small $\eta>0,$
\begin{align*}
&\|(1+|x|^2)^{\frac \alpha 2-\frac 12}\nabla u^{\epsilon}\|^2=\<(1+|x|^2)^{\alpha -1} \nabla u^{\epsilon}, \nabla u^{\epsilon}\>\\
&=-\<(1+|x|^2)^{\alpha -1}u^{\epsilon}, \Delta u^{\epsilon}\>
-2(\alpha-1)\<(1+|x|^2)^{\alpha -2} x \nabla u^{\epsilon}, u^{\epsilon}\>\\
&\le \|u^{\epsilon}\|_{L^2_{\max(2\alpha-2,0)}} \|\Delta u^{\epsilon}\|+ C(\eta)|\alpha-1|\|u^{\epsilon}\|^2_{L^2_{\alpha}}+\eta |\alpha-1|\|\nabla u^{\epsilon}\|^2_{L^2_{\max(\alpha-3,0)}}.
\end{align*}
Combining the above estimates in Proposition \ref{wei-hol-1} and using Young's inequality, we achieve that
\begin{align*}
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}\Big]&\le e^{CT}(1+\epsilon^{-1}),\; \text{if}\; \alpha \in (1,\frac 32],\\
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}\Big]&\le e^{CT}(1+\epsilon^{-\frac 32}),\; \text{if}\; \alpha \in [\frac 32,2),\\
\mathbb E\Big[\sup_{t\in[0,T]}\|u^{\epsilon}(t)\|_{L^2_{\alpha}}^{2}\Big]&\le e^{CT}(1+\epsilon^{-2}),\; \text{if}\; \alpha =2.\\
\end{align*}
\qed
\bibliographystyle{plain}
|
\section{Introduction}
Short text matching (STM) is generally regarded as a task of paraphrase identification or sentence semantic matching.
Given a pair of sentences, the goal of matching models is to predict their semantic similarity.
It is widely used in question answer systems \cite{liu2018improved} and dialogue systems~\cite{gao2019neural,yu2014cognitive}.
Recent years have seen great progress in deep learning methods for text matching \cite{mueller2016siamese,gong2017natural,chen2017enhanced,lan2018neural}.
However, almost all of these models are initially proposed for English text matching. For Chinese language tasks, early work utilizes Chinese characters as input to the model, or first segment each sentence into words, and then take these words as input tokens. Although character-based models can overcome the problem of data sparsity to some degree~\cite{li2019word}, the main drawback is that explicit word information is not fully utilized, which has been demonstrated to be useful for semantic matching \cite{li2019enhancing}.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{lattice.pdf}
\caption{An example of word segmentation and the potential word ambiguity.}
\label{fig:example}
\end{figure}
However, a large number of Chinese words are polysemous, which brings great difficulties to semantic understanding~\cite{xu2016improve}.
Word polysemy in short text is more an issue than that in long text because short text usually has less contextual information, so it is extremely hard for models to capture the correct meaning. As shown in Fig. \ref{fig:example}, the word in red in sentence-1 actually has two meanings: one is to describe bragging (\emph{exaggeration}) and another is \emph{moisture}.
Intuitively, if other words in the context have similar or related meanings, the probability of them will increase.
To integrate semantic information of words, we introduce HowNet~\cite{dong2003hownet} as an external knowledge base. In the view of HowNet, words may have multiple senses/meanings and each sense has several sememes to represent it. For instance, the first sense \emph{exaggeration} indicates some boast information in his words. Therefore, it has sememes {\tt information} and {\tt boast}. Similarly, we can also find the sememe {\tt boast} describing the sense \emph{brag} which belongs to the word ``ChuiNiu (bragging)'' in sentence-2. In this way, model can better determine the sense of words and perceive that two sentences probably have the same meaning.
Furthermore, word-based models often encounter some potential issues caused by word segmentation. If the word segmentation fails to output ``ChuiNiu (bragging)'' in sentence-2, we will lose useful sense information. In Chinese, ``Chui (blowing)'' ``Niu (cattle)'' is a bad segmentation, which deviates the correct meaning of ``ChuiNiu (bragging)''.
To tackle this problem, many researchers propose word lattice graphs~\cite{lai2019lattice,li2020flat,chen2020neural}, where they retain words existing in the word bank so that various segmentation paths are kept. It has been shown that multi-granularity information is important for text matching.
In this paper, we propose a Linguistic knowledge Enhanced graph Transformer (LET) to consider both semantic information and multi-granularity information. LET takes a pair of word lattice graphs as input. Since keeping all possible words will introduce a lot of noise, we use several segmentation paths to form our lattice graph and construct a set of senses according to the word. Based on HowNet, each sense has several sememes to represent it. In the input module, starting from the pre-trained sememe embeddings provided by OpenHowNet~\cite{qi2019openhownet}, we obtain the initial sense representation using a multi-dimensional graph attention transformer (MD-GAT, see Sec. 3.1). Also, we get the initial word representation by aggregating features from the character-level transformer encoder using an Att-Pooling (see Sec. 4.1). Then it is followed by SaGT layers (see Sec. 4.2), which fuse the information between words and semantics. In each layer, we first update the nodes' sense representation and then updates word representation using MD-GAT. As for the sentence matching layer (see Sec. 4.3), we convert word representation to character level and share the message between texts. Moreover, LET can be combined with pre-trained language models, e.g. BERT~\cite{devlin2019bert}. It can be regarded as a method to integrate word and sense information into pre-trained language models during the fine-tuning phase.
Contributions in this work are summarized as: a) We propose a novel enhanced graph transformer using linguistic knowledge to moderate word ambiguity. b) Empirical study on two Chinese datasets shows that our model outperforms not only typical text matching models but also the pre-trained model BERT as well as some variants of BERT. c) We demonstrate that both semantic information and multi-granularity information are important for text matching modeling, especially on shorter texts.
\section{Related Work}
{\bf Deep Text Matching Models} based on
deep learning have been widely adopted for short text matching. They can fall into two categories: representation-based methods~\cite{he2016text,lai2019lattice} and interaction-based methods~\cite{wang2017bilateral,chen2017enhanced}. Most representation-based methods are based on Siamese architecture, which has two symmetrical networks (e.g. LSTMs and CNNs) to extract high-level features from two sentences. Then, these features are compared to predict text similarity. Interaction-based models incorporate interactions features between all word pairs in two sentences. They generally perform better than representation-based methods. Our proposed method belongs to interaction-based methods.
{\bf Pre-trained Language Models}, e.g. BERT, have shown its powerful performance on various natural language processing (NLP) tasks including text matching. For Chinese text matching, BERT takes a pair of sentences as input and each Chinese character is a separated input token. It has ignored word information. To tackle this problem, some Chinese variants of original BERT have been proposed, e.g. BERT-wwm~\cite{cui2019pre} and ERNIE~\cite{sun2019ernie}. They take the word information into consideration based on the whole word masking mechanism during pre-training. However,
the pre-training process of a word-considered BERT requires a lot of time and resources. Thus, Our model takes pre-trained language model as initialization and utilizes word information to fine-tune the model.
\section{Background}
In this section, we introduce graph attention networks (GATs) and HowNet, which are the basis of our proposed models in the next section.
\subsection{Graph Attention Networks}
Graph neural networks (GNNs)~\cite{scarselli2008graph} are widely applied in various NLP tasks, such as text classifcation~\cite{yao2019graph}, text generation~\cite{zhao2020line}, dialogue policy optimization~\cite{ chen2018structured,chen2018policy,chen2019agentgraph,chen2020distributed} and dialogue state tracking~\cite{chen2020schema,zhu2020efficient}, etc.
GAT is a special type of GNN that operates on graph-structured data with attention mechanisms. Given a graph $G = (\mathcal{V}, \mathcal{E})$, where $\mathcal{V}$ and $\mathcal{E}$ are the set of nodes $x_i$ and the set of edges, respectively. $\mathcal{N}^{+}(x_i)$ is the set including the node $x_i$ itself and the nodes which are directly connected by $x_i$.
Each node $x_i$ in the graph has an initial feature vector $\mathbf{h}_i^0 \in \mathbb{R}^d$, where
$d$ is the feature dimension. The representation of each node is iteratively updated by the graph attention operation. At the $l$-th step, each node $x_i$ aggregates context information by attending over its neighbors and itself. The updated representation $\mathbf{h}^l_i$ is calculated by the weighted average of the connected nodes,
\begin{equation}
\begin{split}
\mathbf{h}_i^l
= \sigma\left(\sum_{x_j \in \mathcal{N}^{+}(x_i)} \alpha_{ij}^l \cdot \left( \mathbf{W}^l \mathbf{h}_j^{l-1}\right)\right),
\end{split}
\label{eq:selfatt}
\end{equation}
where $\mathbf{W}^l \in \mathbb{R}^{d \times d}$ is a learnable parameter, and
$\sigma(\cdot)$ is a nonlinear activation function, e.g. ReLU. The attention coefficient $\alpha_{ij}^l$ is the normalized similarity of the embedding between the two nodes $x_i$ and $x_j$ in a unified space, i.e.
\begin{equation}
\begin{split}
\alpha_{ij}^l &= \mbox{softmax}_{j} \ f^l_{sim}\left( \mathbf{h}^{l-1}_i, \ \mathbf{h}^{l-1}_j \right) \\
& = \mbox{softmax}_{j}\left(\mathbf{W}^l_{q} \mathbf{h}_i^{l-1}\right)^T\left(\mathbf{W}^l_{k} \mathbf{h}_j^{l-1}\right),
\label{eq:attention}
\end{split}
\end{equation}
where $\mathbf{W}^l_{q}$ and $\mathbf{W}^l_{k} \in \mathbb{R}^{d \times d}$ are learnable parameters for projections.
Note that, in Eq. (\ref{eq:attention}), $\alpha_{ij}^l$ is a scalar, which means that all dimensions in $\mathbf{h}_j^{l-1}$ are treated equally. This may limit the capacity to model complex dependencies.
Following \citeauthor{shen2018disan}~\shortcite{shen2018disan}, we replace the vanilla attention with multi-dimensional attention.
Instead of computing a single scalar score, for each embedding $\mathbf{h}_j^{l-1}$, it first computes a feature-wise score vector, and then normalizes
it with feature-wised multi-dimensional softmax (MD-softmax),
\begin{equation}
\begin{split}
\bm{\alpha}_{ij}^l
= &\mbox{MD-softmax}_j\left( \hat{\alpha}^l_{ij} + f^l_{m}\left(\mathbf{h}_j^{l-1}\right)\right),
\end{split}
\end{equation}
where $\hat{\alpha}_{ij}^l$ is a scalar calculated by the similarity function $f^l_{sim} ( \cdot)$ in Eq. (\ref{eq:attention}), and $f^l_{m}(\cdot)$ is a vector. The addition in above equation means the scalar will be added to every element of the vector. $\hat{\alpha}_{ij}^l$ is utilized to model the pair-wised dependency of two nodes, while $f^l_{m}(\cdot)$ is used to estimate the contribution of each feature dimension of $\mathbf{h}_j^{l-1}$,
\begin{equation}
f^l_{m}(\mathbf{h}_j^{l-1}) = \mathbf{W}^l_{2} \sigma \left(\mathbf{W}^l_{1} \mathbf{h}_j^{l-1} + \mathbf{b}^l_{1} \right) + \mathbf{b}^l_{2},
\end{equation}
where $\mathbf{W}_{1}^l$, $\mathbf{W}_{2}^l$, $\mathbf{b}^l_{1}$ and $\mathbf{b}^l_{2}$ are learnable parameters.
With the score vector $\bm{\alpha}_{ij}^l$, Eq. (\ref{eq:selfatt}) will be accordingly revised as
\begin{equation}
\begin{split}
\mathbf{h}_i^l
&= \sigma\left(\sum_{x_j \in \mathcal{N}^{+}(x_i)} \bm{\alpha}_{ij}^l \odot \left( \mathbf{W}^l \mathbf{h}_j^{l-1} \right) \right),
\end{split}
\label{eq:mdatt}
\end{equation}
where $\odot$ represents element-wise product of two vectors.
For brevity, we use $\mbox{MD-GAT}(\cdot)$ to denote the updating process using multi-dimensional attention mechanism, and rewrite Eq. (\ref{eq:mdatt}) as follows,
\begin{equation}
\begin{split}
\mathbf{h}_i^l = \mbox{MD-GAT}\left(\mathbf{h}_i^{l-1}, \ \left\{ \mathbf{h}^{l-1}_j | x_j \in \mathcal{N}^{+}(x_i) \right\} \right). \\
\end{split}
\label{eq:mdatt_r}
\end{equation}
After $L$ steps of updating, each node will finally have a context-aware representation $\mathbf{h}_i^L$. In order to achieve a stable training process, we also employ a residual connection followed by a layer normalization between two graph attention layers.
\subsection{HowNet}
\begin{figure}[h]
\centering
\includegraphics[width=0.9\columnwidth]{sagt.pdf}
\caption{An example of the HowNet structure.}
\label{fig:sense}
\end{figure}
HowNet~\cite{dong2003hownet} is an external knowledge base that manually annotates each Chinese word sense with one or more relevant sememes. The philosophy of HowNet regards sememe as an atomic semantic unit. Different from WordNet~\cite{miller1995wordnet}, it emphasizes that the parts and attributes of a concept can be well represented by sememes. HowNet has been widely utilized in many NLP tasks such as word similarity computation~\cite{liu2002word}, sentiment analysis~\cite{xianghua2013multi}, word representation learning~\cite{niu2017improved} and language modeling~\cite{gu2018language}.
An example is illustrated in Fig. \ref{fig:sense}.
The word ``Apple'' has two senses including \emph{Apple Brand} and \emph{Apple}. The sense \emph{Apple Brand} has five sememes including {\tt computer}, {\tt PatternValue}, {\tt able}, {\tt bring} and {\tt SpecificBrand}, which describe the exact meaning of sense.
\section{Proposed Approach}
\label{sec:let}
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{net.pdf}
\caption{The framework of our proposed LET model.}
\label{fig:net}
\end{figure}
First, we define the Chinese short text matching task in a formal way.
Given two Chinese sentences $\mathcal{C}^{a}=\{c_1^a, c_2^a, \cdots, c_{T_a}^a\}$ and $\mathcal{C}^{b}=\{c_1^b, c_2^b, \cdots, c_{T_b}^b\}$, the goal of a text matching model $f(\mathcal{C}^{a},\mathcal{C}^{b})$ is to predict whether the semantic meaning of $\mathcal{C}^{a}$ and $\mathcal{C}^{b}$ is equal. Here, $c_t^a$ and $c_{t'}^b$ represent the $t$-th and $t'$-th Chinese character in two sentences respectively, and $T_a$ and $T_b$ denote the number of characters in the sentences.
In this paper, we propose a linguistic knowledge enhanced matching model. Instead of segmenting each sentence into a word sequence, we use three segmentation tools and keep these segmentation paths to form a word lattice graph $G=(\mathcal{V},\mathcal{E})$ (see Fig. \ref{fig:update} (a)). $\mathcal{V}$ is the set of nodes and $\mathcal{E}$ is the set of edges.
Each node $x_i \in \mathcal{V}$ corresponds to a word $w_i$ which is a character subsequence starting from the $t_1$-th character to the $t_2$-th character in the sentence. As introduced in Sec. 1, we can obtain all senses of a word $w_i$ by retrieving the HowNet.
For two nodes $x_i \in \mathcal{V}$ and $x_j \in \mathcal{V}$,
if $x_i$ is adjacent to $x_j$ in the original sentence, then there is an edge between them.
$\mathcal{N}_{fw}^{+}(x_i)$ is the set including $x_i$ itself and all its reachable nodes in its forward direction, while $\mathcal{N}_{bw}^{+}(x_i)$ is the set including $x_i$ itself and all its reachable nodes in its backward direction.
Thus for each sample, we have two graphs $G^a=(\mathcal{V}^a,\mathcal{E}^a)$ and $G^b=(\mathcal{V}^b,\mathcal{E}^b)$, and our graph matching model is to predict their similarity.
As shown in Fig. \ref{fig:net}, LET consists of four components: input module, semantic-aware graph transformer (SaGT), sentence matching layer and relation classifier. The input module outputs the initial contextual representation for each word $w_i$ and the initial semantic representation for each sense. The semantic-aware graph transformer iteratively updates the word representation and sense representation, and fuses useful information from each other. The sentence matching layer first incorporates word representation into character level, and then matches two character sequences with the bilateral multi-perspective matching mechanism. The relation classifier takes the sentence vectors as input and predicts the relation of two sentences.
\subsection{Input Module}
\subsubsection{Contextual Word Embedding}
\label{sec:cwe}
For each node $x_i$ in graphs, the initial representation of word $w_i$ is the attentive pooling of contextual character representations.
Concretely, we first concat the original character-level sentences to form a new sequence $\mathcal{C} = \{[\text{CLS}], c_1^a, \cdots, c_{T_a}^a, [\text{SEP}], c_1^b, \cdots, c_{T_b}^b, [\text{SEP}]\}$, and then feed them to the BERT model to obtain the contextual representations for each character $\{\mathbf{c}^{\text{CLS}}, \mathbf{c}_1^a$, $\cdots$, $\mathbf{c}_{T_a}^a$, $\mathbf{c}^{\text{SEP}}$, $\mathbf{c}_1^b$, $\cdots$, $\mathbf{c}_{T_b}^b$, $\mathbf{c}^{\text{SEP}}\}$. Assuming that the word $w_i$ consists of some consecutive character tokens $\{c_{t_1}, c_{{t_1}+1}, \cdots, c_{t_2}\}$\footnote{For brevity, the superscript of $c_{k}\ (t_1 \leq k \leq t_2)$ is omitted.}, a feature-wised score vector is calculated with a feed forward network (FFN) with two layers for each character $c_{k} \ (t_1 \leq k \leq t_2)$, and then normalized with a feature-wised multi-dimensional softmax (MD-softmax),
\begin{equation}
\mathbf{u}_{k}=\mbox{MD-softmax}_{k}\left( \text{FFN}(\mathbf{c}_{k}) \right),
\label{eq:attpool-1}
\end{equation}
The corresponding character embedding $\mathbf{c}_{k}$ is weighted with the normalized scores $\mathbf{u}_{k}$ to obtain the contextual word embedding,
\begin{equation}
\mathbf{v}_i = \sum_{k=t_1}^{t_2} \mathbf{u}_{k} \odot \mathbf{c}_{k},
\label{eq:attpool-2}
\end{equation}
For brevity, we use $\mbox{Att-Pooling}(\cdot)$ to rewrite Eq. (\ref{eq:attpool-1}) and Eq. (\ref{eq:attpool-2}) for short, i.e.
\begin{equation}
\mathbf{v}_i=\mbox{Att-Pooling}\left(\{\mathbf{c}_k | t_1 \leq k \leq t_2\}\right).
\end{equation}
\subsubsection{Sense Embedding}
The word embedding $\mathbf{v}_i$ described in Sec. \ref{sec:cwe} contains only contextual character information, which may suffer from the issue of polysemy in Chinese. In this paper, we incorporate HowNet as an external knowledge base to express the semantic information of words.
For each word $w_i$, we denote the set of senses as $\mathcal{S}^{(w_i)}=\{s_{i,1}, s_{i,2}, \cdots, s_{i,K}\}$. $s_{i,k}$ is the $k$-th sense of $w_i$ and we denote its corresponding sememes as $\mathcal{O}^{(s_{i,k})}=\{o_{i,k}^1, o_{i,k}^2, \cdots, o_{i,k}^M\}$.
In order to get the embedding $\mathbf{s}_{i,k}$ for each sense $s_{i,k}$, we first obtain the representation $\mathbf{o}_{i,k}^m$ for each sememe $o_{i,k}^m$ with multi-dimensional attention function,
\begin{equation}
\begin{split}
\mathbf{o}^{m}_{i,k} &= \mbox{MD-GAT}\left(\mathbf{e}_{i,k}^{m}, \ \left\{\mathbf{e}_{i,k}^{m'} | o_{i,k}^{m'} \in \mathcal{O}^{(s_{i,k})} \right\} \right), \\
\end{split}
\end{equation}
where $\mathbf{e}_{i,k}^{m}$ is the embedding vector for sememe $o_{i,k}^m$ produced through
the Sememe Attention over Target model (SAT) \cite{niu2017improved}. Then, for each sense $s_{i,k}$, its embedding $\mathbf{s}_{i,k}$ is obtained with attentive pooling of all sememe representations,
\begin{equation}
\mathbf{s}_{i,k} = \mbox{Att-Pooling}\left(\left\{ \mathbf{o}^{m}_{i,k} | o_{i,k}^{m} \in \mathcal{O}^{(s_{i,k})} \right\} \right).
\end{equation}
\subsection{Semantic-aware Graph Transformer}
\begin{figure}[t]
\centering
\includegraphics[width=1\columnwidth]{update.pdf}
\caption{(a) is an example of lattice graph. (b) shows the process of sense updating. $\text{fw}_2$ and $\text{bw}_2$ refer to the words in forward and backward directions of $\text{w}_2$ respectively. $\text{uw}_2$ is the words that $\text{w}_2$ cannot reach. (c) is word updating; we will not update the corresponding word representation if the word is not in HowNet.}
\label{fig:update}
\end{figure}
For each node $x_i$ in the graph, the word embedding $\mathbf{v}_i$ only contains the contextual information while the sense embedding $\mathbf{s}_{i,k}$ only contains linguistic knowledge. In order to harvest useful information from each other, we propose a semantic-aware graph transformer (SaGT). It first takes $\mathbf{v}_i$ and $\mathbf{s}_{i,k}$ as initial word representation $\mathbf{h}^0_i$ for word $w_i$ and initial sense representation $\mathbf{g}^0_{i,k}$ for sense $s_{i,k}$ respectively, and then iteratively updates them with two sub-steps.
\subsubsection{Updating Sense Representation} At $l$-th iteration, the first sub-step is to update sense representation from $\mathbf{g}^{l-1}_{i,k}$ to $\mathbf{g}^l_{i,k}$. For a word with multiple senses, which sense should be used is usually determined by the context in the sentence. Therefore, when updating the representation, each sense will first aggregate useful information from words in forward and backward directions of $x_i$,
\begin{equation}
\begin{split}
\mathbf{m}^{l,fw}_{i,k} &= \mbox{MD-GAT}\left(\mathbf{g}_{i,k}^{l-1}, \ \left\{\mathbf{h}_j^{l-1} | x_j \in \mathcal{N}_{fw}^{+}(x_i) \right\} \right), \\
\mathbf{m}^{l,bw}_{i,k} &= \mbox{MD-GAT}\left(\mathbf{g}_{i,k}^{l-1}, \ \left\{\mathbf{h}_j^{l-1} | x_j \in \mathcal{N}_{bw}^{+}(x_i) \right\} \right), \\
\end{split}
\end{equation}
where two multi-dimensional attention functions $\mbox{MD-GAT}(\cdot)$ have different parameters. Based on $\mathbf{m}^l_{i,k}=[\mathbf{m}^{l,fw}_{i,k}, \mathbf{m}^{l,bw}_{i,k}]$~\footnote{$[\cdot,\cdot]$ denotes the concatenation of vectors.}, each sense updates its representation with a gate recurrent unit (GRU) \cite{cho2014learning},
\begin{equation}
\mathbf{g}^l_{i,k} = \mbox{GRU}\left( \mathbf{g}_{i,k}^{l-1}, \ \mathbf{m}^l_{i,k}\right).
\label{eq:gru_sense}
\end{equation}
It is notable that we don't directly use $\mathbf{m}_{i,k}^l$ as the new representation $\mathbf{g}_{i,k}^{l}$ of sense $s_{i,k}$.
The reason is that $\mathbf{m}_{i,k}^l$ only contains contextual information, and we need to utilize a gate, e.g. GRU, to control the fusion of contextual information and semantic information.
\subsubsection{Updating Word Representation} The second sub-step is to update the word representation from $\mathbf{h}^{l-1}_{i}$ to $\mathbf{h}^{l}_{i}$ based on the updated sense representations $\mathbf{g}_{i,k}^l \ (1\leq k \leq K)$. The word $w_i$ first obtains semantic information from its sense representations with the multi-dimensional attention,
\begin{equation}
\begin{split}
\mathbf{q}^{l}_{i} = \mbox{MD-GAT}\left(\mathbf{h}_{i}^{l-1}, \ \left\{\mathbf{g}^{l}_{i,k} | s_{i,k} \in \mathcal{S}^{(w_i)} \right\} \right),
\end{split}
\end{equation}
and then updates its representation with a GRU:
\begin{equation}
\mathbf{h}^l_{i} = \mbox{GRU}\left( \mathbf{h}_{i}^{l-1}, \ \mathbf{q}^l_{i}\right).
\end{equation}
The above GRU function and the GRU function in Eq. (\ref{eq:gru_sense}) have different parameters.
After multiple iterations, the final word representation $\mathbf{h}_{i}^L$ contains not only contextual word information but also semantic knowledge. For each sentence, we use $\mathbf{h}_{i}^a$ and $\mathbf{h}_{i}^b$ to denote the final word representation respectively.
\subsection{Sentence Matching Layer}
After obtaining the semantic knowledge enhanced word representation $\mathbf{h}_{i}^{a}$ and $\mathbf{h}_{i}^{b}$ for each sentence, we incorporate this word information into characters. Without loss of generality, we will use characters in sentence $\mathcal{C}^a$ to introduce the process. For each character $c_t^a$, we obtain $\mathbf{\hat{c}}^a_t$ by pooling the useful word information,
\begin{equation}
\mathbf{\hat{c}}^a_t = \mbox{Att-Pooling}\left(\left\{ \mathbf{h}^{a}_{i} | w_i^a \in \mathcal{W}^{(c_t^a)} \right\} \right),
\end{equation}
where $\mathcal{W}^{(c_t^a)}$ is a set including words which contain the character $c_t^a$. The semantic knowledge enhanced character representation $\mathbf{y}_t$ is therefore obtained by
\begin{equation}
\mathbf{y}_t^a = \mbox{LayerNorm}\left( \mathbf{c}_t^a + \mathbf{\hat{c}}^a_t \right),
\end{equation}
where $\mbox{LayerNorm}(\cdot)$ denotes layer normalization, and $\mathbf{c}_t^a$ is the contextual character representation obtained using BERT described in Sec. \ref{sec:cwe}.
For each character $c_t^a$, it aggregates information from sentence $\mathcal{C}^a$ and $\mathcal{C}^b$ respectively using multi-dimensional attention,
\begin{equation}
\begin{split}
\mathbf{m}^{self}_{t} &= \mbox{MD-GAT} \left(\mathbf{y}_{t}^a, \ \{\mathbf{y}_{t'}^{a} | c_{t'}^a \in \mathcal{C}^a \} \right), \\
\mathbf{m}^{cross}_{t} &= \mbox{MD-GAT} \ (\mathbf{y}_{t}^a, \ \{\mathbf{y}_{t'}^{b} | c_{t'}^b \in \mathcal{C}^b \} ) \ . \\
\end{split}
\end{equation}
The above multi-dimensional attention functions $\mbox{MD-GAT}(\cdot)$ share same parameters. With this sharing mechanism, the model has a nice property that, when the two sentences are perfectly matched, we have $\mathbf{m}_t^{self} \approx \mathbf{m}_t^{cross}$.
We utilize the multi-perspective cosine distance \cite{wang2017bilateral} to compare $\mathbf{m}_t^{self}$ and $\mathbf{m}_t^{cross}$,
\begin{equation}
d_k = \text{cosine}\left( \mathbf{w}_k^{cos} \odot \mathbf{m}_t^{self}, \mathbf{w}_k^{cos} \odot \mathbf{m}_t^{cross} \right),
\end{equation}
where $k \in \{1,2,\cdots,P\}$ ($P$ is number of perspectives). $\mathbf{w}_k^{cos}$ is a parameter vector, which assigns different weights to different dimensions of messages. With $P$ distances $d_1, d_2, \cdots, d_P$,
we can obtain the final character representation,
\begin{equation}
\mathbf{\hat{y}}_t^a = \text{FFN} \left(\left[\mathbf{m}_t^{self}, \mathbf{d}_t \right] \right),
\label{eq:ue}
\end{equation}
where $\mathbf{d}_t \triangleq [d_1, d_2, \cdots, d_P]$, and $\mbox{FFN}(\cdot) $ is a feed forward network with two layers.
Similarly, we can obtain the final character representation $\mathbf{\hat{y}}_t^b$ for each character $c_t^b$ in sentence $\mathcal{C}^b$.
Note that the final character representation contains three kinds of information: contextual information, word and sense knowledge, and character-level similarity. For each sentence $\mathcal{C}^a$ or $\mathcal{C}^b$,
the sentence representation vector $\mathbf{r}^a$ or $\mathbf{r}^b$ is obtained using the attentive pooling of all final character representations for the sentence.
\subsection{Relation Classifier}
With two sentence vectors $\mathbf{r}^a$, $\mathbf{r}^b$, and the vector $\mathbf{c}^{\text{CLS}}$ obtained with BERT, our model will predict the similarity of two sentences,
\begin{equation}
p = \text{FFN} \left(\left[\mathbf{c}^{\text{CLS}}, \mathbf{r}^a, \mathbf{r}^b, \mathbf{r}^a \odot \mathbf{r}^b , |\mathbf{r}^a - \mathbf{r}^b| \right] \right),
\end{equation}
where $\mbox{FFN}(\cdot)$ is a feed forward network with two hidden layers and a sigmoid activation after output layer.
With $N$ training samples $\{\mathcal{C}^{a}_i, \mathcal{C}^{b}_i, y_i \}_{i=1}^N$, the training object is to minimize the binary cross-entropy loss,
\begin{equation}
\mathcal{L} = - \sum_{i=1}^N \left( y_i \text{log}\left(p_i\right) + \left( 1 - y_i\right) \text{log}\left(1- p_i\right) \right),
\end{equation}
where $y_i \in \{0,1\}$ is the label of the $i$-th training sample and $p_i \in[0,1]$ is the prediction of our model taking the sentence pair $\{\mathcal{C}^{a}_i, \mathcal{C}_i^{b}\}$ as input.
\section{Experiments}
\subsection{Experimental Setup}
\begin{table*
\renewcommand\arraystretch{1.1}
\centering{
\begin{tabular}{lcccccc}
\specialrule{0.1em}{1pt}{1pt}
\multirow{2}{*}{\textbf{Models}} & \multirow{2}{*}{\textbf{Pre-Training}} & \multirow{2}{*}{\textbf{Interaction}} & \multicolumn{2}{c}{\textbf{BQ}} & \multicolumn{2}{c}{\textbf{LCQMC}} \\
\cline{4-7}
& & & \textbf{ACC.} & \textbf{F1} & \textbf{ACC.} & \textbf{F1} \\
\specialrule{0.1em}{1pt}{1pt}
Text-CNN\cite{he2016text} & $\times$ & $\times$ & 68.52 & 69.17 & 72.80 & 75.70 \\
BiLSTM\cite{mueller2016siamese} & $\times$ & $\times$ & 73.51 & 72.68 & 76.10 & 78.90 \\
Lattice-CNN \cite{lai2019lattice} & $\times$ & $\times$ & 78.20 & 78.30 & 82.14 & 82.41 \\
BiMPM \cite{wang2017bilateral} & $\times$ & $\surd$ & 81.85 & 81.73 & 83.30 & 84.90 \\
ESIM \cite{chen2017enhanced} & $\times$ & $\surd$ & 81.93 & 81.87 & 82.58 & 84.49 \\
\textbf{LET} (Ours) & $\times$ & $\surd$ & \textbf{83.22} & \textbf{83.03} & \textbf{84.81} & \textbf{86.08} \\
\specialrule{0.1em}{1pt}{1pt}
BERT-wwm \cite{cui2019pre} & $\surd$ & $\surd$ & 84.89 & 84.29 & 86.80 & 87.78 \\
BERT-wwm-ext \cite{cui2019pre} & $\surd$ & $\surd$ & 84.71 & 83.94 & 86.68 & 87.71 \\
ERNIE \cite{sun2019ernie} & $\surd$ & $\surd$ & 84.67 & 84.20 & 87.04 & 88.06 \\\specialrule{0.0em}{0pt}{0pt}
BERT\cite{devlin2019bert} & $\surd$ & $\surd$ & 84.50 & 84.00 & 85.73 & 86.86 \\
\textbf{LET-BERT} (Ours) & $\surd$ & $\surd$ & \textbf{85.30} & \textbf{84.98} & \textbf{88.38} & \textbf{88.85} \\
\specialrule{0.1em}{1pt}{1pt}
\end{tabular}
\caption{Performance of various models on LCQMC and BQ test datasets. The results are average scores using 5 different seeds. All the improvements over baselines are statistically significant ($p < 0.05$).}
\label{tab:main-res}
}
\end{table*}
\subsubsection{Dataset} We conduct experiments on two Chinese short text matching datasets: LCQMC \cite{liu2018lcqmc} and BQ \cite{chen2018bq}.
LCQMC is a question matching corpus with large-scale open domain. It consists of 260068 Chinese sentence pairs including 238766 training samples, 8802 development samples and 12500 test samples. Each pair is associated with a binary label indicating whether two sentences have the same meaning or share the same intention. Positive samples are 30\% more than negative samples.
BQ is a domain-specific large-scale corpus for bank question matching. It consists of 120000 Chinese sentence pairs including 100000 training samples, 10000 development samples and 10000 test samples. Each pair is also associated with a binary label indicating whether two sentences have the same meaning. The number of positive and negative samples are the same.
\subsubsection{Evaluation metrics}
For each dataset, the accuracy (ACC.) and F1 score are used as the evaluation metrics. ACC. is the percentage of correctly classified examples. F1 score of matching is the harmonic mean of the precision and recall.
\subsubsection{Hyper-parameters}
The input word lattice graphs are produced by the combination of three segmentation tools: jieba~\cite{sun2012jieba}, pkuseg~\cite{pkuseg} and thulac~\cite{li2009punctuation}. We use the pre-trained sememe embedding provided by OpenHowNet~\cite{qi2019openhownet} with 200 dimensions. The number of graph updating steps/layers $L$ is 2 on both datasets, and the number of perspectives $P$ is 20. The dimensions of both word and sense representation are 128. The hidden size is also 128. The dropout rate for all hidden layers is 0.2.
The model is trained by RMSProp with an initial learning rate of 0.0005 and a warmup rate of 0.1. The learning rate of BERT layer is multiplied by an additional factor of 0.1. As for batch size, we use 32 for LCQMC and 64 for BQ. \footnote{Our code is available at \url{https://github.com/lbe0613/LET}.}
\subsection{Main Results}
We compare our models with three types of baselines: representation-based models, interaction-based models and BERT-based models. The results are summarized in Table \ref{tab:main-res}. All the experiments in Table \ref{tab:main-res} and Table \ref{tab:ablation} are running five times using different seeds and we report the \textbf{average} scores to ensure the reliability of results. For the baselines, we run them ourselves using the parameters mentioned in \citet{cui2019pre}.
\textbf{Representation-based models} include three baselines Text-CNN, BiLSTM and Lattice-CNN. Text-CNN~\cite{he2016text} is one type of Siamese architecture with Convolutional Neural Networks (CNNs) used for encoding each sentence. BiLSTM~\cite{mueller2016siamese} is another type of Siamese architecture with Bi-directional Long Short Term Memory (BiLSTM) used for encoding each sentence. Lattice-CNN~\cite{lai2019lattice} is also proposed to deal with the potential issue of Chinese word segmentation. It takes word lattice as input and pooling mechanisms are utilized to merge the feature vectors produced by multiple CNN kernels over different $n$-gram contexts of each node in the lattice graph.
\textbf{Interaction-based models} include two baselines: BiMPM and ESIM. BiMPM~\cite{wang2017bilateral} is a bilateral multi-perspective matching model. It encodes each sentence with BiLSTM, and matches two sentences from multi-perspectives. BiMPM performs very well on some natural language inference (NLI) tasks. There are two BiLSTMs in ESIM~\cite{chen2017enhanced}. The first one is to encode sentences, and the other is to fuse the word alignment information between two sentences. ESIM achieves state-of-the-art results on various matching tasks.
In order to be comparable with the above models, we also employ a model where BERT in Fig. \ref{fig:net} is replaced by a traditional character-level transformer encoder, which is denoted as LET.
The results of the above models are shown in the first part of Table \ref{tab:main-res}. We can find that our model LET outperforms all baselines on both datasets. More specifically, the performance of LET is better than that of Lattice-CNN. Although they both utilize word lattices, Lattice-CNN only focuses on local information while our model can utilize global information. Besides, our model incorporates semantic messages between sentences, which significantly improves model performance.
As for interaction-based models, although they both use the multi-perspective matching mechanism, LET outperforms BiMPM and ESIM. It shows the utilization of word lattice with our graph neural networks is powerful.
\textbf{BERT-based models} include four baselines: BERT, BERT-wwm, BERT-wwm-ext and ERNIE. We compare them with our presented model LET-BERT. BERT is the official Chinese BERT model released by Google. BERT-wwm is a Chinese BERT with whole word masking mechanism used during pre-training. BERT-wwm-ext is a variant of BERT-wwm with more training data and training steps. ERNIE is designed to learn language representation enhanced by knowledge masking strategies, which include entity-level masking and phrase-level masking. LET-BERT is our proposed LET model where BERT is used as a character level encoder.
The results are shown in the second part of Table~\ref{tab:main-res}. We can find that the three variants of BERT (BERT-wwm, BERT-wwn-ext, ERNIE) all surpass the original BERT, which suggests using word level information during pre-training is important for Chinese matching tasks. Our model LET-BERT performs better than all these BERT-based models. Compared with the baseline BERT which has the same initialization parameters, the ACC. of LET-BERT on BQ and LCQMC is increased by 0.8\% and 2.65\%, respectively. It shows that utilizing sense information during fine-tuning phrases with LET is an effective way to boost the performance of BERT for Chinese semantic matching.
We also compare results with K-BERT~\cite{liu2020k}, which regards information in HowNet as triples \{word, contain, sememes\} to enhance BERT, introducing soft position and visible matrix during the fine-tuning and inferring phases. The reported ACC. for the LCQMC test set of K-BERT is 86.9\%. Our LET-BERT is 1.48\% better than that. Different from K-BERT, we focus on fusing useful information between word and sense.
\subsection{Analysis}
\begin{table}[t]
\centering
\setlength{\tabcolsep}{5mm}{\begin{tabular}{lccc}
\specialrule{0.1em}{1pt}{1pt}
\textbf{Seg.} & \textbf{Sense} & \textbf{ACC.} & \textbf{F1} \\
\specialrule{0.1em}{1pt}{1pt}
jieba & $\surd$ & 87.84 & 88.47 \\
pkuseg & $\surd$ & 87.72 & 88.40 \\
thulac & $\surd$ & 87.50 & 88.27 \\
lattice & $\surd$ & \textbf{88.38} & \textbf{88.85} \\
lattice & $\times$ & 87.68 & 88.40 \\
\specialrule{0.1em}{1pt}{1pt}
\end{tabular}
\caption{Performance of using different segmentation on LCQMC test dataset.}
\label{tab:ablation}}
\end{table}
In our view, both multi-granularity information and semantic information are important for LET. If the segmentation does not contain the correct word, our semantic information will not exert the most significant advantage.
Firstly, to explore the impact of using different segmentation inputs, we carry out experiments using LET-BERT on LCQMC test set. As shown in Table \ref{tab:ablation}, when incorporating sense information, improvement can be observed between lattice-based model (the fourth row) and word-based models: jieba, pkuseg and thulac. The improvements of lattice with sense over other models in Table \ref{tab:ablation} are all statistically significant ($p < 0.05$). The possible reason is that lattice-based models can reduce word segmentation errors to make predictions more accurate.
Secondly, we design an experiment to demonstrate the effectiveness of incorporating HowNet to express the semantic information of words. In the comparative model without HowNet knowledge, the sense updating module in SaGT is removed, and we update word representation only by a multi-dimensional self-attention. The last two rows in Table \ref{tab:ablation} list the results of combined segmentation (lattice) with and without sense information. The performance of integrating sense information is better than using only word representation. More specifically, the average absolute improvement in ACC. and F1 scores are 0.7\% and 0.45\%, respectively, which indicates that LET has the ability to obtain semantic information from HowNet to improve the model's performance.
Besides, compared with using a single word segmentation tool, semantic information performs better on lattice-based model. The probable reason is lattice-based model incorporates more possible words so that it can perceive more meanings.
We also study the role of GRU in SaGT. The ACC. of removing GRU in lattice-based model is 87.82\% on average, demonstrating that GRU can control historical messages and combine them with current information. Through experiments, we find that the model with 2 layers of SaGT achieves the best. It indicates multiple information fusion will refine the message and make the model more robust.
\subsubsection{Influences of text length on performance}
As listed in Table \ref{tab:length}, we can observe that text length also has great impacts on text matching prediction. The experimental results show that the shorter the text length, the more obvious the improvement effect of utilizing sense information. The reason is, on the one hand, concise texts usually have rare contextual information, which is difficult for model to understand. However, HowNet brings a lot of useful external information to these weak-context short texts. Therefore, it is easier to perceive the similarity between texts and gain great improvement.
On the other hand, longer texts may contain more wrong words caused by insufficient segmentation, leading to incorrect sense information. Too much incorrect sense information may confuse the model and make it unable to obtain the original semantics.
\begin{table}[t]
\centering
\resizebox{\linewidth}{!}{
\begin{tabular}{ccccc}
\specialrule{0.1em}{1pt}{1pt}
\multirow{2}{*}{\textbf{text length}} & \multirow{2}{*}{\textbf{\begin{tabular}[c]{@{}c@{}}number of \\ samples\end{tabular}}} & \multicolumn{2}{c}{\textbf{ACC.}}&\multirow{2}{*}{\textbf{RER(\%)}} \\ \cline{3-4}
& & w/o sense & sense & \\ \specialrule{0.1em}{1pt}{1pt}
$<16$ & 2793 & 88.99 & 90.05 &\textbf{9.63} \\\specialrule{0em}{1pt}{1pt}
$16-18$ & 3035 & 88.49 & 89.25 &6.60 \\\specialrule{0em}{1pt}{1pt}
$19-22$ & 3667 & 88.58 & 89.04 &4.03 \\\specialrule{0em}{1pt}{1pt}
$>22$ & 3005 & 84.53 & 85.13 &3.88\\\specialrule{0.1em}{1pt}{1pt}
\end{tabular}
}
\caption{Influences of text length on LCQMC test dataset. Relative error reduction (RER) is calculated by $\frac{\text{sense} - \text{w/o sense}}{100 - \text{w/o sense}} \times 100\%$.}
\label{tab:length}
\end{table}
\subsubsection{Case study}
We compare LET-BERT between the model with and without sense information (see Fig.~\ref{fig:case}). The model without sense fails to judge the relationship between sentences which actually have the same intention, but LET-BERT performs well. We observe that both sentences contain the word ``yuba'', which has only one sense described by sememe {\tt food}. Also, the sense of ``cook'' has a similar sememe {\tt edible} narrowing the distance between texts. Moreover, the third sense of ``fry'' shares the same sememe {\tt cook} with the word ``cook''. It provides a powerful message that makes ``fry'' attend more to the third sense.
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{case.pdf}
\caption{An example of using sense information to get the correct answer.}
\label{fig:case}
\end{figure}
\section{Conclusion}
In this work, we proposed a novel linguistic knowledge enhanced graph transformer for Chinese short text matching. Our model takes two word lattice graphs as input and integrates sense information from HowNet to moderate word ambiguity. The proposed method is evaluated on two Chinese benchmark datasets and obtains the best performance. The ablation studies also demonstrate that both semantic information and multi-granularity information are important for text matching modeling.
\section{Acknowledgments}
We thank the anonymous reviewers for their thoughtful comments. This work has been supported by No. SKLMCPTS2020003 Project.
|
\section{Introduction}
\label{sec:introduction}
The volatility of the financial assets returns, as a risk measure for the financial market, is very crucial for investment decision \citep{markowitz1952portfolio}.
It is also the most important factor in derivatives pricing \citep{Duan1995,Goard2013}, since the uncertainty of the assets returns, reflected as volatility, drives the values of the derivatives.
The volatility of assets returns is not observable and the historical volatility is usually measured by the standard deviation of the historical returns over a fixed time interval.
The returns of financial assets often exhibit heteroscedasticity, in the sense that the volatility of the returns is time-dependent.
In particular, higher volatility tends to be followed by higher volatility, while lower volatility usually is followed by lower volatility, one phenomenon called volatility clustering in the financial market.
In addition, the past positive and negative returns have asymmetric effects on the volatility --- a negative past return will increase the volatility more significantly than a positive past return.
In order to do future investment decision and derivatives pricing, volatility models need to be developed to forecast volatility in the future.
As the volatility is not observable, the biggest challenge in constructing volatility model is to identify the dependence between the unobserved volatility and the observed data (assets returns) so as to capture the above two empirical characteristics of the volatility: volatility clustering and asymmetric effects of positive and negative returns.
Generalized Autoregressive Conditional Heteroscedasticity model (GARCH) \citep{bollerslev1986generalized} is the most popular volatility model, in which the variance of the returns (squared volatility) is assumed to be a linear function of the past squared returns and past squared volatility.
A number of variant GARCH models have been developed, most of which mainly attempt to address the two limitations of the GARCH model: the assumption of a linear relationship between current and past squared volatility and the assumption of symmetric effects of the positive and negative past returns.
Maximum likelihood method can be used to learn the parameters of the GARCH family models.
A detailed review of the GARCH family can be find in \cite{hentschel1995all}.
The GARCH family models are deterministic volatility models, in the sense that given the past returns and past volatility, the future volatility is a deterministic value.
An alternative type of volatility models is the stochastic volatility model, in which the volatility is modeled as a latent stochastic process defined through partial differential equations.
In practice, stochastic volatility models discretized for even spaced data are usually used instead for forecasting \citep{SKim1997,Omori2007}, and are estimated using Monte Carlo Markov Chain methods.
Although theoretically more sound, the stochastic volatility models usually do not outperform the GARCH model in practice \citep{SKim1997,Omori2007}.
Both the above deterministic and stochastic volatility models developed in statistics community impose restricted form on the function of the volatility model.
Recently, volatility modeling has received attention from the machine learning community, where researchers are interested in proposing volatility model with less assumptions on the function form.
For example, \cite{Wu2014} propose to model the dependence of future volatility on past volatility and past return with a flexible non-linear Gaussian process model, which only assumes the volatility model function to be smooth.
Recent developments in deep latent neural network models \citep{krishnan2017structured,fraccaro2016sequential,chung2015recurrent,bayer2014learning,Liu2018} provide a flexible framework to model the dependence between observed and unobservable latent variables.
\cite{Luo2018} propose a volatility model in which a sequence of latent variables are used to extract a compact representation of the original multivariate stock return time series. However, the model does not consider the effects of volatility clustering explicitly, but implicitly through the latent variables. In addition, modeling multivariate stock time series (using more information) entails more effects on data pre-processing, for example, how to choose the group of stocks to be modeled together, which may affect the performance.
In this work, we propose a Deep Stochastic Volatility Model (DSVM) based on the framework of the deep latent variable model using only the information of univariate return time series. The DSVM imposes no restriction on the function form for the volatility model, and explicitly considers the effects of stochastic noise, past volatility (autoregression) and past returns (exogenous variables) on volatility, thus providing a general framework for volatility modeling. A scalable learning and inference method is developed to estimate the DSVM based on variational inference. Real data analysis shows that the DSVM outperforms several popular volatility models in terms of testing negative likelihood based on only the information of the univariate return time series. In addition, the predicted volatility of DSVM gives a more reliable risk measure for financial assets: higher volatility when market is more risky and lower volatility when market is more stable.
In Section 2, we will review several volatility models in detail. The formulation of the DSVM is given in Section 3 and a scalable inference and learning algorithm is presented in Section 4. Section 5 concludes.
\section{Volatility models} \label{sec:volatility}
In this section, we will detail on several volatility models. Considering a sequence of T observations of the stock returns, labeled $r_{1:T}=\{r_1,r_2,\cdots,r_T\}$, we are interested in modeling the distribution $p(r_{1:T})$ as well as the latent volatility process $\sigma_t, t = 1,\cdots, T$. One common approach is to assume the return $r_t$ follows a normal distribution with mean 0 and variance $\sigma_t^2$. Mathematically, we can formulate as
\begin{equation*}
\begin{aligned}
r_{t} &= \sigma_t \epsilon_t \\
\epsilon_t &\sim \text { i.i.d. } \mathcal{N} \left(0, 1\right)
\end{aligned}
\end{equation*}
Volatility models aim to find a function form for the evolution of the volatility $\sigma_t$ or some transformations of the volatility such as the variance $\sigma_t^2$ and the log variance $\log (\sigma_t^2)$.
\subsection{Deterministic volatility models}
The GARCH \citep{bollerslev1986generalized} assumes a linear dependence of the future variance $\sigma_t^2$ on the previous $p$ squared returns and previous $q$ variances, computed as as follows:
\begin{equation*}
\sigma_{t}^{2} =\omega_{0}+\sum_{i=1}^{p} \alpha_{i} r_{t-i}^{2}+\sum_{j=1}^{q} \beta_{j} \sigma_{t-j}^{2}
\end{equation*}
Volatility clustering is captured by the parameters $\alpha_{i}, i=1,\cdots,p$. However, the GARCH model assumes symmetric effects of negative and positive returns. \cite{glosten1993relation} propose the GJR-GARCH model which extends the GARCH model to account for the asymmetry effects of positive and negative returns as follows:
\begin{equation*}
\sigma_{t}^{2}=\omega_{0} +\sum_{i=1}^{p} \alpha_{i} (r_{t-i} ^{2} + \gamma_{i} I\left\{r_{t-i}<0\right\}r_{t-i}^{2} )+\sum_{j=1}^{q} \beta_{j} \sigma_{t-j}^{2}
\end{equation*}
where the asymmetry effect is captured by the indicator function $I\left\{r_{t-i}<0\right\}$ and controlled by the leverage parameters $\gamma_{i}$. Similarly, the Threshold GARCH (TGARCH) proposed by \cite{zakoian1994threshold} also use leverage parameter $\gamma_{i}$ to reflect the asymmetry effects. The difference is that the TGARCH directly works with the volatility $\sigma_{t}$ instead of the variance $\sigma_{t}^2$:
\begin{equation*}
\sigma_{t} =\omega_{0} +\sum_{i=1}^{p} \alpha_{i}(\left|r_{t-i}\right| - \gamma_{i}r_{t-i})+\sum_{j=1}^{q} \beta_{j} \sigma_{t-j}
\end{equation*}
Another popular extension of the GARCH model is the Exponential GARCH (EGARCH, \cite{nelson1991conditional}), which models the evolution of the log variance $\log (\sigma_t^2)$:
\begin{equation*}
\log \left(\sigma_{t}^{2}\right) =\omega_{0}+\sum_{i=1}^{p} (\alpha_{i} \epsilon_{t}+\gamma_i\left(\left|\epsilon_{t}\right|-\mathbb{E}\left|\epsilon_{t}\right|\right))+\sum_{j=1}^{q} \beta_{j} \log \left(\sigma_{t-j}^{2}\right)
\end{equation*}
where the leverage parameter $\gamma_i$ is used to account for the asymmetry effect.
\subsection{Stochastic volatility models}
For analysis convenience, stochastic volatility models are usually formulated with stochastic differential equations in continuous-time \citep{hull1987pricing,heston1993closed}. The Heston model for a univariate process is
\begin{equation*}
\begin{aligned}
\mathrm{d} \sigma(t) &=-\beta \sigma(t) \mathrm{d} t+\delta \mathrm{d} W^{\sigma}(t) \\ \mathrm{d} s(t) &=\left(\mu-0.5 \sigma^{2}(t)\right) \mathrm{d} t+\sigma(t) \mathrm{d} W^{s}(t)
\end{aligned}
\end{equation*}
where $s(t)$ is the logrithm of stock price at $t$, $W^{\sigma}(t)$ and $W^{s}(t)$ are two correlated Wiener processes, with correlation represented as $E[\mathrm{d} W^{\sigma}(t)\mathrm{d} W^{s}(t)] = \rho \mathrm{d}t$. For practical use, stochastic volatility can be formulated in a discrete-time fashion for regularly spaced data. The discrete stochastic volatility without leverage \citep{SKim1997} is defined as
\begin{equation*}
\begin{aligned}
r_t &= \sigma_t \epsilon_t\\
\log (\sigma_t^2) &= \mu + \phi (\log (\sigma_{t-1}^2) -\mu ) + z_t\\
\epsilon_t \sim \text { i.i.d. } \mathcal{N}(0,1) & \quad z_t \sim \text { i.i.d. } \mathcal{N}(0,\sigma_z^2)
\end{aligned}
\end{equation*}
In order to accommodate the asymmetric effect of positive and negative returns, \cite{Omori2007} propose the discrete stochastic volatility with leverage as follows:
\begin{equation*}
\begin{array}{c}
{r_{t}=\sigma_{t} \epsilon_{t}} \\
{\log \left(\sigma_{t}^{2}\right)=\mu+\phi\left(\log \left(\sigma_{t-1}^{2}\right)-\mu\right)+z_{t}} \\
{\left(\begin{array}{c}{\epsilon_{t-1}} \\ {z_{t}}\end{array}\right) \sim \text { i.i.d. } \mathcal{N}_{2}(\mathbf{0}, \Sigma), \quad \Sigma=\left(\begin{array}{cc}{1} & {\rho \sigma_{z}} \\ {\rho \sigma_{z}} & {\sigma_{z}^{2}}\end{array}\right)}\end{array}
\end{equation*}
where the leverage $\rho <0$ is trying to capture the increase in volatility that follows a drop in the return.
\subsection{Machine learning stochastic volatility model}
\cite{Wu2014} proposed a Gaussian process volatility model (GP-Vol):
\begin{equation*}
\begin{aligned}
r_{t} &=\sigma_{t} \epsilon_{t}\\
\log(\sigma_t^2) &= f(\log(\sigma_{t-1}^2), r_{t-1}) + z_t\\
\epsilon_t \sim \text { i.i.d. } \mathcal{N}(0,1) & \quad z_t \sim \text { i.i.d. } \mathcal{N}(0,\sigma_z^2)
\end{aligned}
\end{equation*}
where $f$ can be modeled as a sample path of a Gaussian Process and $z_t$ is the stochastic noise. The model can be trained with Sequential Monte Carlo method.
The neural stochastic volatility model (NSVM) proposed by \citep{Luo2018} can be formulated as follows:
\begin{equation*}
\sigma_t = f(r_{1:t-1}, z_{1:t})
\end{equation*}
where $f$ is a deep neural network model. As will be detailed in the following, one of the main differences of our DSVM model and the NSVM is that we use a different generative probabilistic model to
account for the effect of past volatility and model the volatility as $\sigma_t = f_\sigma(\sigma_{1: t-1},r_{1:t-1}, z_{1:t})$. Another difference is that, \citep{Luo2018} model the volatility utilizing the information from multiple stock return time series which are correlated to each other and thus utilize more information. However, how to properly group multiple stocks is not a easy task in the first place. In this paper, we aim to propose a volatility model using less information, i.e. only based on the knowledge of the univariate stock return series.
Besides, there exist other machine learning volatility models using more exogenous variables, such as sentiment data \citep{xing2019sentiment} and using other advanced deep learning architectures, see \citep{zhang2018benchmarking}.
\section{Deep stochastic volatility model} \label{sec:dsvm}
In this section, we give the formulation for the Deep Stochastic Volatility Model (DSVM).
Recall that $r_t$ is the asset return and $\sigma_t$ is the volatility of the asset at timestep $t$. We introduce $z_t$ as the continuous stochastic noise to the volatility at timestep $t$. The stochastic noise is assumed to be normal distributed, whose parameters depend on past stochastic noise as follows:
\begin{equation*}
z _ { t } \sim N(m_t^{(p)}, (v_t^{(p)})^2)
\end{equation*}
\begin{equation*}
m_ { t }^{(p)} = f _ { 1 } \left( z _ { t - 1 } \right) , \quad v_ { t } ^{(p)} = f_ {2} \left( z _ { t - 1 } \right)
\end{equation*}
where $z _ { t }$ is a Gaussian variable with parameter $m_t^{(p)},v_t^{(p)}$ given by MLP models $f_{1}$, $f_{2}$ that depend on $z _ { t - 1 }$.
In the DSVM, the volatility is assumed to depend on past returns, past volatilities and past stochastic noise through a Recurrent Neural Network (RNN), computed as
\begin{equation*}
\begin{aligned}
h_t &= f_{h}(\sigma_{t-1},r_{t-1},z_{t},h_{t-1}) \\
\sigma_{t} &= f_3(h_t)
\end{aligned}
\label{equ:volatility}
\end{equation*}
where $f_h$ is a RNN model and $f_3$ is a MLP model. Therefore, current volatility depends on previous volatility $\sigma_{t-1}$, previous return $r_{t-1}$ and current stochastic noise $z_{t}$ as well as $\sigma_{1:t-2}, r_{1:t-2}, z_{1:t-1}$ encoded in $h_{t-1}$. This parameterization make $\sigma_{t}$ depends on $\sigma_{1: t-1}, r_{1: t-1}, z_{1:t}$ and the volatility model can then be denoted as $\sigma_{t} = f_\sigma( \sigma_{1: t-1}, r_{1: t-1}, z_{1: t})$.
The return is assumed to follow a Gaussian distribution with mean 0 and standard deviation $\sigma_t$:
\begin{equation*}
r_{t} \sim \mathcal{N}(0,\sigma_t^2)
\end{equation*}
The graphical representation of the DSVM model is given in Figure \ref{fig:deepvolatilitymodel}, and is termed as the generative network which specifies the generating process of the data. The parameters of the generative network include the parameters in the following functions: $\theta =\{f_1,f_2,f_3,f_{h}\}$.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\linewidth]{images/DSVM_Generative}
\caption{The generative network of the DSVM}
\label{fig:deepvolatilitymodel}
\end{figure}
The joint probability of the return $r_{1 : T}$ and the stochastic noise $z_{1 : T}$ can be factorized as
\begin{align}
&\quad p_{\theta}\left(r_{1 : T}, z_{1 : T} |\sigma_0=0, r_0=0,z_0=0\right) \notag\\
&=p_{\theta_{}}\left(r_{1 : T} | z_{1 : T}\right)p_{\theta_{}}\left(z_{1 : T} \right) \notag\\
&=\prod_{t=1}^{T} p_{\theta}\left(r_{t} | r_{1:t-1}, {z}_{1:t} \right) p_{\theta}\left(z_{t} | z_{1:t-1}\right) \notag\\
&=\prod_{t=1}^{T} p_{\theta}\left(r_{t} | \sigma_t \right) p_{\theta}\left(z_{t} | z_{t-1}\right) \notag\\
&=\prod_{t=1}^{T} p_{\theta}\left(r_{t} | \sigma_t \right) p_{\theta}\left(z_{t} | z_{t-1}\right).
\label{equ:jointprobability}
\end{align}
The marginal loglikelihood of the observed returns $r_{1 : T}$, denoted as $\mathcal{L}(\theta) = \log p_{\theta}\left(r_{1 : T}\right)$, can be obtained by averaging out $z_{1 : T}$ in the joint distribution in Equation (\ref{equ:jointprobability}). However, integrating out $z_{1 : T}$ is challenging due to the non-linearility introduced by deep neural networks, resulting intractable $\mathcal{L}(\theta)$. Therefore, maximum likelihood methods can not be used to estimate the parameters of DSVM model. In the next section, we will develop a scalable learning and inference for the DSVM using variational inference.
\section{Scalable learning and inference} \label{sec:learning}
Instead of maximizing the likelihood $\mathcal{L}(\theta)$ with respect to the parameter $\theta$, we instead maximize a variational evidence lower bound $ELBO(\theta,\phi)$ with respect to both $\theta$ and $\phi$, where $\phi$ is the parameter for the approximated posterior.
\subsection{Variational inference for the DSVM}
Denote $q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)$ as any approximated posterior for the true posterior $p_{\theta}\left(z_{1 : T}| r_{1 : T}\right)$, it can be derived that
\begin{equation*}
\begin{aligned}
\mathcal{L}(\theta) &\geq ELBO(\theta,\phi) \\
&= \iint q_{\phi}\left(z_{1 : T} | r_{1 : T}\right) \log \frac{p_{\theta}\left(r_{1 : T}, z_{1 : T} \right)}{q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)} \mathrm{d} z_{1 : T}
\end{aligned}
\end{equation*}
When $q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)$ is equal to the true posterior $p_{\theta}\left(z_{1 : T}| r_{1 : T}\right)$, the lower bound is tight and we have $\mathcal{L}(\theta) = ELBO(\theta,\phi)$. However, the true posterior is intractable. In order to make the lower bound $ELBO$ tight, we design the inference network which represents $q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)$ to best approximate the true posterior $p_{\theta}\left(z_{1 : T}| r_{1 : T}\right)$.
In fact, the true posterior can be factorized as
\begin{equation*}
\begin{aligned}
p_{\theta}\left(z_{1 : T}| r_{1 : T}\right) &= \Pi_{t=1}^Tp_{\theta}\left(z_{t} | {z}_{1 : t-1},r_{1 : T}\right) \\
&= \Pi_{t=1}^T p_{\theta}\left(z_{t}|z_{t-1}, r_{t:T}\right)
\end{aligned}
\end{equation*}
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\linewidth]{images/DSVM_Inference}
\caption{The inference network of the DSVM}
\label{fig:inferencemodel}
\end{figure}
The factorization can be obtained from the Figure \ref{fig:deepvolatilitymodel} using d-seperation \citep{geiger1990identifying}. The factorization shows that, the posterior of ${z_t}$ depends on the past information encoded in $z_{t-1}$ as well as the future information in ${r}_{t : T}$. We can design our inference network according to this factorization form of the true posterior. Specifically,
\begin{equation}
q_{\phi}\left(z_{1 : T}| r_{1 : T}\right)= \prod_{t=1}^{T}q_{\phi}\left(z_{t} | {z}_{ t-1},{r}_{t : T}\right)
\label{equ:inference}
\end{equation}
In addition, we have the following two factorization
\begin{equation*}
p_{\theta}\left(z_{1 : T}\right) =\prod_{t=1}^{T} p_{\theta}\left(z_{t} | z_{t-1}\right)
\end{equation*}
\begin{equation*}
p_{\theta}\left(r_{1 : T} | z_{1 : T}\right) = \prod_{t=1}^{T} p_{\theta}\left(r_{t} | r_{1:t-1}, {z}_{1:t} \right) = \prod_{t=1}^{T} p_{\theta}\left(r_{t} |\sigma_t \right)
\end{equation*}
The past information $r_{1:t-1}, {z}_{1:t}$ is reflected in $\sigma_t$. Therefore, the conditional density of return at time $t$ on the past returns and the stochastic noises can be replaced by the conditional density of return on its simultaneous volatility.
The ELBO then can be derived as
\begin{equation}
\resizebox{1\hsize}{!}{$
\begin{aligned}
& \quad \mathcal{L}(\theta) \\
& \geq ELBO(\theta,\phi)\\
&= \int q_{\phi}\left(z_{1 : T} | r_{1 : T}\right) \log \frac{p_{\theta}\left(r_{1 : T}| z_{1 : T}\right) p_{\theta}\left(z_{1 : T}\right)}{q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)} \mathrm{d} z_{1 : T} \\
&= \int q_{\phi}\left(z_{1 : T} | r_{1 : T}\right) \log p_{\theta}\left(r_{1 : T}| z_{1 : T}\right) \mathrm{d} z_{1 : T} + \int q_{\phi}\left(z_{1 : T} | r_{1 : T}\right) \log \frac{p_{\theta}\left(z_{1 : T}\right)}{q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)} \mathrm{d} z_{1 : T} \\
&=\mathbb{E}_{q_{\phi}\left(z_{1 : T}|r_{1 : T}\right)}\left[\log \prod_{t=1}^{T} p_{\theta}\left(r_{t} | \sigma_t \right) \right] -\int q_{\phi}\left(z_{1 : T} | r_{1 : T}\right) \log \frac{\prod_{t=1}^{T}q_{\phi}\left(z_{t} | z_{t-1},{r}_{t : T}\right) }{\prod_{t=1}^{T} p_{\theta}\left(z_{t} | z_{t-1}\right)}\mathrm{d} z_{1 : T}\\
&= \sum_{t=1}^{T}\mathbb{E}_{q_{\phi}^*(z_{t})}\left[\log p_{\theta}\left(r_{t} | \sigma_t \right) \right] - \sum_{t=1}^{T} \mathbb{E}_{q_{\phi}^*(z_{t-1})} \left[KL[q_{\phi}\left(z_{t} | z_{t-1},{r}_{t : T}\right)|| p_{\theta}\left(z_{t} | z_{t-1}\right)]\right]
\label{equ:elbo_factorized}
\end{aligned}
$}
\end{equation}
where $q_{\phi}^*({z}_t)$ is the marginal probability density of ${z}_t$:
\begin{equation*}
\begin{aligned}
q_{\phi}^{*}\left({z}_{t}\right) &=\int q_{\phi}\left(z_{1 : T} |r_{1 : T}\right) \mathrm{d} {z}_{1 : t-1} \\
&= E_{q_{\phi}^{*}\left(z_{t-1}\right)}\left[q_{\phi}\left({z}_{t}|z_{t-1},{r}_{t : T}\right)\right]
\end{aligned}
\end{equation*}
We approximate the ELBO in Equation (\ref{equ:elbo_factorized}) using the Monte Carlo estimation. Specifically, we samples from $q_{\phi}^{*}\left({z}_{t}\right)$ using ancestral sampling according to the Equation (\ref{equ:inference}). Given a sample $z_{t-1}^{(s)}$ from $q_{\phi}^{*}\left(z_{t-1}\right)$, a sample ${z}_{t}^{(s)}$ from $q_{\phi}\left(z_{t} |z_{t-1}^{(s)},{r}_{t : T}\right)$ follows the distribution of $q_{\phi}^{*}\left({z}_{t}\right)$. After we get the samples $z_t^{(s)} \text{ for } t= 1 \cdots, T$ from the $q_{\phi}\left(z_{1 : T}| r_{1 : T}\right)$, the ELBO in (\ref{equ:elbo_factorized}) can be approximated with the generated samples as:
\begin{align}
ELBO(\theta,\phi) &= \sum_{t=1}^{T}\left[\log p_{\theta}\left(r_{t} | \sigma_t = f_\sigma(\sigma_{1:t-1}, r_{1:t-1}, {z}_{1:t}^{(s)}) \right) \right] \nonumber \\
& \quad - \sum_{t=1}^{T} \left[KL[q_{\phi}\left(z_{t} | z_{t-1}^{(s)},{r}_{t : T}\right)|| p_{\theta}\left(z_{t} | z_{t-1}^{(s)}\right)]\right]
\label{equ:approximatedELBO}
\end{align}
\subsection{Parameterization of the inference network}
Recall that, instead of factorizing $q_{\phi}$ as mean-field approximation across different time step, i.e. $z_t, t = 1,\cdots, T$ are assumed to independent and does not use the information of the dependence structure in the generative network, we choose the variational posterior to have the same factorization as the true posterior in Equation \eqref{equ:inference}, where the posterior of $z_t$ depend on $z_{t-1}$ and $r_{t:T}$.
We additionally assume that the information in $r_{t:T}$ can be encoded with a backward RNN, denoted as $g_A$ with hidden state, denoted as $A_t$, as follows:
\begin{equation*}
\begin{aligned}
{A}_{t} &=g_{A}\left({A}_{t+1},r_{t}\right)\\
q_{\phi}\left(z_{1 : T}| r_{1 : T}\right) &= \prod_{t=1}^{T}q_{\phi}\left(z_{t} | {z}_{ t-1},{r}_{t : T}\right) = \prod_{t=1}^{T}q_{\phi}\left(z_{t} | {z}_{ t-1}, A_t\right)
\end{aligned}
\end{equation*}
where $g_{A}$ is a RNN model. The graphical model of the inference network is shown in Figure \ref{fig:inferencemodel}. $z_{t-1}$ represents the information coming from the past, whereas $A_t$ encodes the information from present and the future, thus the information from all time steps are used to approximate the posterior at each $t$. Further, $q_{\phi_z}\left(z_{t} | z_{t-1}, A_t\right)$ is parameterized to be a Gaussian distribution density:
\begin{equation*}
z _ { t } \sim N(m^{(q)}_t, (v^{(q)}_t)^2)
\end{equation*}
\begin{equation*}
m ^{(q)}_ { t } = g _ { 1 } \left( z _ { t - 1 } , A _ { t } \right) , \quad v^{(q)} _ { t } = g_ {2} \left( z _ { t - 1 } , A _ { t } \right)
\end{equation*}
$z _ { t }$ is a Gaussian variable with parameter $m^{(q)}_t, v^{(q)}_t$ given by MLP models $g_{1}$ and $g_{2}$ that depend on $z _ { t - 1 } , A_t $. The parameters in the inference network include parameters in the following functions: $\phi =\{g_1,g_2,g_{A}\}$.
\subsection{The gradients}
The derivative of the $ELBO(\theta,\phi)$ with respect to $\theta$ can be calculated as:
\begin{equation*}
\begin{aligned}
&\quad \nabla_\theta ELBO(\theta,\phi)\\
&= E_{q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)} \nabla_\theta \log p_{\theta}\left(r_{1 : T}, z_{1 : T} \right)\\
&= \sum_{t=1}^{T} E_{q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)} \nabla_\theta
[\log p_{\theta}\left(r_{t} | \sigma_t \right) p_{\theta}\left(z_{t} | z_{t-1}\right)]
\end{aligned}
\end{equation*}
The derivative of the $ELBO(\theta,\phi)$ with respect to $\phi$ is more tricky as $\phi$ appears in the expectation in $ELBO(\theta,\phi)$. Score function gradient estimator \citep{glynn1987likelilood,williams1992simple} can be used to approximate the gradient, but the obtained gradients suffer from high variance. Reparameterization trick \citep{rezende2014stochastic,kingma2013auto} is often used to obtain a lower variance gradient estimator instead. Specifically, in order to obtain a sample $z_t$ from $ N(m_t^{(q)},(v_t^{(q)})^2)$, we first generate a sample $\eta_t \sim N(0,1)$, and then use $m_t^{(q)} + \eta_t v_t^{(q)}$ as a sample of $z_t$, denote as $z_{t} = k_t(\eta_{t})$. The gradients then can be backpropogated through the gaussian random variables and $\nabla_\phi ELBO(\theta,\phi)$ can be calculated as
\begin{equation*}
\begin{aligned}
&\quad \nabla_\phi ELBO(\theta,\phi) \\
&=\sum_{t=1}^{T} \nabla_\phi E_{q_{\phi}\left(z_{1 : T} | r_{1 : T}\right)} \log \frac{ p_{\theta}\left(r_{t} | \sigma_t \right) p_{\theta}\left(z_{t} | z_{t-1}\right)}{q_{\phi}\left(z_{t} | {z}_{ t-1},{r}_{t : T}\right) }\\
&= \sum_{t=1}^{T} \nabla_\phi E_{\eta_{1:t}\sim N(0,I)} \log \frac{ p_{\theta}\left(r_{t} | \sigma_t \right) p_{\theta}\left(k_t(
\eta_{t}) | k_{t-1}(\eta_{t-1})\right)}{q_{\phi}\left(k_t(\eta_{t}) | k_{t-1}({\eta}_{ t-1}),{r}_{t : T}\right) }\\
&= \sum_{t=1}^{T} E_{\eta_{1:t}\sim N(0,I)} \nabla_\phi \log \frac{ p_{\theta}\left(r_{t} | \sigma_t \right) p_{\theta}\left(k_t(
\eta_{t}) | k_{t-1}(\eta_{t-1})\right)}{q_{\phi}\left(k_t(\eta_{t}) | k_{t-1}({\eta}_{ t-1}),{r}_{t : T}\right) }
\end{aligned}
\end{equation*}
The above inference and learning algorithm for training DSVM are summarized in Algorithm \ref{alg:DSVM}.
\begin{algorithm}
\caption{Structured Inference Algorithm for DSVM}
\label{alg:DSVM}
\begin{algorithmic}
\STATE \textbf{Inputs}: $\{r_{1:T}\}_{i=1}^N$\\
\qquad \qquad Randomly initialized $\phi^{(0)}$ and $\theta^{(0)}$\\
\qquad \qquad Generative network model: $p_{\theta}\left(r_{1 : T}, z_{1 : T}\right)$\\
\qquad \qquad Inference network model: $q_{\phi}\left(z_{1 : T} |r_{1 : T}\right)$
\STATE \textbf{Outputs}: the parameters of $\theta$, $\phi$
\WHILE{$Iter$ $<$ M}
\STATE1. Sample a mini-batch sequences $\{r_{1:T}\}_{i=1}^B$ uniformly from dataset
\STATE2. Generate samples of $\eta_t^{(s)},t=1,\cdots,T$ to obtain samples of $z_t^{(s)},t=1,2,\cdots,T$ sequentially according to Equation (\ref{equ:inference}) to approximate the ELBO in Equation (\ref{equ:approximatedELBO})
\STATE3. Derive the gradient $\nabla_\theta ELBO(\theta,\phi)$ which is approximated with one Monte Carlo sample
\STATE4. Derive the gradient $\nabla_\phi ELBO(\theta,\phi)$ which is approximated with one Monte Carlo sample
\STATE5. Update $\theta^{(Iter)}, \phi^{(Iter)}$ using the ADAM.
\STATE set $Iter = Iter+1$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\subsection{One-step-ahead forecasting}
After training the model, we can use the estimated model to do one-step-ahead forecasting. Specifically, the probability of $r_{T+1}$ conditional on the information of $r_{1:T}$ can be derived as:
\begin{equation}
\resizebox{1\hsize}{!}{$
\begin{aligned}
&\quad \quad p(r_{T+1}|r_{1:T}) \\
&= \int_{z_{1:T+1}} p(r_{T+1},z_{1:T+1}|r_{1:T}) d z_{1:T+1}\\
& = \int_{z_{1:T+1}} p(r_{T+1}|z_{1:T+1},r_{1:T}) p(z_{1:T+1}|r_{1:T})d z_{1:T+1}\\
& = \int_{z_{1:T+1}} p(r_{T+1}|\sigma_{T+1}= f_\sigma(\sigma_{1:T},r_{1:T}, z_{1:T+1})) p(z_{T+1}|z_{T})p(z_{1:T}|r_{1:T})d z_{1:T+1}\\
&\approx \int_{z_{1:T+1}} p(r_{T+1}|\sigma_{T+1}= f_\sigma(\sigma_{1:T},r_{1:T}, z_{1:T+1})) p(z_{T+1}|z_{T})q(z_{1:T}|r_{1:T})d z_{1:T+1}\\
&\approx \frac{1}{S}\sum_{s=1}^S p(r_{T+1}|\sigma_{T+1}= f_\sigma(\sigma_{1:T},r_{1:T}, z_{1:T+1}^{(s)}))
\end{aligned}
$}
\label{equ:predicted}
\end{equation}
We approximate the $p(z_{1:T}|r_{1:T})$ with $q(z_{1:T}|r_{1:T})$ and use ancestral sampling to generate samples from the approximated $p(r_{T+1}|r_{1:t})$. Specifically, we generate $\{z_{1:T}^{(s)}\}_{s=1}^{S}$ from $q(z_{1:T}|r_{1:T})$ and then generate samples $\{z_{T+1}^{(s)}\}_{s=1}^{S}$ from $p(z_{T+1}|z_{T})$ based on $\{z_{T}^{(s)}\}_{s=1}^{S}$. The approximated predictive distribution for $p(r_{T+1}|r_{1:t})$ is then a mixture of $S$ Gaussian models. We can then generate $\{r_{T+1}^{(s)}\}_{s=1}^{S}$ from the predictive distribution (\ref{equ:predicted}) and use the sample standard deviation of $\{r_{T+1}^{(s)}\}_{s=1}^{S}$ as the predicted volatility for $r_{T+1}$.
\section{Real data analysis} \label{sec:rda}
In this section, we present the experiment results of the proposed DSVM model on the US stock market.
\subsection{Data and model setting}
We choose the daily adjusted return of the 39 biggest stocks in terms of market values among SP500 stocks from 2001-01-02 to 2018-12-31.
For each stock, the training and validation data are from 2001-01-02 to 2015-10-18 (in total of 3721 days) and the testing data are from 2015-10-19 to 2018-12-31 (in total of 806 days). We extract sequences of 10 timestep data (corresponding to half of a trading month) from the original univariate time series and the training, validation, testing data size is 105924, 31434 and 31473 respectively (a proportion of around 6:2:2).
We train one model for all the 39 stocks. The dimension of the latent variable $z$ is chosen to be 1. $f_1,f_2,f_3,g_1,g_2$ are all parameterized by two-hidden-layer MLP models with different activation functions at last layer (Linear for mean function $f_1$, $g_1$, Softplus for standard deviation function $f_2,f_3,g_2$). $f_h, g_{{A}}$ are both parameterized by one-layer GRU models with 10 hidden states. Monte Carlo samples size used for forecasting is 1000. The model is trained on a Tesla V100 GPU for 1 hours (300 epochs) before it converges. The final model is chosen by minimizing the validation loss over the 300 epochs. The DSVM is only trained once on the training and validation data, and then is used to do recursive one-day-ahead forecasting without retraining.
\subsection{Alternative models}
We compare the performance of the propsed model with seven popular alternative volatility models described in Section 2.
\begin{enumerate}
\item Deterministic volatility models:\\
We consider four models: GARCH, GJR-GARCH, TGARCH and EGARCH. The parameter estimation is done with a widely used R package ``rugarch" \footnote{https://cran.r-project.org/web/packages/rugarch}.
\item Stochastic volatility models:\\
We choose two models: the discrete stochastic volatility model without leverage and discrete stochastic volatility model with leverage, and use the R package ``stochvo" \footnote{https://cran.r-project.org/web/packages/stochvol} to estimate the parameters.
\item Gaussian process volatility model (GP-Vol) \\
The implementation of the model is obtained from the authors' website\footnote{https://bitbucket.org/jmh233/code-gpvol-nips2014/src/master} and the hyperparameters are chosen as suggested in \cite{Wu2014}.
\end{enumerate}
For each type of the alternatives, we train one model for each stock as the alternatives are designed to be so. We will perform recursive one-day-ahead forecasting and all the alternative models are retrained at every forecasting time step with a rolling window of 1000.
\subsection{Result and discussion}
We first evaluate each model in terms of negative loglikelihood (NLL) for the testing samples for each stock. The smaller the NLL value, the better the model can explain the distribution of the observed testing samples.
The results of NLL evaluations are listed in Table \ref{tab:nll}.
We first apply the Friedman rank sum test and the test indicates that there exist significant differences of the NLL values among different models (the p-value is less than $e^{-35}$). Furthermore, a pairwise pos thoc Nemenyi-Friedman test \citep{Demsar2006} is performed to compare different models, and it confirms that the DSVM performs best among all the methods. Our DSVM is the only stochastic volatility model that is able to beat deterministic volatility models and is the best for 36 out of the 38 stocks.
\begin{table}[!t]
\footnotesize
\centering
\caption{Negative log-likelihood (NLL) evaluated on the test samples of 39 stock return time series for different volatility models. The best performing model are highlighted in bold. The last line shows the mean NLL for the 39 stocks. NA values for TGARCH are due to non-convergence of the parameter estimation (MLE).}
{\resizebox{1.02\linewidth}{!}{
\begin{tabular}{lrrrrrrrr}
\toprule
\toprule
\multicolumn{1}{c}{\textbf{Stock}} & \multicolumn{1}{c}{\textbf{DSVM}} & \multicolumn{1}{c}{\textbf{GARCH}} & \multicolumn{1}{c}{\textbf{GJRGARCH}} & \multicolumn{1}{c}{\textbf{TGARCH}} & \multicolumn{1}{c}{\textbf{EGARCH}} & \multicolumn{1}{c}{\textbf{GPVol}} & \multicolumn{1}{c}{\textbf{StoVol}} & \multicolumn{1}{c}{\textbf{StoVolL}} \\
\midrule
\textbf{ORCL} & \textbf{-3.03} & -2.87 & -2.87 & -2.87 & -2.88 & -2.82 & -2.25 & -2.23 \\
\textbf{MSFT} & \textbf{-2.96} & -2.84 & -2.86 & -2.86 & -2.87 & -2.87 & -2.44 & -2.45 \\
\textbf{KO} & -3.35 & -3.36 & -3.36 & -3.38 & \textbf{-3.38} & -3.27 & -3.14 & -3.15 \\
\textbf{XOM} & \textbf{-3.11} & -3.10 & -3.10 & \textbf{-3.11} & \textbf{-3.11} & -2.99 & -2.96 & -2.97 \\
\textbf{GE} & \textbf{-2.87} & -2.81 & -2.83 & -2.83 & -2.80 & -2.72 & -2.46 & -2.46 \\
\textbf{IBM} & \textbf{-3.06} & -2.92 & -2.92 & NA& -2.93 & -2.96 & -2.49 & -2.51 \\
\textbf{PEP} & -3.29 & \textbf{-3.30} & -3.30 & NA & -3.30 & -3.21 & -3.13 & -3.14 \\
\textbf{MO} & \textbf{-3.08} & -3.00 & -2.99 & -3.01 & -3.03 & -2.97 & -2.71 & -2.73 \\
\textbf{COP} & \textbf{-2.52} & -2.51 & \textbf{-2.52} & \textbf{-2.52} & \textbf{-2.52} & -2.32 & -2.40 & -2.42 \\
\textbf{AMGN} & \textbf{-2.92} & -2.85 & -2.86 & -2.86 & -2.86 & -2.81 & -2.67 & -2.70 \\
\textbf{SLB} & \textbf{-2.82} & \textbf{-2.82} & \textbf{-2.82} & \textbf{-2.82} & \textbf{-2.82} & -2.73 & -2.73 & -2.77 \\
\textbf{CVX} & \textbf{-2.95} & -2.93 & -2.94 & -2.93 & -2.93 & -2.80 & -2.80 & -2.84 \\
\textbf{AAPL} & \textbf{-2.89} & -2.82 & -2.85 & -2.85 & -2.85 & -2.77 & -2.52 & \textbf{-2.59} \\
\textbf{UTX} & \textbf{-3.16} & -3.12 & -3.13 & -3.13 & -3.12 & -3.05 & -2.85 & -2.89 \\
\textbf{PG} & \textbf{-3.27} & -3.25 & -3.25 & -3.26 & -3.26 & -3.18 & -2.94 & -2.98 \\
\textbf{BMY} & \textbf{-2.85} & -2.67 & -2.66 & NA & -2.67 & -2.71 & -2.07 & -2.08 \\
\textbf{BA} & \textbf{-2.86} & -2.77 & -2.77 & -2.76 & -2.76 & -2.71 & -2.50 & -2.50 \\
\textbf{ABT} & \textbf{-3.00} & -2.91 & -2.91 & -2.92 & -2.90 & -2.82 & -2.51 & -2.57 \\
\textbf{PFE} & \textbf{-3.18} & -3.15 & -3.16 & -3.15 & -3.15 & -3.07 & -2.96 & -2.97 \\
\textbf{JNJ} & \textbf{-3.27} & -3.20 & -3.21 & -3.19 & -3.19 & -3.16 & -2.83 & -2.83 \\
\textbf{MMM} & \textbf{-3.17} & -3.09 & -3.08 & -3.11 & -3.09 & -3.06 & -2.67 & -2.71 \\
\textbf{MRK} & \textbf{-3.08} & -2.99 & -3.01 & NA & -2.99 & -2.92 & -2.71 & -2.75 \\
\textbf{DIS} & \textbf{-3.12} & -3.07 & -3.07 & -3.08 & -3.08 & -3.04 & -2.86 & -2.88 \\
\textbf{WFC} & \textbf{-2.93} & -2.88 & -2.87 & -2.88 & -2.88 & -2.82 & -2.67 & -2.74 \\
\textbf{MCD} & \textbf{-3.20} & -3.07 & -3.04 & -3.06 & -3.13 & -3.03 & -2.71 & -2.75 \\
\textbf{JPM} & \textbf{-2.98} & -2.93 & -2.93 & -2.93 & -2.93 & -2.83 & -2.66 & -2.74 \\
\textbf{WMT} & \textbf{-3.09} & -2.90 & -2.87 & -2.94 & -2.94 & -2.88 & -2.34 & -2.34 \\
\textbf{INTC} & \textbf{-2.85} & -2.73 & -2.76 & -2.76 & -2.76 & -2.65 & -2.33 & -2.33 \\
\textbf{BAC} & \textbf{-2.74} & -2.71 & -2.71 & -2.71 & -2.71 & -2.61 & -2.53 & -2.58 \\
\textbf{VZ} & \textbf{-3.09} & -3.05 & -3.04 & -3.04 & -3.05 & -2.96 & -2.81 & -2.80 \\
\textbf{T} & \textbf{-3.09} & -3.03 & -3.03 & -3.04 & -3.04 & -2.91 & -2.66 & -2.65 \\
\textbf{HD} & \textbf{-3.12} & -3.10 & -3.11 & -3.11 & -3.11 & -3.02 & -2.94 & -2.98 \\
\textbf{AIG} & \textbf{-3.02} & -2.91 & -2.91 & -2.90 & -2.89 & -2.76 & -2.60 & -2.63 \\
\textbf{C} & \textbf{-2.82} & -2.78 & -2.79 & -2.80 & -2.80 & -2.71 & -2.58 & -2.65 \\
\textbf{CSCO} & \textbf{-3.01} & -2.91 & -2.93 & -2.95 & -2.95 & -2.86 & -2.50 & -2.56 \\
\textbf{QCOM} & \textbf{-2.72} & -2.52 & -2.52 & -2.50 & -2.53 & -2.45 & -1.51 & -1.50 \\
\textbf{BRK} & \textbf{-3.20} & -3.16 & -3.16 & -3.15 & -3.15 & -3.09 & -2.93 & -2.91 \\
\textbf{AMZN} & \textbf{-2.77} & -2.58 & -2.60 & NA & -2.65 & -2.69 & -2.24 & -2.29 \\
\textbf{UNH} & \textbf{-3.06} & -3.00 & -3.01 & -3.01 & -3.02 & -2.91 & -2.81 & -2.86 \\
\midrule
\textbf{Mean} & \textbf{-3.01} & -2.94 & -2.94 & -2.95 & -2.95 & -2.88 & -2.63 & -2.65 \\
\bottomrule
\bottomrule
\end{tabular}%
}}
\label{tab:nll}%
\end{table}%
We further plot the predicted volatility of DSVM for training data and testing data and compare with the volatility produced by the EGARCH model (the best performing alternative model in terms of mean NLL).
We use the two largest stocks (in terms of market value), the MSFT and the AAPL, for illustration.
Other stocks perform similarly and are omitted in the manuscript due to page limit.
The predicted volatility for training and testing data are displayed in Figure \ref{fig:insample} (with a randomly selected zoom-in period) and Figure \ref{fig:outsample} respectively.
In general, the changes of the DSVM is more adaptive to the market than the EGARCH model, which is reflected by several spikes in the predicted volatility of DSVM.
In addition, the DSVM produces higher predicted volatility when the market is more volatile (see the circle areas) and gives lower volatility when the market is more stable (see the rectangle areas) for both training and testing data.
In contrast, the predicted volatility of EGARCH is more smooth and the changes of the market status could not be quickly captured.
Take the AAPL stock for example (see Figure \ref{fig:outsample}), the volatility of the EGARCH decays more slowly around 2016 June when the market was calm, while it increases more slowly on 2018 September when the market became more volatile.
In summary, the DSVM produces predicted volatility more responsive to the market status, which is more desirable as a risk measure for the financial market.
\begin{figure}
\centering
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{images/stock1_insample}
\subcaption{Stock MSFT }
\label{fig:Stock1Insample}
\end{subfigure}
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{images/stock1_insample_zoom0}
\subcaption{Stock MSFT, zoom in for a subsample }
\label{fig:Stock1InsampleZoom}
\end{subfigure}
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{images/stock12_insample}
\subcaption{Stock AAPL}
\label{fig:Stock2Insample}
\end{subfigure}
\begin{subfigure}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{images/stock12_insample_zoom0}
\subcaption{Stock AAPL, zoom in for a subsample }
\label{fig:Stock2InsampleZoom}
\end{subfigure}
\caption{Training data}
\label{fig:insample}
\end{figure}
\begin{figure}
\begin{minipage}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{images/stock1_outsample}
\subcaption{Stock MSFT}
\label{fig:Stock1Outsample}
\end{minipage}
\begin{minipage}{1\linewidth}
\centering
\includegraphics[width=1\linewidth]{images/stock12_outsample}
\subcaption{Stock AAPL }
\label{fig:Stock2Outsample}
\end{minipage}
\caption{Testing data with recursive one-day-ahead forecasting}
\label{fig:outsample}
\end{figure}
\section{Conclusion} \label{sec:conclusion}
In this work, we propose a deep stochastic volatility model based on the framework of deep latent variable models.
The proposed volatility model does not impose restrictions on the function form on the volatility model, instead, it uses the flexible deep learning models to automatically detect the dependence of the future volatility on past returns, past volatilities and the stochastic noise. We developed a scalable inference and learning algorithm based on the variational inference, in which the parameters of the generative network and the inference network are jointly trained.
In addition, the results of real data analysis on the U.S. stock market show that the DSVM model using only the information from the univariate return time series outperforms several alternative volatility models. Furthermore, the
the predicted volatility of the deep stochastic volatility model is more adaptive to the market status, and thus provides a more reliable risk measure to better reflect the risk in the financial market.
\clearpage
\bibliographystyle{dcu}
|
\section{Introduction}
The cognitive map, a concept raised by Edward C. Tolman in 1930s \cite{tolman1948cognitive}, describes that the mental representation of a physical space could be built by integrating knowledge gained from the environmental features (e.g., goals, landmarks, and intentions). We used the Webots robot simulation environment \cite{michel2004cyberbotics} to investigate cognitive map behavior observed in rats during a spatial and working memory task, known as the triple T-maze \cite{olson2017subiculum,olson2020secondary}. We suggest that similar behavior could be observed in a robot that had a biologically plausible neural network evolved to solve such a task. In this task, the rat or the robot must take one of four paths to receive a reward. If it repeated a path, there would be no additional reward. It would eventually learn to quickly reach each of the four rewards with minimal repeats. This requires knowledge of where it is now, where it has been, and where it should go next.
In our Webots setting, we designed a 3-D environment that resembled the rat experiment. The proximity sensors, the linear accelerometer and the grayscale camera pixels of a simulated e-puck robot \cite{mondada09theepuck} provided sensory input for a recurrent neural network (RNN). The RNN output directly manipulated the motor speed of the e-puck. Using neuroevolution, the input weights into the RNN, the recurrent weights within the RNN, and the output weights from the RNN were evolved based on an objective designed to replicate the rat behavior.
Our results show that the evolved RNN was capable of guiding the robot through the triple-T maze with similar behavior to that observed in the rat. Our analysis of the RNN activity indicated that the behavior was not dependent on any one sensory projection type but rather relied on the evolved RNN dynamics. Furthermore, the population of neurons in the RNN were not only sufficient to predict the robot's current location but also carried a predictive code of future intended reward paths. Furthermore, the present method for evolving neural networks for robot controllers may be applicable to other memory tasks.
\section{Methods}
\label{method}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{images/full_maze.PNG}
\caption{The maze visualization in Webots. As shown in the figure, there were corridors that the robot could traverse and landmarks on the wall. The e-puck robot is denoted by the small green circle. The red circles denote the reward locations. Note that these rewards were not visible to the robot's sensors.}
\label{fig:landmarks}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.25\textwidth]{images/epuck.png}
\caption{The 3D simulated e-puck robot. The picture is adapted from \citet{simulatedepuck}.}
\label{fig:e-puck}
\end{figure}
We picked Webots \cite{michel2004cyberbotics} as our virtual robotic environment. Inside this 3D simulator, a triple T-maze was constructed that closely followed the dimensions and landmarks used in the rat experiment \cite{olson2017subiculum}. Figure \ref{fig:landmarks} shows the maze simulation environment. The red circles, which denote the location of the rewards, were not observable by the robot and are only included in the figure for illustrative purposes. The agent was an e-puck robot \cite{mondada09theepuck} which has an accelerometer, a front camera, 8-direction proximity sensors, several LEDs, and 2 wheel motors (see Figure \ref{fig:e-puck}). The e-puck needed to learn by neuroevolution to find four rewards (and return home after each reward visit) with minimal repeats before timeout. Its actuation was updated every 64 milliseconds. The timeout threshold was tuned to 5000 steps (i.e., 320 seconds in real-time) per trial to guarantee enough time to visit all four rewards with minimal repeats and some tolerance for slight movement variations. For each trial, the robot would always start from the home position in the bottom middle part of the maze and move upward (see the e-puck's location in Figure \ref{fig:maze}). After reaching a T-intersection, a door behind the robot would close to prevent backtracking. Since the robot could only move forward in a reward path, it could neither revisit places on the same path nor switch to a different one before completion of the previous path. After the robot moved within 6 cm from the center position of a novel reward in a trial, the reward was added to the objective function given by Equation \ref{eq:fitness1}. After the robot passed through the third T-intersection right above any reward position, an additional door on the right/left would close to enforce its usage of the closer return path to home before exploring the next reward path. It should be noted that although the doors prevented backtracking, the robot still needed to evolve its ability to move smoothly and efficiently through the corridors to receive all four rewards with minimal repeats prior to the timeout.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{images/maze.jpg}
\caption{A closer look of the maze, with 4 rewards labeled in red circles and 7 T-intersections labeled in blue font. Home was located at the e-puck's current position (bottom-middle) in this figure. }
\label{fig:maze}
\end{figure}
The rotation speed of an e-puck was ranged between -3.14 rad/s and 6.28 rad/s. The evolved output weights in the RNN adjusted the speed and the turning rate of the robot to navigate the task with optimal and stable performance. To prevent from being stuck at corners or T-intersections, the robot had a default obstacle avoidance algorithm that used the 8 proximity sensors, which only influenced movement when the robot was very close to an obstacle, to move away from the closest point of contact \cite{fajen2003behavioral}. The obstacle avoidance motor signal was added to the rotational speed of the motors dictated by the RNN output.
We conducted 5 evolutionary runs and selected the best performing agent from each run for further analysis. An evolutionary run was composed of 200 generations to achieve optimal performance. During each generation, there were 50 genotypes generated according to the evolutionary algorithm described in Section \ref{sec:ea}. For each genotype during the evolutionary process, the fitness value was recorded as an average over 5 trials to improve robustness in the selection. In each test scenario afterwards, each of the 5 best performing agents from these runs was utilized to run 20 demo trials with the same task setting and timeout threshold. For each demo trial, the activities of 50 recurrent neurons and the robot positions were recorded at each time step for further analysis.
\subsection{Network Architecture}
\label{sec:architecture}
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{images/network.png}
\caption{The neural network architecture for controlling the e-puck robot in Webots. Sensors were converted into input neural activities. The input weights ($W_{xr}$), recurrent weights ($W_{rr}$), and output weights ($W_{ry}$) were evolved concurrently. The output weights dictated the left and right rotational wheel speed of the e-puck.}
\label{fig:architecture}
\end{figure*}
The neural network architecture received inputs from the e-puck's 8-direction proximity sensors, 3D linear accelerometer values, and normalized pixel values from its $10\times 8$ grayscale camera frame (Figure \ref{fig:architecture}). These 91 input neurons were fully connected to 50 recurrent neurons, which were fully connected with one another. This recurrent layer was then fully connected with the two neurons in the output layer that controlled the rotational speed of the two wheel motors separately.
\subsubsection{Recurrent neural network}
For each recurrent neuron $i$, its recurrent activity $R_i$ was updated at every time step $t$ according to Equation \ref{eq:recur}:
\begin{equation}
\left.\begin{aligned}
& R_i^{(t=0)} = 0.0,\\
& \text{synIn}_i^{t} = \sum_{k \in in} W_{ki}\cdot x_k^{t} + \sum_{j \in rec}^{j\neq i} W_{ji}\cdot R_j^{(t-1)},\\
& R_i^{t} = (1-p)\cdot \tanh{\left(\text{synIn}_i^{t} \right)} + p \cdot R_i^{(t-1)}
\end{aligned}
\right\}
\label{eq:recur}
\end{equation}
Here $x$ denoted the input sensor value and $W_{ki}$ was the weight from Neuron $k$ to Neuron $i$. In other words, the synaptic input for each recurrent neuron ($\text{synIn}_i$) contained (1) the summation of the product of each sensor value and the corresponding input weight plus (2) the summation of the product of every other recurrent neuron's previous activity and the corresponding recurrent weight. The $\tanh$ wrap ensured the recurrent activity between -1.0 and 1.0. A small $p$ value of 0.01 helped to avoid any abnormal performance of the computed recurrent activity. The recurrent activity was then used to compute the rotational speed of each wheel motor by multiplying with the output weight.
\subsubsection{Evolutionary algorithm}
\label{sec:ea}
An evolutionary algorithm was used to evolve the input, recurrent, and output weights ($W_{xr}$, $W_{rr}$, and $W_{ry}$ in Figure \ref{fig:architecture}). Because of all-to-all connections, there were a total of 7150 genes for each genotype. The fitness value of each genotype was the average over 5 trials. The evolutionary algorithm used a population of 50 genotypes selected by linear ranking. Two-point crossover and mutation (with a decaying mutation standard deviation) were applied to reproduce the non-elite 90\% of the population. The mutation rate was 0.06 and the mutation standard deviation decreased throughout a run via the function:
\begin{equation}
\text{mutation\_std} = 0.3\times \frac{50.0}{50.0+m}
\label{eq:mutstd}
\end{equation}
Here $m$ denotes the generation index.
The fitness function is shown in Equation \ref{eq:fitness1}:
\begin{equation}
\text{fitness} = \text{ num\_obtainedRwds} + \text{portion\_routes\_completed} {-} 0.2 \times \text{num\_repeats}
\label{eq:fitness1}
\end{equation}
During each trial, we would reward (1) each non-repetitive visit of any reward arm (i.e., $0\sim 4$) and (2) the portion of reward path visits for which home was returned afterwards (i.e., $0\sim 1$); meanwhile, we would penalize every repeated visit.
\subsection{Bin-based Recurrent Activities}
\label{sec:method-bin-analysis}
To analyze the performance of the robot after evolution, we divided the 1.6m-by-1.25m maze into 0.08m-by-0.10m sized bins, which was close to the e-puck’s diameter (0.074m) (Figure \ref{fig:bins}). We computed the average activity of each recurrent neuron for each bin in the maze. For each of the 5 best performing agents (from 5 evolutionary runs), we used 15 demo trials to generate the expected bin-based activity matrix and 5 demo trials for bin occupancy prediction. The results are shown in Sections \ref{sec:results_spatial} and \ref{sec:results_prospective}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/maze_bin_w_segments}
\caption{The maze layout with 110 bins of size 0.08m-by-0.10m. Segments used for the trajectory-dependent coding analysis are labeled as seg1, seg3-1, seg3-2, seg8-1 and seg8-2 in yellow. }
\label{fig:bin_segments}
\end{figure}
\subsubsection{Spatial memory analysis}
We used the RNN activity to predict the robot's location in the maze. Since there were 50 recurrent neurons, for each demo trial, we generated a bin-based recurrent activity matrix of size $110\times50$. The activity at each bin for each neuron was averaged over all steps spent on that bin in a trial. Then we obtained an expected bin-based activity matrix of size $110\times50$ by taking the average over all the matrices for the first 15 demo trials. The remaining 5 trials were used to analyze the RNN's ability to encode location. For each of these 5 test trials, we compared each bin's RNN activity vector, which had a length of 50, with all 110 expected bin-based activity vectors using a Euclidean distance metric. The predicted bin was the smallest Euclidean distance to that actual bin. Thus, the Euclidean distance denotes the prediction error in bins. For example, if the Euclidean distance was 6, there was a perfect prediction error of 6 bins or approximately 0.5m (see Figure \ref{fig:bin_occupancy}).
\subsubsection{Trajectory-dependent analysis}
After traversing some of the maze's vertical (South-to-North) segments, the robot would decide to turn left or right at a T-intersection. We analyzed if the RNN activity during traversal of a vertical segment could predict the robot's future path or the robot's prior path. As shown in Figure \ref{fig:bin_segments}, Segment 1 is the vertical segment right before the first T-intersection on the path for any of the four rewards. Segment 3-1 (or Segment 3-2) denotes the vertical segment right before the second T-intersection on the path for Reward 1 or 2 (or for Reward 3 or 4). Segment 8-1 (or Segment 8-2) represents the vertical segment for returning from Reward 1 or 2 (or from Reward 3 or 4). Similar to the method described above, we computed an expected bin-based activity matrix for each of these segments from the first 15 demo trials for each reward path. We then used the remaining 5 demo trials to test whether the RNN activity could predict which path the robot was taking (i.e., \textit{Prospective}; seg1, seg3-1, seg3-2) or which path the robot was returning from (i.e., \textit{Retrospective}; seg8-1 and seg8-2).
\section{Results}
\subsection{Evolutionary Runs}
\label{sec:results_evolve}
\begin{figure}
\centering
\includegraphics[width=0.95\textwidth]{images/evolved_trend}
\caption{The evolutionary performance (left: fitness, right: number of elapsed time steps) for the best-so-far agent evolved in each generation. Each subplot was averaged over 5 runs with 200 generations per run. The shaded area denotes the 70\% confidence level. }
\label{fig:evolved_trend}
\end{figure}
Successful behavior, similar to that observed in rats, emerged from the evolutionary process. We ran all simulations using Webots (version R2020a) on a desktop with one GPU (Nvidia GeForce GTX 1080 Ti). Figure \ref{fig:evolved_trend} shows the best-so-far evolutionary performance and the number of elapsed steps for five runs. Each run lasted 200 generations, with 50 genotypes per generation. In each generation, the fitness value of a genotype was averaged over 5 trials. By the end of each run, the best-so-far fitness curve reached a plateau close to the maximal fitness value of 5, whereas the number of steps taken to complete the task dropped below 3500 steps per trial on average. An example perfect trial trajectory (with no repeated visits of any reward path and a fitness of 5) can be observed in Figure \ref{fig:demo_trajectory}.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{images/demo_trajectory}
\caption{The trajectory of a perfect trial, which covered reward paths 1, 4, 3, 2 in order with no repeated visits of any path. }
\label{fig:demo_trajectory}
\end{figure}
\subsection{Ablation Performance}
\label{sec:results_ablation}
\begin{table}
\caption{Ablation performance (mean $\pm$ the 95\% confidence level) over 20 trials per ablation test for the best performing agent in each of the 5 evolutionary runs. The values highlighted with bold fonts and asterisks denote ablations that had a significant impact on the performance. Significance threshold was a p-value < 0.01/6 = 0.0017 using the Wilcoxon Rank Sum test.}
\label{tab:ablation}
\centering
\begin{tabular}{l|l|l}
\toprule
& Fitness & Elapsed Steps\\
\midrule
No Ablation & 3.65$\pm$0.14 & 4644$\pm$169 \\
Proximity Sensors & 3.29$\pm$0.16 & 4774$\pm$128 \\
Linear accelerometer & 3.49$\pm$0.16 & 4640$\pm$131 \\
Grayscale Vision & 3.21$\pm$0.17 & 4785$\pm$67 \\
\textbf{Input Weights} & 2.93$\pm$0.085 & \textbf{4981$\pm$16$\ast$} \\
\textbf{Recurrent Weights} & \textbf{2.95$\pm$0.11$\ast$} & \textbf{4946$\pm$43$\ast$} \\
\textbf{Output Weights} & \textbf{2.84$\pm$0.10$\ast$} & \textbf{4978$\pm$22$\ast$} \\
\bottomrule
\end{tabular}
\end{table}
We carried out a set of ablation simulations to test whether performance was dependent on any sensory projection type or just evolved weights in the neural network (Figure \ref{fig:architecture}). To test this, we either shuffled different sensor input values or shuffled the RNN input ($W_{xr}$,) recurrent ($W_{rr}$), or output ($W_{ry}$) weights. For each of the 6 ablations, we ran the best performing agent from each of 5 evolutionary runs in demo trials. The results were averaged over 20 demo trials for each shuffle test. Random shuffle sequences occurred at each time step for each demo trial.
The ablation studies show that the dynamics of the RNN was critical for performance (Table \ref{tab:ablation}). We compared the control (no ablation) with the 6 ablation groups. Since there were 6 comparisons, the significance threshold for the p-value is 0.01/6 = 0.0017 based on a Bonferroni correction. Interestingly, none of the sensory projection ablations had a significant impact on performance. However, ablating the evolved weights (input, recurrent, and output) all had a significant impact. That suggests that it was the recurrent neural network dynamics that were key to performance. Moreover, that performance was not dependent on any one sensory projection type.
\subsection{Order Effects}
\label{sec:results_order}
We wanted to test if the robot evolved strategies to solve the triple-T maze task. Rats tend to show idiosyncratic behavior in the same maze setting \cite{olson2020secondary,olson2021complementary}. For instance, a rat alternated by going to the left side of the maze towards Rewards 1 and 2, and then the right side of the maze towards Rewards 3 and 4. Idiosyncratic behavior did emerge in our evolved robots. Although we did not observe the robots alternating between sides of the maze, each best performing agent for an evolutionary run exhibited a unique strategy for traversing the maze. Figure \ref{fig:transition} shows the probability of transitioning from one reward path to the next. Only transitions probabilities that are greater than 0.33 are shown. We did find some generalities between genotypes. For example, the agents evolved in Genotypes 2, 3, and 5 tended to transition from the path for Reward 4 to the path for Reward 1. The agents evolved in Genotypes 2, 3, and 4 tended to transition from the Reward 3 path to the Reward 1 path. In both cases, the robot was navigating the right side of the maze before transitioning to the left side of the maze. These strategies that emerged in our robot and in the rat may simplify the task by breaking down the problem into chunks (e.g., first go to the right, and then go to the left).
\begin{figure}
\centering
\includegraphics[width=0.68\textwidth]{images/transition.png}
\caption{The trend of transitioning from one reward path to the next was different for each best performing agent (with a different genotype evolved) after 200 generations for each of the five evolutionary runs. The numbered circles denote the reward path, and the labeled arrows denote the probability of transitioning from one reward path to another.}
\label{fig:transition}
\end{figure}
\subsection{Spatial Coding in the RNN}
\label{sec:results_spatial}
We investigated if the RNN activity was sufficient to predict the robot's position. If the activities of the 50 recurrent neurons could accurately encode the position, then the robot might be using this piece of information to solve the maze task. Borrowing techniques from neuroscience \cite{olson2017subiculum, Wilson1993, ferbinteanu2003prospective}, we tested whether the RNN contained spatial information with a population code.
Individual neurons in the recurrent layer did not seem to have place information. For example, Figure \ref{fig:bins} shows the average bin-based recurrent activities across the entire maze from 15 demo trials of a top performing agent. Each bin's activity per trial divided by the number of time steps spent on that bin. The activity of each neuron is noisy with some neurons being highly active, quiescent, or oscillating.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{images/bin_recurrent_activities}
\caption{The average bin-based activities for all 50 recurrent neurons on 110 bins across the entire maze for a top performing agent.}
\label{fig:bins}
\end{figure}
However, the population of 50 RNN neurons was able to predict the robot's location throughout the maze. With the method described in Section \ref{sec:method-bin-analysis}, Figure \ref{fig:bin_occupancy} shows the location prediction for each bin in all 25 test trials (with 5 best performing agents from 5 evolutionary runs and 5 trials per agent). It is apparent from the figure that the RNN activity was sufficient to predict the robot's position in the maze. The robot's position was predicted with perfect accuracy on 58\% of the bins, and the predicted error had an average distance of 3.1 bins (i.e., 0.25 meters).
\begin{figure}
\centering
\includegraphics[width=0.68\textwidth]{images/avg_bin_occupancy_color}
\caption{The average predicted bin occupancy over all 25 test trials of the best performing agent from each of the five evolutionary runs. A dark blue bin (if existing) would have perfect prediction right at itself (distance = 0), whereas a dark red bin (if existing) would have a farthest prediction (across the diagonal of the entire maze). }
\label{fig:bin_occupancy}
\end{figure}
\subsection{Trajectory-Dependent Coding in the RNN}
\label{sec:results_prospective}
Solving the triple-T maze task requires the agent, whether it is a robot or an animal, to remember which path it has already taken, as well as to decide which path to take next. We hypothesized that the dynamic activity of the RNN carried such information, which is known as retrospective coding (i.e., where it has been) and prospective coding (i.e., where it intends to go).
To test whether there was retrospective coding and/or prospective coding in the RNNs, we analyzed the RNN's ability to encode trajectory-dependent information at a population level. Table \ref{tab:prospective} shows how well the RNN could predict the robot's future path based on the activity of Segments 1, 3-1 and 3-2, and how well the RNN could predict the robot's past path based on the activity of Segments 8-1 and 8-2. The probability of correct path prediction on Segment 1, where the robot could take one of 4 paths, was well above chance level (t-test; p < 0.0001). The correctness on Segment 3-1 or 3-2, where the robot could take one of two paths, was also well above chance level (t-test; p < 0.0001 for Segment 3-1 and p < 0.005 for Segment 3-2). This suggests that the RNN carried a prospective code of where the robot intent to go next. The probabilities of correct path prediction on Segments 8-1 and 8-2 were not above chance (t-test > 0.05), which indicates they did not predict if the robot came from one of the 2 prior paths.
Taken together, these results suggest that the evolved RNN had prospective information in that it could be discerned which direction the robot would take before turning left and right. It is interesting that we were not able to observe retrospective information in the RNN population since some knowledge of where the robot had already visited was necessary for the observed performance.
\begin{table}
\caption{Average prospective path prediction for different segments. }
\label{tab:prospective}
\centering
\begin{tabular}{l|ccccc}
\toprule
&Seg1 & Seg3-1 & Seg3-2 & Seg8-1 & Seg8-2\\
\midrule
correctness & 41\% & 77\% & 69\% & 47\% & 55\% \\
bins off & 0.9 & 0.2 & 0.2 & 2 & 2 \\
\bottomrule
\end{tabular}
\end{table}
\section{Discussion}
The present work demonstrated that a robot controlled by an evolved RNN could solve a spatial and working memory task where the robot needed to navigate a maze and remember not to repeat paths it had taken previously. The RNN population activity carried spatial information sufficient to localize robot, and the RNN population activity carried predictive information of which path robot intended on taking.
Behavior was dependent on RNN dynamics and not any particular sensory channel. The present method shows that complex robot behavior, using a detailed robot simulation, could be realized by evolving all weights of a RNN.
\subsection{Evolved RNN with Spatial and Working Memory}
We evolved a RNN to control a robot in a spatial and working memory task that replicated behavior and neural activities observed in rats \cite{olson2017subiculum}. The robot was able to navigate the triple T-maze efficiently by reaching all four rewards with minimal repeats. Successful performance required the robot to have spatial knowledge and working memory of which rewards it had already visited. Prospective information, in which RNN activity predicted the robot's intention, emerged in the simulations. Although not observed in the analysis, the neural network must have had retrospective information to minimize repeating previously traversed paths.
At the population level, the evolved RNN activity has similar characteristics as the hippocampus. It has been observed that the population activity of the CA1 region in the hippocampus can accurately predict the rat's location in a maze \cite{Wilson1993,olson2021complementary}. Furthermore, journey-dependent CA1 neurons have been observed in the rat that can predict the upcoming navigational decision \cite{ferbinteanu2003prospective}. Similar to CA1, the RNN received speed, direction, and visual information as input, and combined these types of sensory information to construct a journey-dependent place code \cite{taube1990head,sargolini2006conjunctive,kropff2015speed,o1976place,hafting2005microstructure,sun2019ca1,potvin2007contributions,frost2020anterior}.
In \citet{olson2021complementary} and in other rodent studies, it has been observed that rats acquire individual strategies to solve navigational tasks. For example, in the triple-T task, many rats alternated between the left and right arms of the maze. Similarly, our robot demonstrated this idiosyncratic behavior. Each best performing agent for an evolutionary run had an order-dependent pattern for taking different paths in the maze. It suggests that the RNN evolved a strategy breaking down the complex maze task into simpler pieces, which may also be how animals solve tough problems.
\subsection{Behavioral Dependence on RNN Dynamics}
To investigate the dependence of the robot behavior on different components of the RNN neural architecture, we conducted an ablation study (see Figure \ref{fig:architecture}). Results in Section \ref{sec:results_ablation} show that ablating a given sensory channel had no significant impact on performance. However, ablating the evolved weights (input, recurrent, and output) all had a significant impact. This suggests that it was the recurrent neural network dynamics that were key to performance, and that performance was not dependent on any one sensory projection type. The results justify evolving all weights in the RNN structure, rather than evolving only the readout weights as is often done in Liquid State Machines (LSM) or Echo State Networks \cite{maass2002real}. Furthermore, the present results demonstrate the potential of extending our RNN controller to other types of robots that use different sensory inputs.
\subsection{Capability of Evolving Complex Robotics Behavior}
One advantage of our method is the simple design of our fitness function. It only includes rewards for each visit of a novel reward and the completeness of each reward path plus a small penalty for repeated visits. We also tried with an additional reward term for the portion of time steps left before timeout, but finally excluded it from the fitness function. Instead, the time cost would be automatically influenced by rewarding non-repetitive reward path visits as demonstrated in Figure \ref{fig:evolved_trend}. It is not a fitness function only for a certain type of robot, because it is independent of robot properties (e.g., sensors, speed, motor structure, etc.). Therefore, it could be easily generalize to fit other task settings or robots.
The key to performance in our experiments was the evolved weights into, within, and from the RNN. Generally a RNN has connections between internal neurons form a directed cycle. Arbitrary sequences of inputs could be processed by using the internal state of the RNN as the memory. RNNs have been applied to a broad range of domains such as terrain classification, motion prediction, and speech recognition \cite{zou2020terrain,kashyap2018recurrent,graves2013speech}. A reservoir-based approach, such as an LSM \cite{maass2002real}, can tractably harness such recurrence.
Similar to an LSM, we also tried a reservoir-based approach to evolve only the output weights while keeping randomly initialized input and recurrent weights fixed throughout generations. In addition, we attempted to evolve both output weights along with input or recurrent weights (but not with both). However, their evolutionary performance could not generate fitness values as well as evolving all three types of weights. Evolving all three types of weights allow maximal utilization of neural activities to create the dynamics needed to solve a sequential memory task, such as the triple-T maze one. This process of evolving input, recurrent, and output weights could be readily transferred to other complex robotic settings.
Because of the accurate representation of many popular robot designs in the Webots simulator, we could always run much faster than real-time (e.g., $\times30$ faster on average on our desktop with one GPU) to evolve for enough generations before applying a well-performed genotype to the RNN controller for real-world robotic navigation tasks. The power of using a detailed simulator, such as Webots, is that the evolved controller should transfer to the real e-puck with minimal adjustments.
\subsection{Comparison with Prior Work}
\subsubsection{Evolutionary robotics}
Evolutionary robotics is a method for building control system components or the morphology of a robot \cite{bongard2013evolutionary,nolfi2016evolutionary}. The biological inspiration behind this field is Darwin's theory of evolution, which was constructed with three principles. (1) \textit{Natural Selection}: genotypes that can be well adapted to their environments are more likely to survive and reproduce. (2) \textit{Heredity}: the fittest genotypes from the previous generation can be directly kept in the next generation; moreover, the new offspring in each generation is generated based on the selected genes from two parents. (3) \textit{Variation}: the new offspring goes through mutations and crossover with a certain probability and thus differs from both parents.
Our method follows these three principles and falls into the common category of evolving the control system. However, what we evolved is novel compared to other work on evolutionary robotics. There has been work to evolve robots in cognitive tasks. For example, one group evolved virtual iCub humanoid robots to investigate the spontaneous emergence of emotions. Their populations were evolved to decide whether to ``keep'' or ``discard'' different visual stimuli \cite{pacella2017basic}. There has also been work on evolving robot controllers capable of navigating mazes. For example, Floreano's group evolved neural network controllers to navigate a maze without colliding into wall. Their neural networks were directly tested on a Khepera robot that had proximity sensors and two wheel motor system that was similar to the e-puck used in our present studies. Their evolved neural networks developed a direct mapping from the proximity sensors to the motors \cite{floreano2010evolution}. Our present work extends this prior work by evolving an RNN capable of navigating mazes, as well as demonstrating cognitive behavior.
Rather than evolving a direct mapping from sensors to motors, we instead evolved the weights from sensory inputs to the RNN, within the RNN, and from the RNN to the robot's motors as the genes for each genotype evolved in our network. As is discussed below, RNNs such as the ones we discuss here are neurobologically plausible and allow for comparisons with neuroscience and cognitive science data \cite{wang2018prefrontal, yang2019task}. Moreover, the RNN architecture is more generalizable to different types of robots working in complex scenarios (e.g., the triple T-maze with multiple rewards and landmarks), which results in optimal performance independent of any projection type.
\subsubsection{Evolved RNNs}
Although there are only a few studies that have RNNs evolved directly in robotic experiments, evolving RNNs has been more frequently applied to virtual task settings. For example, \citet{akinci2019evolving} used either a steady-state genetic algorithm (SSGA) or an evolutionary strategy (ES) to evolve weights of the Long Short-Term-Memory (LSTM) network or RNN for the Lunar Lander game provided by the OpenAI gym \cite{brockman2016openai}. In their case, the ES developed more dynamic behavior than the SSGA, whereas the SSGA kept good genotypes and re-evaluated them with different configurations.
\citet{li2018opponent} evolved poker agents called ASHE with various types of evolutionary focus, such as learning diversified strategies from strong opponents, learning weakness from less competitive ones, learning opposite strategies at the same time, or a mix in-between. The genes in their GA covered all parameters in the estimators, including the LSTM weights, per-block initial states, and the estimator weights.
A more biologically inspired example is related to the recent work by \citet{wieser2020eo}. Inspired by the neuroplasticity and functional hierarchies in the human neocortex, they proposed to use a network called EO-MTRNN to optimize neural timescales and restructure itself when training data underwent significant changes over time.
All these related works have their unique perspectives that could inspire us to build a more robust and potentially faster evolutionary process for RNN systems in the future. For instance, we may consider to experiment with features in different evolutionary algorithms or co-evolve different neural regions which have different strategies or focus on a cognitive task.
\subsubsection{Comparison with NEAT/HyperNEAT robot controllers}
The evolutionary algorithms were utilized to evolve only weights in our RNN system and many other groups' work as mentioned earlier. Another popular evolutionary mechanism is NEAT, which has its unique feature of evolving the network topology together with the weights \cite{stanley2002evolving}. HyperNEAT extends NEAT by evolving connective CPPNs that generate patterns with regularities (e.g., symmetry, repetition, repetition with variation, etc.) \cite{stanley2009hypercube}. In the case of quadruped locomotion investigated by \citet{clune2009evolving}, HyperNEAT could evolve common gaits by exploiting the geometry to generate front-back, left-right, or diagonal symmetries. Our current model was tuned to have a fixed number of recurrent neurons since the first generation and have all-to-all connections between two layers or within the recurrent layer. It may be of interest to combine NEAT/HyperNEAT with topologies of RNNs to solve more complex problems and scenarios. For example, NEAT might be utilized in the beginning of the evolutionary process to efficiently derive a morphology for a more standard evolutionary algorithm to use in later generations. This hybrid approach has similarities to \citet{akinci2019evolving}.
\subsubsection{Working memory}
Our evolved RNN encoded not only spatial information but also working memory to remember which paths had been traversed recently and which paths remained to be explored. Working memory helps to connect what happened earlier with what occurs later. It can be thought of as a general purpose memory system that can generalize, integrate and reason over information related to decision making or executive control \cite{diamond2013executive, vyas2020computation}. For example, \citet{yang2019task} trained single RNNs to perform 20 tasks simultaneously. Clustering of recurrent units emerged in their compositional task representation. Similar to biological neural circuits, their system could adapt to one task based on combined instructions for other tasks. Furthermore, individual units in their network exhibited different selectivity in various tasks.
Working memory usually relies on the prefrontal cortex (PFC) for information maintenance and manipulation \cite{baddeley1994developments, eldreth2006evidence, smith1999storage}. \citet{wang2018prefrontal} investigated such brain functioning with a meta-reinforcement learning (meta-RL) system. Their model trained the weights of an RNN centered on PFC through a reward prediction error signal driven by dopamine (DA). This RNN ``learned to learn'', which means it had the ability to learn new tasks via its trained activation dynamics with no further tuning of its connection weights.
With further investigation and utilization of working memory, we also would like to have our evolved RNN generalize over multiple cognitive tasks and demonstrate cognitive functions observed in different brain regions.
\section{Conclusions}
In this paper, we introduced a recurrent neural network (RNN) model that linked the robot sensor values to its motor speed output. By evolving weights from sensory inputs to the RNN, within the RNN, and from the RNN to the robot's motors, the evolved network architecture achieved the goal of successfully performing a cognitive task that required spatial and working memory. The RNN population carried spatial information sufficient to localize robot in the triple T-maze. It also carried predictive information of which path robot intended on taking. Moreover, the robotic behavior was dependent on RNN dynamics rather than a sensor-to-motor mapping. Our method shows that complex robot behavior, similar to which being observed in animal models, can be evolved and realized in RNNs.
\section*{Acknowledgment}
This material is based upon work supported by the United States Air Force Award \#FA9550-19-1-0306. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the United States Air Force.
\bibliographystyle{plainnat}
|
\section{Introduction}\label{sec:introduction}
\input{sections/introduction.tex}
\section{Related Work}\label{sec:relatedwork}
\input{sections/relatedwork.tex}
\section{Problem Setup}
\input{sections/problemstatement.tex}
\input{sections/method.tex}
\section{Analysis}\label{sec:analysis}
\input{sections/proofs.tex}
\section{Evaluation}\label{sec:evaluation}
We have implemented a proof-of-concept simulation of our proposed approach, as shown in~\cref{fig:pushing,fig:cluster,fig:flowdirections}.
In our supplementary material, we include a video demonstrating how robots move, resolve contention for space, and push in packed environments when abiding by our motion rules.
Each robot in our simulation creates and follows their own plan without knowledge of other robots' planning details.
We show examples of several possible polygonal environments partitioned as per~\cref{sec:method.spacetypes}.
\section{Conclusion}\label{sec:conclusion}
In this work we present a method for ensuring progress for every robot in a shared multi-robot space via a \emph{motion pact}.
Our motion pact is comprised of a partitioning of the environment freespace along with a set of motion rules that allow robots to plan independently and flexibly, with protection from being blocked (either intentionally or inadvertently) by other robots' behavior.
We prove that the rules guarantee progress for every robot, and demonstrate a proof-of-concept implementation of our system.
This work assumes that all robots are holonomic and capable of instantaneous bounded acceleration.
Beyond relaxing this assumption, this work leaves several exciting open questions.
For example, one could consider re-imagining the form of goals to encompass team-based tasks, where multiple robots must collaborate to satisfy their objectives rather than individually reach a location. With such an expansion, robots can work in teams, or even switch between teams, to more flexibly move and work in a shared space.
\appendices
\renewcommand{\thesectiondis}[2]{\Alph{section}:}
\input{sections/appendixa.tex}
\input{sections/appendixb.tex}
\renewcommand*{\bibfont}{\small}
\printbibliography{}
\end{document}
\section{Pushing Logic}\label[app]{appendix:A}
We now describe, in \cref{algo:bkwd}, the logic for computing the pushing angles and velocity vectors for each robot in the set \ensuremath{\mathcal{R}}{} at a timestep $t$.
\Cref{algo:bkwd} depends on the following subroutines:
\begin{itemize}
\item \texttt{activeSR(\ensuremath{r_\idx})}: Returns the first non-null space request for \ensuremath{r_\idx}, in the order $T_f$, $S_f$, $ T_p$, $S_p$.
\item \texttt{adoptpriority($r_j$, \ensuremath{r_\idx}{})}: Copies the timestamp $\tau$ from \ensuremath{r_\idx}'s priority and the identity $r_p$ of the original pushing robot for \ensuremath{r_\idx}{} to $r_j$.
\item \texttt{arc($push_i$)}: Returns the set of angles within $\pm\frac{\pi}{2}$ radians of $push_i$.
\item \texttt{overlaps($d_1, d_2$)}: Returns \texttt{true} iff the discs $d_1$ and $d_2$ intersect. If called with robots $r_i, r_j$ as arguments, $d_1$ and $d_2$ are the \ensuremath{B_{\ensuremath{\rho}}}-sized discs centered at the positions of $r_i$ and $r_j$, respectively.
\end{itemize}
\begin{algorithm}[h]
\SetAlgoLined
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\SetKwProg{Initialization}{Initialization}{}{}
\SetKw{Continue}{continue}
\SetKw{And}{and}
\DontPrintSemicolon
\While{any $r_i\in\ensuremath{\mathcal{R}}$'s velocity changes}{
\For{$\ensuremath{r_\idx}, r_j\in \ensuremath{\mathcal{R}}{} \times \ensuremath{\mathcal{R}}{}$}{
\lIf{$\ensuremath{L}_i < \ensuremath{L}_j$ or $I_{p_i} = \ensuremath{\ell}_j$}{
\Continue{}
}
\ElseIf{$priority(r_i) = priority(r_j)$ \And $(r_j(t) - \ensuremath{\robot(t)}) \in arc(push_j)$}{
$push_j = \frac{push_i + push_j}{2}$\;
\Continue{}
}
\lIf{$overlaps(\ensuremath{\robot(t)}, r_j(t))$}{ $p = \ensuremath{\robot(t)}$ }
\If{overlaps(activeSR(\ensuremath{r_\idx}{}),\ensuremath{r_\idx}{}) \And overlaps(activeSR(\ensuremath{r_\idx}{}), $r_j(t)$) }{
$p = centroid(activeSR(\ensuremath{r_\idx}{}))\;$
}
$\Theta = cos^{-1}(\frac{push_i \cdot (r_j(t)-p)}{\|push_i\| \|r_j(t)-p \|})$\label{line:theta}\;
\If{$\Theta \in arc(push_i)$}{
$adoptpriority(r_j, \ensuremath{r_\idx}{})$\;
\lIf{$r_j \in \ensuremath{\Phi}$}{$v_j = d_{r_j(t)}$}
\lElse{$v_j= r_j(t)-\ensuremath{\robot(t)}$}
}
}
}
\caption{Computing pushing vectors}
\label{algo:bkwd}
\end{algorithm}
At a high level, \cref{algo:bkwd} iterates through each pair of robots $r_i, r_j \in \ensuremath{\mathcal{R}}$ checking for each $r_j$ that is close enough to a higher-priority $r_i$, or the active space request of a higher-priority $r_i$, for pushing to engage.
We compute the angle between the velocity vector $push_i$ of $r_i$ and the vector pointing from $r_i$ (or its active space request) to $r_j$ (\cref{line:theta}), and, if this angle is within the acceptable $\pm\frac{\pi}{2}$ range of the original pushing vector, propagate the pushing priority of $r_i$ to $r_j$ and set $r_j$'s velocity vector to move out of the way.
This process repeats until no robot's velocity vector changes, at which time the robots in \ensuremath{\mathcal{R}}{} move accordingly.
\subsection{Warehouse Robotics}
\section{Region Generation}\label[app]{appendix:B}
As described in \cref{sec:method.spacetypes}, our approach requires partitioning the free space \ensuremath{\mathcal{F}}{} of a bounded environment \ensuremath{\mathcal{E}}{} into three sets: the open space \ensuremath{\Omega}{}, the flow space \ensuremath{\Phi}{}, and the passage region space \ensuremath{\Psi}.
In this section, we describe constructing such a partition.
We will refer to the set of all obstacles and the internal perimeter of the environment as \emph{flow generating objects}.
Our algorithm proceeds through the following high-level steps (with further detail in corresponding subsections):
\begin{enumerate}
\item Construct \emph{initial flow regions} around each flow generating object to capture a space wide enough for a line of robots around each object's perimeter (\cref{sec:initial.flow.regions}).
\item Separately construct \emph{inflated flow regions} around each flow generating object (\cref{sec:initial.flow.regions}).
These regions capture subsets of the free space too narrow to hold three robots side by side.
\begin{enumerate}
\item In areas where two inflated flow regions overlap, split the region lengthwise between corresponding flow generating objects.
\end{enumerate}
\item Construct \emph{single-lane regions} from areas where two initial flow regions overlap (\cref{sec:single.lane.regions}) to capture flow space areas too narrow for two robots side by side.
\item Place \emph{passage regions} (\cref{sec:pasage.regions}) to connect different flow regions in \ensuremath{\Phi}{}.
\item Smooth the perimeter of open space (\cref{sec:smooth}).
\item Assign directions to each flow region (\cref{sec:direction.assignment}).
\item Return the flow regions as \ensuremath{\Phi}, the passage regions as \ensuremath{\Psi}, and the remainder of the free space as \ensuremath{\Omega}
\end{enumerate}
\newcommand{\appsec}[1]{\refstepcounter{subsection}\noindent\textbf{\Alph{subsection}. #1}}
\appsec{Initial Flow Regions:}\label[app]{sec:initial.flow.regions}
The initial flow regions are constructed by taking the Minkowski sum of each object $O_i$ and a \ensuremath{B_{\ensuremath{\rho}}}-sized disc.
Each flow region is $f_i = (O_i \oplus \ensuremath{B_{\ensuremath{\rho}}}) \setminus O_i$.
If any $f_i$ intersects itself or a flow generating object, the environment is invalid and we return an error.
\appsec{Inflated Flow Regions:}\label[app]{sec:inflated.flow.regions}
The inflated flow regions are constructed the same as the initial flow regions, but using a $1.5 * \ensuremath{B_{\ensuremath{\rho}}}$-sized disc.
If two inflated flow regions overlap, generating initial flow regions without modification creates a subset of \ensuremath{\Omega}{} too narrow to hold a robot.
Instead, we split the inflated region overlaps along their central axis and add each divided-up area onto their corresponding initial flow regions.
\appsec{Single-lane Regions:}\label[app]{sec:single.lane.regions}
Anywhere that two initial flow regions overlap is a subset of the flow space too narrow to hold two robots: both initial flow regions cannot exist in the same space.
We construct single-lane regions by joining the two generating objects at the ends of each overlap and creating a new flow region between these bounds.
\appsec{Passage Regions:}\label[app]{sec:pasage.regions}
We place a passage region (per~\cref{sec:method.spacetypes}) on both ends of each single-lane region and at each point where three or more flow regions meet.
This
\begin{enumerate*}[label=(\arabic*)]
\item ensures that transitioning for single-lane regions is controlled by space requests and
\item covers ``sharp corners'' in flow regions.
\end{enumerate*}
\appsec{Smoothing:}\label[app]{sec:smooth}
Convolve a \ensuremath{B_{\ensuremath{\rho}}}-sized disc with the inner perimeter of the open space (i.e.\ $\ensuremath{\mathcal{F}} \setminus \ensuremath{\Phi}$).
Anywhere that this disc is in contact with the inner perimeter at two points, we expand the corresponding flow regions out to meet the disc.
Smoothing ensures that no reachable point in \ensuremath{\mathcal{F}}{} is made unreachable by the flow/open space partition.
\appsec{Direction Assignment:}\label[app]{sec:direction.assignment}
We assign directions to each flow region by winding either clockwise or counterclockwise around the centroid of its generating object (for single-lane regions, we arbitrarily choose one of the generating objects).
We then return the flow, single-lane, and passage regions.
\section{Space Partitioning: Regions}\label{sec:method.spacetypes}
We partition the free space of an environment into three classes of space: \emph{flow space}, \emph{open space}, and \emph{passage space} according to the algorithm described in~\cref{appendix:B}.
These classes are composed of, respectively, \emph{flow regions}, \emph{open regions}, and \emph{passage regions}.
\label{definition:spot}%
In the following, a \emph{``spot''} refers to a \(\ensuremath{\mathcal{B}}(\ensuremath{\rho}) \subseteq \ensuremath{\mathcal{F}}\), an exactly robot-shaped disc in the free space of \ensuremath{\mathcal{E}}{}.
As disc-modeled robots in arbitrary environments can become stuck when the space is too densely occupied, we impose a global density cap:
\refstepcounter{assume}
\begin{tcolorbox}[title=\textbf{Assumption: Global Density Cap}]\label[assume]{assumption:density}
Let $N$ be the total number of open, flow, and passage regions in $\ensuremath{\mathcal{F}}$ for some environment \ensuremath{\mathcal{E}}{}. Let the maximal sphere packing\footnote{Sphere packing is difficult and well-studied \cite{scott1969density}. Similar sphere-modeled multi-agent robotics work assumes conservative estimates of density, often in the range $50-60\%$
\cite{chinta_coordinating_motion_2018} or lower. Identifying $M$, the upper bound, in practice is non-trivial, but determining a number below $M$ that generously upholds the density assumption while outperforming the density of $50\%$ is achievable.} of a space be $M$ Then the global cap on the number of robots allowed in \ensuremath{\mathcal{E}}{} is $M$ - $N$. In other words, there is always $N$ spots' worth of unoccupied space.
\end{tcolorbox}
In open space, robots can move freely; in flow space, there is an associated directional constraint on robot motion; passage regions help robots move between flow regions.
We define the following: given a bounded environment \ensuremath{\mathcal{E}}{} with free space \ensuremath{\mathcal{F}}{} and robots of radius \ensuremath{\rho}{}, we partition \ensuremath{\mathcal{F}}{} into sets \ensuremath{\Omega}{} (the open space), \ensuremath{\Phi}{} (the flow space), and \ensuremath{\Psi}{} (the passage regions) such that:
\begin{enumerate}
\item \ensuremath{\Phi}{} contains a minimally one-robot-wide corridor around every obstacle in \ensuremath{\mathcal{E}}{} and the interior of \ensuremath{\mathcal{E}}'s perimeter.
\item Every point in \ensuremath{\mathcal{F}}{} that can be reached by a robot without following any rules can be reached by a robot following our proposed motion rules in \cref{sec:method.motionrules} in \ensuremath{\Omega}{}, \ensuremath{\Phi}{}, or \ensuremath{\Psi}{}.
\item Every point in \ensuremath{\Phi}{} constrains robot motion to move in a particular direction (clockwise or counter-clockwise) around the perimeter and each obstacle in \ensuremath{\mathcal{E}}{}, and that these directions are \emph{compatible} with each other --- that is, that there are no two adjacent points directing a robot to move in opposing directions.
\item \ensuremath{\Psi}{} is composed of passage regions \ensuremath{P}{}, where each passage region is a spot (a robot-shaped disc) that connects adjacent flow regions.
\end{enumerate}
\section{Rules of Motion}\label{sec:method.motionrules}
At any time, a robot is exclusive either in an open, flow, or passage region, or is transitioning between two regions.
These conditions are disjoint --- a robot may not be in two types of region at the same time without transitioning.
To transition between regions, a robot must use the ``space request'' mechanism defined in~\cref{sec:god}.
In the following, let $\ensuremath{r_\idx} \in \ensuremath{\mathcal{R}}$ be a robot, $\ensuremath{\robot(t)}$ its position at time $t$, and $\ensuremath{\vec{v}_i(t)}$ its velocity vector at time $t$.
We identify \ensuremath{r_\idx}{} with its geometry, i.e.\ saying ``all of \ensuremath{r_\idx}{} is contained by a region'' means that the \ensuremath{\rho}-radius disc representing \ensuremath{r_\idx}{} is contained by the region.
\subsection{Determining a robot's region and transition state}
We determine whether \ensuremath{r_\idx}{} is transitioning between regions or in a particular region type at time $t$ by the following sequence of checks:
\begin{enumerate}
\item If $\ensuremath{\robot(t)} \cap P \ne \emptyset$ for a passage region $P \in \ensuremath{\Psi}$, \ensuremath{r_\idx}{} is in a passage region.
\item Else, if \ensuremath{\robot(t)}{} is \emph{completely} contained in a flow region $F \subset \ensuremath{\Phi}$, \ensuremath{r_\idx}{} is in a flow region.
\item Else, if \ensuremath{\robot(t)}{} is \emph{completely} contained by an open region $O \subset \ensuremath{\Omega}$ (i.e.\ $\ensuremath{\mathcal{B}}_\ensuremath{\rho}(\ensuremath{\robot(t)}) \subseteq O$), \ensuremath{r_\idx}{} is in an open region.
\item Otherwise, if none of the other conditions hold, \ensuremath{r_\idx}{} is transitioning between regions.
\end{enumerate}
\subsection{Constraints on robot motion}
If \ensuremath{\robot(t)}{} is in a flow region, its velocity \ensuremath{\vec{v}_i(t)}{} must satisfy $\ensuremath{\vec{v}_i(t)} \cdot \vec{d}_{c_i} \ge 0$, for $c_i$ the centroid of \ensuremath{\robot(t)}{} and $\vec{d}_{c_i}$ the flow direction at $c_i$.
In any region type, the pushing rules defined in~\cref{sec:god} apply and may constrain the motion of \ensuremath{r_\idx}.
If \ensuremath{r_\idx}{} is transitioning between regions, it must abide by all of the motion constraints for the region types it is transitioning between (e.g.\ a robot moving into a flow region cannot move against the direction of flow at its transition point).
\begin{figure}\label{fig:regions2}
\centering
\includegraphics[width = .8\linewidth]{figures/regionexamplev2.png}
\caption{An example of an environment partitioned into open regions (white), flow regions (periwinkle and lilac), and passage regions (green circles) with arrows indicating direction of motion. }
\label{fig:flowdirections}
\end{figure}
\subsection{Space Requests}\label{sec:method.spacerequests}
Space requests are the core of our adaptive centralization mechanism.
Robots use space requests to
\begin{enumerate*}[label=(\arabic*)]
\item transition between regions and
\item resolve spatial conflicts (i.e.\ conflicting attempts to occupy a space)
\end{enumerate*}.
Space requests are tuples $(S, d, T)$ comprising:
\begin{description}
\item[The primary spot $S$:] The primary spot represents the goal space, if any, that the robot wishes to access. $S$ is a radius \ensuremath{\rho}{} ball\footnote{\label{footnote:capsule}Or, in the case of a goal in or transition through a passage region, the convex hull of two adjacent radius \ensuremath{\rho}{} balls (\cref{note:exception}).} lying entirely in one region if the robot has a goal, or \texttt{null} otherwise.
\item[The duration $d$:] A real number representing the amount of time the robot would like to spend in the primary spot.
\item[The transition spot $T$:] The transition spot represents the space, if any, that the robot may need to access to move between two regions. $T$ is a radius \ensuremath{\rho}{} ball\cref{footnote:capsule} in a region adjacent to the robot's current location if the robot is moving between regions, or \texttt{null} otherwise.
\end{description}
\noindent{}To use space requests, each robot \ensuremath{r_\idx}{} maintains the following:
\begin{description}
\item[Priority:] A tuple $\ensuremath{L}_i = (\ensuremath{\tau}_i, \ensuremath{\ell}_i)$ where $\ensuremath{\tau}_i$ is a timestamp and $\ensuremath{\ell}_i \in \mathbb{N}$ is a unique label assigned to \ensuremath{r_\idx}{}.
``Timestamp priority'' refers to $\ensuremath{\tau}_i$ alone.
If \ensuremath{r_\idx}{} does not have an active space request at some time $t$, then $\ensuremath{\tau}_i = \infty$.
\item[Personal space request:] $SR_{p_i} = (S_{p_i}, d_{p_i}, T_{p_i})$, a space request used for the robot's unforced goals and transitions, as in~\cref{sec:method.srs.personal}.
\item[Forced space request:] $SR_{f_i} = (S_{f_i}, d_{f_i}, T_{f_i})$, a space request used for the robot's forced goals and transitions, as in~\cref{sec:method.srs.pushing}.
$d_f$ is always zero.
\item[Pusher identity:] $I_p$, a record of the robot (if any) which originally initiated the pushing ``network'' containing \ensuremath{r_\idx}{}.
If \ensuremath{r_\idx}{} is not being pushed, $I_p = \ensuremath{\ell}_i$.
\end{description}
\subsection{Moving with personal space requests}\label{sec:method.srs.personal}
Space requests for voluntary movement are a way of reserving access to a desired space.
A robot \ensuremath{r_\idx}{} communicates its intent to move into a space by submitting a space request, the request is checked to ensure that it is valid and won't cause a collision, and then, once granted access to the requested space, \ensuremath{r_\idx}{} can safely move in.
There are three stages to using a personal space request $(S_{p_i}, d_{p_i}, T_{p_i})$:
\begin{enumerate*}[label=(\arabic*)]
\item making the request,
\item moving into the request once it is granted, and
\item completing the request
\end{enumerate*}.
\noindent\textbf{Making a space request:}
A robot \ensuremath{r_\idx}{} can make a new space request if $S_{p_i}$, the primary spot of its personal space request, is currently \texttt{null}.
To do so, \ensuremath{r_\idx}{} sets $S_{p_i}$ to the spot it wishes to move into, sets $d_{p_i}$ to the duration it wishes to spend in $S_{p_i}$, and communicates $SR_{p_i}$ to the space request arbiter \ensuremath{\chi}{} (\cref{sec:god}).
\ensuremath{r_\idx}{} then waits for the arbiter to notify it that its request has been granted and set its timestamp priority to the time of approval.
While it waits, \ensuremath{r_\idx}{} cannot move unless forced (per~\cref{sec:method.srs.pushing}).
\ensuremath{r_\idx}{} can cancel its personal space request at any time by setting $S_{p_i}$ to \texttt{null} and communicating $SR_{p_i}$ to \ensuremath{\chi}.
\noindent\textbf{Moving into a granted space request:}
Once $SR_{p_i}$ is granted, \ensuremath{r_\idx}{} must move into $S_{p_i}$ as quickly as possible, following the shortest collision-free path to the requested space (abiding by motion rules).
If this path crosses between regions, \ensuremath{r_\idx}{} must use the transition spot $T_{p_i}$ from $SR_{p_i}$, by setting $T_{p_i}$ to a spot
\begin{enumerate*}[label=(\arabic*)]
\item tangent to \ensuremath{r_\idx}{} and
\item completely in the region \ensuremath{r_\idx}{} wants to enter
\end{enumerate*}, and communicating the modified $SR_{p_i}$ to \ensuremath{\chi}{} for approval of the transition per the above.
Once \ensuremath{r_\idx}{} changes regions, the transition spot is reset to \texttt{null}.
\noindent\textbf{Completing a space request:} Once \ensuremath{r_\idx}{} has moved into $S_{p_i}$ (i.e. \ensuremath{r_\idx}{} and $S_{p_i}$ align) for $d_{p_i}$ seconds, the space request is completed and $SR_{p_i}$ is reset to $(\texttt{null}, 0.0, \texttt{null})$.
\begin{figure}\label{fig:spacerequest}
\centering
\includegraphics[width = .3\linewidth]{figures/cluster4.png}
\caption{An example of multiple robots (green, gray, and brown) satisfying their space requests (green with polka dots, empty gray circle, and brown with polka dots) at the same time.}
\label{fig:cluster}
\end{figure}
\subsection{Forced space requests and pushing}\label{sec:method.srs.pushing}
To guarantee progress for all robots, we sometimes need to make one or more robots move out of the space another robot wants to occupy.
We do so by introducing a notion of ``pushing'': a mechanism by which a high priority robot can clear its path.
At a high level, pushing is triggered when a higher priority robot encounters a lower priority robot in its way.
If that lower priority robot can move out of the higher priority robot's path, it must; otherwise, if its own path is blocked, it transitively pushes out, borrowing the higher priority to do so.
This iterative pushing, combined with a total ordering of robot priorities, ensures that the original higher priority robot will be able to proceed toward its goal in finite time (proving this is the primary focus of~\cref{sec:analysis}).
In more detail: a robot \ensuremath{r_\idx}{} can (in different circumstances) both \emph{push} and \emph{be pushed}.
Each of these processes uses each robot's priority $\ensuremath{L}_i$ and forced space request $SR_{f_i}$.
\noindent\textbf{Comparing priorities:}
We impose a total order on robot priorities by saying that, for $i \ne j$, $\ensuremath{L}_i < \ensuremath{L}_j$ iff $\ensuremath{\tau}_i$ is later than $\ensuremath{\tau}_j$ or $\ensuremath{\tau}_i = \ensuremath{\tau}_j$ and $\ensuremath{\ell}_i < \ensuremath{\ell}_j$.
This order ensures that earlier requests have higher priority, and thus that a space request's priority monotonically increases until it is completed.
\noindent\textbf{Pushing:}
When a robot \ensuremath{r_\idx}{} is moving to complete its space request (personal or forced) and comes tangent to\footnote{\label{footnote:tangent}Note that this does not imply physical contact between robots; \ensuremath{\rho}{} can be chosen to provide a buffer of space around each robot.} a robot $r_j$ with lower timestamp priority, then $r_i$ pushes $r_j$.
$r_j$ first sets $I_{p_j} = I_{p_i}$ and adopts the pusher's (\ensuremath{r_\idx}{}) timestamp priority for its forced space request $SR_{f_j}$.
We then define a ``pushing vector'' recursively to derive how $r_j$ has to move from $\ensuremath{r_\idx}{}$'s pushing vector.
If \ensuremath{r_\idx}{} is not being pushed (i.e. it's the pushee), then its pushing vector is the vector from \ensuremath{\robot(t)}{} to $r_j(t)$.
If \ensuremath{r_\idx}{} is in a flow region, then its pushing vector is the flow direction.
Otherwise, $\ensuremath{r_\idx}{}$'s pushing vector also abides by the following recursive definition.
\hypertarget{pushrules}{When \ensuremath{r_\idx}{} pushes robot $r_j$, $r_j$ selects a unit vector $p_\theta$ with an angle ($\theta$) within $\pi/2$ of $\ensuremath{r_\idx}{}$'s pushing vector ($p_i$) and averages $p_\theta$ and $p_i$ to generate $p_j$.}
\noindent\textbf{Being pushed:}
A robot $r_j$ is pushed when it comes tangent to\cref{footnote:tangent} a robot \ensuremath{r_\idx}{} with equal or higher priority.
There is one notable exception: to ensure that no robot can transitively push itself, if the timestamp priorities of $r_j$ and $r_i$ are equal, but $I_{p_i} = \ensuremath{\ell}_j$ (i.e.\ $r_j$ originated the pushing network containing \ensuremath{r_\idx}), then $r_j$ is not considered pushed.
When $r_j$ \textit{is} pushed, it sets $T_{p_j} = \texttt{null}$ (cancelling any pending requests to transition between regions for voluntary movement), and adopts \ensuremath{r_\idx}'s timestamp priority.
$r_j$ must immediately attempt to move away from \ensuremath{r_\idx}{} into a spot adjacent to its current location within $\pm\frac{\pi}{2}$ radians of \ensuremath{r_\idx}'s pushing vector.
If there exists an unoccupied adjacent spot within the allowed angle range, $r_j$ must move into it.
If $r_j$ needs to place a space request to move (i.e.\ there are no unoccupied adjacent spots within the allowed angle range), then it makes and completes a \emph{forced space request} using $SR_{f_j}$ in the angle range, following the same process as for personal space requests with \ensuremath{r_\idx}'s timestamp priority for the request.
A forced space request is immediately canceled if the robot is no longer being pushed or transitioning to satisfy the pushed request. $r_j$ updates its pushing vector as described \hyperlink{pushrules}{above}.
When equal priority robots push each other and neither originated the pushing network, their pushing vectors are averaged, ensuring a consistent overall pushing direction.
Pushing is re-evaluated every timestep, so pre-established redundant forced space requests are automatically canceled.
\begin{figure}
\centering
\includegraphics[width = .5\linewidth]{figures/pushFlowArrows.png}
\caption{An example of how pushing occurs when TR (gray) is moving to granted space request (gray circle), the desired motion is shown as a red arrow. Each pushed robot's direction is indicated by an arrow (in particular, observe that the purple robot overlaps the requested spot, so it must move forward in the flow to get out of the way). }
\label{fig:pushing}
\end{figure}
Pushing is covered in more detail in~\cref{appendix:A}.
\refstepcounter{note}
\begin{tcolorbox}[title=\textbf{The Passage Region Exception}]\label[note]{note:exception}
Space requests work differently when a robot \ensuremath{r_\idx}{} wants to occupy a passage region \ensuremath{P}.
\begin{enumerate}
\item The primary or transition spot (depending on why \ensuremath{r_\idx}{} wants to occupy \ensuremath{P}) becomes a \textit{capsule} region defined by $\ensuremath{P}\bigcup B_{EXIT}$, where $B_{EXIT}$ is a radius \ensuremath{\rho}{} ball adjacent to \ensuremath{P}{} in the flow region \ensuremath{r_\idx}{} wishes to exit into.
\item A robot pushed by \ensuremath{r_\idx}{} can pass through the \textit{capsule} to fulfill pushing, though it must clear the \textit{capsule} as quickly as possible.
\end{enumerate}
This allows robots to ``escape'' packed flow regions to allow \ensuremath{r_\idx}{} in and prevents deadlock caused by \ensuremath{r_\idx}{} blocking other robots from moving out of its way (see \cref{fig:figure8}).
\end{tcolorbox}
\subsection{Global Space Request Arbiter}\label{sec:god}
The global space request arbiter \ensuremath{\chi}{} is in charge of space requests.
We refer to \ensuremath{\chi}{} as an adaptive centralization mechanism because it coordinates precisely the set of robots which must be moved for a given robot to make progress and determines this set online.
In practice, the arbiter could be implemented as a centralized entity or through local communication directly between robots.
The arbiter has an initial role of randomly assigning a unique ID $\ensuremath{\ell}_i$ to each robot \ensuremath{r_\idx}.
It may reassign $\ensuremath{\ell}_i$ for all robots at fixed intervals to provide better overall fairness in access to resources (as $\ensuremath{\ell}_i$ breaks ties when comparing priorities).
However, the arbiter's main job is to keep track of space requests and their priorities.
Every time a robot makes a space request, it is submitted to the global arbiter with the timestamp \ensuremath{\tau}{} when it was placed.
The arbiter keeps track of the priorities $(\ensuremath{\tau}_i, \ensuremath{\ell}_i)$ of all robots that have made a space request, removes cancelled requests, and uses this information to send pushing information to the robots and grant requests.
\ensuremath{\chi}{} must approve all primary and transition spots for all space requests.
The approval process first checks that the requested spot is valid (i.e.\ that it is contained in the free space).
The arbiter immediately grants space requests that have valid spots and do not require transitioning into a flow or passage region.
Requests to transition into a flow or passage region are approved once the requested spot is clear (i.e.\ does not contain any robots).
When a request is made to transition into a flow or passage region, the arbiter informs nearby robots; robots with lower timestamp priorities must stay clear of the spot or exit it if they are currently occupying it.
If two conflicting requests (i.e.\ the requested spots overlap) with the same timestamp priority are ready to be granted, \ensuremath{\chi}{} grants the request with the higher-priority value of \ensuremath{\ell}.
|
\section{Introduction}
Enhanced flow transport in confined spaces is important in many processes, from catalytic-based chemical conversion \cite{Kortunov2005}, porous membrane separation \cite{Tao2018}, nanomedicine \cite{Radha2019, Sanhai2008,Sparreboom2009}, and water treatment \cite{Wu2017}, to enhanced oil recovery \cite{ZhangY2017, Druetta2020} and CO$_2$ sequestration \cite{Gao2017,Feng2019}. In confined spaces, liquid mixing is dominated by slow mutual diffusion \cite{Tartakovsky2019, Frediksen2018}, influenced by the physical properties of liquids. Extensive studies have been performed to understand the effects of wettability of the wall surface \cite{Frediksen2018, Aziz2020, Wu2017}, fluid viscosity \cite{Ho2019, Guillen2012}, or interfacial tension of immiscible phases \cite{Feng2019}. On an industrial scale, the liquid displaced from confined spaces usually consists of multiple chemical constituents that can interact with the displacing liquid \cite{Xu2014, Guillen2012, Zhou2019}. For example, in enhanced oil recovery, chemical flooding processes use a displacing liquid to recover the entrapped oil phase from porous rocks by altering the interfacial tension or by reducing the viscosity of the pore fluid \cite{Xu2014, Mandal2010}. It is important for both fundamental understanding and practical applications to achieve fast fluid movement in confined spaces.
Enhanced fluid transport and autonomous motion of droplets and colloidal particles driven by chemical concentration gradients in multicomponent liquid mixtures \cite{Lohse2020} are topics of current research interest. Microparticles or microdroplets move as a result of diffusiophoretic or solutal driven Marangoni phenomena \cite{Lohse2020}. Concentration gradients are often created and sustained by chemical reaction \cite{Shi2016}, dissolution of surfactants \cite{Nery2017} and surface chemistry interactions \cite{Banarjee2016}. Manipulating the electrolyte concentration in the displacing fluid has been demonstrated to drive colloid transport into and out of microchannels dead-ends \cite{Kar2015, Shin2016, Ault2018}.
In addition to chemical concentration gradients present in multicomponent single-phase liquid mixtures, subphase formation from phase separation in multicomponent systems is also advantageous in displacing liquid in confined spaces. Even from a practical application perspective, several studies have indicated improvements in oil recovery efficiency from porous media when microdroplet emulsions are formed during displacement compared to cases without emulsion formation. In these cases, the enhancement effect is attributed to the desired rheological properties and seepage characteristics of surfactant-stabilized emulsions \cite{Zhou2019}. More fundamentally, microdroplets form spontaneously from the simple addition of a poor solvent into a ternary solution via the Ouzo effect. When the ternary solution is diluted by a poor solvent that is miscible with the solvent but immiscible with droplet liquid, the mixture becomes oversaturated, leading to the formation of many micro-sized droplets dispersed over the entire continuous liquid phase without the use of mechanical agitation or surfactant \cite{Vitale2003, Lu2017, Zemb2016}. In confinement, the spontaneous formation of microdroplets leads to enhanced transport of the penetrating fluid \cite{Lu2017}. While the effect has been demonstrated, it has not been explored in detail. It remains unclear how the displacement rate is influenced by the relative and absolute compositions of phases formed during phase separation - even for a single mixture.
In this work, we use a simple ternary liquid mixture as a model system to explore the displacement process in quasi-2D confined spaces. More specifically, the confined liquid comprises a ternary solution of 1-octanol, ethanol, and water. The displacing liquid is water. The displacement process is shown to be strongly dependent on the composition of the confined liquid. We demonstrate that over one range of compositions, the confined liquid is displaced. Over another range of compositions, we demonstrate the selective separation and displacement of 1-octanol. As the composition trajectories and hence the intersection points with the one-phase to two-phase boundary in the phase diagram are known, our findings suggest that an effective approach to enhance mass transport in confined spaces is to use displacement liquid composition as a design variable.
\section{Experimental Section}
\subsection{Chemicals and solution preparation}
Chemicals were used as received without further purification. Solution A consisted of 1-octanol (ACS grade $>$95\%, Fischer Scientific), ethanol (Histological grade, Fischer Scientific) and (Milli-Q) water. The compositions of Solution A were varied from 2\% 1-octanol to 50\% 1-octanol as listed in Table \ref{Experimental Conditions Table}. Solution B consisted of (Milli-Q) water. Additional experiments were performed with variations to both confined and displacing liquids, as summarized in the Table \ref{Experimental Conditions Additional Table} below.
The surfaces used for the experiments were hydrophilic glass slides (Fisherbrand Microscope Slides) and hydrophobic silicon wafers. Before use, both substrates were initially washed with piranha solution consisting of 70\%-volume $H_2SO_4$ (ACS Plus grade, Fischer Scientific) and 30\%-volume $H_2O_2$ (30\% ACS grade, Fischer Scientific) at 85$\degree$ C for 20 minutes (caution: piranha solution is highly caustic). They were further cleaned and sonicated with water and ethanol each for 15 minutes. To render the silicon wafers hydrophobic, they were coated with octadecyltrichlorosilane (OTS-Si), as has been previously documented \cite{ZhangX2008}. In brief, 0.5\%-volume of OTS in hexane mixture was used to soak the wafers for 12 hours at room temperature. The OTS-Si substrates were then sonicated with hexane, acetone, and ethanol for 10 minutes each to remove any excessive OTS on the surface.
For fluorescence experiments, Nile Red (Fischer Scientific) was used due to its high solubility in hydrocarbon-rich phases where the oil-rich domains become fluorescent, which is useful for identifying liquid regions rich in 1-octanol.
\begin{table}[H]
\captionsetup{font = {small}}
\caption{Evaluated experimental compositions of solution A (displaced liquid) by mass percentage.}
\centering
\begin{tabular}{|c| c| c| c|}
\hline
Composition & 1-Octanol & Ethanol & Water \\
& Mass \% & Mass \% & Mass \%\\
\hline
1 & 50 & 40 & 10 \\
\hline
2 & 40 & 45 & 15 \\
\hline
3 & 35 & 48 & 17 \\
\hline
4 & 30 & 48 & 22 \\
\hline
5 & 25 & 48 & 27 \\
\hline
6 & 20 & 48 & 32 \\
\hline
7 & 15 & 48 & 37 \\
\hline
8 & 10 & 48 & 42 \\
\hline
9 & 8 & 47 & 45 \\
\hline
10 & 5 & 48 & 47 \\
\hline
11 & 2 & 48 & 50 \\
\hline
\end{tabular}
\label{Experimental Conditions Table}
\end{table}
\begin{table}[H]
\captionsetup{font = {small}}
\caption{Compositions of solution A (displaced liquid) and solution B (confined liquid) by mass percentage in additional experiments.}
\begin{tabular}{|c|c|c|c|c|c|c}
\hline
\multicolumn{1}{|l|}{} & \multicolumn{3}{c|}{Solution A} & \multicolumn{3}{c|}{Solution B} \\ \hline
\multicolumn{1}{|l|}{Composition} &
\multicolumn{1}{l|}{1-octanol} &
\multicolumn{1}{l|}{Ethanol} &
\multicolumn{1}{l|}{Water} &
\multicolumn{1}{l|}{1-octanol} &
\multicolumn{1}{l|}{Ethanol} &
\multicolumn{1}{l|}{Water} \\ \hline
E1 & 50 & 40 & 10 & 0 & 25 & \multicolumn{1}{c|}{75} \\ \hline
E4 & 10 & 48 & 42 & 0 & 25 & \multicolumn{1}{c|}{75} \\ \hline
R1 & 0 & 0 & 100 & 50 & 40 & \multicolumn{1}{c|}{10} \\ \hline
R4 & 0 & 0 & 100 & 10 & 48 & \multicolumn{1}{c|}{42} \\ \hline
\end{tabular}
\label{Experimental Conditions Additional Table}
\end{table}
\subsection{Dimensions and setup of the microchamber}
The flow chamber, as sketched in Figure \ref{experimentalsketch}, consisted of a polycarbonate base (8.5 $\times$ 4.5 $\times$ 1 cm), with the hydrophilic cover glass as the top sealed with a silicon rubber spacer. Before any experimentation, every component from the chamber was sonicated in both water and ethanol for 10 minutes each. The chamber was assembled by connecting the tubing to the polycarbonate base as the inlet and outlet streams. A droplet of Solution A was placed on the OTS-Si substrate. The substrate was held by a capillary bridge to the cover glass. A small channel height of around 20 to 30 $\mu m$ was kept. The cover glass was then held in position by a spacer and was clamped to the base of the chamber with large binder clips.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.6\columnwidth]{./figs/experimentalsketch}
\caption{Sketch of the fluid microchamber with top view and cross-section of the main and side channels.
}
\label{experimentalsketch}
\end{figure}
\subsection{Process of the liquid displacement}
Solution B was pumped using a syringe pump (Fisherbrand Single Syringe Pump) to the channel. The displacement process as water diffused transversely into the narrow channel was captured in-situ using a camera (Nikon DS-Fi3) connected to an optical microscope (Nikon Eclipse Ni-U). The magnification under normal conditions was 10$\times$ with a F.O.V. of 22 $mm$, and videos were captured using an auto-exposure setting leading to captured footage with framerates ranging from 10-15 frames per second (FPS) at a fixed resolution of 2880 $\times$ 2048. Fluorescence footage required longer exposure times to capture the fluorescent features leading to low frame rates of 0.9 to 1 FPS.
\subsection{Analysis of the displacement rates}
The videos were processed through ImageJ by converting them to image sequences. If required, the contrast and brightness levels were adjusted, and a bleach correction was applied. The tracking of the boundaries was achieved by the manual tracking plugin within the program. The calibration used for the captured footage at 10x magnification was 0.24 px/$\mu$m. The tracked data were saved as a raw CSV file, including the number of frames, x and y positions in pixels. The displacement rate was calculated using the x and y position using the first entry in the data set as the initial position.
\section{Results and Discussion}
The phase diagram for the model ternary mixture 1-octanol (oil), ethanol (good solvent) and water (poor solvent) is illustrated in Figure \ref{overview}A. Confined liquid compositions labelled from 1 to 11 (Table \ref{Experimental Conditions Table}), are shown in the phase diagram. The phase boundary was adapted from the literature \cite{Lopian2016}, tie lines not shown were computed using the UNIFAC model and inform the discussion. The Dortmund UNIFAC model was implemented using the parameter values noted in the Supporting Information (Tables S1 $\sim$ S4).
The mixing between solution A and water as water diffused into the narrow channel induced liquid-liquid phase separation. Below we show the displacement process along with phase separation for the wide range of solution A compositions listed in Table \ref{Experimental Conditions Table}. The 1-octanol-rich liquid is revealed using a hydrophobic dye in fluorescence images. The displacement dynamics are categorized into four regimes, as shown in Figure \ref{overview}B-E.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=1\columnwidth]{./figs/overview}
\caption{Overview of different dynamic regimes. (A) Ternary phase diagram of 1-octanol, ethanol, and water system in terms of mass fractions with evaluated experimental compositions annotated. The phase diagram is adapted from ref \cite{Lopian2016}. Displacement regimes are indicated by colors: Regime 1 (yellow), Regime 2 (orange), Regime 3 (green), and Regime 4 (blue). (B-E) Representative optical images of the phase separation regimes are shown in colour-coded rectangles as in (A).
}
\label{overview}
\end{figure}
\subsection{Regime 1: Receding interface}
Regime 1 corresponds to a high 1-octanol concentration, above 50\% by mass in solution A. The composition is labelled in the phase diagram in Figure \ref{Regime 1}A. As water diffused from the side channel, the resulting images in Figure \ref{Regime 1}B show a clear interface separating solution A from water in the narrow channel.
A few 1-octanol microdomains pinched off from the receding phase boundary at several locations and remained attached to the hydrophobic wall on the water side of the boundary. Their composition was confirmed using fluorescence imaging (Figure \ref{Regime 1}C). This fluorescence image of a representative domain formed from solution A doped with an oil-soluble fluorescent dye (Nile red). The high fluorescence intensity from the microdomain indicates the high concentration of the dye in the microdomain, in contrast to the absence of a fluorescence signal from the surrounding liquid. The spatial distribution of the dye is consistent with the chemical composition of the 1-octanol-rich microdomains surrounded by water-rich liquid.
In this regime, water diffusing into the narrow channel did not lead to liquid-liquid phase separation along the boundary - a behaviour consistent with that of an immiscible displacing liquid. However, liquid-liquid phase separation did occur within the stranded microdomains. Many nanodroplets (with a diameter of approximately 4 $\mu m$) formed, as shown in Figure \ref{Regime 1}D. The formation of these water-rich nanodroplets is attributed to subphase formation arising from diffusion of water into the microdomains followed by liquid-liquid phase separation.
As water diffuses into a microdomain and 1-octanol and ethanol diffuse out, the global composition of the microdomain changes. To a first approximation, the dilution path that is followed is indicated by the dashed line in the phase diagram in Figure \ref{Regime 1}A. The mixture becomes unstable as the composition intersects the solubility curve and spontaneously forms water-rich and 1-octanol-rich subphases. Although the global composition of the microdomain at separation is unknown, the compositions of the 1-octanol-rich and water-rich subphases created are approximately 80 and 19\% 1-octanol by mass. It is interesting that phase separation occurred in the microdomains of stranded solution A, but not at the boundary between water and solution A. The large surface to volume ratio of the small microdomains permits them to become supersaturated with water quickly. By contrast, at the boundary, water diffuses into the bulk 1-octanol-rich phase, and 1-octanol and ethanol diffuse into the bulk water-rich phase. Consequently, the liquids on both sides of the boundary do not supersaturate within the time frame of measurements. At longer times, one would certainly expect the microdomains to dissolve into the water-rich phase.
While the shape and size of stranded solution A microdomains varied between experiments (Figure \ref{Regime 1}D), possibly due to the nature of the interface instability - subphase formation within them over time was common.
If we now focus on the displacement of the boundary between solution A and water, in the absence of microdomain formation, we can see that the x-location displacement rate is essentially time-invariant with an average value of 10.7 $\pm$ 0.3 $\mu m/s$ (Figure \ref{Regime 1}E). Figure \ref{Regime 1}F shows the fluctuating x-location displacement rate where microdomains form and detach over time. The interface progressively displaced solution A in the positive x-direction. During the formation of a microdomain, the interface remains stationary or moves backward. When a microdomain detaches from the interface, the interface accelerates in the positive x-direction. Such cycles were observed at multiple locations on the surface.
The blue and red dashed curves in Figure \ref{Regime 1}F are 1D diffusion-based displacement curves, $l=(2Dt)^{1/2}$, fitted to the interface displacement data before microdomain detachment to obtain an effective diffusion constant \cite{Lu2017}. The effective diffusivity varies as the boundary progresses. The values range from 5.3 $\times 10^{-9}$ $m^2/s$ for a part of the boundary without microdomain formation (Figure \ref{Regime 1}E) and 2.8 $\times 10^{-10}$ to 7.9 $\times 10^{-10}$ $m^2/s$ for a location where there is microdomain formation. The mutual diffusivities of water and ethanol, and water and octanol range from 1.08 $\times 10^{-10}$ to 1.23 $\times 10^{-9}$ $m^2/s$ and 2.0 $\times 10^{-10}$ to 7.3 $\times 10^{-10}$ $m^2/s$, respectively \cite{Pratt1974, Kinoshita2017, Su2010}. So the diffusion rates are consistent with expectations. In locations with microdomain formation, the boundary displacement rate was variable and ranged from 4.3 to 14.0 $\mu m/s$.
The influence of solution A microdomains on the local motion of the boundary may be attributed to the pinning effect and mass transfer to microdomains. Imperfections on the wall surface pin the boundary, leading to microdomain formation once the boundary becomes unstable. As the microdomain is formed, the water-rich phase close to the microdomain experiences interfacial tension opposite to its flow direction. The acceleration experienced after the detachment of the microdomain may be attributed to the release of built-up tension that comes from the stretched interface of the microdomain.
In Regime 1, water largely displaces solution A (with a high initial concentration of 1-octanol) in a confined space with one hydrophobic wall. The large scale stability of the boundary between solution A and water contributes to the effective displacement. In contrast, local instability of the boundary and microdomain formation leads to lower and fluctuating displacement rates indicated in Figure \ref{Regime 1}F, as well as less efficient 1-octanol-rich liquid displacement.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.80\columnwidth]{./figs/Regime_1}
\caption{Overview of Regime 1. (A) Ternary phase diagram indicating the dilution path with the two macroscopic subphases formed. P is the plait point of the ternary mixture. Red regions within the phase envelope are the Ouzo and reverse Ouzo regions. Adapted from ref \cite{Lopian2016}. The initial solution A composition is given by a red diamond. An illustrative unstable mixture composition during dilution is given as a blue diamond. The red stars represent the 1-octanol-rich and water-rich compositions formed along the green tie line. (B) Time-based optical images of early film and domain development. (C) Side by side fluorescence and bright-field images of representative microdomains. (D) Time-based images of a deformed domain narrowing with time. (E) Plot showing local boundary displacement without microdomain formation. (F) Plot showing local boundary displacement at a location where microdomains form. Blue data set was taken along the x-direction indicated by white dashed line in (B) with the same reference position of the side channel. Time window 1 is for a period without a microdomain present. During time period 2, a microdomain forms at the boundary. Time period 3 starts as soon as the microdomain detaches from the boundary. The formation of microdomains slows boundary displacement. Their detachment accelerates boundary displacement.}
\label{Regime 1}
\end{figure}
\subsection{Regime 2: Moving microdroplet zone}
Regime 2 corresponds to intermediate 1-octanol concentrations, about 40\% by mass in solution A. The overall behaviour in Regime 2 differs markedly from Regime 1, despite the similarity of the solution A compositions. Starting from this composition, the subphase formed by mixing with water is well approximated by the trajectory shown in Figure \ref{Regime 2}A. From the tie line shown in the phase diagram, much more 1-octanol-rich subphase is still expected than water-rich subphase following liquid-liquid phase separation. Unlike Regime 1, a mobile 1-octanol-rich microdroplet zone separates solution A and water throughout the narrow channel, as shown in the fluorescence images in Figure \ref{Regime 2}B. A dark zone containing water, a zone rich in microdroplets and a clear zone of solution A are clearly present.
Figure \ref{Regime 2}C shows the level of fluorescence intensity crossing the three zones. The high fluorescence intensity in the intermediate zone confirms that it is rich in 1-octanol. The boundary zone appears dark in bright field images, possibly due to scattering from small 1-octanol-rich drops.
The time-based images in Figure \ref{Regime 2}D show oscillation of the boundary zone, but this oscillation does not lead to the formation of stranded microdomains. Instead, the boundary zone thickness appears to remain constant at 18-20 ${\mu}m$ in the time interval of 12 $s$ as the boundary moves forward, as shown in the high magnification inset (Figure \ref{Regime 2}D). The time invariance of the boundary zone thickness may indicate that the mass transfer occurring through the boundary is balanced, even though the compositions of the water-rich and 1-octanol-rich phases in the zone vary along the x-direction.
Lines of stranded 1-octanol-rich droplets form behind the receding boundary zone, mirroring the shape of the boundary as it progresses (Figure \ref{Regime 2}E). The interval between two lines of droplets on the surface may be related to the time required for phase separation to occur in the boundary zone or the time required for the droplets inside the boundary zone to accumulate on the contact line with water.
The rate of the boundary displacement is quantified in Figure \ref{Regime 2}F along a fixed-line in the x-direction. The rate of displacement fluctuates with time due to the oscillatory advancement of the boundary zone. Again, the data can be fitted with a diffusion model \cite{Lu2017}. Compared to the displacement of the boundary in Regime 1 (for a case where microdomain formation does not occur), the displacement versus square root of time is slower in Regime 2. This is evidenced by the lower effective diffusivity obtained from the fit, which ranged from 5.8 $\times 10^{-10}$ to 6.7 $\times 10^{-10}$ $m^2/s$, and a lower displacement rate of 5.8 $\mu m/s$.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.9\columnwidth]{./figs/Regime_2}
\caption{Overview of Regime 2. (A) Ternary phase diagram indicating the approximate dilution path and the compositions of the subphases formed. P is the plait point of the ternary mixture. Red regions within the phase envelope are the Ouzo and reverse Ouzo regions. Adapted from ref \cite{Lopian2016}. The initial solution composition is given by a red diamond. The unstable mixture composition after dilution is approximated by a blue diamond. The red stars represent the 1-octanol-rich and water-rich subphases that are formed based on the green tie line. (B) Time-based fluorescence imagery at 40\% 1-octanol within Regime 2. (C) Normalized fluorescence intensity as a function of x-location within the frame of (B) at t = 15 s. The red area in the graph indicates a boundary zone. (D) Time-based optical images of moving boundary within Regime 2 with an initial solution A composition of 40\% 1-octanol. Inset contains magnified images of the moving boundary indicating zone thickness. (E) Time-based optical images of droplet formation in the boundary zone. (F) Boundary displacement as a function of the square root of time assuming diffusion-based dynamics. Regime 1 smooth interface displacement (black) and Regime 2 (blue) comparison.
}
\label{Regime 2}
\end{figure}
\newpage
\subsection{Regime 3: Moving three-zone configuration}
Regime 3 arises for the 1-octanol concentration range 20\% to 30\% by mass in solution A. Data for the 30\% case are presented in detail and are representative of this regime. Figure \ref{Regime 3}A illustrates the dilution process within the phase diagram. On separation, approximately equal amounts of water-rich and 1-octanol-rich subphases form because the dilution line intersects the phase envelope close to the plait point.
Fluorescence images in Figure \ref{Regime 3}B reveal three distinct zones along the x-direction as labelled in Figure \ref{Regime 3}B at t = 30 s. The normalized fluorescence intensity plotted in Figure \ref{Regime 3}C shows that zone 1 consists of a 1-octanol-rich phase revealed by partition of a dye \cite{Li2018}, that zone 2 consists of a water-rich subphase, and that zone 3 contains numerous immiscible water-rich droplets (dark features in fluorescence images within a 1-octanol-rich phase).
Figure \ref{Regime 3}D illustrates the early development of the three-zone configuration as water enters from the side to the narrow channel. A 1-octanol-rich phase forms from phase separation, and water droplets appear simultaneously within the 1-octanol-rich subphase and later coalesce and merge into a water-rich zone 2.
The displacement of solution A proceeds as the entire zone 2 moves into the narrow channel. The width of zone 2 appears to increase slightly with time at a fixed y-position, as shown in Figure \ref{Regime 3}E. The most noticeable aspect of the entire zone 2 is its stability in space over minute long time scales. We suspect that this stability is related to the near-constant width of zone 2, which we in turn attribute to the near equivalence of the influx and outflux of liquid to zone 2 over time. Liquid may enter zone 2 in the form of water-rich droplets from zone 3 or by water entering from the side channel, as evidenced in the Figure S2. Concurrently the 1-octanol-rich liquid leaves zone 2 to zone 3. The balancing of mass fluxes in and out of zone 2 accounts for its stability.
The displacement of the zone 1 to zone 2 and zone 2 to zone 3 boundaries versus the square root of time are shown in Figure \ref{Regime 3}F. These boundaries move with an oscillatory behaviour in the positive x-direction with a slower overall rate than the previous regime 2 at 0.71 $\mu m/s$. The fluctuations were found to be related to the water-rich droplets entering zone 2. At locations were droplets merged with the boundary between zone 2 and 3, the boundary retracted in the negative x-direction. The boundary successively becomes deformed, leading to the acceleration of the boundary in the positive x-direction as the shape of the boundary starts to recover. The boundary displacement curves from the two locations in Figure \ref{Regime 3}F show similar trends in pinning and depinning transitions. Regardless of all the fluctuations in boundary motion, the three-zone configuration remained stable with time due to water replenishment.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=1\columnwidth]{./figs/Regime_3}
\caption{Overview of Regime 3 and dynamic data for 30\% 1-octanol in solution A. (A) Ternary phase diagram showing the approximate dilution path and subphases compositions. P is the plait point of the ternary mixture. (B) Illustrative time-based fluorescence images for Regime 3. Zone 1 is a 1-octanol-rich phase, zone 2 is a water-rich phase, and zone 3 is 1-octanol-rich liquid with dispersed water-rich phase drops. (C) Normalized fluorescence intensity as a function of x-location within the frame of (B) at t = 0 s. (D) Time-based fluorescence images stages as water begins to enter the narrow channel. (E) Water-rich zone 2 width as a function of time (dashed line indicating slight increasing trend). (F) Boundary displacement as a function of the square root of time illustrating diffusion-dominated movement. (Blue) corresponds to the boundary between zones 1 and 2. (Red) corresponds to boundary between zones 2 and 3. (Orange) corresponds to Regime 2 displacement data. Inset depicts the slopes, from 2 $<$ $t^{1/2}$ $<$ 6 for the boundaries in Regime 3.}
\label{Regime 3}
\end{figure}
\newpage
\subsection{Regime 4: Diffusive boundary}
In Regime 4, the composition ranges from 2-20\% 1-octanol by mass in solution A. In this regime, a diffusive boundary forms and the dynamics differ from the other three regimes. The results depicted in Figure \ref{Regime 4 part 1} are for 1-octanol for 10\% by mass.
Figure \ref{Regime 4 part 1}A shows the approximate dilution path and subphase compositions in the phase diagram for the 10\% 1-octanol case. The diluting mixture intersects the phase envelope below the plait point, and following phase separation, the water-rich subphase fraction is expected to an order of magnitude greater than the 1-octanol-rich subphase mass fraction.
Figure \ref{Regime 4 part 1}B shows the diffusive boundary at an early transition stage consisting of numerous triangular protrusions of solution A that extend into the water phase. Droplets formed from the tips of these protrusions in the water-rich phase were larger than those formed at other positions along the boundary because, at the tip of the protrusions, the concentration gradients are sharper than at other positions \cite{Lu2017, Vitale2003}. At longer times, protrusions approach one another and collapse into a line-shaped boundary region, as shown in Figure \ref{Regime 4 part 1}C At longer times, droplets are formed along the entire boundary in the water-rich phase.
Fluorescence images in Figures \ref{Regime 4 part 1}(D, E) show the 1-octanol-rich protrusions with a 1-octanol concentration greater than in solution A. Ethanol transfers prefers preferentially to the water-rich phase. After 30 seconds, the composition of the protrusions becomes consistent with that of the 1-octanol rich droplets formed, as indicated by the fluorescence intensity.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=1\columnwidth]{./figs/regime_4_part_1}
\caption{Overview of Regime 4 at low 1-octanol compositions. (A) Ternary phase diagram indicating the approximate dilution path and subphase compositions. P is the plait point of the ternary mixture. (B) Time-based optical images showing boundary development and (C) boundary for the 10\% 1-octanol solution A case. (D) Time-based fluorescence imagery of protrusion and change to (E) line-shaped boundary for the 10\% 1-octanol case.}
\label{Regime 4 part 1}
\end{figure}
Figure \ref{Regime 4 part 2}A provides an overview and dynamic data for solution A with 2\% 1-octanol. Figure \ref{Regime 4 part 2}A shows the approximate dilution path and subphase compositions. The mass ratio of water-rich to 1-octanol-rich subphase is expected to be around 25 to 1 in this case. Fluorescence images in Figure \ref{Regime 4 part 2}C again show triangular protrusions of 1-octanol-rich liquid into the boundary. While the outcomes are similar to those described in Figure \ref{Regime 4 part 1}, the triangular protrusions are more stable (Figure \ref{Regime 4 part 2}B). This outcome is consistent with earlier work showing that droplets form branches from a diffusion-limited nucleation and growth process in a 2D confinement at low concentration \cite{Lu2015}.
Once formed, numerous droplets travelled toward the side channels in the negative x-direction before becoming immobilized on the wall surface or coalescing with other droplets. The lines with low fluorescent intensity indicated in the yellow boxes in Figure \ref{Regime 4 part 2}C show induced droplet flow. The droplets appear as lines due to the exposure time of the camera when taking fluorescence imagery. This induced flow may explain the relatively fast boundary displacement rate associated with Regime 4.
The optical images in Figures \ref{Regime 4 part 2}D and E indicate the qualitative similarity of droplet formation and boundary movement phenomena for 5\% and 8\% 1-octanol by mass in solution A. Differences arise from the number of droplets formed and droplet size with time.
The displacement of the boundary is plotted against the square root of time for the 2\%, 5\% and 10\% initial 1-octanol wt-\% cases. As indicated by the slopes added to each curve in Figure \ref{Regime 4 part 2}F, a slower followed by a more rapid diffusive displacement process is clearly evident for the 2\% and 10\% conditions. While qualitatively similar, the first diffusion-controlled process slows and has a longer duration as the initial 1-octanol wt-\% increases. In Regime 4, boundary displacement is faster than Regime 3 compositions, and there are seldom oscillations in displacement versus time plots.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.85\columnwidth]{./figs/Regime_4_part_2}
\caption{Regime 4 dynamics with a focus on low 1-octanol wt\%. (A) Ternary phase diagram indicating approximate the dilution path and subphases compositions for solution A with an initial 1-octanol composition of 2\% by mass. (B) Time-based optical images of triangular shaped boundary movement using 2\% by mass in solution A. (C) Time-based fluorescence images of 2\% 1-octanol by mass indicating the 1-octanol-rich and water-rich regions. Yellow boxes show droplets recirculating back to the boundary region. Time-based optical images comparing the 5\% 1-octanol (D) and 8\% 1-octanol (E) conditions. (F) Boundary displacement as a function of the square root of time plot comparing the 2\% octanol (blue), 5\% (green) and 10\% (red) conditions. Dashed lines show anticipated two-tiered diffusion-based displacement dynamics.}
\label{Regime 4 part 2}
\end{figure}
\subsection{Comparison of the displacement rates in the four regimes}
In Figure \ref{Summary Displacement}A, we compare the displacement rates in all four regimes presented. An effective diffusivity was calculated for each regime, as shown in Figure \ref{Summary Displacement}B. The highest effective diffusivity value was obtained in Regime 1 at 50\% of 1-octanol for a smooth boundary, which has an enhanced effective diffusivity of $8 \times 10^{-9}$ $m^2/s$. The effective diffusivity obtained is thus higher than the expected range of the mutual diffusivities between water and ethanol (1.1 $\times 10^{-10}$ to 1.2 $\times 10^{-9}$ $m^2/s$), and water and 1-octanol (2.0 $\times 10^{-10}$ to 7.3 $\times 10^{-10}$ $m^2/s$) showing the enhanced diffusion. In both high and low 1-octanol ratios of Regimes 1 and 4, the displacement rate of the boundary was faster at $\sim$ 15 ${\mu}m/s$ and $\sim$ 4 ${\mu}m/s$. In contrast, in the intermediate Regime 3, the rate is only $\sim$ 0.7 ${\mu}m/s$ as shown by the average velocities in Figure \ref{Summary Displacement}C. Boundary displacement in Regime 3 is slower than in the other regimes but remains 2 times the displacement rate expected for a purely-diffusive single-phase process obtained from the effective diffusivity.
To understand the dependence of displacement rate on 1-octanol concentration, including a shallow minimum at intermediate 1-octanol concentrations, we must consider competing effects. Differences in the energy released during phase separation with 1-octanol concentration are illustrated in Figure \ref{Summary Displacement}D. The energy released during the transition from a non-equilibrium single-phase state to an equilibrium two-phase state increases with 1-octanol concentration. To the extent that displacement rate depends on this thermodynamic driving force, one would anticipate an increase in displacement rate with 1-octanol concentration. This effect is clearly evident at high 1-octanol concentration. However, the displacement rate at low 1-octanol concentrations is somewhat higher than at intermediate 1-octanol concentrations. We attribute this apparent enhancement in displacement at low 1-octanol concentrations to droplet propulsion and the flow induced in the continuous phase from the self-propelling microdroplets in this concentration range, i.e., to how the energy of phase transition is dissipated.
Impacts of relative viscosity of the liquid phases, and the relative wettability of the liquid phases on the wall surface, which approach one at intermediate 1-octanol concentrations (near the plait point), are seen as secondary. Viewed from this perspective, the strong dependence of displacement rate on the initial composition is expected to be a general feature for displacing multi-component liquids undergoing phase separation. Further, compositions transitioning through Regimes 1 and 2 are preferred for displacing liquid, while compositions transitioning through Regimes 3 and 4 are preferred for promoting the separation and extraction of a component or a component category (1-octanol a model oil in this case) from the solution at lower overall displacement rates.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.90\columnwidth]{./figs/overview_of_regimes_displacement}
\caption{(A) Summary of displacement versus square root of time for 1-octanol compositions spanning the four regimes. (B) Effective diffusivity and average velocity (C) calculated from displacement versus time data in all four regimes. (D) Calculated enthalpy of transition from a non-equilibrium single phase state to two equilibrium states based on the Dortmund UNIFAC method.}
\label{Summary Displacement}
\end{figure}
\newpage
\subsection{Using water + ethanol binary mixtures to displace solution A}
Additional experiments were performed at high and low 1-octanol composition regime conditions using a displacing liquid containing 25\% ethanol by mass in water. Figure \ref{Variations to displacing liquid}A shows the approximate trajectory and associated tie line from the dilution process with this displacing liquid for a Regime 1 composition when water alone is used to displace it (see E1 in Table \ref{Experimental Conditions Additional Table}). The resulting water-rich phase formed by phase separation, in this case, has a composition corresponding to a Regime 4 solution composition, while the composition of the 1-octanol-rich phase remains qualitatively similar.
Figure \ref{Variations to displacing liquid}B shows an observed insoluble moving boundary and stranded 1-octanol-rich microdomains similar to the Regime 1 behaviour (Figure \ref{Regime 1}B) with fewer undulations. The smoother boundary displacement may be attributed to the reduction in the interfacial tensions and the pinning effect from the substrate with ethanol in the displacing liquid. Higher solubility of 1-octanol in ethanol solution may attribute to smaller microdomains.
Consistent results were also observed in Figure \ref{Variations to displacing liquid}C, where solution A was 10\% 1-octanol by mass, and the displacing liquid was 25\% ethanol by mass solution (see E4 in Table \ref{Experimental Conditions Additional Table}). In this case, droplets formed behind the moving boundary and dissolved slowly due to the high solubility of 1-octanol in displacing liquid.
The boundary displacement as a function of time summarized in Figure \ref{Variations to displacing liquid}D shows a time-invariant displacement rate for binary displacing liquid. Conditions with water as the displacing liquid, the boundary displacement rate for Regime 4 is 10 times greater, and for Regime 1, the rate is 0.9 times lower. The faster displacement rate for Regime 1 may be attributed to improvements in solubility of 1-octanol in the displacing liquid.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.8\columnwidth]{./figs/variationsdisplacingliquid}
\caption{(A) Ternary phase diagram indicating the approximate dilution path using 25\% ethanol in the displacing liquid at Regime 1 condition. (B) Optical image of Regime 1 condition using 25\% ethanol by mass in the displacing liquid. (C) Time-based optical images of Regime 4 condition using 25\% ethanol by mass in the displacing liquid. (D) Boundary displacement as a function time for ethanol in displacing liquid comparing the 10\% 1-octanol (blue) and 50\% (yellow) conditions. Dashed lines indicate linear regression from data. Data for pure water as displacing liquid in both regimes was added for comparison (triangles).}
\label{Variations to displacing liquid}
\end{figure}
\subsection{Displacing Water with Solution A}
By displacing water with solution A (a 1-octanol-containing ternary mixture), we are able to consider impacts of reverse flow arising in large scale porous media and to understand the impacts of relative wettability of the displacing liquid. Two cases, linked to experiments described in detail above, are illustrative.
Figure \ref{Variations to displacing liquid}A shows the insoluble boundary arising when the confined liquid is water and the displacing liquid is solution A. The solution A composition corresponds to Regime 1 (see R1 in Table \ref{Experimental Conditions Additional Table}) shown in Figure \ref{Regime 1}A. Droplets did not form within microdomains left behind by the boundary. Instead, small 1-octanol droplets were confined in the boundary zone, similar to Regime 2 conditions in Figure \ref{Regime 2}. The boundary displacement rate was 9 times faster compared to normal Regime 1 conditions.
Figure \ref{Variations to displacing liquid}B shows the displacement boundary when solution A composition corresponds to Regime 4 (see R4 in Table \ref{Experimental Conditions Additional Table}). In this case, 1-octanol-rich droplets and domains form at the boundary . The overall displacement rate of the boundary (Figure \ref{Variations to displacing liquid}C) is time-invariant (77 $\mu m/s$) and 50 times faster than normal Regime 4 conditions. This dramatic difference in displacement rate demonstrates that using displacing liquids with better wettability (lower interfacial tension) on the wall surface may provide an effective route to enhance fluid displacement rate in confinement. Such wettability manipulation is a common professional practice for enhanced oil recovery \cite{Aziz2020}.
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=1\columnwidth]{./figs/variationsdisplacingliquid2}
\caption{(A) Optical image showing confined water being displaced by solution A with a 50\% 1-octanol by weight composition. (B) Optical image showing confined water being displaced by solution A with a 10\% 1-octanol by weight composition. (C) Boundary displacement as a function time for reverse conditions comparing the 10\% octanol (blue) and 50\% (yellow) conditions. Dashed lines indicate linear regression of data. Data for water as displacing liquid and ternary solution as confined liquid in both regimes is added for comparison (triangles).}
\label{Variations to displacing liquid 2}
\end{figure}
\section{Conclusions}
We investigate the displacement process for phase-separating ternary liquid mixtures in 2D confinement. The model mixtures contain 1-octanol (a model oil), water (a model poor solvent) and ethanol (a model good solvent), where water (the poor solvent) is the displacing fluid. One of the confining walls is hydrophobic, and the other is hydrophilic. Four composition-dependent displacement regimes are identified. While the details of the displacement mechanism corresponding to each regime differ, order of magnitude enhancements over single-phase diffusive displacement arise. The displacement rate was also shown to be further enhanced with the addition of ethanol (good solvent) to the displacing liquid, improving the wettability on the wall surface. Reverse flow cases where confined water is displaced with the ternary mixture further illustrate the importance of the impact of relative wettability of coexisting phases on boundary displacement. The findings in this work are readily generalized using phase diagrams to provide insights and guidelines for the design of solution formulations for confined liquid displacement. Such displacement processes are important in geological, chemical, and biological processes in enhanced oil recovery, CO$_2$ sequestration, catalytic reactions and drug delivery systems.
\begin{acknowledgement}
This project is supported by the Natural Science and Engineering Research Council of Canada (NSERC) and Future Energy Systems (Canada First Research Excellence Fund). This research was undertaken, in part, thanks to funding from the Canada Research Chairs program.
\end{acknowledgement}
\newpage
\subsection{Thermodynamic calculations of excessive free energy from phase separation}
The calculations were performed to approximate the excess energy that was released from the liquid-liquid phase separation. The ternary liquid system was modelled with an activity coefficient model to predict the phase behaviour. The modified UNIFAC model or Dortmund UNIFAC model \cite{Weidlich1987,Gmehling1993} was used as the main thermodynamic model. For the simulations, we utilized both Symmetry Process Simulation software (mass balances) and a MATLAB code (energy balances) based on two works by Fredenslund and Weidlich. The van der Waals properties of the functional groups shown in Figure \ref{Thermodynamic Calculations}A, such as the group volume $R_k$, the surface area $Q_k$ are in Table \ref{vdW UNIFAC}, and the interaction parameters ($a_{mn}$, $b_{mn}$, $c_{mn}$) are to be found in Table \ref{UNIFAC interaction a}, \ref{UNIFAC interaction b} and \ref{UNIFAC interaction c}.
The lowest 1-octanol composition achieved by simulations was the 5\% wt-\% condition by Symmetry. At 5\% condition, we chose the water-rich phase to be around 75 times the amount of the created 1-octanol-rich phase. This was done to have a non-equilibrium composition close to the binodal curve that may approximate the true condition as phase separation occurs in confinement. An example of two initial compositions are depicted in Figure \ref{Thermodynamic Calculations}B. Symmetry is a process design software, meaning every calculation is in terms of a flowrate. To simulate our conditions, we had to fix the initial solution A feed to be a specific mass flowrate quantity. Only solution B was changed for the different initial conditions. The changes in the mass flowrates of B were all decided by a proportionality factor based on an initial optimized value. The proportions are all tabulated with the corresponding phase compositions in Table \ref{sim parameters}. The mass balance simulations were performed using a mixer unit in the program at a temperature of 25 $\degree$C and a pressure of 101.3 kPa.
The compositions from the Symmetry simulations in Table \ref{sim compositions} were then used in the MATLAB code that was created to only calculate the excess enthalpy as a function of the composition of each component. With this, we were thus able to calculate the excess enthalpy of each bulk phase, water-rich and 1-octanol-rich phases. The equilibrium energy states were calculated by summation of each water and 1-octanol-rich phase with their specific fraction in the final state. The non-equilibrium energy states were calculated using the bulk-phase composition. The calculated excess enthalpy values are shown in Figure \ref{Thermodynamic Calculations}C. With these values, it was possible to calculate the energy released for each initial composition, which should provide a positive value for every case.
\begin{table}[H]
\captionsetup{font = {small}}
\caption{Modified UNIFAC van der Waals parameters of functional groups used in simulation for 1-octanol system.}
\centering
\begin{tabular}{c| c c c}
\hline
Functional Group & Classification & Group Volume, & Surface Area, \\
& & $R_k$ & $Q_k$ \\
\hline
Methyl & Main group (n): 1 (-$CH_2$), & 0.6325 & 1.0608 \\
& Sub-group (m): 1 (-$CH_3$) & & \\
\hline
Methylene & Main group (n): 1 (-$CH_2$), & 0.6325 & 0.7081 \\
& Sub-group (m): 2 (-$CH_3$) & & \\
\hline
Hydroxide & Main group (n): 5 (-OH), & 1.2302 & 0.8927 \\
& Sub-group (m): 14 (Primary) & & \\
\hline
Water & Main group (n): 7 (-OH), & 1.7334 & 2.4561 \\
& Sub-group (m): 16 (Primary) & & \\
\hline
\end{tabular}
\label{vdW UNIFAC}
\end{table}
\begin{table}[H]
\captionsetup{font = {small}}
\caption{UNIFAC modified interaction parameters $a_{mn}$ used in simulation for 1-octanol system.}
\centering
\begin{tabular}{c| c c c c}
\hline
& $CH_{3}$ & $CH_{2}$ & OH & $H_{2}O$ \\
\hline
$CH_{3}$ & 0 & 0 & 2777 & 1391.3 \\
\hline
$CH_{2}$ & 0 & 0 & 2777 & 1391.3 \\
\hline
OH & 1606 & 1606 & 0 & -801.9 \\
\hline
$H_{2}O$ & -17.253 & -17.253 & 1460 & 0 \\
\hline
\end{tabular}
\label{UNIFAC interaction a}
\end{table}
\begin{table}[H]
\captionsetup{font = {small}}
\caption{UNIFAC modified interaction parameters $b_{mn}$ used in simulation for 1-octanol system.}
\centering
\begin{tabular}{c| c c c c}
\hline
& $CH_{3}$ & $CH_{2}$ & OH & $H_{2}O$ \\
\hline
$CH_{3}$ & 0 & 0 & -4.674 & -3.6156 \\
\hline
$CH_{2}$ & 0 & 0 & -4.674 & -3.6156 \\
\hline
OH & -4.746 & -4.746 & 0 & 3.824 \\
\hline
$H_{2}O$ & 0.8389 & 0.8389 & -8.673 & 0 \\
\hline
\end{tabular}
\label{UNIFAC interaction b}
\end{table}
\begin{table}[H]
\captionsetup{font = {small}}
\caption{UNIFAC modified interaction parameters $c_{mn}$ used in simulation for 1-octanol system.}
\centering
\begin{tabular}{c| c c c c}
\hline
& $CH_{3}$ & $CH_{2}$ & OH & $H_{2}O$ \\
\hline
$CH_{3}$ & 0 & 0 & 0.001551 & 0.001144 \\
\hline
$CH_{2}$ & 0 & 0 & 0.001551 & 0.001144 \\
\hline
OH & 0.0009181 & 0.0009181 & 0 & -0.007514 \\
\hline
$H_{2}O$ & 0.0009021 & 0.0009021 & 0.01641 & 0 \\
\hline
\end{tabular}
\label{UNIFAC interaction c}
\end{table}
\begin{table}[H]
\captionsetup{font = {small}}
\caption{The multiplier factors were used with the initial flowrate of the 5 \% 1-octanol condition as basis, which was found to be 1.9 $m^{3}/h$ from optimization with a fixed Solution A flowrate of 1 $m^{3}/h$. The optimization was based on having the set initial condition of 75 times the mass amount of water-phase to 1-octanol-rich phase.}
\centering
\begin{tabular}{c c c c}
\hline
Condition & Multiplier & Solution A & Solution B \\
& & $m^3/h$ & $m^3/h$\\
\hline
5/48/47 & 1 & 1.00 & 1.90 \\
\hline
8/47/45 & 0.625 & 1.00 & 1.19 \\
\hline
10/48/42 & 0.5 & 1.00 & 0.95 \\
\hline
15/48/37 & 0.333 & 1.00 & 0.63 \\
\hline
20/48/32 & 0.25 & 1.00 & 0.48 \\
\hline
30/48/22 & 0.167 & 1.00 & 0.32 \\
\hline
40/45/15 & 0.125 & 1.00 & 0.24 \\
\hline
50/40/10 & 0.1 & 1.00 & 0.19 \\
\hline
\end{tabular}
\label{sim parameters}
\end{table}
\begin{table}[H]
\captionsetup{font = {small}}
\caption{Compositions obtained from Symmetry for simulations described in terms of mass percentage 1-octanol/ethanol/water.}
\centering
\begin{tabular}{c c c c c}
\hline
Condition & Bulk Phase & Water-rich Phase & 1-Octanol-rich Phase & W/O Fraction \\
\hline
5/48/47 & 1.58/15.18/83.24 & 0.64/15.12/84.23 & 71.37/19.11/9.52 & 98.7/1.3 \\
\hline
8/47/45 & 3.40/19.95/76.66 & 0.94/19.76/79.29 & 64.16/24.47/11.37 & 96.12/3.88 \\
\hline
10/48/42 & 4.78/22.95/72.27 & 1.19/22.64/76.16 & 59.72/27.63/12.65 & 93.87/6.13 \\
\hline
15/48/37 & 8.65/27.68/63.67 & 1.71/26.99/71.31 & 53.08/32.13/14.79 & 86.48/13.52 \\
\hline
20/48/32 & 12.85/30.85/56.29 & 2.14/17.20/80.26 & 48.95/34.77/16.28 & 77.1/22.90 \\
\hline
30/48/22 & 21.78/34.85/43.36 & 2.74/32.63/64.63 & 44.43/37.50/18.07 & 54.33/45.67 \\
\hline
40/45/15 & 31.10/34.99/33.91 & 2.54/31.73/65.73 & 45.84/36.67/17.49 & 34.03/65.97 \\
\hline
50/40/10 & 40.64/32.51/26.85 & 1.93/28.46/69.61 & 50.85/33.58/15.57 & 20.86/79.14 \\
\hline
\end{tabular}
\label{sim compositions}
\end{table}
\section{Supporting figures}
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=1\columnwidth]{./figs/thermo_calc}
\caption{(A) Chemical structure of compounds in system with color coded functional groups used in modified UNIFAC calculations of enthalpy of mixing. Blue (methyl), red (methylene), green (hydroxide), and purple (water). (B) Phase diagram of 1-octanol system showing two of the bulk compositions that were used to calculate the excess enthalpy energy of the different initial compositions one at low 1-octanol concentration (blue) and high 1-octanol concentration (red). Initial compositions (diamond shape), 1-octanol-rich and water-rich subphases (star), and bulk compositions (circles) are all shown in this figure. (C) Calculated excess enthalpy with Dortmund UNIFAC method. Blue data indicates excess enthalpy of mixture at non-equilibrium state. Red data indicates excess enthalpy of mixture at equilibrium state obtained by summation of the enthalpy state of both 1-octanol-rich subphase and water-rich subphase in equilibrium.}
\label{Thermodynamic Calculations}
\end{figure}
\begin{figure}[H]
\includegraphics[trim={0cm 0cm 0cm 0cm}, clip, width=0.95\columnwidth]{./figs/supplementary_figure}
\caption{Optical images showing early development of the water-rich zone 2 within Regime 3 conditions.}
\label{supporting}
\end{figure}
\section{Supporting videos}
\noindent Video S1: Bright-field imagery of boundary displacement at Regime 1 (50\% 1-octanol by mass in Solution A) condition. Video shows boundary undulation and microdomain formation.
\noindent Video S2: Bright-field imagery of boundary displacement at Regime 2 (40\% 1-octanol by mass in Solution A) condition. Video shows boundary undulation and stagnant droplets left behind by the boundary.
\noindent Video S3: Boundary displacement at Regime 3 (30\% 1-octanol by mass in Solution A) condition. Video from 0 to 8 seconds shows the fluorescence imagery of the three-zone configuration at the boundary. From 8 to 20 seconds the fluorescence mode was changed to bright-field imaging.
\noindent Video S4: Boundary displacement at Regime 4 (10\% 1-octanol by mass in Solution A) condition. Video showing the transition from a triangular protrusion in the boundary to a line shaped protrusion. From 0 to 5 the video was in fluorescence mode, and from 5 to 25 it was in bright-field mode.
\newpage
|
\section{Introduction}
As of February 2021, the COVID-19 virus has claimed 2.4 million lives and infected 110 million individuals worldwide~\cite{whoCoronavirus}. Lack of effective treatments, high contagion rates~\cite{mohapatra2020recent}, long incubation periods~\cite{backer2020incubation, guan2020clinical, li2020early, lauer2020incubation}, and asymptomatic cases~\cite{byambasuren2020estimating, chang2020time, mizumoto2020estimating,ing2020covid} pose significant challenges in containing and eradicating pandemics. Recent pandemics, including gonorrhea~\cite{lajmanovich1976deterministic}, Ebola~\cite{dike2017susceptible}, and COVID-19~\cite{calafiore2020modified}, have accelerated the development of infection models. The main goal of epidemic model development is to identify conditions to eradicate the pathogen, and leverage the knowledge of these conditions to design mitigation strategies. Various infection models have been proposed, based on characteristics of individual pathogens, and studied in the literature, including susceptible-infected-susceptible (SIS), susceptible-infected-removed (SIR), and susceptible-infected-removed-susceptible (SIRS)~\cite{rock2014dynamics, mei2017dynamics}.
In this paper, we focus on the SIR epidemic model. We aim to expand on the SIR model, by exploring mutating viruses over networks, estimation of the underlying states, and distributed eradication strategies.
The patchwork response to COVID-19~\cite{haffajee2020thinking} gives rise to susceptible community subpopulations, with heterogeneous time-varying factors not previously explored by the SIR model. Extensions on the SIS model, studied in~\cite{van2009virus, ahn2013global, liu2019networked}, augment the compartmental epidemic models originated in~\cite{bailey1975mathematical} to include interactions between subpopulations of susceptible communities. Additionally, various advanced epidemic models consider
time-varying factors~\cite
pascual2005seasonal, gracy2019mutating, pare2015stability, pare2018epidemic, bokharaie2010spread, gracy2020asymptotic, prakash2010virus, liu2016threshold}. In this paper, we establish sufficient conditions for the set of healthy states of a networked time-varying SIR model to be globally exponentially stable (GES). These equilibrium states are not unique, as the final susceptible and removed states are dependent on the initial conditions and the time-varying infectious and healing parameters.
The delay in onset of COVID-19 symptoms~\cite{backer2020incubation, guan2020clinical, li2020early, lauer2020incubation}, large asymptomatic populations estimated between $17-81\%$~\cite{byambasuren2020estimating, chang2020time, mizumoto2020estimating, ing2020covid}, and delay in test results~\cite{bergquist2020covid} compromise the ability for accurate estimation of current infection states.
An estimation algorithm that incorporated a constant delay between the change in infection proportion and testing data was introduced in \cite{hota2020closed}. They studied the inference problem by using a Bayesian approach. Inspired by the delay characterization suggested in \cite{hota2020closed}, we propose a stochastic delay to model the unpredictability of the COVID-19 virus and testing strategies. We have developed methods for estimating the underlying epidemic states from testing data with a delay sampled from a geometric distribution, which cannot be completely filtered by the method suggested in \cite{hota2020closed}. The geometric delay model accounts for the stochastic effect of individuals failing to get tested immediately after exposure. We study the aggregated effect of each individual delay on the trajectory of confirmed cases and devise a method for estimating the underlying epidemic states of an SIR model from these delayed measurements. We also investigate the proposed method's estimation error, which provides insights for achieving an accurate estimation of the system states. We then employ this more realistic estimation to strategically eradicate a disease.
As proven in~\cite{hota2020closed}, the SIR epidemic model converges to a healthy state, however, an exponential convergence is not shown. Combining the modeling and inference approach allows us to develop two distributed control strategies capable of eradicating epidemic spread exponentially, at an equilibrium with a higher proportion of susceptible population. Decreasing the removed (recovered) and increasing the susceptibility proportion is of exceptional importance for the COVID-19 pandemic, as long term and severe health complications have been documented in the recovered populations, including impaired cognition~\cite{liotta2020frequent} and damage to cardiac tissue~\cite{mitrani2020covid}.
Our main result shows that by applying the estimated and the true susceptible states of each node, our proposed eradication strategies will guarantee global exponential stability of a healthy state of the overall network.
\subsection{Paper Contributions}
We summarize the main contributions of this paper as follows:
\begin{itemize}
\item We establish sufficient conditions for global exponential stability of the set of healthy states; see Theorem~\ref{thm:GES}.
\item We propose a stochastic framework which estimates the trajectories of the system states of the networked SIR model from testing data. Furthermore, we provide analytical expressions for the error of the estimation algorithm we propose; see Prop.~\ref{prop:estimation_error}
\item We propose two distributed eradication strategies for adjusting healing rates, one that is based on the true system states and the other one is based on the inferred system states. Both methods guarantee that the virus is eradicated within exponential time; see Theorem~\ref{thm:control_datadriven} and Corollary~\ref{coro:control_inferred}.
\end{itemize}
\subsection{Paper Outline}
We organize this paper as follows:
Section \ref{sec:model} lays down some basic assumptions and restates the well-known SIR model in the networked fashion,
and it presents the main problems studied in this paper.
Section \ref{sec:stability} first recalls preliminary results that are essential for stability analysis and then discusses the sufficient conditions for global exponential stability of a healthy state of the networked time-varying SIR models.
Section~\ref{sec:inference} covers the proposed techniques of estimating hidden epidemics states with the stochastic delay of tested individuals and testing data.
Section \ref{sec:eradication_strategies} covers the two distributed control strategies which ensure that the system converges to a healthy state in at least exponential time.
Section~\ref{sec:simulation} illustrates the results from Section \ref{sec:inference} and \ref{sec:eradication_strategies} with numerical simulations. Finally, in Section~\ref{sec:conclusion}, we summarize the main conclusions of this paper and discuss potential future directions.
\subsection{Notation}
We denote the set of real numbers, the non-negative integers, and the positive integers as $\mathbb{R}$, $\mathbb{Z}_{\geq 0}$, and $\mathbb{Z}_{\geq 1}$, respectively. For any positive integer $n$, we have $[n] =\{1,2,...,n \}$. The spectral radius of a matrix $A \in \mathbb{R}^{n\times n}$ is $ \rho(A)$. A diagonal matrix is denoted as diag$(\cdot)$. The transpose of a vector $x\in \mathbb{R}^n$ is $x^\top$. The Euclidean norm is denoted by $\lVert \cdot \rVert$. We use $I$ to denote identity matrix. We use $\mathbf{0}$ and $\mathbf{1}$ to
denote the vectors whose entries all equal 0 and 1, respectively. The dimensions of the vectors are determined by context. Given a matrix $A$, $A \succ 0$ (resp. $A \succeq 0$) indicates that $A$ is positive definite (resp. positive semidefinite), whereas $A \prec 0$ (resp. $A \preceq 0$) indicates that A is negative definite (resp. negative semidefinite). Let $G =(\mathbb{V},\mathbb{E})$ denote a graph or network where $\mathbb{V} = \{ v_1, v_2,..., v_n\}$ is the set of subpopulations, and $\mathbb{E} \subseteq \mathbb{V}\times \mathbb{V} $ is the set of edges. We denote the expectation of a random variable as $ {\rm I\!E}[\cdot]$
\section{Model and Problem Formulation}\label{sec:model}
Consider a time-varying epidemic network of $n$ subpopulations, where the size of subpopulation $v_i$ is $N_i \in \mathbb{Z}_{>0}$, and
the infection rates and healing rates could be time-varying. We denote $\beta_{ij}(t)$ as the infection rate from node $v_j $ to node~$v_i$ at time $t$, we denote $\gamma_i(t)$ as the healing rate of node~$v_i$ at time $t$. The proportions of the subpopulation at node~$v_i$
which are susceptible, infected, and recovered at time $t$ are denoted by $s_i(t),x_i(t)$ and $r_i(t)$, respectively. The deterministic continuous-time evolution of the SIR epidemic is given by
\begin{subequations}
\IEEEyesnumber\label{eq:ct_SIR}
\begin{align}
\dot{s}_i(t)&= -s_i(t) [\sum_{j=1}^n \beta_{ij}(t)x_j(t)], \label{eq:ct_SIRsub1}\\
\dot{x}_i(t) &= s_i(t)[\sum_{j=1}^n \beta_{ij}(t)x_j(t)] -\gamma_i(t) x_i(t), \label{eq:ct_SIRsub2}
\\
\dot{r}_i(t) &= \gamma_i(t) x_i(t), \;\ \forall i \in [n]
\label{eq:ct_SIRsub3}
\end{align}
\end{subequations}
We now state the discrete-time SIR epidemic dynamics obtained through Euler discretization of~\eqref{eq:ct_SIR}. For a small sampling time $h>0$, the discrete-time evolution of the SIR epidemic is given by
\begin{subequations}
\IEEEyesnumber\label{eq:dt_SIR}
\begin{align}
s_i(k+1) &= s_i(k) +h[-s_i(k)\sum_{j=1}^n \beta_{ij}(k) x_j(k)], \label{eq:dt_SIRsub1}\\
x_i(k+1) &= x_i(k) + h[s_i(k)\sum_{j=1}^n \beta_{ij}(k) x_j(k)-\gamma_i(k)x_i(k)], \label{eq:dt_SIRsub2}
\\
r_i(k+1) &= r_i(k) +h\gamma_i(k) x_i(k). \label{eq:dt_SIRsub3}
\end{align}
\end{subequations}
Equation \eqref{eq:dt_SIRsub2} can be rewritten as
\begin{equation}\label{eq:SIR_dynamic}
x(k+1) =x(k) +h [S(k)B(k) - \Gamma(k)]x(k),
\end{equation}
where $S(k) = \text{diag}(s(k))$, $B(k)$ is the matrix with $(i,j)$th entry $\beta_{ij}(k)$, and $\Gamma(k) = \text{diag}(\gamma_i(k))$. The spread of a virus over a network can be captured using a graph $G= (\mathbb{V},\mathbb{E})$,
where $\mathbb{E} = \{ (v_i, v_j) | \beta_{ij}(k) \neq 0 \}$ is the set of directed edges.
We make the following assumptions
in order for
the system in~\eqref{eq:dt_SIR}
to be
well defined.
\begin{assumption}\label{assume:one}
For every $i\in [n]$, $h \gamma_i(k) > 0 $ and $\forall j\in [n], \beta_{ij}(k) \geq 0$, for every $k\in \mathbb{Z}_{\geq 0}$.
\end{assumption}
\begin{assumption}\label{assume:two}
For every $i\in [n]$, $h \gamma_i(k) \leq 1 $ and $h \sum_j \beta_{ij}(k) \leq 1$, for every $k\in \mathbb{Z}_{\geq 0}$.
\end{assumption}
We have the following result which shares the same idea as the time-invariant model, proven in~\cite{hota2020closed}.
\begin{lemma}\label{lemma:one}
Suppose $s_i(0), x_i(0), r_i(0) \in [0,1]$, $s_i(0)+x_i(0)+r_i(0) =1$, and Assumptions~\ref{assume:one} and~\ref{assume:two} hold. Then, for all $k \in \mathbb{Z}_{\geq0}$,
\begin{enumerate}
\item $s_i(k), x_i(k), r_i(k) \in [0,1]$,
\item $s_i(k)+x_i(k)+r_i(k) =1$, and
\item $s_i(k+1) \leq s_i(k)$.
\end{enumerate}
\end{lemma}
\begin{definition}
We define the set of healthy states of~\eqref{eq:dt_SIR} as $\{s_i^*(k), x_i^*(k), r_i^*(k): i\in [n], k \in \mathbb{Z}_{\geq 0}\}$, where $x_i^*(k)=0$, $s_i^*(k)\in [0,1],$ and $ r_i^*(k)\in [0,1]$ for all $i \in [n]$.
\end{definition}
Given a network that is infected by a virus, our goal is to guarantee that each subpopulation $v_i$ converges to the set of healthy states in exponential time regardless of the initial conditions of the each subpopulations.
We now officially state the questions being studied in this paper:
\begin{enumerate}[label=(\roman*)]
\item \label{itm:first} For the system with dynamics given in~\eqref{eq:SIR_dynamic}, under what condition is the set of healthy states, i.e.,
$x(k)=\mathbf{0}$,
global exponentially stable (GES)?
\item \label{itm:second_a} Given the testing data, how can the stochastic framework be constructed to estimate
the susceptible,
infected,
and recovered proportions,
denoted by
$\widehat{s_i}(k)$, $\widehat{x_i}(k)$, and $\widehat{r_i}(k)$, respectively,
for each subpopulation $v_i$ in the network?
\item \label{itm:second_b} What is the estimation error of the stochastic framework that we proposed?
\item \label{itm:third_a} Given the knowledge of the conditions that ensure GES of a healthy state, i.e.,
$x(k)=\mathbf{0}$ and
$\widehat{s_i}(k)$ inferred from testing data,
how can we devise dynamic control algorithms which apply new healing rates $\widehat{\gamma_i}(k)$ to each agent in~\eqref{eq:SIR_dynamic} so that the epidemic is eradicated with a faster rate of convergence than the rate of exponential?
\end{enumerate}
\section{Stability Analysis} \label{sec:stability}
This section
presents
conditions that ensure
global exponential stability of the set of healthy states.
First, we introduce some preliminaries and then we present our main analysis results.
\subsection{Preliminaries}
In this subsection, we
recall
results that are crucial for understanding the rest of the paper.
\begin{lemma}\label{lemma:diagonal}
\cite{rantzer2011distributed}
Suppose that $M$ is a nonnegative matrix which satisfies $\rho(M)<1$. Then there exists a diagonal matrix $P \succ 0$ such that $M^\top P M -P \prec 0$.
\end{lemma}
Consider a system described as follows:
\begin{equation}\label{eq:lip}
x(k+1) = f(k,x(k))\phil{.
\end{equation}
\begin{definition}\label{def:GES}
An equilibrium point of \eqref{eq:lip} is GES is there exist positive constants $\alpha$ and $\omega$, with $0\leq \omega <1$, such that
\begin{equation}
\lVert x(k) \rVert \leq \alpha \lVert x(k_0) \rVert \omega^{(k-k_0)}, \forall k, k_0 \geq 0, \forall x(k_0) \in \mathbb{R}^n.
\end{equation}
\end{definition}
We recall a sufficient condition for GES of an equilibrium of \eqref{eq:lip}
from~\cite{vidyasagar2002nonlinear}.
\begin{lemma}\label{lemma:GES}
\cite[Theorem 28]{vidyasagar2002nonlinear}
Suppose there exists a function $V: \mathbb{Z}_+ \times \mathbb{R}^n \rightarrow \mathbb{R}$, and constants $a,b,c>0$ and $p>1$ such that $a\lVert x \rVert^p \leq V(k,x) \leq b \lVert x \rVert^p$, $\Delta V(k,x):= V(x(k+1))-V(x(k)) \leq -c\lVert x \rVert^p, \forall k \in \mathbb{Z}_{\geq0}$, and $\forall x (k_0)\in \mathbb{R}^n$, then $x(k)=\mathbf{0}$ is a globally exponential stable equilibrium of \eqref{eq:lip}.
\end{lemma}
\begin{lemma}\label{lemma:rate_GES}
\cite[Theorem 23.3]{rugh1996linear} Under the assumption of Lemma~\ref{lemma:GES}, the rate of convergence to
the origin
is upper bounded by an exponential rate of $\sqrt{1-(c/b)} \in [0,1)$, where $b$ and $c$ are defined in Lemma~\ref{lemma:GES}.
\end{lemma}
Note that a healthy state of the system in \eqref{eq:dt_SIR} is GES if Assumptions \ref{assume:one} and \ref{assume:two} hold and the condition in Definition~\ref{def:GES} and Lemma~\ref{lemma:GES} are satisfied for all $x(k_0) \in [0,1]^n$, since this is the domain where the model is well defined.
\subsection{Global Exponential Stability of the Healthy States}
In this subsection,
we present sufficient conditions for the global exponential stability of the set of healthy states of the system. We find the conditions by analyzing the spectral radius of the state transition matrix of~\eqref{eq:dt_SIRsub2}.
We define
\begin{align}
M(k) &= I-h\Gamma(k) +hB(k), \label{eq:M} \\
\hat{M}(k) &= I+ h[S(k)B(k) - \Gamma(k)].
\label{eq:Mhat}
\end{align}
Notice that $\hat{M}(k)$ is the state transition matrix of~\eqref{eq:dt_SIRsub2} and it can be written that
\begin{equation}\label{eq:M_M_hat}
\hat{M}(k) = M(k)-h(I-S(k))B(k).
\end{equation}
We use $M(k)$ and $\eqref{eq:M_M_hat}$ to illustrate the sufficient conditions for the GES of the set of healthy states in the subsequent theorem
\begin{theorem}\label{thm:GES}
Given Assumptions \ref{assume:one} and \ref{assume:two}, suppose for all $k \in \mathbb{Z}_{\geq0}$, $B(k)$ is symmetric.
If $\text{sup}_{k\in \mathbb{Z}_{\geq 0}} \rho(M(k))< 1$,
then the set of healthy states of \eqref{eq:dt_SIR} is GES.
\end{theorem}
\begin{proof}
See Appendix.
\end{proof}
Recall from the previous result that $M(k)$ is a nonnegative matrix which satisfies $\text{sup}_{k\in \mathbb{Z}_{\geq 0}} \rho(M(k))< 1$, such that $M^\top(k) Q(k+1) M(k) - Q(k)\prec 0$, where $Q(k)$ is a diagonal matrix defined in the Lyapunov function:
\begin{equation}
V(k,x) = x^\top Q(k)x.
\end{equation}
\begin{corollary}\label{coro:rate_exp}
Under the assumptions of Theorem~\ref{thm:GES}, the rate of convergence to a healthy state is upper bounded by an exponential rate of
$\sqrt{1-\frac{\sigma_3}{\sigma_2}}$, where $\sigma_2=\max_{k\in \mathbb{Z}_{\geq 0}} \lambda_{\text{max}}(Q(k))$, $\sigma_3 =\max_{k\in \mathbb{Z}_{\geq 0}} \lambda_{\text{min}}[Q(k)-M(k)^\top Q(k+1)M(k)]$
\end{corollary}
\begin{proof}
See Appendix.
\end{proof}
\begin{remark}
Notice that in Theorem~\ref{thm:GES}, $\text{sup}_{k\in \mathbb{Z}_{\geq 0}} \rho(M(k))< 1$ is the key condition which ensures the set of healthy states of~\eqref{eq:dt_SIR} is GES. We can interprete $\rho(M)$ in the context of epidemiology as the basic reproduction number of the virus over the network. Particularly, Theorem~\ref{thm:GES} affirms that given the time-varying parameters of the network satisfy the condition provided,
the mutating virus will exponentially converge to the set of healthy states.
\end{remark}
In this section, we found the conditions that ensure exponential convergence to the set of healthy states of~\eqref{eq:dt_SIRsub2} in Theorem~\ref{thm:GES}. This answers question~\ref{itm:first} in Section~\ref{sec:model}.
The stability condition can help the policymakers reallocate the medical resources, staff etc. which leads to modifying the parameters in~\eqref{eq:dt_SIR} so that the spreading of the virus stops completely.
One of the other factors that will assist in decision making is the COVID-19 observed testing data.
\section{State Estimation from Testing Data} \label{sec:inference}
In this section, we study how to estimate the epidemic states $(s(k), x(k), r(k))$ from testing data in order to design a feedback controller in the following section. One of the challenges of estimating the underlying system states is that the testing data on a given day does not capture the new infections on the same day. Instead, the testing data is a delayed representation of the change in the system. Characterizing the delay of each individual is difficult, because the delay is determined by numerous factors such as the incubation period of COVID-19, the duration of obtaining test results, the willingness of each individual to get tested, etc. Therefore, we propose a stochastic framework in this section to capture the factors which cause the testing delay.
\begin{definition}
The testing delay $\tau_i$ is the length of time between when an individual from subpopulation $v_i$ is infected and when their positive test result is reported.
\end{definition}
In our discrete-time model, we assume that $\tau_i \in \mathbb{Z}_{\geq 0}$.
We model the testing delay of each infected individual $\tau_i$ as two aggregate components to represent the uncertainty in the testing process:
\begin{equation}\label{eq:testing_delay}
\tau_i =\eta_i+\mathcal{Y}_i,
\end{equation}
where $\eta_i\in \mathbb{Z}_{\geq 0}$ is a constant and $\mathcal{Y}_i$ is sampled from a discrete time random variable
whose measurable space is~$\mathbb{Z}_{\geq 0}$.
\begin{remark}
In~\eqref{eq:testing_delay}, the constant component $\eta_i$ can be interpreted as the length of time needed to acquire testing results.
The random variable can be interpreted as the incubation period and/or the amount of time that it takes an individual to get tested after becoming infected.
\end{remark}
\begin{figure}
\centering
\begin{overpic}[width = \columnwidth]{Inference_Diagram_Blank.png}
\put(21,47){{\parbox{0.75\linewidth}{\small
$\Omega_i(k) = \big(C_i(k), D_i(k)\big)$ $\rightarrow \big( c_i(k), d_i(k)\big)$
}}}\normalsize
\put(21,22){{\parbox{0.75\linewidth}{\small
$\widehat{\Theta}_i(k-\tau_i)=\big(\widehat{s_i}(k-\tau_i),\widehat{x_i}(k-\tau_i),\widehat{r_i}(k-\tau_i)\big)$}}}\normalsize
\put(21,1){{\parbox{0.75\linewidth}{\small
$\widehat{\Theta}_i(k)=\big(\widehat{s_i}(k),\widehat{x_i}(k),\widehat{r_i}(k)\big)$}}}\normalsize
\end{overpic}
\caption{Estimation of System States from Testing Data
}
\label{fig:Diagram_Inference}
\end{figure}
First, we denote the set of estimated system states for subpopulation $v_i$ at time $k$ as $\widehat{\Theta}_i(k)=\big(\widehat{s_i}(k),\widehat{x_i}(k),\widehat{r_i}(k)\big)$, we denote the set of testing data recorded at time $k$ to be $\Omega_i(k)= \big(C_i(k), D_i(k)\big)$, where $C_i(k)$ is the number of confirmed cases at time $k$, and $D_i(k)$ represents the number of removed (recovered) cases at time $k$. In addition, the cumulative number of confirmed and removed cases at node $v_i$ are written as $\mathbb{C}_i(k) = \sum_{j=0}^k C_i(j)$ and $\mathbb{D}_i(k) = \sum_{j=0}^k D_i(j)$, respectively. Therefore, the number of active cases is calculated by $\mathbb{A}_i(k) =\mathbb{C}_i(k) - \mathbb{D}_i(k)
$.
Recall that the size of each subpopulation is $N_i$; we define $c_i(k)= \frac{C_i(k)}{N_i}$ and $d_i(k)= \frac{D_i(k)}{N_i}$ as the proportion of daily confirmed cases and removal, respectively. Note that $c_i(k), d_i(k) \in [0,1]$, for all $i\in[n], k\in \mathbb{Z}_{\geq 0}$.
The estimation procedure is illustrated in Fig.~\ref{fig:Diagram_Inference}.
We then study how to relate $c_i(k)$ to the underlying states. We define a vector space $\Pi_{T_1}$
as the space of all the proportions of daily number of confirmed cases from time step $k=T_1$ to time step $k=T_2+1$.
We define $\Xi_{T_1}$ as the vector of all the decreases in the proportion of susceptible individuals, $-\Delta s_i(k)$, from time step $k=T_1$ to time step $k=T_2+1$.
We denote $\Phi(T_1, T_2)$ as the transfer matrix which results in
\begin{equation}\label{eq:transfer_matrix}
\Pi_{T_1} = \Phi(T_1, T_2) \Xi_{T_1},
\end{equation}
where $\Phi(T_1, T_2)$ is a $(T_2-T_1+2) \times (T_2-T_1+2)$ matrix, which depends on the SIR dynamics, the testing strategies, and the delay.
When the testing delay is a constant, i.e., $\tau_i =\eta_i > 0$, the only non-zero entries in $\Phi(T_1, T_2)$ are: $ \Phi_{l+\eta_i, l} =1, l\in [T_2-T_1+2-\eta_i]$.
Since for all $k \in [T_1+ \eta_i,T_2+1 ]$, we can write $c_i(k)$ as
\begin{equation}
c_i(k) = -\Delta s_i(k-\eta_i).
\end{equation}
When the delay $\eta_i =0$, the transfer matrix $\Phi(T_1, T_2)=I$.
Because for all $k \in [T_1,T_2+1 ]$, we can write that
\begin{equation}
c_i(k) = -\Delta s_i(k).
\end{equation}
We now propose a stochastic testing framework to capture the delay between when an individual is infected and when they receive a positive test result. We first let $\eta_i =0$ in~\eqref{eq:testing_delay}, without the loss of generality. Furthermore, we assume that each infected individual at node $v_i$ has an equal probability $p_i^x \in (0,1]$
of receiving a diagnostic test each day starting from the day after they are infected.
Therefore, we model $\mathcal{Y}_i$ in~\eqref{eq:testing_delay} as a random variable following the geometric distribution,
with the probability of an infected individual acquiring a positive test $\delta$ days after infection being:
\begin{equation}\label{eq:geo_dist_single}
P(\mathcal{Y}_i =\delta) =p_i^x(1-p_i^x)^{\delta -1}
\end{equation}
for $\delta \in \mathbb{Z}_{\geq 1}$.
The geometric distribution of the testing delay models the number of days before an infected individual obtains a diagnostic test which represents the incubation period of COVID-19 and/or the unwillingness of each individual getting a test. We assume that the delay of each infected individual's distribution is i.i.d. (independent and identically distributed) from others. Furthermore, we assume that an infected individual can be tested only once. Based on~\cite{an2020clinical} and~\cite{abbasi2020promise}, we assume that even if an individual recovers from COVID-19, their antibody tests will still give positive results. We also assume that all the tests generate accurate results.
We now relate the proportion of confirmed cases $c_i(k)$ with the underlying states of the system. We define a binary random variable $\mathcal{X}_i(\nu)$ with $\mathcal{X}_i(\nu )=1$ (resp. $\mathcal{X}_i(\nu) =0$) if a randomly chosen individual from subpopulation $v_i$ became infected at time $\nu$. It can be written that
\begin{equation} \label{eq:infection_nu}
\mathcal{X}_i (\nu) = \begin{cases}
1 \ \ w.p. \ \ -\Delta s_i(\nu) \\
0 \ \ w.p. \ \ 1+\Delta s_i(\nu),
\end{cases}
\end{equation}where, from~\eqref{eq:dt_SIR}, $-\Delta s_i(\nu) =s_i(\nu-1)-s_i(\nu)=hs_i(\nu-1)\sum_{j}\beta_{ij}x_j(\nu-1)\geq 0$ for all $\nu \geq 0$.
We define the binary random variable $\mathcal{T}_i(\mu,\delta)$, with $\mathcal{T}_i(\mu,\delta)=1$
if a randomly chosen individual acquired a positive test at time $\mu$ and was infected $\delta$ days before $\mu$.
Now we rewrite $\nu $, in \eqref{eq:infection_nu}, as $\mu - \delta$.
From~\eqref{eq:geo_dist_single}, the conditional probability $P(\mathcal{T}_i(\mu,\delta)=1 |\mathcal{X}_i(\mu - \delta)=1)$
is given by the geometric probability mass function (pmf) $p_i^x(1-p_i^x)^{\delta-1}$ and represents the probability of an infected individual acquiring a positive test specifically $\delta$ days after being infected.
Hence, the joint pmf of the two random variables $\mathcal{X}_i(\mu -\delta) , \mathcal{T}_i(\mu,\delta) $ is written as:
\begin{equation}
P_{\mathcal{X}_i,\mathcal{T}_i}(\mu -\delta,\mu) = P(\mathcal{X}_i(\mu -\delta)
\cap
\mathcal{T}_i(\mu,\delta)),
\end{equation}
which is interpreted as the probability that
a randomly chosen individual
became
infected
at time $\mu -\delta$ and
acquired
a positive test at time $\mu$, where $\mu -\delta , \mu \in [T_1, T_2]$.
Therefore, the joint pmf $P_{\mathcal{X}_i,\mathcal{T}_i}(\mu-\delta,\mu)$ is calculated as:
\begin{align}
& P_{\mathcal{X}_i, \mathcal{T}_i} (\mathcal{X}_i(\mu-\delta)=1 \cap \mathcal{T}_i(\mu,\delta)=1) \nonumber \\ &= P(\mathcal{T}_i(\mu,\delta)=1 |\mathcal{X}_i(\mu - \delta)=1)P(\mathcal{X}_i(\mu -\delta)=1) \nonumber \\ &= p_i^x(1-p_i^x)^{\delta-1}[-\Delta s_i(\mu- \delta)], \label{eq:joint_pmf}
\end{align}
\begin{align}
& P_{\mathcal{X}_i, \mathcal{T}_i} (\mathcal{X}_i(\mu-\delta)=0 \cap \mathcal{T}_i(\mu,\delta)=1) \nonumber \\ &= P(\mathcal{T}_i(\mu,\delta)=1 |\mathcal{X}_i(\mu - \delta)=0)P(\mathcal{X}_i(\mu -\delta)=0) \nonumber \\ &=0[1+\Delta s_i(\mu-\delta)]= 0, \label{eq:zero}
\end{align}
since we assume that the test results are accurate. Similarly,
\begin{align}
& P_{\mathcal{X}_i, \mathcal{T}_i} (\mathcal{X}_i(\mu-\delta)=1 \cap \mathcal{T}_i(\mu,\delta)=0) \nonumber \\ &= P(\mathcal{T}_i(\mu,\delta)=0 |\mathcal{X}_i(\mu - \delta)=1)P(\mathcal{X}_i(\mu -\delta)=1) \nonumber \\ &= [1-p_i^x(1-p_i^x)^{\delta-1}][-\Delta s_i(\mu- \delta)], \nonumber
\end{align}
\begin{align}
& P_{\mathcal{X}_i, \mathcal{T}_i} (\mathcal{X}_i(\mu-\delta)=0 \cap \mathcal{T}_i(\mu,\delta)=0) \nonumber \\ &= P(\mathcal{T}_i(\mu,\delta)=0 |\mathcal{X}_i(\mu - \delta)=0)P(\mathcal{X}_i(\mu -\delta)=0) \nonumber \\ &= 1+\Delta s_i(\mu- \delta). \nonumber
\end{align}
Let $\mathcal{W}_i(\mu)$ be the marginal distribution of $\mathcal{T}_i(\mu,\delta)$ over the set of feasible delays, $\delta$, with its pmf being
the
probability of a random
individual acquiring
a positive test at time $\mu$:
\begin{align}
&P_{\mathcal{W}_i}(\mathcal{W}_i(\mu)=1) \nonumber \\
& =\sum_{\delta =1}^{\mu -T_1} P_{\mathcal{X}_i,\mathcal{T}_i}(\mathcal{X}_i(\mu-\delta) \cap \mathcal{T}_i(\mu,\delta)=1) \nonumber \\
& = \sum_{\delta =1}^{\mu-T_1} p_i^x (1-p_i^x)^{\delta-1}[ -\Delta s_i(\mu-\delta)], \nonumber
\end{align}
by combining~\eqref{eq:joint_pmf} and~\eqref{eq:zero}.
Therefore, the number of confirmed cases at time $k$ is calculated as \begin{align}
C_i(k) &=
{\rm I\!E}[\sum_{l=1}^{N_i} \mathcal{W}_i(k) ] \nonumber \\
&=\sum_{l=1}^{N_i} {\rm I\!E}[ \mathcal{W}_i(k) ] \label{eq:indicator} \\
& =N_i \sum_{\delta=1}^{k-T_1} p_i^x (1-p_i^x)^{\delta-1}[ -\Delta s_i(k-\delta)], \label{eq:C_i_k
\end{align}
where~\eqref{eq:indicator} holds because of the linearity of expectation and since the testing delays are i.i.d.
Hence, by combining $c_i(k) =\frac{C_i(k)}{N_i}$ and~\eqref{eq:C_i_k} for all $k\in [T_1+1,T_2+1]$, the transfer matrix $\Phi(T_1, T_2)$
in~\eqref{eq:transfer_matrix} is written as
\begin{align}
&\Phi(T_1, T_2) \nonumber \\ &= \begin{bmatrix}
0 & 0 & 0 & 0 &\hdots & 0 \\
p_i^x & 0 & 0 & 0 & \hdots & 0 \\
p_i^x (1-p_i^x) & p_i^x & 0 & 0 & \hdots & 0 \\
p_i^x (1-p_i^x)^2 &p_i^x (1-p_i^x) & p_i^x & 0 & \hdots & 0 \\
\vdots & \vdots & \ddots & \ddots & \ddots &\vdots \\
p_i^x(1-p_i^x)^{q-2} & p_i^x(1-p_i^x)^{q-3} &\hdots &\hdots & p_i^x & 0
\end{bmatrix},\label{eq:transfer_geo}
\end{align}
where $q = T_2-T_1+1$.
By combining~\eqref{eq:transfer_matrix},~\eqref{eq:transfer_geo},
we obtain that
\begin{equation}\label{eq:model_daily}
c_i(k) = p_i^x (-\Delta s_i(k-1))+ (1-p_i^x)c_i(k-1),
\end{equation}
for all $k \in [T_1+1,T_2+1]$. Meanwhile, we set $c_i(k) =0$ for all $k \notin [T_1+1, T_2+1]$, since no testing occurs.
\begin{remark}
The proportion of daily confirmed cases $c_i(k)$ in \eqref{eq:model_daily} consists of two terms: the first term $p_i^x (-\Delta s_i(k-1))$ can be interpreted as an infected individual's urgency in obtaining a test, and the second term $(1-p_i^x)c_i(k-1)$ captures the unwillingness/unlikeliness of an infected individual acquiring a test.
\end{remark}
Finally, we relate the proportion of the daily number of recoveries, i.e., $d_i(k)$, with the underlying states. In the data collected, $d_i(k)$ corresponds to the change in the proportion of recovered individuals and the total number of known active cases $\mathbb{A}_i(k-1)$. We assume
\begin{equation}\label{eq:recovered_data}
d_i(k) \sim \mathtt{Bin}\Big(\frac{\mathbb{A}_i(k-1)}{N_i}, h\gamma_i(k-1)\Big).
\end{equation}
Namely, each known active case recovers with healing rate $h \gamma_i(k-1)$. From~\cite{hota2020closed}, when the number of active
cases is large, $d_i(k)$ is approximately equal to $\frac{h\gamma_i(k-1) \mathbb{A}_i(k-1)}{N_i}$.
The above analysis links the collected data proportions with the underlying states of the system. If we acquire the parameter: $p_i^x$,
we will be able to estimate the state systems as follows:
\begin{definition}\label{def:initial}
We assume that: $\widehat{x_i}(k) =\widehat{x_i}(0)$, $\widehat{r_i}(k) =\widehat{r_i}(0)$, and $\widehat{s_i}(k)=\widehat{s_i}(0)$, where $\widehat{x_i}(0), \widehat{r_i}(0), \widehat{s_i}(0) \in [0,1]$ for all $i\in [n], k< T_1$. Given the testing data set $\Omega_i(k)$ collected from time step $T_1+1$ to $T_2+1$, according to~\eqref{eq:model_daily}, we define the estimated proportion of new infections at node $v_i$ as
\begin{equation}\label{eq:infer_sus_geo}
-\widehat{\Delta s}_i(k) =\frac{c_i(k+1)-(1-p_i^x)c_i(k)}{p_i^x } , k\in [T_1,T_2].
\end{equation}
\end{definition}
Notice that when $p_i^x =1$,~\eqref{eq:infer_sus_geo} becomes:
\begin{equation*}
-\widehat{\Delta s}_i(k) = c_i(k+1), k\in [T_1,T_2],
\end{equation*}
which can be interpreted as: every infected individual will be tested the day after being infected. Hence, the estimated change in proportion of infection on a given day $k$ exactly equals to the fraction of the number of positive cases on the next day $k+1$.
Moreover, we let $\widehat{\Delta r}_i(k) =0$ for $k =T_1$. Note that the following equality holds from the formulation of the SIR model:
\begin{align*
\widehat{\Delta s}_i(k) + \widehat{\Delta x}_i(k) + \widehat{\Delta r}_i(k) = 0.
\end{align*}
We further define that
\begin{IEEEeqnarray}{CC}
\IEEEyesnumber\label{eq:infer_SIR_st}
\widehat{s_i}(k) &= \widehat{s_i}(k-1)+ \widehat{\Delta s}_i(k), \nonumber \\
\widehat{x_i}(k) &= \widehat{x_i}(k-1)+ \widehat{\Delta x}_i(k),
\\
\widehat{r_i}(k) &= \widehat{r_i}(k-1)+ \widehat{\Delta r}_i(k),\nonumber \label{eq:infer_SIRsub3}
\end{IEEEeqnarray}
for $k \in [T_1,T_2]$.
According to~\eqref{eq:recovered_data} and~\cite{hota2020closed}, the change in the proportion of recovered individuals at node $v_i$ can be inferred as
\begin{equation}\label{eq:infer_recovered_change}
\widehat{\Delta r}_i(k) = \frac{N_id_i(k)}{\mathbb{A}_i(k-1)}\widehat{x_i}(k-1), k \in [T_1+ 1, T_2],
\end{equation}
where $\widehat{x_i}(k-1)$ is calculated from \eqref{eq:infer_sus_geo} and~\eqref{eq:infer_SIR_st}. When $\mathbb{A}_i(k-1)=0$, we assume $\widehat{\Delta r}_i(k) =0$.
Therefore, if the testing data $\Omega_i(k)$ is available over an interval $k\in [T_1+1, T_2+1]$, we can estimate the states of the system by repetitively applying~\eqref{eq:infer_sus_geo},~\eqref{eq:infer_SIR_st}, and~\eqref{eq:infer_recovered_change} with the initial conditions, i.e., $\widehat{s_i}(0),$ $\widehat{x_i}(0),$ and $ \widehat{r_i}(0)$, assumed for the geometric distribution model. This addresses question~\ref{itm:second_a} in Section~\ref{sec:model}.
\begin{assumption}\label{assum:inf}
We assume that $c_i(k) =0$ for all $k \in [T_1]\cup \{0\}$
and the initial inferred susceptible proportion is $\widehat{s_i}(0)$.
\end{assumption}
\begin{remark}
When estimating the system states, we first assume an initial condition for the system based on reality. We also assume that outside of the testing period, the proportion of positive cases collected is zero.
\end{remark}
\begin{proposition}\label{prop:estimation_error}
Under Assumption~\ref{assum:inf}, the error of the inference method in~\eqref{eq:infer_sus_geo}
\eqref{eq:infer_recovered_change}
at time $k$
is given by
\begin{equation}\label{eq:infer_error_bound}
| \widehat{s_i}(k) -s_i(k) | = \left|\widehat{s_i}(0)-s_i(0)-\sum_{l=1}^{T_1-1}\Delta s_i(l)
\right|,
\end{equation}
for all $k \geq T_1$
\end{proposition}
\begin{proof}
From~\eqref{eq:dt_SIRsub1}, we first represent $s_i(k)$ by:
\begin{equation}\label{eq:sus_true_sum}
s_i(k) = s_i(0) + \sum_{l=1}^k \Delta s_i(l).
\end{equation}
Now, we characterize $\widehat{s_i}(k)$:
\begin{align}
\widehat{s_i}(k) &= \widehat{s_i}(0) + \sum_{l=1}^k \widehat{\Delta s}_i(l) \nonumber \\ \label{eq:sum_hat_sus_2}
&= \widehat{s_i}(0) + \sum_{l=T_1}^k \widehat{\Delta s}_i(l) \\\label{eq:sum_hat_sus_3}
& = \widehat{s_i}(0) - \frac{c_i(k+1)}{p_i^x}-\sum_{l=T_1+1}^k c_i(l),
\end{align}
where~\eqref{eq:sum_hat_sus_2} is written because $-\widehat{\Delta s}_i(l)=0$ for all $l\leq T_1-1$ in~\eqref{eq:infer_sus_geo}. We acquired~\eqref{eq:sum_hat_sus_3} through representing each $\widehat{\Delta s}_i(l)$, $l\geq T_1$ by~\eqref{eq:infer_sus_geo} and following Assumption~\ref{assum:inf}.
By applying~\eqref{eq:model_daily}, we calculate the $\sum_{l=T_1+1}^k c_i(l)$ on the R.H.S. of~\eqref{eq:sum_hat_sus_3} as
\begin{align}\label{eq:sum_c_i(j)}
&\sum_{l=T_1+1}^k c_i(l) =-p_i^x \sum_{l=T_1}^{k-1} \Delta s_i(l) +(1-p_i^x)\sum_{l=T_1+1}^{k-1} c_i(l) \\\label{eq:sum_c_i(j)_2}
& \ \ \ \ =-p_i^x \sum_{l=T_1}^{k-1} \Delta s_i(l) + (1-p_i^x) \Big[\sum_{l=T_1+1}^k c_i(l) -c_i(k) \Big],
\end{align}
since $\sum_{l=T_1+1}^{k-1} c_i(l)= \sum_{l=T_1+1}^k c_i(l) -c_i(k)$.
We can
reorganize~\eqref{eq:sum_c_i(j)_2} and acquire:
\begin{equation}\label{eq:sum_c_i_clearversion}
\sum_{l=T_1+1}^k c_i(l) = -\sum_{l=T_1}^{k-1} \Delta s_i(l) - \frac{1-p_i^x}{p_i^x} c_i(k).
\end{equation}
Hence, we replace $\sum_{l=T_1+1}^k c_i(l)$ on the R.H.S. of~\eqref{eq:sum_hat_sus_3} with~\eqref{eq:sum_c_i_clearversion} and obtain:
\begin{align}
\widehat{s_i}(k) & = \widehat{s_i}(0) - \frac{c_i(k+1)}{p_i^x}+\sum_{l=T_1}^{k-1} \Delta s_i(l)
+ \frac{1-p_i^x}{p_i^x} c_i(k) \label{eq:s_hat_1} \\ \label{eq:s_hat_2}
&= \widehat{s_i}(0) + \sum_{l=T_1}^{k} \Delta s_i(l)
\end{align}
where~\eqref{eq:s_hat_2}
follows from
writing $c_i(k+1)$ in~\eqref{eq:s_hat_1}
as $p_i^x (-\Delta s_i(k))+(1-p_i^x)c_i(k)$, using~\eqref{eq:model_daily}. Therefore, we can calculate $ | \widehat{s_i}(k) -s_i(k) |$ by comparing~\eqref{eq:sus_true_sum} with~\eqref{eq:s_hat_2} and yield the result.
\end{proof}
Prop.~\ref{prop:estimation_error} provides an analytical expression of the estimation error given the initial susceptible level assumed and the start testing time. Hence, Prop.~\ref{prop:estimation_error} solves question~\ref{itm:second_b} in Section~\ref{sec:model}.
\begin{corollary}\label{coro:est_error}
In Prop.~\ref{prop:estimation_error}, if we assume that $\widehat{s_i}(0)=1$, then we can write that $\widehat{s_i}(k)\geq s_i(k)$, for all $k \geq T_1$. Moreover, if we assume the inferred initial conditions
correctly, i.e., $\widehat{s_i}(0)=s_i(0)$, and $T_1= 1$, then
the algorithm will estimate the susceptible state perfectly.
\end{corollary}
\begin{remark}
The result of Prop.~\ref{prop:estimation_error} consists of two parts: $\widehat{s_i}(0)-s_i(0)$ and $-\sum_{l=1}^{T_1-1}\Delta s_i(l)$. The first component depends on the difference between the inferred initial susceptible level and true initial susceptible level. The second component depends on the start testing date. Therefore, the accuracy of the estimation algorithm corresponds to the estimated initial condition and how early the testing data is collected.
\end{remark}
\noindent
We will explore this error via simulations in Section \ref{sec:simulation}.
By estimating the proportion of infected individuals in a subpopulation of a network, we are able to acquire the estimation of the infection prevalence in the whole system. These inferred states provide an understanding of the epidemic and important factors for designing eradication schemes for infectious diseases.
\section{Distributed Eradication Strategy}\label{sec:eradication_strategies}
In this section, we propose two distributed strategies that employ the true states and the estimated states, respectively, and guarantee the eradication of the virus in at least exponential time.
We propose the following healing rate to control the epidemic spread over the network:
\begin{equation}\label{eq:control_scheme_healing_rate}
\widetilde{\gamma}_i(k) =s_i(k) \sum_{j=1}^n \beta_{ij}(k) +\epsilon_i, \;\ i \in [n],
\end{equation}
where
$\epsilon_i>0$, for each $i \in [n]$.
This algorithm can be understood as boosting the healing rate of each subpopulation separately by providing effective medication, medical supplies, and/or healthcare workers.
\begin{theorem}\label{thm:control_datadriven}
Consider the system in~\eqref{eq:dt_SIR} and assume that
\begin{enumerate}
\item $0\leq h \sum_j \beta_{ij}(k) < 1$,
$\forall i \in [n]$ and $\forall k \in \mathbb{ Z}_{\geq 0}$,
\item $B(k)$ is symmetric and irreducible $\forall k \in \mathbb{ Z}_{\geq 0}$,
\item $\exists \epsilon_i$ small enough that $h\widetilde{\gamma}_i(k) <1$,
$\forall i \in [n],
k \in \mathbb{ Z}_{\geq 0}$.
\end{enumerate}
Then the algorithm~\eqref{eq:control_scheme_healing_rate} guarantees GES of the set of healthy states and $x(k)$ converges to $\mathbf{0}$ with at least an exponential rate.
\end{theorem}
\begin{proof}
By substituting \eqref{eq:control_scheme_healing_rate} into \eqref{eq:dt_SIR}, we obtain
\begin{align}\label{eq:infected_control}
&x_i(k+1) = x_i(k) + \nonumber \\ & h\{s_i(k)\sum_{j=1}^n \beta_{ij}(k) x_j(k) - [s_i(k) \sum_{j=1}^n \beta_{ij}(k) +\epsilon_i]x_i(k)\}.
\end{align}
The state transition matrix of \eqref{eq:infected_control} can be written as
\begin{equation}
\widetilde{M}(k) = I+h[S(k)B(k) -(S(k)\text{diag}(B(k) \mathbf{1}_{n\times 1}) + \text{diag}(\epsilon_i))].
\end{equation}
For any $i
, j\in [n], j\neq i$, the entries of the $i$-th row of $\widetilde{M}(k)$ are
\begin{equation}
\widetilde{m}_{ii}(k) = 1-h[s_i(k)\sum\limits_{j\neq i}^n \beta_{ij}(k)+\epsilon_i],
\end{equation}
\begin{equation}
\widetilde{m}_{ij}(k) = hs_i(k)\beta_{i,j}(k) ,
\end{equation}
which satisfies the following inequality
\begin{equation}
\widetilde{m}_{ii}(k)+ \sum\limits_{j\neq i}^n\widetilde{m}_{ij}(k) \leq 1- h\min \{\epsilon_i\}, \forall i \in n.
\end{equation}
Therefore, by Gershgorin circle theorem, the spectral radius of $\widetilde{M}(k) $ is upper bounded by $1- h\min \{\epsilon_i\}$:
\begin{equation}
\rho(\widetilde{M}(k)) \leq 1- h\min \{\epsilon_i\}.
\end{equation}
Since we have $x(k+1) = \widetilde{M}(k) x(k)$ and $x(k) \geq 0$ for all $k$, we can write that $\lVert x(k+1) \rVert \leq[1- h\min \{\epsilon_i\}] \lVert x(k) \rVert$ for all $k$. Since $\epsilon_i >0, \forall i\in n$, we obtain that, for all $x_i(0)\in [0,1]^n$,
\begin{equation}\label{eq:converge}
\lVert x(k) \rVert \leq [1- h\min \{\epsilon_i\}]^k \lVert x(0) \rVert \leq e^{-k h\min \{\epsilon_i\}} \lVert x(0) \rVert,
\end{equation}
where the second inequality holds by
Bernoulli's inequality~\cite{carothers2000real}, \begin{equation}\label{eq:bernoulli}
e^x= \lim_{n\rightarrow \infty}(1+\frac{x}{n})^n \geq 1+x.
\end{equation}
Hence, $x(k)$ converges to $\mathbf{0}$ with an exponential rate of at least $h\min \{\epsilon_i\}$.
Therefore,
the set of healthy states is GES
\end{proof}
\begin{remark}
The control strategy proposed in Theorem~\ref{thm:control_datadriven} can be interpreted as follows: if the healing rate of each subpopulation is appropriately increased according to its susceptible proportion, for example by distributing effective medication, medical supplies, and/or healthcare workers to each subpopulation,
then the epidemic will be eradicated with at least an exponential rate. This theorem provides decision makers
insight into, given sufficient resources, how to allocate
medical supplies and healthcare workers
to different subpopulations
so that the epidemic can be eradicated quickly. Furthermore, Theorem~\ref{thm:control_datadriven} provides sufficient conditions for guaranteeing an exponentially decreasing $\lVert x(k) \rVert$ for all $k$ when the conditions apply. In other words, implementing the control strategy in Theorem~\ref{thm:control_datadriven} at full length will prevent the potential upcoming waves of the epidemic in the 2-norm sense of $x(k)$.
\end{remark}
Using the estimation results from Section~\ref{sec:inference}, we consider the following healing rate:
\begin{equation}\label{eq:control_scheme_healing_rate_inferred}
\widehat{\widetilde{\gamma}}_i(k) =\widehat{s_i}(k) \sum_{j=1}^n \beta_{ij}(k) +\epsilon_i, \;\ i \in [n],
\end{equation}
where $\widehat{s_i}(k)$ is the estimated susceptible rate from~\eqref{eq:infer_SIR_st}.
\begin{corollary}\label{coro:control_inferred}
Consider the system in~\eqref{eq:dt_SIR} and assume that
\begin{enumerate}
\item $0\leq h \sum_j \beta_{ij}(k) \leq 1$ , $\forall i \in [n]$ and $\forall k \in \mathbb{ Z}_{\geq 0}$,
\item $B(k)$ is symmetric and irreducible $\forall k \in \mathbb{ Z}_{\geq 0}$.
\item $\exists \epsilon_i$ small enough that $h\widehat{\widetilde{\gamma}}_i(k) <1$, $\widehat{s_i}(0)=1$
$\forall i \in [n]$ and $\forall k \in \mathbb{ Z}_{\geq 0}$.
\end{enumerate}
Then the algorithm~\eqref{eq:control_scheme_healing_rate_inferred} guarantees GES of the set of healthy states and $x(k)$ converges to $\mathbf{0}$ with at least an exponential rate.
\end{corollary}
\begin{proof}
Similar to the proof of Theorem~\ref{thm:control_datadriven}, we substitute \eqref{eq:control_scheme_healing_rate_inferred} into \eqref{eq:dt_SIR} and obtain the state transition matrix for $x_i(k)$:
\begin{equation}
\widehat{\widetilde{M}}(k) = I+h[S(k)B(k) -(\widehat{S}(k)\text{diag}(B(k) \mathbf{1}_{n\times 1}) + \text{diag}(\epsilon_i))],
\end{equation}
where $\widehat{S}(k) =\text{diag}(\widehat{s_i}(k))$.
For any $i
, j\in [n]$, $j\neq i$, the entries of the $i$-th row of $\widehat{\widetilde{M}}(k)$ satisfy:
\begin{align}
\widehat{\widetilde{m}}_{ii}(k)+ \sum\limits_{j\neq i}^n\widehat{\widetilde{m}}_{ij}(k)
& \leq 1- h\min \{\epsilon_i\},
\end{align}
since from Corollary~\ref{coro:est_error} we know that when we assume that $\widehat{s_i}(0)=1$,
we obtain $\widehat{s_i}(k)\geq s_i(k)$ for all $i\in [n]$. Consequently, by the Gershgorin circle theorem, we obtain that the spectral norm of $\widehat{\widetilde{M}}(k)$ is upper bounded by $ 1- h\min \{\epsilon_i\}$.
Therefore, by referring to~\eqref{eq:converge} and~\eqref{eq:bernoulli}, we acquire that
the set of healthy states is GES
\end{proof}
Theorem~\ref{thm:control_datadriven} (resp. Corollary~\ref{coro:control_inferred}) has proven that given the true (resp. estimated) susceptible state the distributed eradication strategy proposed eradicates the virus with at least an exponential rate. Therefore, question~\ref{itm:third_a} from Section~\ref{sec:model} has been addressed here.
In this section, we have presented two distributed eradication strategies based on the true and estimated system states. Both strategies ensure that the SIR epidemics converge to the sets of healthy states exponentially.
We compare the two strategies with numerical simulations in Section~\ref{sec:simulation}, and study how a system will react if the eradication strategies are removed too early
\section{Simulations} \label{sec:simulation}
\begin{figure}
\centering
\includegraphics[width=.18\textwidth]{figs/Topo_Graph_Indiana.png}\includegraphics[width=.28\textwidth]{figs/Infection_Indiana_new.png}
\caption{Graph topology in the map of the state of Indiana~\cite{map2013} analyzed and the evolution of infected proportion in each city}
\label{fig:Graph_Topo_Indiana}
\end{figure}
In this section, we simulate a virus spreading over a static network
with 5 nodes in Fig~\ref{fig:Graph_Topo_Indiana} to illustrate our results. The nodes are modeled after the five metropolitan areas with a population over 150,000 in northern Indiana, U.S.: Gary (G), Lafayette (L), Indianapolis (I), Fort Wayne (F) and South Bend (S). Two nodes are neighbors if there is a major highway connecting
them.
We set the initially infected proportion to be $0.02$ at node I and $0.01$ at node G and 0 elsewhere. The infection rates, healing rates, and the size of each subpopulation
are static and presented in Table.~\ref{table:parameters}.
The evolution of the infected proportion
for each city
is shown in Fig.~\ref{fig:Graph_Topo_Indiana}.
\begin{table}[h!]
\centering
\begin{tabular}{|c | c c c c c|}
\hline
$\beta_{ij}$ & G & L & I & F & S \\ [0.5ex]
\hlin
G & 0.08 & 0.15 & 0.24 & 0 & 0.06 \\
L & 0.15 & 0.12 & 0.13 & 0 & 0 \\
I & 0.24 & 0.13 & 0.25 & 0.05 & 0.04 \\
F & 0 & 0 & 0.05 & 0.11 & 0.15 \\
S & 0.06 & 0 & 0.04 & 0.14 & 0.09 \\ \hline\hline
$\gamma_i$ & 0.075 & 0.115 & 0.085 & 0.125 & 0.1 \\
$N_i$ & 500000 & 160000 & 900000 & 350000 & 300000 \\[1ex]
\hline
\end{tabular}
\caption{Network Parameters of Fig.~\ref{fig:Graph_Topo_Indiana}}
\label{table:parameters}
\end{table}
Considering the stochastic framework,
we simulate testing data using~\eqref{eq:model_daily}
and~\eqref{eq:recovered_data}, with $p_i^x=0.2,$
$\forall i \in \{\text{G, L, I, F, S}\}$ from $T_1=6$ to $T_2=300$.
The number of daily and cumulative confirmed cases and removed (recovered) cases over time at node L are shown in Fig.~\ref{fig:Cases_Indiana_geo_p=0.3}.
When $k\geq 80$, the proportion of infected individuals at node L begins to decrease in Fig.~\ref{fig:Graph_Topo_Indiana}, which leads to the decline of the number of active cases in Fig.~\ref{fig:Cases_Indiana_geo_p=0.3}.
\begin{figure}
\centering
\begin{overpic}[width = 0.48\columnwidth]{figs/daily_cases_sto_p02.png}
\put(-4,28){{\parbox{0.75\linewidth}\footnotesize \rotatebox{90}{\footnotesize$C_{\text{L}}$, $D_{\text{L}}$}
}}
\put(50,-3){\footnotesize{\parbox{0.75\linewidth}\footnotesize $k
}}\normalsize
\end{overpic}
\begin{overpic}[width = 0.48\columnwidth]{figs/total_cases_sto_p02.png}
\put(-1,23){{\parbox{0.75\linewidth}\footnotesize \rotatebox{90}{\small$\mathbb{C}_{\text{L}}$, $\mathbb{D}_{\text{L}}$, $\mathbb{A}_{\text{L}}$}
}}\normalsize
\put(50,-3){\footnotesize
$k
}
\end{overpic}
\caption{Simulated daily and cumulative number of cases for node L with $p_i^x=0.2$}
\label{fig:Cases_Indiana_geo_p=0.3}
\end{figure}
We now use the method proposed in Section~\ref{sec:inference} to estimate the susceptible proportion at node I. We assume that the initial condition of the recovered state is $\widehat{r}_{\text{I}}(0)=0$. Hence, the initial infected state is written as: $\widehat{x}_{\text{I}}(0) =1-\widehat{s}_{\text{I}}(0)$. In Fig.~\ref{fig:Estimation_error_sus},
we plot the absolute value of the estimation error of the susceptible state at $k=100$ versus the start testing time $T_1$ and initial condition assumed, $\widehat{s}_{\text{I}}(0)$. It can be seen in Fig.~\ref{fig:Estimation_error_sus} (top) that the estimation error increases linearly with the initial susceptible level assumed. When the initial condition is assumed correctly for node I, with a later start testing date, the estimation error at $k=100$ builds up from $0$ to $r_{\text{I}}(k)$ eventually. The increase in the estimation error with $T_1$ signifies the importance of an early testing during an outbreak: with appropriate initial conditions assumed, we should initiate testing as quickly as possible to improve the accuracy of the state estimation. Meanwhile, Fig.~\ref{fig:Estimation_error_sus} (bottom) indicates that with a later start testing date, we must assume a lower initial susceptible level accordingly to achieve accurate estimation. The intuition behind this finding is that
since,
by Definition \ref{def:initial}, $\widehat{s_i}(k)=\widehat{s_i}(0)$, for all $k< T_1$, the lower initial condition can compensate for missed tests from $k\in [0,T_1-1]$, captured by the last term in \eqref{eq:infer_error_bound}. However, guessing $\widehat{s_i}(0)$ correctly, namely $\widehat{s_i}(0)= s_i(0)+\sum_{l=1}^{T_1-1}\Delta s_i(l)$, for $T_1>0$ is quite difficult.
Additionally, if we assume that $\widehat{s_i}(0)=1$, the estimated $\widehat{s_i}(k)$ is always larger than the true susceptible state in Fig.~\ref{fig:Estimation_error_sus}. The overestimation of susceptible level encourages us to design a stronger strategy to eradicate the virus, as will be seen in the subsequent simulations.
\begin{figure}
\centering
\begin{overpic}[width = 0.9 \columnwidth]{figs/Est_Error_abs_3D.png}
\put(2,19){{\parbox{0.75\linewidth}\tiny \rotatebox{90}{\small$| \widehat{s}_{\text{I}}(k) -s_{\text{I}}(k) | $}
}}\normalsize
\put(24,10){{\parbox{0.75\linewidth}\small
$T_1$ }
}\normalsize
\put(72,12){{\parbox{0.75\linewidth}\small
$\widehat{s}_{\text{I}}(0)$ }
}\normalsize
\end{overpic}
\begin{overpic}[width = 0.91 \columnwidth]{figs/Est_Error_abs_2D.png}
\put(2,34){{\parbox{0.75\linewidth}\small \rotatebox{90}{$\widehat{s}_{\text{I}}(0)$}
}}\normalsize
\put(50,0){{\parbox{0.75\linewidth}\small
$T_1$ }
}\normalsize
\end{overpic}
\caption{The absolute value of susceptible state estimation error at $k=100$ with respect to start testing date and the initial susceptible level assumed at node I, where the red points represent the estimation error: $| \widehat{s}_{\text{I}}(k) -s_{\text{I}}(k)| <0.01$. Both of the plots are illustrations of Prop.~\ref{prop:estimation_error}}
\label{fig:Estimation_error_sus}
\end{figure}
We simulate three scenarios over the network in Fig.~\ref{fig:Graph_Topo_Indiana} with the parameters of Table.~\ref{table:parameters}: no control, the distributed eradication strategy in~\eqref{eq:control_scheme_healing_rate}, and the
distributed eradication strategy utilizing estimated states in~\eqref{eq:control_scheme_healing_rate_inferred}. The inferred states were produced by the algorithm in Section~\ref{sec:inference} with $p_i^x=0.5$ $\forall i \in \{\text{G, L, I, F, S}\}$.
The average states for each scenario are plotted in Fig.~\ref{fig:Control_strategies}. It can be seen that both eradication strategies are able to eliminate the virus at a much higher speed than with no control. Furthermore, when $k\geq 200$, the healthy states with the two eradication strategies applied achieve higher susceptible fractions than the healthy state without control.
We can interpret the higher susceptible proportion as fewer individuals in the network becoming sick during the entire outbreak.
The control algorithm from~\eqref{eq:control_scheme_healing_rate_inferred} converges to a healthy state faster than
the algorithm in~\eqref{eq:control_scheme_healing_rate}, and
both eradication strategies prevent resurgences of the virus over the network.
In Fig.~\ref{fig:Lift_Control_strategies}, we remove both eradication strategies when $k=50$ and $k=100$ and do not reinstate them. It can be seen that both of the infection curves rise up when $k\geq 50$ (resp. $k\geq 100$), and reach peaks before they slowly die out. Fig.~\ref{fig:Lift_Control_strategies} can be interpreted as removing the allocation of resources and healthcare workers from a subpopulation too early during a pandemic, resulting in the increase in infection level and a potential outbreak. In Fig.~\ref{fig:Control_strategies_inter}, we only enforce our eradication strategies within time interval: $k\in [20,50]$ and $k \in [20,150]$, respectively. We can see that although the control strategies reduce the infection level significantly, a resurgence of the outbreak occurs immediately upon the removal of the eradication strategies. Hence, policy makers are suggested to enforce the eradication strategies during the entire outbreak to avoid the upcoming wave of epidemic
\begin{figure}
\begin{overpic}[width = 0.49 \columnwidth]{figs/Control_strategies_compare.png}
\put(-3,23){{\parbox{0.75\linewidth}\small \rotatebox{90}{\small$\frac{1}{n}\sum_i^n x_i(k)$}
}}\normalsize
\put(50,-3){{\parbox{0.75\linewidth}\small $k
}}\normalsize
\end{overpic
\begin{overpic}[width = 0.49 \columnwidth]{figs/Control_strategies_sus.png}
\put(-3,23){{\parbox{0.75\linewidth}\small \rotatebox{90}{\small$\frac{1}{n}\sum_i^n s_i(k)$}
}}\normalsize
\put(50,-3){{\parbox{0.75\linewidth}\small $k
}}\normalsize
\end{overpic}
\caption{Average system states over time}
\label{fig:Control_strategies}
\end{figure}
\begin{figure}
\begin{overpic}[width = 0.49 \columnwidth]{figs/Lift_Control_Strategy.png}
\put(-4,20){{\parbox{0.75\linewidth}\small \rotatebox{90}{\small$\frac{1}{n}\sum_i^n x_i(k)$}
}}\normalsize
\put(50,-1.5){{\parbox{0.75\linewidth}\small $k
}}\normalsize
\end{overpic}
\begin{overpic}[width = 0.49 \columnwidth]{figs/Lift_Control_Strategy_100.png}
\put(-4,20){{\parbox{0.75\linewidth}\small \rotatebox{90}{\small$\frac{1}{n}\sum_i^n x_i(k)$}
}}\normalsize
\put(50,-1.5){{\parbox{0.75\linewidth}\small $k
}}\normalsize
\end{overpic}
\caption{Average infection proportion of the virus over time with the eradication strategies enforced at $k\in [0,50]$ (left) and at $k \in [0,100]$ (right)}
\label{fig:Lift_Control_strategies}
\end{figure}
\begin{figure}
\begin{overpic}[width = 0.49 \columnwidth]{figs/Control_Strategy_inter_20_50.png}
\put(-4,20){{\parbox{0.75\linewidth}\small \rotatebox{90}{\small$\frac{1}{n}\sum_i^n x_i(k)$}
}}\normalsize
\put(50,-1.5){{\parbox{0.75\linewidth}\small $k
}}\normalsize
\end{overpic}
\begin{overpic}[width = 0.49 \columnwidth]{figs/Control_Strategy_inter_20_150.png}
\put(-4,20){{\parbox{0.75\linewidth}\small \rotatebox{90}{\small$\frac{1}{n}\sum_i^n x_i(k)$}
}}\normalsize
\put(50,-1.5){{\parbox{0.75\linewidth}\small $k
}}\normalsize
\end{overpic}
\caption{Average infection proportion of the virus over time with both of the eradication strategies imposed at $k\in [20,50]$ (left) and at $k\in [20,150]$ (right)}
\label{fig:Control_strategies_inter}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
This paper studied the stability, inference and control of discrete time, time-varying SIR epidemics over networks. We established the sufficient condition for GES of the set of healthy states. In addition, we proposed a stochastic framework for estimating the
underlying epidemic states
from collected testing data. We provided analytic expressions for the error of the estimation algorithm.
We also proposed two distributed control strategies that are able to eradicate the virus in at least exponential time.
The control strategies provide
insights for decision makers on how to eliminate an ongoing outbreak.
In future work, we plan to study the stability and control of models with more states than SIR such as SEIRS (susceptible-exposed-infected-recovered-susceptible) and SAIR (susceptible-asymptomatic-infected-recovered) as they can possibly capture the characteristics of COVID-19 better. In our stochastic testing framework, we did not consider the existence of
inaccurate testing kits,
which appear frequently and cause confusion for
policy
makers. Hence, we plan to include false positive/negative test results into our testing and estimation model and investigate the new model's estimation accuracy in the future. Furthermore, we aim to apply our model on the real data to identify the system parameters.
\section*{Acknowledgment}
The authors would like to thank Ashish Hota (IIT Kharagpur) and Baike She (Purdue University) for useful conversations related to
Section~\ref{sec:inference}.
\normalem
\bibliographystyle{IEEEtran}
|
\subsubsection{#1}\vspace{-3\baselineskip}\color{black}\bigskip{\noindent \bf \thesubsubsection. #1.}}
\newcommand{\myparagraph}[1]{\needspace{1\baselineskip}\medskip\noindent {\it #1.}}
\newcommand{\myparagraphtc}[1]{\needspace{1\baselineskip}\medskip\noindent {\it #1.}\addcontentsline{toc}{subsubsection}{\qquad\qquad\quad#1}}
\section{Introduction}
\label{S:intro}
\input{introduction}
\section{Problem Formulation}
\label{S:problem_formulation}
\input{problem_formulation}
\section{Constrained Reinforcement Learning for Safety}
\label{S:crl}
\input{crl}
\section{Improvement Directions}
\label{S:theory}
\input{theory}
\section{Numerical Results}
\label{S:sims}
\input{numerical_results}
\section{Conclusions}
\label{S:conclusions}
\input{conclusions}
\bibliographystyle{IEEEtran}
|
\section*{Introduction}
The purpose of this paper is to introduce the reader to a new area of mathematical physics, called
theory of generalizations of wave equations to multidimensional variational problems. This theory arose
in connection with numerous attempts of scientists to give a rigorous mathematical sense to constructions
of quantum field theory, string theory and, more generally, theory of $D$-branes. Physical intuition behind these
areas of physics already gave rise to many remarkable mathematical results, and so providing a mathematical
background to these disciplines can be very useful for mathematics as well as for physics.
The formal derivation of the generalized wave equations (see \S1 below)
is a direct generalization of formal derivation of
equations of quantum mechanics and wave optics (the Schr\"odinger equation and the wave equation)
from the equations of classical mechanics and geometric optics (the Hamilton--Jacobi equation and the eikonal equation).
The natural examples of generalized wave equations, called the generalized Schr\"odinger equation and the
quantum Plato problem, are given by interesting formulas including determinants.
However, giving a mathematical sense to the generalized wave equations meets problems. It is unclear
what are solutions of these equations. The situation here is similar to the situation at the birth
of theory of partial differential equations and the related famous discussion of scientists in the XVIII century
on what is the mathematical sense of the notion of function. The generalized wave equations are linear
differential equations on infinite dimensional function spaces. But the infinite dimensional
differential operators in these equations, as a rule, {\it cannot be multiplied}. In order to give a mathematical
sense to these equations, one can try to use regularization of the corresponding differential operators. The regularized
functional differential equations have a ``naive'' sense as equations for unknown functionals.
However, in such approach, removing regularization
leads to divergencies. A way to overcome this difficulty essentially worked out by physicists is essentially to subtract
the divergent part of the evolution operator, and consider the remaining finite part; this is
heuristically equivalent to adding an infinite summand to the differential operator.
Thus, in fact, physicists refuse to consider the initial equation, adding to it infinite ``counterterms''.
This construction was usually justified by the argument that in the framework of perturbation theory with respect to
the interaction constant, the differential operator for zero interaction and its corrections in perturbation theory
have no physical sense, and hence may be not considered. However, in our approach to the equations, there is no
perturbation theory at all, and one considers directly the theory with interaction. Therefore this physical
argument looses sense.
The paper consists of two Sections. \S1 is devoted to formal derivation of generalized wave equations.
In \S2 we give an abstract definition of solutions of generalized
wave equations, and discuss the arising problem of constructing appropriate space of
distributions on infinite dimensional function space.
{\bf Acknowledgments.} I thank V. V. Dolotin and V. P. Maslov for illuminating discussions.
\section{Formal derivation of generalized wave equations}
This construction first appeared in the papers [1,2,3] and in the book [4] which can serve as a general introduction
to the ideas of the present paper. The generalized wave equations appeared in these references as purely formal ones,
without definition of their solutions.
\subsection{Generalized Hamilton--Jacobi equation}
Consider the action functional of the form
\begin{equation}
J=\int_D F(x^0,\ldots,x^n,\varphi^1,\ldots,\varphi^m,\varphi^1_{x^0},\ldots,
\varphi^m_{x^n})\,dx^0\ldots dx^n,
\end{equation}
where $x=(x^0,\ldots,x^n)$ is a space-time point, $\varphi^i=\varphi^i(x)$ are real smooth field
functions, $\varphi^i_{x^j}=\partial\varphi^i/\partial x^j$,
and integration goes over an $(n+1)$-dimensional smooth surface $D$
(the graph of the functions $\varphi(x)=(\varphi^i(x))$) with the boundary $\partial D=C$ in the space $\R^{m+n+1}$.
We shall need the well known formula for variation of action on an extremal surface
with moving boundary, which is used, for instance, in derivation of the Noether theorem. Let us recall this
formula. Let
\begin{equation}
x=x(s),\ \ \varphi=\varphi(s),\ \ s=(s^1,\ldots,s^n)
\end{equation}
be a local parameterization of the boundary surface $C$. Assume that for each $n$-dimensional surface $C$
sufficiently close to certain fixed surface, there exists a unique $(n+1)$-dimensional surface $D$
with the boundary $\partial D=C$ which is an extremal of the variational principle (1), i.~e., the graph of a
solution of the Euler--Lagrange equations. Denote by $S=S(C)$ the value of integral (1) over the surface $D$.
Then one has the following formula for variation of the functional $S(C)$:
\begin{equation}
\delta S=\int_C\left(\sum\pi_i(s)\delta\varphi^i(s)-\sum H_j(s)\delta x^j(s)\right)ds,
\end{equation}
where
\begin{equation}
\begin{aligned}{}
\pi_i&=\sum_l F_{\varphi^i_{x^l}}(-1)^l\frac{\partial(x^0,\ldots,\widehat{x^l},\ldots,x^n)}
{\partial(s^1,\ldots,s^n)},\\
H_j&=\sum_{l\ne j}\left(\sum_i F_{\varphi^i_{x^l}}\varphi^i_{x^j}\right)
(-1)^l\frac{\partial(x^0,\ldots,\widehat{x^l},\ldots,x^n)}
{\partial(s^1,\ldots,s^n)}\\
&+\left(\sum_i F_{\varphi^i_{x^j}}\varphi^i_{x^j}-F\right)
(-1)^j\frac{\partial(x^0,\ldots,\widehat{x^j},\ldots,x^n)}
{\partial(s^1,\ldots,s^n)}.
\end{aligned}
\end{equation}
Here $\frac{\partial(x^1,\ldots,x^n)}
{\partial(s^1,\ldots,s^n)}$ is the Jacobian, the hat over a variable means that the variable is omitted, and
$F_{\varphi^i_{x^j}}$ is the partial derivative of the function $F$ with respect to its argument $\varphi^i_{x^j}$.
Let us derive the equations satisfied by the functional
$$
S(C)=S(\varphi^i(s),x^j(s)).
$$
To this end, note that
by formula (4), the quantities
\begin{equation}
\pi_i(s)=\frac{\delta S}{\delta\varphi^i(s)},\ \ H_j(s)=-\frac{\delta S}{\delta x^j(s)}
\end{equation}
depend not only on the functions $\varphi(s)$, $x(s)$, but also on the derivatives $\varphi^i_{x^j}$
characterizing the tangent plane to the surface $D$ at the point $(x(s),\varphi(s))$. These $m(n+1)$ derivatives
are related by $mn$ equations
\begin{equation}
\sum_j\varphi^i_{x^j}x^j_{s^k}=\varphi^i_{s^k},\ \ i=1,\ldots,m,\ \ k=1,\ldots,n.
\end{equation}
Therefore, $m+n+1$ quantities (5) depend on $m(n+1)-mn=m$ free parameters. Hence they are related by $n+1$ equations.
$n$~of these equations are easy to find:
\begin{equation}
\sum_i\pi_i\varphi^i_{s^k}-\sum_j H_jx^j_{s^k}=0,\ \ k=1,\ldots,n.
\end{equation}
These equations express the fact that the value of the functional $S(C)$ does not depend on a parameterization
of the surface $C$.
The remaining $(n+1)$-th equation depends on the form of the function $F$. Denote it by
\begin{equation}
\HH\left(x^j(s),\varphi^i(s),x^j_{s^k}(s),\varphi^i_{s^k}(s),\pi_i(s),-H_j(s)\right)=0.
\end{equation}
From equalities (5) we obtain the following system of equations on the functional $S(C)$, called
the {\it generalized Hamilton--Jacobi equation} (or the {\it generalized eikonal equation}):
\begin{equation}
\begin{aligned}{}
&\sum_i\frac{\delta S}{\delta\varphi^i(s)}\varphi^i_{s^k}+\sum_j\frac{\delta S}{\delta x^j(s)}
x^j_{s^k}=0,\ \ \ k=1,\ldots,n,\\
&\HH\left(x^j(s),\varphi^i(s),x^j_{s^k}(s),\varphi^i_{s^k}(s),
\frac{\delta S}{\delta\varphi^i(s)},\frac{\delta S}{\delta x^j(s)}\right)=0.
\end{aligned}
\end{equation}
\medskip
{\bf Examples.}
\nopagebreak
\medskip
1) (Scalar field with self-action) [3,4] Let $m=1$ and
\begin{equation}
\begin{aligned}{}
F(x^j,\varphi,\varphi_{x^j})&=\frac12\left(\varphi_{x^0}^2-\sum_{j\ne0}\varphi_{x^j}^2\right)-V(x,\varphi)\\
&=\frac12\varphi_{x^\mu}\varphi_{x_\mu}-V(x,\varphi),
\end{aligned}
\end{equation}
where we have introduced Greek indices $\mu$ instead of $j$, and raising and lowering indices goes using the
Lorentz metric
$$
dx^2=(dx^0)^2-\sum_{j\ne 0}(dx^j)^2.
$$
Then the generalized Hamilton--Jacobi equation has the following form:
\begin{equation}
\begin{aligned}{}
&x^\mu_{s^k}\frac{\delta S}{\delta x^\mu(s)}+
\varphi_{s^k}\frac{\delta S}{\delta\varphi(s)}=0,\ \ k=1,\ldots,n,\\
&D_\mu\frac{\delta S}{\delta x^\mu(s)}+
\frac12\left(\frac{\delta S}{\delta\varphi(s)}\right)^2
+ D^\mu D_\mu\left(\frac12 d\varphi(s)^2+V(x(s),\varphi(s))\right)=0,
\end{aligned}
\end{equation}
where
$$
D^\mu=(-1)^\mu \frac{\partial(x^0,\ldots,\widehat{x^\mu},\ldots,x^n)}
{\partial(s^1,\ldots,s^n)},
$$
and $d\varphi(s)^2$ is the scalar square of the differential $d\varphi(s)$ of the function $\varphi(s)$ on the surface.
\medskip
2) (The Plato problem) [4] Assume that one considers $(n+1)$-dimensional surfaces $D$ with the boundary $C:x=x(s)$
in the Euclidean space $\R^N$
with coordinates $(x^1,\ldots,x^N)$, and the role of integral (1) is played by the area of the surface $D$.
Then the generalized eikonal equation reads
\begin{equation}
\begin{aligned}{}
&\sum_j x^j_{s^k}\frac{\delta S}{\delta x^j(s)}=0,\ \ 1\le k\le n,\\
&\sum_j\left(\frac{\delta S}{\delta x^j(s)}\right)^2=
\sum_{j_1<\ldots<j_n}\left(\frac{\partial(x^{j_1},\ldots,x^{j_n})}
{\partial(s^1,\ldots,s^n)}\right)^2.
\end{aligned}
\end{equation}
In particular, in the case $n=1$, $N=3$, the equation reads
\begin{equation}
\begin{aligned}{}
&x_s\frac{\delta S}{\delta x(s)}+y_s\frac{\delta
S}{\delta y(s)}+z_s\frac{\delta S}{\delta z(s)}=0,\\
&\left(\frac{\delta S}{\delta x(s)}\right)^2+\left(\frac{\delta
S}{\delta y(s)}\right)^2+\left(\frac{\delta S}{\delta z(s)}\right)^2=x_s^2+y_s^2+z_s^2.
\end{aligned}
\end{equation}
\medskip
The generalized Hamilton--Jacobi equation was first systematically studied, for two-dimensional variational
problems, in the book [5]. Regarding the general theory of integration of this equation and, in particular,
regarding the characteristics equations which are the generally covariant generalized
Hamilton canonical equations equivalent to the Euler--Lagrange equations, see [2, 4].
Let us note that in the case of one-dimensional variational problems, integration of the Hamilton--Jacobi equation
is the most powerful method of integration of the canonical Hamilton equations. It would be interesting to check
whether one can integrate the Euler--Lagrange equations for some multidimensional variational problems
(nonlinear partial differential equations, for example, the Einstein or Yang--Mills equations) using integration
of the corresponding generalized Hamilton--Jacobi equations.
\subsection{Generalized wave equations}
Generalized wave equations are obtained from the generalized Hamilton--Jacobi equation (9) by formal substitution
\begin{equation}
\begin{aligned}{}
&\frac{\delta S}{\delta x^j(s)} \to -ih\frac{\delta}{\delta x^j(s)},\\
&\frac{\delta S}{\delta\varphi^i(s)} \to -ih\frac{\delta}{\delta\varphi^i(s)}.
\end{aligned}
\end{equation}
We obtain the following system of linear functional differential equations:
\begin{equation}
\begin{aligned}{}
&\sum_i\varphi^i_{s^k}\frac{\delta\Psi}{\delta\varphi^i(s)}+\sum_j x^j_{s^k}\frac{\delta\Psi}{\delta x^j(s)}
=0,\ \ \ k=1,\ldots,n,\\
&\HH\left(x^j(s),\varphi^i(s),x^j_{s^k}(s),\varphi^i_{s^k}(s),
-ih\frac{\delta}{\delta\varphi^i(s)},-ih\frac{\delta}{\delta x^j(s)}\right)\Psi=0.
\end{aligned}
\end{equation}
The first $n$ of these equations express independence of the ``functional'' $\Psi(C)$ on a parameterization
of the surface $C:x=x(s)$, $\varphi=\varphi(s)$.
\medskip
{\bf Examples.}
\nopagebreak
\medskip
1) In the case of scalar field with self-action (see Subsect. 1.1), we have the following generalized
wave equation which we call the {\it generalized Schr\"odinger equation}:
\begin{equation}
\begin{aligned}{}
&x^\mu_{s^k}\frac{\delta\Psi}{\delta x^\mu(s)}+
\varphi_{s^k}\frac{\delta\Psi}{\delta\varphi(s)}=0,\ \ k=1,\ldots,n,\\
&ihD_\mu\frac{\delta\Psi}{\delta x^\mu(s)}=
-\frac{h^2}2\frac{\delta^2\Psi}{\delta\varphi(s)^2}
+D^\mu D_\mu\left(\frac12 d\varphi(s)^2+V(x(s),\varphi(s))\right)\Psi.
\end{aligned}
\end{equation}
\medskip
2) In the case of the Plato problem, we obtain the following generalized wave equation which we call the
{\it quantum Plato problem}:
\begin{equation}
\begin{aligned}{}
&\sum_j x^j_{s^k}\frac{\delta\Psi}{\delta x^j(s)}=0,\ \ 1\le k\le n,\\
&\sum_j-h^2\frac{\delta^2\Psi}{\delta x^j(s)^2}=
\sum_{j_1<\ldots<j_n}\left(\frac{\partial(x^{j_1},\ldots,x^{j_n})}
{\partial(s^1,\ldots,s^n)}\right)^2\Psi.
\end{aligned}
\end{equation}
In particular, in the case $n=1$, $N=3$, the equation reads
\begin{equation}
\begin{aligned}{}
&x_s\frac{\delta\Psi}{\delta x(s)}+y_s\frac{\delta\Psi}{\delta y(s)}+z_s\frac{\delta\Psi}{\delta z(s)}=0,\\
&-h^2\left(\frac{\delta^2\Psi}{\delta x(s)^2}+\frac{\delta^2\Psi}{\delta y(s)^2}+\frac{\delta^2\Psi}{\delta z(s)^2}\right)=
(x_s^2+y_s^2+z_s^2)\Psi.
\end{aligned}
\end{equation}
\section{On the solutions of generalized wave equations}
In this Section we give a definition of solutions of generalized wave equations.
In this definition, the space-time variables and the field variables are considered in equal rights. This material is new.
Denote by $V$ the Schwartz space
(or any nuclear space [6]) of functions ($\varphi^i(s)$, $x^j(s)$, $\pi_i(s)$, $-H_j(s)$)
(in the case of generalized wave equations (15)), or functions $(x^j(s),-H_j(s))$ (in the case of
quantum Plato problem (17)).
The variables $\pi_i(s)$ are called {\it canonically conjugate} to the variables $\varphi^i(s)$, and the variables
$-H_j(s)$ are called canonically conjugate to $x^j(s)$.
But, for shortness, we shall call these functions simply by $(\varphi(s),\pi(s))$, and omit discrete
indices. Let us introduce the following notation:
$SV$ is the space of symbols of regular differential operators (the topological symmetric algebra of the space $V$); $SV'$ is the
space of symbols of all polynomial differential operators (polynomial functionals on the space $V$, i.~e., the topological
symmetric algebra of the space $V'$).
Define the {\it Weyl--Moyal} algebra
$SV$ with the Moyal product
\begin{equation}
\begin{aligned}{}
&H^{(1)}*_{Moyal}H^{(2)}(\varphi,\pi)=\\
&\exp\frac{ih}2\int\left(\frac{\delta}{\delta\pi_1(s)}\frac{\delta}{\delta\varphi_2(s)}-
\frac{\delta}{\delta\pi_2(s)}\frac{\delta}{\delta\varphi_1(s)}\right)\ ds\\
&\times H^{(1)}(\varphi_1,\pi_1)H^{(2)}(\varphi_2,\pi_2)|_{\varphi_1=\varphi_2=\varphi,\pi_1=\pi_2=\pi}.
\end{aligned}
\end{equation}
\medskip
{\bf Definition.} Let us call by a {\it general space of distributions} a representation of the Weyl--Moyal
algebra $(SV,*_{Moyal})$
in a topological vector space $\EE$. Let us call by a {\it solution} of a system of differential equations
$(\widehat D_\alpha\Psi=0)$, $D_\alpha\in SV'$,
a distribution $\Psi\in\EE$ such that for any regularization $D_{\alpha,\e}\in SV$, $\e>0$ of symbols
$D_\alpha$, $D_{\alpha,\e}\to D_\alpha$ as $\e\to0$, we have $\widehat D_{\alpha,\e}\Psi\to0$ in the space $\EE$
as $\e\to0$.
\medskip
Numerous examples of general spaces of distributions can be obtained from the unitary representations of the infinite
dimensional Heisenberg group constructed in the book [6] (they are called in [6] by representations of
canonical commutation relations) by passing to the dual space of the space of smooth vectors. Another
example of a general space of distributions is constructed in the paper [7]. See also the book [8]
and references therein. However, up to now no examples (or proofs of existence) of nontrivial solutions
of generalized wave equations are known.
|
\section{Introduction}
Microcavity polaritons, the novel quasiparticles resulting from the
coherent strong coupling between quantum well excitons and cavity
photons~\cite{kavokin_laussy}, have unique mixed matter-light
properties that none of their constituents displays on its
own. Because of their energy dispersion and their strong non-linearity
inherited from the excitonic components, polaritons continuously
injected by an external laser into a pump state with suitable
wavevector and energy can undergo coherent stimulated scattering into
two conjugate states~\cite{ciuti00:prb,ciuti01,ciuti03}, the signal
and the idler, a process known as optical parametric oscillator (OPO).
Since their first
realisation~\cite{stevenson00,savvidis00:prl,savvidis00:prb,baumberg00:prb,saba01},
the interest in microcavity optical parametric phenomena has involved
several fields of fundamental and applicative
research~\cite{edamatsu2004,savasta05:prl,lanco2006,abbarchi2011,ardizzone2012,xie2012,lecomte2013}.
Recently, considerable resources have been invested in exploring the
fundamental properties of parametric processes, including the
possibility of macroscopic phase coherence and superfluid
behaviour~\cite{Carusotto2013a}.
In spite of the coherent nature of the driving laser pump, the OPO
process belongs to the class of non-equilibrium phase transitions, in
which a $U(1)$ phase symmetry is spontaneously
broken~\cite{wouters06b}.
While the phase of the pumped mode is locked to the incident laser,
the ones of signal and idler are free to be simultaneously rotated in
opposite directions.
Because of this phase freedom, recent experiments~\cite{sanvitto10}
have tested the OPO superfluid properties by exploring the physics of
the signal-idler order parameter, demonstrating the existence and
metastability of vortex configurations. As the order parameter
involves both signal and idler, their phase winding have opposite
signs~\cite{sanvitto10,marchetti10,marchetti_review}. Crucially, this
makes both OPO fluids to display quantized flow metastability
simultaneously.
While in equilibrium condensates different aspects of superfluidity
are typically closely related~\cite{LeggettRMP_99}, this is no longer
true in a non-equilibrium context as for microcavity
polaritons~\cite{Carusotto2013a}.
In particular, those aspects of superfluidity related to the
frictionless flow around defects are expected to be much more involved
in OPO condensates than for any other investigated polariton
condensates, such as for the case of incoherent
pumping~\cite{kasprzak06:nature,wouters2010}, and single-state
resonantly pumped microcavities~\cite{amo09_b}.
Independently on the pumping scheme, the driving and the polariton
finite lifetime prompt questions about the meaning of superfluid
behaviour, when the spectrum of collective excitations is complex
rather than real, raising conceptual interrogatives about the
applicability of a Landau criterion~\cite{wouters2010}.
Yet, an additional complexity characterises the OPO regime, where the
simultaneous presence of three oscillation frequencies and momenta for
pump, signal and idler correspondingly multiplies the number of
collective excitation branches~\cite{wouters06b}.
Note that from the experimental point of view, pioneering
experiments~\cite{amo09} have observed a ballistic non-spreading
propagation of signal/idler polariton wavepackets in a triggered-OPO
configuration.
However, given the complexity of the dynamics as well as the nonlinear
interactions involved in this time-dependent
configuration~\cite{szymanska2010}, theoretical understanding of these
observations is not complete yet.
This Article reports a joint theoretical and experimental study of an
OPO configuration where a wide and steady-state condensate hits a
stationary localised defect in the microcavity.
Contrary to the criterion for quantized flow metastability for which
signal and idler display simultaneous locked responses, we find that
their scattering properties when the OPO hits a static defect are
different.
In particular we investigate the scattering properties of all three
fluids, pump, signal and idler, in both real and momentum space. We
find that the modulations generated by the defect in each fluid are
determined not only by its associated Rayleigh scattering ring, but
each component displays additional rings because of the cross-talk
with the other components imposed by nonlinear and parametric
processes.
We single out three factors determining which one of these rings
influences the most each fluid response: the coupling strength between
the three OPO states, the resonance of the ring with the blue-shifted
lower polariton dispersion, the values of each fluid group velocity
and lifetime together establishing how far each modulation can
propagate from the defect.
The concurrence of these effects implies that the idler strongly
scatters inheriting the same modulations as the pump, while the
modulations due to its own ring can propagate only very close to the
defect and cannot be appreciated. Yet, the modulations in the signal
are strongly suppressed, and not at all visible in experiments,
because the slope of the polariton dispersion in its low momentum
component brings all Rayleigh rings coming from pump and idler out of
resonance.
Note that the kinematic conditions for OPO are incompatible with the
pump and idler being in the subsonic regime. Thus, the coupling
between the three components always implies some degree of scattering
in the signal. In practice, the small value of the signal momentum
strongly suppresses its actually visible modulations, something
confirmed by the experimental observations.
\begin{figure}[h!]
\centering
\includegraphics[width=1.0\columnwidth]{figure1.png}
\caption{(Color online) OPO mean-field blue-shifts and fluctuation
Rayleigh rings in the linear response scheme for homogeneous
pumping. Left panel: Signal $s$ ([blue] upper triangles), pump $p$
([red] circles), and idler $i$ ([green] lower triangles) mean-field
energy blue-shifts $\epsilon_{n=s,p,i}$ (in units of $\gamma_p =
\gamma_{\vect{k}_p}$) versus the rescaled pump intensity $I_p$ (in
units of $\gamma_p^3$) in the optical limiter regime. Parameters are
$\Omega_R=5$~meV, zero cavity-exciton detuning, $\gamma_X = \gamma_C
= 0.12$~meV, $\omega_p - \omega_0^X = -1.25$~meV, $k_p=1.6{\mu
m}^{-1}$, $k_s \simeq 0$, and $k_i=3.2{\mu m}^{-1}$. The shaded
area is stable OPO region, while the vertical dashed line
corresponds to the pump power value chosen for plotting the right
panel. Right panel: Blue-shifted LP dispersion~\eqref{eq:blues} with
superimposed Rayleigh curves $\Gamma_{p,i,(u,v), \tilde{\vect{k}} +
\vect{k}_{p, i}}$ evaluated within the linear response
approximation (same symbols as left panel). The two rings
corresponding to the signal state, $\Gamma_{s,(u,v),
\tilde{\vect{k}}}$, are shrinked to zero because $k_s \simeq 0$.}
\label{fig:spect}
\end{figure}
\section{Model}
The dynamics of polaritons in the OPO regime, and their hydrodynamic
properties when scattering against a defect, can be described via a
classical driven-dissipative non-linear Gross-Pitaevskii equation
(GPE) for the coupled exciton and cavity fields $\psi_{X,C}
(\vect{r},t)$ ($\hbar=1$)~\cite{whittaker2005_b,Carusotto2013a}:
\begin{equation}
i\partial_t \begin{pmatrix} \psi_X \\ \psi_C \end{pmatrix} =
\hat{H} \begin{pmatrix} \psi_X \\ \psi_C \end{pmatrix}
+ \begin{pmatrix} 0 \\ F_p(\vect{r},t) \end{pmatrix} \; .
\label{eq:gpequ}
\end{equation}
The dispersive $X$- and $C$-fields decay at a rate $\gamma_{X,C}$ and
are coupled by the Rabi splitting $\Omega_R$, while the non-linearity
is regulated by the exciton coupling strength $g_X$:
\begin{equation}
\hat{H} = \begin{pmatrix} \omega^{X}_{-i\nabla} - i
\frac{\gamma_X}{2} + g_X |\psi_X|^2 & \Omega_R/2 \\ \Omega_R/2 &
\omega^C_{-i\nabla} - i \frac{\gamma_C}{2} + V_d \end{pmatrix} \;
.
\end{equation}
We describe the defect via a potential $V_d (\vect{r})$ acting on the
photonic component; this can either be a defect in the cavity mirror
or a localised laser field~\cite{amo09_b,amo10,zajac2012}.
In the conservative, homogeneous, and linear regime ($\gamma_{X,C}=0=
V_d (\vect{r})= g_X$), the eigenvalues of $\hat{H}$ are given by the
lower (LP) and upper polariton (UP) energies, $2
\omega_{\vect{k}}^{LP,UP} = \omega_{\vect{k}}^{C} +
\omega_{\vect{k}}^{X} \mp \sqrt{(\omega_{\vect{k}}^{C} -
\omega_{\vect{k}}^{X})^2 + \Omega_R^2}$.
The cavity is driven by a continuous-wave laser field $F_p(\vect{r},t)
= \mathcal{F}_p(\vect{r}) e^{i (\vect{k}_p \cdot \vect{r} - \omega_p
t)}$ into the OPO regime: Here, polaritons are continuously injected
into the pump state with frequency $\omega_p$ and momentum
$\vect{k}_p$ and, above a pump strength threshold, undergo coherent
stimulated scattering into the signal $(\omega_s, \vect{k}_s)$ and
idler $(\omega_i, \vect{k}_i)$ states.
As a first step, it is useful to get insight into the system behaviour
in the simple case of a homogeneous pump of strength
$\mathcal{F}_p(\vect{r}) = f_p$. A numerical study of the coupled
equations~\eqref{eq:gpequ} for the more realistic case of a
finite-size top-hat pump profile $\mathcal{F}_p(\vect{r})$ will be
presented later.
To further simplify our analysis, we assume here that the UP
dispersion does not get populated by parametric scattering processes
and thus, by means of the Hopfield coefficients $2X_{\vect{k}}^2,
2C_{\vect{k}}^2 = 1 \pm (\omega_{\vect{k}}^{C} -
\omega_{\vect{k}}^{X})/\sqrt{(\omega_{\vect{k}}^{C} -
\omega_{\vect{k}}^{X})^2 + \Omega_R^2}$, we project the
GPE~\eqref{eq:gpequ} onto the LP
component~\cite{ciuti01,wouters07:prb} $\psi_{\vect{k}}^{} =
X_{\vect{k}} \psi_{X,\vect{k}} + C_{\vect{k}} \psi_{C,\vect{k}}$,
where $\psi(\vect{r},t) = \sum_{\vect{k}} e^{i\vect{k}\cdot \vect{r}}
\psi_{\vect{k}}^{} (t)$:
\begin{multline}
i\partial_t \psi_{\vect{k}}^{} = \left[\omega_{\vect{k}}^{LP} -
i\Frac{\gamma_{\vect{k}}}{2}\right]\psi_{\vect{k}}^{} +
C_{\vect{k}} \sum_{\vect{q}} C_{\vect{q}} V_d(\vect{k} - \vect{q})
\psi_{\vect{q}}^{}\\ + \sum_{\vect{k}_1, \vect{k}_2} g_{\vect{k},
\vect{k}_1, \vect{k}_2} \psi^*_{\vect{k}_1 + \vect{k}_2-\vect{k}}
\psi_{\vect{k}_1}^{} \psi_{\vect{k}_2}^{} + \tilde{f}_p(t)
\delta_{\vect{k},\vect{k}_p}\; .
\label{eq:efflp}
\end{multline}
Here, $\gamma_{\vect{k}}=\gamma_X X_{\vect{k}}^2 + \gamma_C
C_{\vect{k}}^2$ is the effective LP decay rate, the interaction
strength is given by $g_{\vect{k}, \vect{k}_1, \vect{k}_2}=g_X
X_{\vect{k}} X_{\vect{k}_1 + \vect{k}_2-\vect{k}} X_{\vect{k}_1}
X_{\vect{k}_2}$, and the pumping term by $\tilde{f}_p(t)=
C_{\vect{k}_p} f_p e^{-i\omega_p t}$.
\section{Linear response theory}
In the limit where the homogeneously pumped system is only weakly
perturbed by the external potential $V_d(\vect{r})$, we apply a linear
response analysis~\cite{astrakharchik04}: The LP field is expanded
around the mean-field terms for the three $n=1,2,3=s,p,i$ OPO
states~\cite{whittaker05},
\begin{equation}
\psi_{\tilde{\vect{k}}}^{} = \sum_{n=1}^3 e^{-i \omega_n t}
\left[\psi_{n}^{} \delta_{\tilde{\vect{k}},0} +
u_{n,\tilde{\vect{k}}}^{} e^{-i\omega t} +
v^*_{n,-\tilde{\vect{k}}} e^{i\omega t} \right]\; ,
\label{eq:expan}
\end{equation}
where $\tilde{\vect{k}} = \vect{k} - \vect{k}_n$. Eq.~\eqref{eq:efflp}
is expanded linearly in both the fluctuations terms,
$u_{n,\tilde{\vect{k}}}^{}$ and $v_{n,\tilde{\vect{k}}}^{}$, as well
as the defect potential. At zero-th order, the three complex uniform
mean-field equations can be solved to obtain the dependence of the
signal, pump and idler energy blue-shifts, $\epsilon_n = g_X
X^2_{\vect{k}_n} |\psi_{n}^{}|^2$ on the system
parameters~\cite{wouters07:prb,SM}. A typical behaviour of
$\epsilon_n$ as a function of the rescaled pump intensity $I_p = g_X
C_{\vect{k}_p}^2 f_p^2/X_{\vect{k}_p}^2$ in the optical limiter regime
is plotted in the left panel of Fig.~\ref{fig:spect}.
At first order, one obtains six coupled equations diagonal in momentum
space~\cite{wouters06b}
\begin{equation}
\omega \vect{w}_{\tilde{\vect{k}}} = \mathcal{L}_{\tilde{\vect{k}}}
\vect{w}_{\tilde{\vect{k}}} + \Frac{1}{2} \vectgr{\Psi}_d\; ,
\label{eq:fluct}
\end{equation}
for the $6$-component vector $\vect{w}_{\tilde{\vect{k}}} =
(u_{n,\tilde{\vect{k}}}^{} , v_{n,\tilde{\vect{k}}}^{})^{T}$ and for
the potential part, $\vectgr{\Psi}_d = (\psi_n^{} C_{\vect{k}_n}
C_{\vect{k} + \vect{k}_n} V_d(\vect{k}),-\psi_n^* C_{\vect{k}_n}
C_{\vect{k}_n - \vect{k}} V_d(-\vect{k}))^{T}$. In ~\eqref{eq:fluct}
we have only kept the terms oscillating at the frequencies $\omega_n
\pm \omega$ and neglected the other terms in the expansion (i.e.,
$2\omega_n - \omega_m \pm \omega$), which are oscillating at
frequencies far from the LP band, and thus with negligible amplitudes.
In the particle-like and the hole-like channels, the Bogoliubov matrix
determining the spectrum of excitations can be written
as~\cite{wouters06b}
\begin{equation}
\mathcal{L}_{\vect{k}} = \begin{pmatrix} M_{\vect{k}}^{} & Q_{\vect{k}}^{}
\\ -Q_{-\vect{k}}^* & -M_{-\vect{k}}^* \end{pmatrix} \; ,
\end{equation}
where the 3 OPO states components are
\begin{align}
(M_{\vect{k}}^{})_{mn} &= \left[\omega^{LP}_{\vect{k}_m + \vect{k}}
- \omega_m - i\Frac{\gamma_{\vect{k}_m + \vect{k}}}{2}\right]
\delta_{m,n} \\
\nonumber & + 2\sum_{q,t=1}^3
g_{\vect{k}_m+\vect{k},\vect{k}_n+\vect{k},\vect{k}_t} \psi_q^*
\psi_t^{} \delta_{m+q,n+t}\\
(Q_{\vect{k}}^{})_{mn} &= \sum_{q,t=1}^3
g_{\vect{k}_m+\vect{k},\vect{k}_q,\vect{k}_t} \psi_q^{} \psi_t^{}
\delta_{m+n,q+t}\; .
\end{align}
\begin{figure}[h!]
\centering
\includegraphics[width=0.85\columnwidth]{figure2.png}
\caption{Linear responses to a static defect of the three OPO states
in real and momentum space. Rescaled filtered OPO emissions (signal
[top panels], pump [middle], and idler [bottom]) in real space
$|\psi (\vect{r},\omega_n)|^2/|\psi_n|^2$ (left panels in linear
scale) and momentum space $|\psi_{\tilde{\vect{k}}} (\omega_n)|^2$
(right panel in logarithmic scale) obtained within the linear
response approximation. The parameters are the same ones of
Fig.~\ref{fig:spect}. For the top left panel of the signal emission
in real space, Gaussian filtering is applied to enhance the short
wavelength modulations, which amplitudes are otherwise roughly $1\%$
of the average signal intensity and about a factor of 10 times
weaker than the modulation amplitudes in the pump fluid.}
\label{fig:ereal}
\end{figure}
In absence of a defect potential ($\vectgr{\Psi}_d =0$),
Eq.~\eqref{eq:fluct} is the eigenvalue equation for the spectrum of
excitations of a homogeneous OPO, i.e.,
$\det(\mathcal{L}_{\tilde{\vect{k}}} - \omega)=0$. The spectrum has 6
branches, $\omega_{n,(u,v),\tilde{\vect{k}}}$, labeled by $n=s,p,i$
and $(u,v)$. Even though these degrees of freedom are mixed together,
at large momenta, one recovers the LP dispersions shifted by the three
states energies and momenta, i.e.,:
\begin{equation}
\lim_{\tilde{k} \gg \sqrt{2m_C \Omega_R}} \omega_{n,(u,v),\tilde{\vect{k}}} = \pm
(\omega^{LP}_{\vect{k} - \vect{k}_n} - \omega_n)\; ,
\label{eq:large}
\end{equation}
where $+$ ($-$) corresponds to the $u$ ($v$) particle- (hole-)like
branch.
The OPO solution is stable (shaded area in Fig.~\ref{fig:spect}) as
far as $\Im \omega_{n,(u,v),\tilde{\vect{k}}} < 0$.
The shape of the patterns, or Cherenkov-like waves, resulting from the
elastic scattering of the OPO 3-fluids against the static ($\omega=0$)
defect can be determined starting from the spectrum, and in particular
evaluating the closed curves $\Gamma_{n,(u,v), \tilde{\vect{k}}}$ in
$\vect{k}$-space, or ``Rayleigh rings''~\cite{Rousseaux2011} defined
by condition $\Re \omega_{n,(u,v), \tilde{\vect{k}}} =
0$~\footnote{Even if they do not appear to be relevant here, note that
the presence of a non-vanishing imaginary part of the excitation
spectrum $Im\omega_{n,(u,v),\tilde{\vect{k}}}\neq 0$ introduces some
complications: Even in the absence of any Rayleigh ring, the drag
force can be non-vanishing and the standard Landau criterion may
fail identifying a critical velocity~\cite{wouters2010}.}.
The modulations propagate with a direction
$\hat{\eta}_{n,(u,v),\tilde{\vect{k}}}$ orthogonal to each curve
$\Gamma_{n,(u,v), \tilde{\vect{k}}}$, a pattern wavelength given by
the corresponding $|\tilde{\vect{k}}|$, and a group velocity
$\vect{v}_{n,(u,v),\tilde{\vect{k}}}^{(g)}=\nabla_{\tilde{\vect{k}}}
\Re \omega_{n,(u,v), \tilde{\vect{k}}}$, where
$\xi_{n,(u,v),\tilde{\vect{k}}} =
|\vect{v}_{n,(u,v),\tilde{\vect{k}}}^{(g)} / \Im \omega_{n,(u,v),
\tilde{\vect{k}}}|$ determines the distance, at any given direction
$\hat{\eta}_{n,(u,v),\tilde{\vect{k}}}$, over which the perturbation
extends away from the defect. For a single fluid under a coherent
pump, the qualitative shape of the modulation pattern generated in the
fluid by the defect is mostly determined by the excitation
spectrum~\cite{carusotto06_prl,carusotto04}.
For OPO, the spectrum of excitation on top of each of the three,
$n=1,2,3$, states (see~\cite{SM}) generates six identical Rayleigh
rings $\Gamma_{n,(u,v), \tilde{\vect{k}}}$ for the 3 states.
The Rayleigh rings for the OPO conditions specified in
Fig.~\ref{fig:spect} are clearly visible in the right panels of
Fig.~\ref{fig:ereal}, where we plot the $\vect{k}$-space
photoluminescence filtered at each state energy, i.e.,
$|\psi_{\tilde{\vect{k}}} (\omega_n)|^2= |\psi_{n}^{}
\delta_{\tilde{\vect{k}},0} + u_{n,\tilde{\vect{k}}}^{} +
v^*_{n,-\tilde{\vect{k}}}|^2$.
We have here chosen a $\delta$-like defect potential, $V_d(\vect{k}) =
g_d$, but we have however checked that our results do not depend on
its exact shape~\cite{SM}.
For the OPO conditions considered here, the signal momentum is at $k_s
\simeq 0$, and thus only four of the six rings are present. The same
rings are also plotted in the right panel of Fig.~\ref{fig:spect},
shifted at each of the three OPO state momentum $\vect{k}_n$,
$\Gamma_{n,(u,v), \tilde{\vect{k}}+\vect{k}_n}$ and energies
$\omega_n$.
It is important to note that, even though the three OPO states have
locked responses because the three states display the same spectrum of
excitations, only one of the rings $\Gamma_{n,(u,v),
\tilde{\vect{k}}+\vect{k}_n}$ is the most resonant at $\omega_n$
with the interaction blue-shifted lower polariton dispersion,
\begin{equation}
\bar{\omega}_{\vect{k}}^{LP} = \omega_{\vect{k}}^{LP} + 2
X_{\vect{k}}^2 \sum_{n=1}^3 \epsilon_n\; ,
\label{eq:blues}
\end{equation}
where $\epsilon_n = g_X X_{\vect{k}_n}^2 |\psi_n^{}|^2$ are the
mean-field energy blue-shifts (measured in Fig.~\ref{fig:spect} in
units of $\gamma_p = \gamma_{\vect{k}_p}$). This implies that the
most visible modulation for each fluid should be the most resonant
one, with superimposed weaker modulations coming from the other two
state rings.
In the specific case of Fig.~\ref{fig:spect}, the signal is at $k_s
\simeq 0$ and thus produces no rings in momentum space. The other four
rings are very far from being resonant with the blue-shifted LP
dispersion~\eqref{eq:blues} at $\omega_s$, and thus the signal
displays only an extremely weak modulation coming from the next
closer ring, which is the one associated with the pump state,
$\Gamma_{p,u, \tilde{\vect{k}}+\vect{k}_s}$.
We estimate that the signal modulation amplitudes are roughly $1\%$ of
the average signal intensity and about a factor of 10 times weaker
than the modulation amplitudes in the pump fluid.
In order to show that the signal has weak modulations coming from
the pump, we apply a Gaussian filter to the real space images (see
inset of the left top panel of Fig.~\ref{fig:ereal}).
As explained in more details in~\cite{SM}, Gaussian filtering consists
of subtracting from the original data a copy which has been
convoluted with a Gaussian kernel, thus getting rid of the
long-wavelength modulations.
This procedure reveals that indeed the pump imprints its modulation
also into the signal, even though these are extremely weak, thus
leaving the signal basically insensitive to the presence of the
defect.
Pump and idler states are each mostly resonant with their own
rings, i.e., $\Gamma_{p,u, \tilde{\vect{k}}+\vect{k}_p}$ at $\omega_p$
and $\Gamma_{i,u, \tilde{\vect{k}}+\vect{k}_i}$ at $\omega_i$,
respectively. Thus one should then observe two superimposed
modulations in both pump and idler filtered emissions, the stronger
one for each being the most resonant one.
However, the modulations associated to the idler only propagate very
close to the defect, at an average distance
$\overline{\xi_{i,u,\tilde{\vect{k}}}} \sim 1.7~\mu$m before getting
damped, and thus are not clearly visible. For the OPO conditions
considered, this is due to the small idler group velocity
$\vect{v}_{i,u,\tilde{\vect{k}}}^{(g)}$, as the dispersion is almost
excitonic at the idler energy.
We can conclude that, for the typical OPO condition with a signal at
$k_s \simeq 0$, considered in Figs.~\ref{fig:spect}
and~\ref{fig:ereal}, the signal fluid does not show modulations and
the extremely weak scattering inherited from the pump state can be
appreciated only after a Gaussian filtering procedure of the image. In
contrast, the idler has a locked response to the one of the pump
state.
Note that, for the conditions shown in Fig.~\ref{fig:spect}, as well
as the other cases considered in Ref.~\cite{SM}, the subsonic to
supersonic crossover of the pump-only state~\cite{amo09_b} happens
well above the region of stability of OPO. Thus it is not possible to
study a case where the pump is already subsonic and at the same time
promotes stimulated scattering.
\begin{figure}[h!]
\centering \includegraphics[width=0.8\columnwidth]{figure3.png}
\caption{(Color online) Experimental OPO spectrum and filtered
emissions of signal, pump and idler in presence of a structural
defect. The six panels show the filtered emission profiles in real
space of signal (top), pump (middle) and idler (bottom) $I_{s,
p,i}(\vect{r})$. A Gaussian filtering to enhance the short
wavelength modulations is applied in the right column panels. Here,
the extracted wave-crests from the idler emission (yellow contours
in the bottom panel) are also superimposed to the pump profile
(middle) by applying a $\pi$-phase shift. The (orange) dot indicates
the position of the defect. The lower panel shows the experimental
OPO spectrum. Energy and momentum of the three OPO states are
labeled with a [blue] upper triangle (signal), a [red] circle
(pump), and a [green] lower triangle (idler), while the localised
state, clearly visible just below the bottom of the LP dispersion,
is indicated with the symbol $d$. The bare LP dispersion is
extracted from an off-resonant low pump power measurement, as well
as the emission of the exciton reservoir (X) and the one of the UP
dispersion (each in a different scale).}
\label{fig:exper}
\end{figure}
\section{Experiments}
We now turn to the experimental analysis. We use a continuous-wave
laser to drive a high quality ($Q=14000$) GaAs microcavity sample into
the OPO regime --- details on the sample can be found in a previous
publication~\cite{ballarini2013,dominici14}.
The polariton dispersion is characterised by a Rabi splitting
$\Omega_R=5.4$~meV, the exciton energy $\omega_0^{X}=1485.26$~meV and
we choose a sample region where the cavity-exciton detuning is
slightly negative, $-1$~meV. We pump at $k_p=0.89~\mu$m$^{-1}$ and
$\omega_p - \omega_0^{X}=-2.43$~meV, and, at pump powers $1.5$-times
above threshold, we obtain an OPO with signal at small wavevector
$k_s=0.21~\mu$m$^{-1}$ and $\omega_s - \omega_0^{X}=-2.95$~meV, and
idler at $k_i=1.57~\mu$m$^{-1}$ and $\omega_i -
\omega_0^{X}=-1.91$~meV.
The defect we use in the sample is a localized inhomogeneity naturally
present in the cavity mirror. Note that the exact location of the
defect can be extracted from the emission spectrum and is indicated
with a dot (orange) symbol in the profiles of Fig.~\ref{fig:exper}.
In order to filter the emission at the three states energies,
$I_{s,p,i}(\vect{r}=x,y)$, and obtain 2D spatial maps for the OPO
three states, we use a spectrometer and, at a fixed position $x_0$,
obtain the intensity emission as a function of energy and position,
$I(\epsilon,x_0,y)$. By changing $x_0$ we build the full emission
spectrum as a function of energy and 2D position,
$I(\epsilon,\vect{r})$. The filtered emission for each OPO state is
obtained from the integrals $I_{n=s,p,i}(\vect{r}) =
\int_{\omega_n-\sigma}^{\omega_n+\sigma} d\epsilon I(\epsilon,
\vect{r})$, with $\sigma=0.08$~meV. The results are shown in
Fig.~\ref{fig:exper} for respectively the signal (top panel), pump
(middle) and the idler (bottom) profiles.
The signal profile shows no appreciable modulations around the defect
locations, nor these could be observed after applying a Gaussian
filtering procedure of the image.
In contrast, in agreement with the theoretical results, both filtered
profiles of pump and idler show the same Cherenkov-like pattern. We
extract the wave-crests from the idler profile ([yellow] contours in
the bottom panel) and superimpose them to the pump profile (middle
panel) with an added $\pi$-phase-shift, revealing that the only
modulations visible in the idler state are the ones coming from the
pump state.
\section{Numerical analysis}
The agreement between the results obtained experimentally and within
the linear response approximation is additionally confirmed by an
exact full numerical analysis of the coupled
equations~\eqref{eq:gpequ} for a finite size pump via a
5$^{\text{th}}$-order adaptive-step Runge-Kutta algorithm.
Details are given in~\cite{SM}.
The pumping conditions are very similar to those previously considered
in the linear response approximation of Figs.~\ref{fig:spect}
and~\ref{fig:ereal}, while the pump profile $\mathcal{F}_p(\vect{r})$
is now finite-size top-hat. In particular, we consider the case of
zero cavity-exciton detuning, $k_p=1.6$~$\mu$m$^{-1}$,
$\omega_p-\omega_X^0 = -0.44$~meV and the pump power strength is fixed
just above threshold, so that to produce a stable steady state OPO
with, in absence of the defect, signal at small wavevector
$k_s=-0.2$~$\mu$m$^{-1}$ and idler at $k_i=3.4$~$\mu$m$^{-1}$.
When adding a localised defect potential, the steady state OPO
develops Rayleigh rings in momentum space, yet, as shown in~\cite{SM},
the spectrum continues to be $\delta$-like in energy, allowing to
easily filter in energy the emission of the three OPO states.
Results are shown in Fig.~\ref{fig:numer}, where real-space emissions
$|\psi_C (\vect{r},\omega_n)|^2$ are plotted in the left panels, while
the ones in momentum space $|\psi_C (\tilde{\vect{k}},\omega_n)|^2$ in
the right panels. We observe a very similar phenomenology to that one
obtained in the linear approximation shown in
Fig.~\ref{fig:ereal}. The signal now is at slightly negative values of
momenta $k_s=-0.2$~$\mu$m$^{-1}$, thus implying a very small Rayleigh
ring associated with this state. Thus we observe that only the
modulations associated with the pump are the ones that are weakly
imprinted in the signal state and that can be observed by means of a
Gaussian filtering (inset of top-left panel). We have fitted the
upstream wave-crests and obtained the same modulation wave-vector as
the pump one ([blue] upper triangles). Similarly to the linear
response case, we also find here that the most visible perturbation in
the emission filtered at the idler energy is the one due to the pump
Rayleigh ring. As before, the modulations due to the idler Rayleigh
ring cannot propagate far from the defect because of the small group
velocity associated with this state.
\begin{figure}[h!]
\centering
\includegraphics[width=0.85\columnwidth]{figure4.png}
\caption{(Color online) Full numerical responses to a static defect of
the three OPO states in real and momentum space. Filtered OPO
emissions (signal [top panels], pump [middle], and idler [bottom])
in real space $|\psi_C (\vect{r},\omega_n)|^2$ (left panels in
linear scale) and momentum space $|\psi_C
(\tilde{\vect{k}},\omega_n)|^2$ (right panel in logarithmic scale)
obtained by a full numerical evaluation of~\eqref{eq:gpequ}. For the
top left panel of the signal space emission, Gaussian filtering is
applied to enhance the short wavelength modulations of this state,
revealing that the modulations corresponding to the pump state are
also imprinted (though weakly) into the signal. The symbols indicate
the pump ring diameter extracted from fitting the upstream
modulations and resulting in a density wave wave-vector coinciding
with the one of the pump $k_p=1.6$~$\mu$m$^{-1}$.}
\label{fig:numer}
\end{figure}
\section{Conclusions}
To conclude, we have reported a joint theoretical and experimental
study of the superfluid properties of a non-equilibrium condensate of
polaritons in the so-called optical parametric oscillator
configuration by studying the scattering against a static defect.
We have found that while the signal is basically free from
modulations, pump and idler lock to the same response. We have
highlighted the role of the coupling between the OPO components by
non-linear and parametric processes. These are responsible for the
transfer of the spatial modulations from one component to the
other. This process is most visible in the clear spatial modulation
pattern that is induced by the non-superfluid pump onto the idler,
while the same modulations are only extremely weakly transferred into
the signal, because of its low characteristic wavevector, so much that
experimentally cannot be resolved.
The main features of the real- and momentum-space emission patterns
are understood in terms of Rayleigh scattering rings for each
component and a characteristic propagation length from the defect; the
rings are then transferred to the other components by nonlinear and
parametric processes.
Our theoretical and experimental results stress the complexities and
richness involved when looking for superfluid behaviours in
non-equilibrium multicomponent condensates such as the ones obtained
in the optical parametric oscillation regime.
\acknowledgments We are grateful to M. Wouters, C. Tejedor and M. De
Giorgi for stimulating discussions.
Financial support from the ERC POLAFLOW (grant n. 308136) is
acknowledged.
FMM acknowledges financial support from the Ministerio de Econom\'ia y
Competitividad (MINECO, contract No. MAT2011-22997), the Comunidad
Autonoma de Madrid (CAM, contract No. S-2009/ESP-1503), and the
European Science Foundation (ESF) program Intelbiomat.
IC acknowledges financial support by the ERC through the QGBE grant
and by the Autonomous Province of Trento, partly through the ``On
silicon chip quantum optics for quantum computing and secure
communications'' (``SiQuro'').
MHS acknowledges support from EPSRC (grants EP/I028900/2 and
EP/K003623/2).
\newpage
\pagenumbering{gobble}
\setcounter{secnumdepth}{2}
\renewcommand{\figurename}{\textsc{S.~Fig.}}
\setcounter{equation}{0}
\setcounter{figure}{0}
\onecolumngrid
\begin{center}
\large
\textbf{Supplemental Material for ``On multicomponent polariton
superfluidity in the optical parametric oscillator regime''}
\end{center}
\vspace{3ex}
\twocolumngrid
In the manuscript, we carry on a theoretical and experimental analysis
of the response of microcavity polaritons in the optical parametric
oscillator (OPO) regime to a static defect.
For the theoretical calculations we follow two independent approaches:
In the first approach, we exactly numerically solve the dynamics of
the two coupled Gross-Pitaevskii equations (GPEs) for the exciton and
cavity fields for the realistic case of a finite-size top-hat pump
profile $\mathcal{F}_p(\vect{r})$. In the second approach, we apply a
perturbative linear response approximation for the lower polariton
(LP) state which leads to semi-analytical results in the limiting case
of a spatially homogeneous pump profile. Both methods have been
already successfully used in the literature to explore several
properties of the OPO operation.
While all fundamental formulae and information has been given in the
main text, in this Supplementary Material (SM) we present some
additional details on both approaches that might be of interest to the
specialized reader. In addition to that, we make use of the linear
response approach to examine some regimes that are hardly accessed
experimentally or within a full numerical approach, which however
allow to put the conclusions of our work into a broader perspective.
\section{Full numerics}
\label{sec:detun}
The classical driven-dissipative non-linear Gross-Pitaevskii equation
(GPE) for the coupled exciton and cavity fields $\psi_{X,C}
(\vect{r},t)$ ($\hbar=1$)
\begin{equation}
i\partial_t \begin{pmatrix} \psi_X \\ \psi_C \end{pmatrix} =
\hat{H} \begin{pmatrix} \psi_X \\ \psi_C \end{pmatrix}
+ \begin{pmatrix} 0 \\ F_p(\vect{r},t) \end{pmatrix}\; ,
\label{eq:numer}
\end{equation}
where
\begin{equation}
\hat{H} = \begin{pmatrix} \omega^{X}_{-i\nabla} - i
\frac{\gamma_X}{2} + g_X |\psi_X|^2 & \Omega_R/2 \\ \Omega_R/2 &
\omega^C_{-i\nabla} - i \frac{\gamma_C}{2} + V_d \end{pmatrix} \;
,
\end{equation}
is solved numerically on a 2D grid of $N\times N=2^8\times 2^8$ points
and a separation of $0.47$~$\mu$m (i.e., in a box $L\times L =
121$~$\mu$m$\times 121$~$\mu$m) by using a 5$^{\text{th}}$-order
adaptive-step Runge-Kutta algorithm. Convergence has been checked both
with respect the resolution in space $L/N$ as well as in momentum
$\pi/L$, without~\cite{marchetti10,marchetti_review} as well as in
presence of the defect.
The same approach has been already used in previous publications from
some of the authors (for a review, see
Refs.~\cite{marchetti10,marchetti_review}).
As for the system parameters, we have considered a LP dispersion at
zero photon-exciton detuning, $\omega^C_0 = \omega^X_0$, a
dispersionless excitonic spectrum, $\omega^X_{\vect{k}} = \omega^X_0$
and a quadratic dispersion for photons $\omega^C_{\vect{k}} =
\omega^C_0 + k^2/2m_C$, with the photon mass $m_C=2.3 \times 10^{-5}
m_e$, where $m_e$ is the bare electron mass. The LP dispersion, $2
\omega_{\vect{k}}^{LP} = \omega_{\vect{k}}^{C} + \omega_{\vect{k}}^{X}
- \sqrt{(\omega_{\vect{k}}^{C} - \omega_{\vect{k}}^{X})^2 +
\Omega_R^2}$ is characterised by a Rabi splitting $\Omega_R =
4.4$~meV. Further, the exciton and cavity decay rates are fixed to
$\gamma_X=\gamma_C=0.53$~meV.
For the defect we choose a $\delta$-like defect potential
\begin{equation}
V_d(\vect{r}) = g_d \delta(\vect{r} - \vect{r}_0)\; ,
\end{equation}
where its location $\vect{r}_0$ is fixed at one of the $N \times N$
points of the grid.
Note that in a finite-size OPO, local currents lead to inhomogeneous
OPO profiles in within the pump spot, despite the external pump having
a top-hat profile with a completely flat inner
region~\cite{marchetti10,marchetti_review} --- as shown later, this
can be observed in the filtered OPO profiles evaluated in absence of a
defect shown as dashed lines in S.~Fig.~\ref{fig:rafull}.
We have thus chosen the defect location so that it lies in the
smoothest and most homogeneous part of the OPO profiles.
Also we have checked that our results do not qualitatively depend on
the strength $g_d$ (nor on the sign) of the defect potential, as far
as this does not exceed a critical value above which it destabilises
the OPO steady-state regime.
The pump, $F_p(\vect{r},t) = \mathcal{F}_p(\vect{r}) e^{i (\vect{k}_p
\cdot \vect{r} - \omega_p t)}$, has a smoothen and rotationally
symmetric top-hat profile, $\mathcal{F}_p(\vect{r}) = \mathcal{F}_p(r)
= \frac{f_p}{2}[\tanh(\frac{r+\sigma_p}{r_0}) -
\tanh(\frac{r-\sigma_p}{r_0})]$ with strength $f_p = 1.23
f_p^\text{th} = 0.053$~meV/$\mu$m and parameters $r_0 = 8.68~\mu$m,
$\sigma_p = 34.72~\mu$m.
We pump at $k_p=1.6$~$\mu$m$^{-1}$ in the $x$-direction, $\vect{k}_p =
(k_p,0)$, and at $\omega_p-\omega_X^0=-0.44$~meV, i.e., roughly
$0.5$~meV above the bare LP dispersion. By increasing the pump
strength $f_p$, we find the threshold $f_p^{\text{th}}$ above which
OPO switches on, leading to two conjugate signal and idler states.
We then fix the pump strength just above this threshold ($f_p=1.23
f_p^{\text{th}}$), where we find a steady state OPO solution which is
stable (see Ref.~\cite{marchetti_review} for further details). In
absence of the defect, this condition corresponds to a signal state at
$k_s=-0.2$~$\mu$m$^{-1}$ and $\omega_s-\omega_0^X = -1.64$~meV and an
idler at $k_i=3.4$~$\mu$m$^{-1}$ and $\omega_i-\omega_0^X =
0.76$~meV.
It is interesting to note that already very close to the
lower pump power threshold for OPO, the selected signal momentum is
very close to zero. This contrasts with what one obtains in the linear
approximation scheme, where instead just above the lower OPO threshold
there exists a broad interval of permitted values for $k_s$ (and thus
$k_i$) --- we will discuss further this ``selection problem'' for
parametric scattering later in the section ``Linear response''.
\begin{figure}[h!]
\includegraphics[width=1.0\columnwidth]{GP_disp.png}
\caption{(Color online) OPO spectrum obtained by full numerics. Left
panel: Photonic component of the OPO spectrum in presence of a
point-like defect, $|\psi_C(k_x,0,\omega)|^2$ (logarithmic scale),
as a function of the rescaled energy $\omega - \omega_0^X$ versus
the $x$-component of momentum $k_x$ (cut at $k_y=0$) for a top-hat
pump (see text for the space profile and parameter values), with
intensity $f_p=1.23 f_p^{\text{th}}$ above the OPO threshold, pump
wave-vector $k_p=1.6$~$\mu$m$^{-1}$ in the $x$-direction and
$\omega_p-\omega_X^0=-0.44$~meV. The symbols indicate the signal
([blue] upper triangle), pump ([red] circle), and idler ([green]
lower triangle) energies, as well as the two momenta $k_x$ on each
state Rayleigh ring at $k_y=0$. Note that the logarithmic scale
results in a fictitious broadening in energy of the spectrum, which
is in reality $\delta$-like (see right panel). The bare LP
dispersion, including its broadening due to finite lifetime, is
plotted as a shaded grey region, while the bare UP dispersion as a
(black) dot-dashed line. Right panel: Momentum integrated spectrum,
$\sum_{\vect{k}} |\psi_C(\vect{k},\omega)|^2$ (linear scale) as a
function of the rescaled energy $\omega - \omega_0^X$, where it can
be clearly appreciated that the emission is $\delta$-like in
energy.}
\label{fig:spectGP}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.8\columnwidth]{rangesGP.png}
\caption{Real space signal, pump and idler OPO one-dimensional
filtered profiles derived from finite size numerics with and without
a defect. Rescaled OPO filtered emissions along the $y=0$ direction,
$|\psi_C(x,y=0,\omega_n)|^2 \frac{\Omega_R}{2g_X}$, obtained by
numerically solving the GPE Eq.~(1) of the main manuscript. While
the dashed lines represent the filtered emissions of signal (top
panel), pump (middle) and idler (bottom) for a top-hat pump without
a defect, the solid lines are the same OPO conditions but now for a
defect positioned at $(x_d, y_d) = (9.5, -0.5)$~$\mu$m corresponding
to the vertical dotted lines. The system parameters are the same
ones as those of Fig.~4 of the main manuscript.}
\label{fig:rafull}
\end{figure}
We evaluate the time dependent full numerical solution
of~\eqref{eq:numer} $\psi_{X,C} (\vect{r},t)$, until a steady state
regime is reached. Here, both its Fourier transform to momentum
$\vect{k}$ and energies $\omega$ can be evaluated numerically.
We plot on the left panel of S.~Fig.~\ref{fig:spectGP} a cut at $k_y=0$
of the photonic component of the OPO spectrum in presence of a
point-like defect, $|\psi_C(k_x,0,\omega)|^2$, as a function of the
rescaled energy $\omega - \omega_0^X$ versus the $x$-component of
momentum $k_x$ (cut at $k_y=0$). In the right panel we plot instead
the corresponding momentum integrated spectrum, $\sum_{\vect{k}}
|\psi_C(\vect{k},\omega)|^2$.
Here, we can clearly see that the presence of the defect does not
modify the fact that the OPO emission for the OPO signal ([blue] upper
triangle), pump ([red] circle), and idler ([green] lower triangle)
states has a completely flat dispersion in energy, thus indicating
that a stable steady state OPO solution has been reached. Note that in
the spectrum map of the left panel of Fig.~\ref{fig:spect}, the
logarithmic scale results in a fictitious broadening in
energy. However, from the integrated spectrum plotted in linear scale
in the right panel of Fig.~\ref{fig:spect} one can clearly appreciate
that this emission is $\delta$-like, exactly as it happens for the
homogeneous OPO case~\cite{marchetti_review}.
Thus the effect of the defect is to induce only elastic (i.e., at the
same energy) scattering; now the three OPO states emit each on its own
Rayleigh ring (given each by the symbols on the left panel of
S.~Fig.~\ref{fig:spectGP} which represent the rings at a cut for
$k_y=0$). This makes it rather difficult to extract the separated
signal, pump and idler profiles by filtering in momentum, as done
previously for the homogeneous case, but it still allows to filter
those profiles very efficiently in energy. In fact, because the
emission is $\delta$-like, it is enough to fix a single value of the
energy $\omega$ to the one of the three states $\omega_{n=s,p,i}$,
thus extracting the filtered profiles either in real space
$|\psi_C(\vect{r},\omega_n)|^2$ or in momentum space
$|\psi_C(\vect{k},\omega_n)|^2$ --- we have however checked that
integrating in a narrow energy window around $\omega_n$ does not
quantitatively change the results.
The results of the above described filtering are shown in Fig.~4 of
the manuscript.
In S.~Fig.~\ref{fig:rafull} we instead plot the corresponding
one-dimensional profiles in the $y=0$ direction both in presence
(solid line) and without (dashed line) a defect. Here, we can observe
that, even if the pump has a top-hat flat profile, as also commented
previously, in absence of a defect, the finite-size OPO is
characterised by inhomogeneous profiles of signal, pump and idler
because of localised currents. Further, we observe that the presence
of a defect induces strong modulations in pump and idler.
In order to reveal that the pump also imprints its modulation into the
signal, even though these are extremely weak (and hardly visible in
both the top main panel of Fig.~4 of the manuscript and the top panel
of S.~Fig.~\ref{fig:rafull}), we apply a Gaussian filtering to the
signal images, which result is shown in the inset of the top-left
panel of Fig.~4 in the manuscript.
This consists of the following procedure. The original data for the
real space profile $\psi(\vect{r})$ are convoluted with a Gaussian
kernel $K(\vect{r} - \vect{r}')$, obtaining a new profile,
$\tilde{\psi}(\vect{r}') = \int d\vect{r} \psi (\vect{r}) K(\vect{r} -
\vect{r}')$, where short wavelengths features are smoothen out. The
convoluted image $\tilde{\psi}(\vect{r}')$ is then subtracted from the
original data, giving $\psi(\vect{r}) - \tilde{\psi}(\vect{r})$, and
effectively filtering out all long wavelength details.
The same procedure of Gaussian filtering is also applied to the signal
emission profile obtained within the linear response analysis and
shown in the inset of the top left panel of Fig.~2 of the manuscript.
\begin{figure}[h!]
\includegraphics[width=0.9\columnwidth]{fig_excitation_ks_0_000.png}
\caption{Spectrum of collective excitations. Cut at $k_y=0$ of the
real part of the quasiparticle energy dispersion $\Re
\omega_{n,(u,v),\vect{k}-\vect{k}_n}$ plotted versus $k_x -
k_n$. The spectrum is evaluated within the linear approximation
scheme and the parameters are the same ones used for Fig.~2 of the
main manuscript.}
\label{fig:bogol}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=1.0\columnwidth]{3d0_7.png}
\includegraphics[width=0.85\columnwidth]{ks0_7.png}
\caption{(Color online) OPO for finite and positive signal momentum in
the linear response approximation for homogeneous pumping. Rescaled
OPO emission within the linear response approximation in real $|\psi
(\vect{r},\omega_n)|^2/|\psi_n|^2$ (left panels in linear scale) and
momentum space $|\psi_{\tilde{\vect{k}}} (\omega_n)|^2$ (right
panels in linear scale) filtered at the energies of the three OPO
states: signal (top panels), pump (middle), and idler (bottom). The
parameters are the same as those used in Fig.~2 of the main
manuscript, with the exception of the signal and idler momenta,
which are now at $k_s = 0.7$~$\mu$m$^{-1}$ and $k_i =
2.5$~$\mu$m$^{-1}$ respectively. Each scale of variation for the
profiles is plotted in the color-box next to the left panels. A cut
in the $y=0$ direction for the three profiles is plotted in the
bottom panel of S.~Fig.~\ref{fig:range}.}
\label{fig:ksp07}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=1.0\columnwidth]{3d-0_4.png}
\includegraphics[width=0.85\columnwidth]{ks-0_4.png}
\caption{(Color online) OPO for finite and negative signal momentum in
the linear response approximation for homogeneous pumping. Rescaled
OPO emission within the linear response approximation in real $|\psi
(\vect{r},\omega_n)|^2/|\psi_n|^2$ (left panels in linear scale) and
momentum space $|\psi_{\tilde{\vect{k}}} (\omega_n)|^2$ (right
panels in linear scale) filtered at the energies of the three OPO
states: signal (top panels), pump (middle), and idler (bottom). The
parameters are the same as those used in Fig.~2 of the main
manuscript, with the exception of the signal and idler momenta,
which are now at $k_s = -0.4$~$\mu$m$^{-1}$ and $k_i =
3.6$~$\mu$m$^{-1}$ respectively. A cut in the $y=0$ direction for
the three profiles is plotted in the top panel of
S.~Fig.~\ref{fig:range}.}
\label{fig:ksm04}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=0.8\columnwidth]{ranges.png}
\caption{(Color online) Real space signal, pump and idler OPO
one-dimensional filtered profiles in the linear response
approximation. OPO filtered emissions along the $y=0$ direction, $I
(x, y=0, \omega_n) = |\psi(x,y=0, \omega_n)|^2/|\psi_n|^2$ rescaled
by the mean-field solution $\psi_n$ in absence of the defect, for
the three cases analysed above within the linear response
approximation: the case of a signal at $k_s = -0.4$~$\mu$m$^{-1}$
(top panel) corresponds to the same conditions as
S.~Fig.~\ref{fig:ksm04} , a signal at $k_s = 0.0$~$\mu$m$^{-1}$
(middle panel) corresponds to Fig.~2 of the main manuscript, and a
signal at $k_s = 0.7$~$\mu$m$^{-1}$ (bottom panel) corresponds to
S.~Fig.~\ref{fig:ksp07}. In each panel we plot, on the same linear
scale, the filtered profiles of signal ([blue] dashed line), pump
([red] solid), and idler ([green] dotted), while the horizontal gray
dashed lines represent the values of the mean-field emission prior
to adding a defect (rescaled here to $1$).}
\label{fig:range}
\end{figure}
\begin{figure}[h!]
\includegraphics[width=1.0\columnwidth]{uncoupled.png}
\caption{Real space profiles of three uncoupled signal, pump and idler
fluids. The rescaled profiles $|\psi
(\vect{r},\omega_n)|^2/|\psi_n|^2$ are obtained by setting all
off-diagonal couplings in Eq.~(6) of the main manuscript to zero,
resulting in three uncoupled signal (left column), pump (middle
column), and idler (right column) fluids. The three rows correspond
to the three different cases previously analysed within the linear
response approximation: the case of a signal at $k_s =
-0.4$~$\mu$m$^{-1}$ (top row) corresponds to the same conditions as
S.~Fig.~\ref{fig:ksm04} , a signal at $k_s = 0.0$~$\mu$m$^{-1}$
(middle row) corresponds to Fig.~2 of the main manuscript, and a
signal at $k_s = 0.7$~$\mu$m$^{-1}$ (bottom row) corresponds to
S.~Fig.~\ref{fig:ksp07}.}
\label{fig:uncou}
\end{figure}
\section{Linear response}
\label{sec:analy}
In the manuscript we also make use of the linear response
approximation to analyse the OPO response to a static defect, valid
for a homogeneous pump scheme. In order to do so, we first evaluate
the Bogoliubov matrix given in Eq.~(6) of the manuscript, whose
eigenvalues determine the spectrum of collective excitations. We plot
in S.~Fig.~\ref{fig:bogol} a typical collective dispersion (here we
consider the same system parameters as the ones used in Fig.~2 of the
manuscript), by plotting the real part of the Bogoliubov matrix
eigenvalues $\Re \omega_{n,(u,v),\vect{k}-\vect{k}_n}$ as a function
of $k_x - k_n$ (cut at $k_y=0$). Note that the Rayleigh rings can be
found by finding the intersections $\Re
\omega_{n,(u,v),\vect{k}-\vect{k}_n}=0$.
As also done for the full numerics, we
consider there a $\delta$-like disorder potential $V_d(\vect{r}) = g_d
\delta(\vect{r} - \vect{r}_0)$, and Figs.~1 and~2 of the manuscript,
as well as S.~Figs.~\ref{fig:ksp07} and~\ref{fig:ksm04} below, are
plotted for this case.
We have however checked that our results do not depend on the specific
shape of the defect potential: In particular we have also considered
the response to defects with smoothen Gaussian-like profiles, whose
effect is only to partially weaken the upstream modulations emission
in real space.
We have seen in the manuscript that both experimentally as well as in
the full numerical analysis, one obtains OPO conditions where the
signal momentum is very close to zero, $\vect{k}_s \sim 0$. The reason
why OPO selects almost zero momentum signal conditions is still
awaiting a theoretical explaination.
In contrast, the linear response approximation allows to access very
different OPO mean-field conditions, for which, by leaving unaltered
the value of the pump momentum, the signal can appear at finite
momentum values, either positive or negative.
This is a peculiarity of this approximation scheme, where one can show
that, at mean-field level, one has the possibility of choosing
different values of the signal momentum $\vect{k}_s$ (and thus the one
of the idler $\vect{k}_i$), and that this choice range is quite broad
particularly when the pump power is close to the lower pump threshold
for OPO --- the stability region for OPO is plotted as a shaded grey
region in the top left panels of S.~Figs.~\ref{fig:ksp07}
and~\ref{fig:ksm04}. This is a well known ``selection problem'' for
parametric scattering (see Ref.~\cite{wouters07:prb}): The reason why
in full numerical calculations, parametric scattering processes select
a signal with a momentum close to zero, already very close to the
lower pump thresold for OPO, is still awaiting for an explaination. In
particular, this quest cannot be addressed within a spatially
homogeneous approximation where the three mean-field solutions for
pump, signal, and idler states can be describes by plane waves.
One can only show that, within the same mean-field approximation
scheme, when increasing the pump power towards the upper threshold for
OPO, the blue-shift of the LP polariton dispersion due to the
increasing mean-field polariton density, causes the signal momentum to
converge towards zero~\cite{whittaker05} $\vect{k}_s \to 0$.
\paragraph{OPO conditions with a finite momentum signal ---}
Given the freedom of choice for the signal momentum $\vect{k}_s$ close
to the lower OPO threshold, we consider here two additional cases,
that could not be studied neither experimentally, nor within a full
numerical approach, but that instead we can easily analyse within the
linear response theory.
In particular, we have left fixed the pumping conditions
$k_p=1.6$~$\mu$m$^{-1}$ and $\omega_p-\omega_X^0=-1.25$~meV and
considered two opposite situations.
In the first case, the signal has a finite and positive momentum $k_s
= 0.7$~$\mu$m$^{-1}$ and thus the idler is at low momentum, $k_i
=2.5$~$\mu$m$^{-1}$. The results are shown in S.~Fig.~\ref{fig:ksp07}.
Here we see that all six Rayleigh rings are clearly visible and, in
addition, as the idler is at lower momentum compared to the case
considered in Figs.~1 and~2 of the manuscript, and thus its dispersion
steeper, the idler group velocity is large enough to appreciate the
modulation of the Rayleigh ring associated to this state. As a result,
each of the three filtered OPO emissions exhibits as the strongest
modulation the one coming from its own Rayleigh ring, included the
signal which is now at finite and large momentum. In this case, the
OPO response of each filtered state profile looks completely
independent from the other, as if we were pumping each state
independently.
In the second case shown in S.~Fig.~\ref{fig:ksm04} , the signal is
finite and negative, $k_s = -0.4$~$\mu$m$^{-1}$, and the idler is now
at very large momentum, $k_i = 3.6$~$\mu$m$^{-1}$, where its
dispersion is very exciton-like and flat, and thus the idler has a
very small group velocity and its own modulations are visible only
very close to the defect. For this case, we can appreciate in the
idler filtered profile overlapped modulations from all the three state
Rayleigh rings (note that because the signal is at negative momentum,
its modulations have an opposite direction compared to the ones of
pump and idler), while in the signal we can mostly see the signal long
wavelength modulations and only very weakly the pump one.
We can compare the different modulation strengths of the three OPO
profiles by looking at the color bars plotted next to the profiles. In
order to better compare them on the same plot, we show in
S.~Fig.~\ref{fig:range} the one-dimensional OPO filtered emissions
along the $y=0$ direction for the three cases analysed above. While
for the OPO conditions with a signal at $k_s = 0.0$~$\mu$m$^{-1}$
(middle panel, corresponding to Fig.~2 of the main manuscript), the
imprinted modulations from the pump are hardly visible, and can only
be appreciated after a Gaussian filter manipulation, both OPO cases
with a finite momentum signal result in modulations in the signal with
an amplitude of the same order of magnitude of both pump and idler
fluids.
In S.~Fig.~\ref{fig:uncou} we instead show the real space profiles
obtained by setting all off-diagonal couplings in the Bogoliubov
matrix $\mathcal{L}_{\vect{k}}$ of Eq.~(6) of the main manuscript to
zero, resulting in three uncoupled signal (left column), pump (middle
column), and idler (right column) fluids. This underlines the
importance of the coupling between the three fluids in the three
different OPO regimes previously analysed within the linear response
approximation. In particular, for the OPO conditions such that the
signal has a finite and positive momentum $\vect{k}_s$ (bottom row of
S.~Fig.~\ref{fig:uncou}, which corresponds to the conditions shown in
S.~Fig.~\ref{fig:ksp07}), the coupling has little effect and the three
fluid respond to the defect in practice in a independent way (each
Rayleigh ring influences its own fluid). The OPO condition with a
finite and negative momentum $\vect{k}_s$, which corresponds to the
conditions shown in S.~Fig.~\ref{fig:ksm04}, are shown in the top row
of S.~Fig.~\ref{fig:uncou}: Here, we can see that the coupling between
the three fluids plays an effect. The modulations of the pump can be
appreciated both in the signal profile (though weakly), as well as in
the idler profile. In the idler fluid its own modulations can only
propagate very close to the defect and thus the only really
appreciable modulations are the ones inherited from the pump. Finally,
for the experimentally relevant case of an OPO with a signal at zero
momentum (middle row of S.~Fig.~\ref{fig:uncou}, which corresponds to
the conditions shown in Fig.~2 of the manucript), there is also a role
played by the coupling, though, as already thoroghly analysed, the
modulations in the signal inhereted from the pump can only be
appreciated after a Gaussian filtering manipulation.
Finally note that, as it also happens for the OPO conditions shown in
Fig.~1 of the manuscript, the subsonic to supersonic crossover of the
pump-only state~\cite{amo09_b} happens at pump intensities well above
the region of stability of OPO --- the shaded gray regions of
S.~Figs.~\ref{fig:ksp07} and~\ref{fig:ksm04} Thus, it is not possible
to study a case where the pump is already subsonic and at the same
time promotes stimulated scattering.
|
\section{Introduction}
Solid state systems inevitably contain impurities. Study of the impurity effects is of special importance in topological insulators (for a review see, e.g.\ Ref.~\onlinecite{hasa2010, hasa2011, fran2013, qi2011}), because their surface states are expected to be fundamentally robust with respect to certain types of disorder. Understanding how impurities affect these systems also has practical implications because
the fabrication of topological insulators (TIs) requires fine tuning of the doping concentration to place the Fermi energy within the band gap\cite{hsie2009, chen2009,xia2009} or to induce superconductivity\cite{wray2011,sasa2011,hsie2012}.
Impurities that are located at or near the TI surface are especially interesting. One relevant example is
the treatment with $\mathrm{NO}_2$ (see supplement of Ref.~\onlinecite{hsie2009}) on the surface of as-grown topological insulator $\mathrm{Bi}_{2-x}\mathrm{Cu}_x\mathrm{Se}_3$ which is necessary to prevent the surface band bending caused by the adsorption of residual atoms present in the vacuum chamber. As the manifestations of topological systems -- such as the protected surface states or Majorana fermions -- are localized near the surface or the edge of the system \cite{mour2012,das2012, alic2012, been2013, elli2015}, the effect of surface impurities on the surface spectral and transport properties are of both theoretical and practical interest\cite{schu2012, pien2012, rain2013}.
Local density of states (LDOS) can be efficiently probed by scanning tunneling spectroscopy while the surface spectral function can be extracted from
angle resolved photoemission spectroscopy (ARPES) measurements. Various techniques have been employed to study the transport properties of TI surfaces.
Recently, an exact diagonalization (ED) approach has been applied to non-interacting TIs with surface disorder\cite{schu2012,quei2014}. This study found evidence for a crossover between a nearly ballistic response of the surface electrons at weak disorder, localization physics at intermediate disorder, and then a restored nearly ballistic surface state hiding in the second layer when disorder is very strong. Further investigations of this phenomenon using ED are going to prove very challenging: the computational cost for analyzing the surface physics can be very high because the bulk of the system---although gapped---must be treated on equal footing with the surface degrees of freedom. The system sizes accessible to ED analysis are, therefore, rather small.
In particular, we demonstrate here that when strong surface impurities are present in the system, the lower bound of system size required to clearly resolve the bulk electronic properties and their effect on the surface states becomes large in proportion to the impurity potential strength (at least in two-dimensions). Treating sufficiently large systems using ED becomes computationally challenging in this limit and new techniques to address the problem are required. In this manuscript we develop such a technique. We also emphasize the need to treat sufficiently large systems in order to distinguish system properties
at the thermodynamic limit versus those of the ``quantum dot'' regime, where finite size effects dominate.
In this manuscript, we introduce a new technique that allows us to efficiently extract the surface state properties of surface-disordered TIs. We obtain properties such as the surface spectral function, LDOS and transport properties of the surface channels, by essentially ``integrating out'' the clean bulk degrees of freedom analytically and obtain the effective surface-Hamiltonian describing a TI surface with arbitrarily strong impurities. This approach not only allows us to reduce the computational difficulty by one dimension (e.g.\ for a 3D TI with surface disorder we only need to solve a 2D problem), but also allow us to map a strong disordered problem into a weak disordered one where perturbation theory is valid. By constructing a self-consistent transfer-matrix approach, we are able to recover the exact energies and wavefunctions of the surface states both at the disordered layer and in the remaining bulk layers.
The manuscript is organized as follows. In Sec.~\ref{sec:gene_fram}, we first explain how to integrate out the clean bulk degrees of freedom.
In Sec.~\ref{sec:appl_to} we then introduce a generic model Hamiltonian of a 2D TI with impurities on one of its edges. While our method is applicable to any layered system in arbitrary dimension, we choose to concentrate on the 2D case since it is both simple to present and analyze. Next, we report a series of results that clearly differentiate the 0D quantum-dot regime from the bulk regime (where finite-size errors are suppressed) and specify the lower bound on the system size to observe the latter for magnetic and non-magnetic edge impurities. We conclude the section with a discussion of the surface properties.
Lastly, in Sec.~\ref{sec:tran_beha}, the conductance through the 2D TI edge channels is computed. The latter provides complementary information to the spectral properties discussed in Sec.~\ref{sec:gene_fram}. We conclude with a discussion in Sec.~\ref{sec:conc_and}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=85mm]{picture/layer_moti3.pdf}
\caption{A pictorial description of a system with impurities on the surface only. We are interested in studying the local density of surface state as the impurity strength increases. Our strategy is to decompose the system into clean layers, coupled through matrices $B$ and $B^{\dagger}$, and a surface containing impurities, and then analytically integrate out the former. }
\label{fig:3d_lattice}
\end{center}
\end{figure}
\section{General framework: effective single layer Hamiltonian}
\label{sec:gene_fram}
In this section, we provide a general derivation of our approach. Our goal is to exactly reduce diagonalization of a $D$-dimensional system with surface impurities to the diagonalization of an effective Hamiltonian describing just the $D-1$ dimensional surface. Taking the top surface to be disordered, we introduce a way to integrate out the clean layers from the bottom all the way to the top layer. This leaves us with a single layer effective Hamiltonian which includes the impurity potential and a self energy which accounts for the entire clean bulk.
\subsection{Layered Schr\"odinger equation and self energy}
\label{sec:tran_matr}
We begin our analysis with the Schr\"odinger equation for layers parallel to the disordered surface
\begin{eqnarray}
B\psi_{n-1} + [H_{0}+ V_\text{imp} \delta_{n,1}-E]\psi_{n} + B^\dagger \psi_{n+1}=0, \label{scho}
\end{eqnarray}
where $\psi_{0}=0$, $\psi_n$ is a wavefunction on $n_{th}$ layer parallel to the impurity surface, $H_{||}$ is an in-layer Hamiltonian, $V_\text{imp}$ is an impurity potential in the first layer, $B$ and $B^\dagger$ are hopping terms between layers, $n > 0$ is the layer index.
For notational convenience, we set $\psi_0=0$.
For $n=N$, the last layer of the system, we can write exactly
\begin{eqnarray}
B\psi_{N-1} + [H_0-E] \psi_N = 0 \label{scho_last}
\end{eqnarray}
Using the Schr\"odinger equation for the $n=N-1$ layer, $B\psi_{N-2} + [H_0-E]\psi_{N-1} + B^\dagger \psi_N = 0$, and substituting Eq.~\eqref{scho_last}, we can``integrate" out the last ($N$'th) layer
\begin{eqnarray}
B\psi_{N-2} + \left[H_0-E + B^\dagger \frac{1}{E-H_0} B \right] \psi_{N-1}= 0. \label{scho_inte}
\end{eqnarray}
Eliminating $\psi_N$ introduces for $\psi_{N-1}$ the effective potential $\Sigma_{N-1}=B^\dagger \frac{1}{E-H_0} B $. By repeating this process, we can integrate out all layers up to the first layer and the following recursion relation can be found,
\begin{eqnarray}
\Sigma_n = B^\dagger\frac{1}{E-H_0-\Sigma_{n+1}} B \label{recu_back}
\end{eqnarray}
with a boundary condition $\Sigma_N = 0$. Recall that $B^\dagger$ is a hopping to the next layer and $B$ is a hopping to the prior layer. And the effective potential $\Sigma_n$ is obtained by sandwiching the Green's function in $(n+1)_{th}$ layer by $B^\dagger$ and $B$, describing a scattering process of hopping to the next layer, propagating, and hopping back to the original layer.
Let us next write an effective Hamiltonian in the top layer in the following way:
\begin{eqnarray}
[E- H_0 - \Sigma_1 ] \psi_1 = V_\text{imp} \psi_1 \\
\left[ B \Sigma_0 ^{-1} B^\dagger \right] \psi_1 = V_\text{imp} \psi_1, \label{scho_S},
\end{eqnarray}
where the recursion relation \eqref{recu_back} is again used to further simplify the clean part of the Hamiltonian. In the next section, we introduce a way to solve the recursion relation exactly.
\subsection{`Holographic' mapping of the self energy}
The recursion relation can be straightforwardly solved by mapping the effective potential to a matrix $M$ which obeys the same Schr\"odinger equation as the layer wavefunctions
\begin{eqnarray}
BM_{n-1} + [H_{0}-E]M_{n} + B^\dagger M_{n+1}=0, \label{scho_M}
\end{eqnarray}
where the matrix $M_n$ has the same dimension as the Hamiltonian $H_{||}$, and it is invertible by construction.
With Eq.~\eqref{scho_M}, the recursion relation for the self-energy is easily solved
\begin{eqnarray}
\Sigma_n = B^\dagger M_{n+1} M_{n}^{-1}. \label{sigm_M}
\end{eqnarray}
One can directly verify that this is a solution of the recursion relation for $M_n$ satisfying boundary condition $M_{N+1}=0$. For a clean bottom-surface, we can exactly construct $M_n$ for a system with finite thickness (the calculation is detailed in the Appendix).
The last step involves writing a close-form equation for the wavefunction of the top (disordered) layered. Note that as $M_n$ is also a solution of the Schr\"odinger equation, an element of $M_n$ scales with eigenvalues of transfer matrix of Schr\"odinger equation \eqref{scho_M}: $(M_n)_{ii'} \sim \rho_j^{N-n}$. Using the exact expression of $M_n$'s, we construct the left side of Eq.~\eqref{scho_S}. Then, we obtain the Schr\"odinger equation expressed in terms of $M_n$'s
\begin{eqnarray}
\left[ B M_0 M_1^{-1} \right] \psi_1 = V_\text{imp} \psi_1. \label{Hami_redu}
\end{eqnarray}
This is the effective single layer Hamiltonian. The left-hand side contains only elements from the clean part of the Hamiltonian, and involves the self energy from all subsequent layers; the right-hand side is simply the surface impurity potential operating on the top-layer wavefunction. $M_0$ and $M_1$ are a function of energy, and one can find all eigenvalues of a system by finding the energies that satisfy $\operatorname{det}\left[ B M_0 M_1^{-1} - V_\text{imp}\right] =0$. The surface wavefunction can be subsequently found from Eq.~\eqref{Hami_redu}, which is identical to the result from exact diagonalization. To obtain whole wavefunction in the layers beneath the top layer, we apply the transfer matrix which is also obtained in terms of $M_n$ as shown in the next section.
\subsection{Transfer matrix of wavefunctions}
The first layer wavefunction can be exactly obtained from Eq.~\eqref{Hami_redu}, therefore, the computational complexity is essentially reduced by one dimension. To obtain a full profile of the wavefunction in the subsequent layers, we construct an approach similar to the transfer matrix approach in this section. We will use the term ``transfer matrix'' quite liberally in what follows.
$\Sigma_n$ plays the role of effective potential in the $n_{th}$ layer, induced by integrating out the $(n+1)_{th}$ layer up to $N_{th}$ layer. In the Schr\"odinger equation \eqref{scho}, such a contribution is accounted by the third term on the left side. Therefore, we have the following equality:
\begin{eqnarray}
B^\dagger \psi_{n+1} = \Sigma_n \psi_n.
\end{eqnarray}
One can explicitly show this relation by the elimination method introduced in Eq.~\eqref{scho_inte}. The transfer matrix is conveniently expressed in terms of the $M_n$'s using Eq.~\eqref{sigm_M}
\begin{eqnarray}
\psi_{n+1} =\left[ M_{n+1} M_n^{-1} \right] \psi_n.
\end{eqnarray}
Or, more generally, using the relation between the wavefunctions in the $m$ and $n$ layers
\begin{eqnarray}
\psi_{n} =\left[ M_{n} M_m^{-1} \right] \psi_m,\label{tran},
\end{eqnarray}
where the exact expression of $T_{n,m}\equiv M_{n} M_m^{-1}$ is known. Note that the expression for the transfer matrices is disorder-free, which implies that disordered wavefunction in the first layer propagates into the subsequent layers just as a clean wavefunction would. This, of course, makes sense since only the top layer contains impurities. A conventional transfer matrix constructed from the top surface, however, would always contain impurity potentials, and therefore the construction of the whole wavefunction would not be as straightforward.
\subsection{`Holographic' mapping of the impurity potential}
For completeness, we address another question of interest: what is the effective impurity potential experienced by an electronic state in the bulk due to the surface impurity. This question can be answered in the same formalism introduced in earlier sections. To compute the effective impurity potential, we integrate out the first $(n-1)$ layers. The recursion relation for effective potential is
\begin{eqnarray}
V_{n+1} = B \frac{1}{E-H_0-V_{n}} B^\dagger \label{recu_forw},
\end{eqnarray}
with boundary condition $V_1 = V_\text{imp}$. The `holographic' mapping helps us to analytically derive a scaling behavior of the effective potential
\begin{eqnarray}
V_n = B \tilde{M}_{n-1} \tilde{M}_n ^{-1},
\end{eqnarray}
where the $\tilde{M}_n$'s are similarly constructed to satisfy the Schr\"odinger equation for the individual layers, \eqref{scho_M}, and to be invertible. However, their boundary condition is different from the previous clean case. $\tilde{M}_n$ has to be constructed such that the following condition is satisfied:
\begin{eqnarray}
V_\text{imp} = B \tilde{M}_{0} \tilde{M}_1 ^{-1} .
\end{eqnarray}
Since $V_\text{imp}$ is a random matrix, it is nontrivial to determine $\tilde{M}_0$ in general. But, we know the object $\tilde{M}_n$ propagates just like a clean wavefunction. Therefore, it is possible to deduce the scaling of $(V_n)_{ij}$ with respect to layer index n; which includes the contribution from surface impurities as well as clean layers from the top to $(n-1)^{th}$ layer.
\section{Application to a 2D topological insulator}
\label{sec:appl_to}
In the previous section, we introduced a general transfer-matrix framework for computing the full wavefunctions of layered systems with surface impurities. In this section, a 2D topological insulator model\cite{bern2006} is employed to explicitly show how the local density of states can be computed in a system with edge impurities.
Our main results are presented in Fig.~\ref{fig:vmap}, \ref{fig:width} and \ref{fig:width_mag}, where we use the formalism developed earlier to compute the LDOS of the first and second layers of the TI varying the disorder strength $W$.
\subsection{Model Hamiltonian}
\label{sec:mode_hami}
Consider the toy model of a 2D topological insulator. In momentum space,
\begin{eqnarray} H(\vec{k}) = [m- 2b(2-\cos k_x-\cos k_y)]\tau_z \nonumber \\ + A[ \tau_x s_z \sin k_x+ \tau_y \sin k_y] \end{eqnarray}
where $\tau_i$ is a Pauli matrix in orbital basis, $s_i$ a Pauli matrix in spin basis.
The lattice spacing is set to $a=1$ such that the momenta $k_x$ and $k_y$ lie within the interval $[-\pi,\pi]$.
To introduce an edge state, open boundary conditions are introduced in the $y$-direction, and periodic boundary conditions are applied to the $x$-direction.
The intra-layer Hamiltonian and the hopping term between layers described in Eq.~\eqref{scho} are:
\begin{align}\begin{split}
H_{0} &= [m- 2b(2-\cos k_x)]\tau_z +\tau_x s_z A\sin k_x, \\ B &= b \tau_z- i \frac{A}{2} \tau_y.
\end{split}\end{align}
The system is in the topological phase if the bands are inverted for some range of momentum: $\operatorname{sign}(mB)>0$. For this case, the dispersion of the top and bottom edge states are given by $E=\pm A\sin k_x$ \cite{imur2010}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=75mm]{picture/2d_lattice_4.pdf}
\caption{A 2D lattice model for a TI with edge impurities. Along $x$-direction we have a periodic boundary condition that $x_1=x_{N+1}$, and along y-direction there are N-layers and the subscript $n$ indicates the $n_{th}$ layer along the $y$-direction. Different colors and sizes of dots on the top layer represent the random on-site impurity potentials. We study how the effect of the top edge impurities propagate down into the bulk layers. }
\label{fig:2d_lattice}
\end{center}
\end{figure}
\subsection{A single layer effective Hamiltonian}\label{subsection:half_syst}
The system is equivalent to a set of parallel 1D wires coupled by a hopping matrix $B$ in spin and orbital basis.
We want to construct the matrix $M_n$ which is an essential building block of a single layer Hamiltonian [Eq.~\eqref{Hami_redu}] and the transfer matrix [Eq.~\eqref{tran}].
To demonstrate the method, we construct $M_n(k_x=0, E=0, s_z=1)$ here, and other $(k_x,E)$ cases will be shown in the appendix. Note that because the clean Hamiltonian is diagonal in momentum and spin space, we only need to analyze the orbital space. The Schr\"odinger equation we need to solve is
\begin{eqnarray}
\left[ \tau_z \left( b\left( \frac{1}{\rho} +\rho\right) + m-2b \right) -i\tau_y
\frac{A}{2}\left( \frac{1}{\rho}-\rho\right) \right] \psi_n =0, \nonumber
\end{eqnarray}
where we use $\psi_{n+m} = \rho^m\psi_n$. This is one section of Schr\"odinger equation at $k_x=0$, $E=0$, and $s_z=1$. There are four transfer eigenvalues and corresponding eigenvectors. By taking the determinant of the terms in the square bracket, we get $\rho_1=\lambda_1$, $\rho_2=1/\lambda_1$, $\rho_3=\lambda_2$, and $\rho_4=1/\lambda_2$, where
\begin{eqnarray}
\lambda_1 &=& \frac{-(m-2b)- \sqrt{m^2-4mb+A^2}}{A+2b} \nonumber \\
\lambda_2 &=& \frac{-(m-2b) + \sqrt{m^2-4mb+A^2}}{A+2b}
\nonumber
\end{eqnarray}
where $\lambda_{1,2}$ is chosen to be $|\lambda_{1,2}|<1$ for $0<m<4b$ and corresponding eigenvectors are $\phi_1=\phi_3=\ket{+}$ and eigenvectors of $1/\lambda_{1,2}$ are $\phi_2=\phi_4=\ket{-}$, where
\begin{eqnarray}
\ket{+} = \frac{1}{\sqrt{2}}\begin{pmatrix} 1 \\ 1 \end{pmatrix}, \ket{-} = \frac{1}{\sqrt{2}}\begin{pmatrix} 1 \\ -1 \end{pmatrix} .
\end{eqnarray}
Considering only one spin section, this implies that a state $\ket{+}$ is localized at the top ($n=1$) and a state $\ket{-}$ is localized at the bottom ($n=N$).
The interlayer hopping operator can be expressed in terms of the eigenvectors: $B= (b-A/2)\ket{-}\bra{+} + (b+A/2)\ket{+}\bra{-}$. For this given set of eigenvalues and vectors, we can construct an invertible $M_n$ in the following manner:
\begin{eqnarray}
M_{l+N+1}= (\rho_1^l - \rho_3^l) \ket{+}\bra{+} -(\rho_2^l - \rho_4^l)\ket{-}\bra{-} \nonumber ,
\end{eqnarray}
where $l=n-N-1$. $M_{n}$ is invertible for $n\neq N+1$ and satisfies the homogeneous boundary condition $M_{N+1} = 0$. If $m^2-4mb+A^2<0$, eigenvalues are $\rho_1=\lambda e^{i\theta}$ and $\rho_3=\lambda e^{-i\theta}$ with $\lambda= |\frac{2b-A}{2b+A}|$. Thus, the effective Hamiltonian in the first layer is as follows\eqref{Hami_redu}:
\begin{align}
&E-H^\text{eff}_1 = B M_0 M_1^{-1} \nonumber \\
&= B \left[ \frac{\rho_1^{-N-1}-\rho_3^{-N-1}}{\rho_1^{-N}-\rho_3^{-N}} \ket{+}\bra{+} + \frac{\rho_2^{-N-1}-\rho_4^{-N-1}}{\rho_2^{-N}-\rho_4^{-N}} \ket{-}\bra{-} \right] \nonumber \\
&= (b-A/2) \frac{1}{\lambda} \frac{\sin [(N+1)\theta]}{\sin (N\theta)} \ket{- }\bra{+} \notag \\
&\quad + (b+A/2) \lambda \frac{\sin [(N+1)\theta]}{\sin (N\theta)} \ket{+}\bra{-} ,
\label{Heff}
\end{align}
where $N$ is the number of parallel wires. The effective Hamiltonian contains off-diagonal elements only in $\ket{\pm}$ basis.
Note that the effective single-layer Hamiltonian at $(k_x,E)=(0,0)$ depends on the width of the system, and if a wavefunction contains a component at $(k_x,E)=(0,0)$, it will also be system-size dependent. Because we cannot think of a localized wavefunction dependent of the system size for large enough $N$, we can say no eigenstate localized to an edge sits at $(k_x,E)=(0,0)$. More relevant Hamiltonian sections at zero energy will be $k_x \neq 0$, which is system size independent in the large $N$ limit.
More generally, the Hamiltonian sectors not at $(k_x,E)=(2\pi l/N, A \sin k_x)$ is expressed in the following way:
\begin{align}\begin{split}
\bra{+}H_{1}^\text{eff} \ket{+} &= \frac{b+A/2}{1/r_3-1/r_1} \left[ \rho^{-1}_1-\rho^{-1}_3 \right] , \\
\bra{+}H_{1}^\text{eff} \ket{-} &= \frac{b+A/2}{1/r_3-1/r_1} \left[ \rho^{-1}_1 /r_1 - \rho^{-1}_3/ r_3 \right],\\
\bra{-}H_{1}^\text{eff} \ket{+} &= \frac{b-A/2}{r_1-r_3} \left[ \rho_1^{-1} r_1 - \rho_3^{-1} r_3 \right] , \\
\bra{-}H_{1}^\text{eff} \ket{-} &= \frac{b-A/2}{r_1-r_3} \left[ \rho^{-1}_1- \rho^{-1}_3 \right].
\end{split}\end{align}
Here $\rho_{1,3}$'s are the eigenvalues of the transfer matrix with magnitude smaller than unity, and $r_i = \dirac{+}{\phi_i}/ \dirac{-}{\phi_i}$'s are the ratios of the overlaps between the transfer-matrix eigenstates corresponding to $\rho_i$ with the $\ket{+}$ and $\ket{-}$ states. We can see that as the $(k_x, E)$ approaches to the on-shell condition, $r_i$ approaches zero and the wavefunctions have infinitesimal overlap with $\ket{-}$ since $\bra{-}H_{1}^\text{eff} \ket{-} $ component is huge. In other words, the Hamiltonian expressed in this way can be interpreted as a projection to the on-shell eigenstates.
\begin{figure}[b]
\begin{center}
\includegraphics[width=90mm]{picture/ldosratio_hist-eps-converted-to.pdf}
\caption{For disorder strength $W=20$, histograms of the LDOS ratio for different system sizes $N_x=10,20,40,80$ are shown. When system size $N_x > 80$ the distribution converges and the ratio of the LDOS in the first and the second wire becomes system size independent, and the average of the LDOS ratio in the thermodynamic limit can be estimated. On the other hand, system sizes $N_x<40$ are in quantum-dot regime and not proper to compute the bulk electronic properties because of their system-size dependence.}
\label{fig:ldos_Nx}
\end{center}
\end{figure}
With a set of impurities on the top wire (Fig.~\ref{fig:2d_lattice}), to obtain eigenenergies we find energies where the determinant of effective single layer Hamiltonian is zero (see Eq.~\eqref{Hami_redu}).
In the strong impurity regime we must use a large enough system size to correctly see the size independent behavior of bulk electronic properties. Here, we distinguish the quantum-dot regime from the bulk regime by the dependence of physical observables on the system size. Figure \ref{fig:ldos_Nx} shows the histogram of the ratio of the local density of state in the first and the second wire for impurity strength $W=20$ with increasing system sizes $N_x=10,20,40,80$. The series of histograms shows size-dependence for $N_x \lesssim 40$, the histogram becomes Gaussian shape and size-independent for $N_x \gtrsim 40$. Therefore, if one wants to numerically obtain physical observables in the thermodynamic limit, it is important to use system size larger than $N_x=80$ for $W=20$ non-magnetic edge impurities.
The large-size requirement is less stringent for smaller impurity strengths, as evident from Fig. \ref{fig:ldosratio}.
\begin{figure}[t]
\includegraphics[width=90mm]{picture/cond_Nx3-eps-converted-to.pdf}
\caption{The LDOS ratio of the first and second layer is plotted over the impurity strength $W$ for system size $N_x=20, 50, 120 , 200$. The edge state at zero energy is primarily populated in the first layer from weak to moderate impurity strength, $W<5$, and then the edge state moves to the second and following layers at strong impurity strength. The $N_x=200$ curve shows the behavior in thermodynamic limit as it becomes size-independent, while $N_x=20$ curve shows quantum-dot behavior for $W>5$, meaning the system size is not big enough to see bulk properties.}
\label{fig:ldosratio}
\end{figure}
Figure \ref{fig:ldosratio} shows the ratio of the LDOS at zero energy with increasing impurity strength. This quantity tells us where the edge-state wavefunctions actually reside. A ratio below 1 indicates edge states rooted in the first layer. But ratios greater than 1 indicate edge states expelled to the the second layer, which are therefore increasingly less immune to the disorder.
The edge states in weak and moderate disorder edge state are dominated by the first layer (also see Figure \ref{fig:vmap}). This comes hand in hand with a spread of the edge function Fourier transform: it has broad support away from $k_x=0$ due to impurity scattering. Once the impurity strength is comparable to or larger than the band width, the edge state is populated less in the impurity layer, and it moves to the second and following layers. All curves in Figure \ref{fig:ldosratio} show this behavior with a dip at $W=5$. While we believe that $N_x=200$ curve properly describes the system-size independent LDOS ratio in thermodynamic limit, $N_x=20$ curve is only good for $W<5$ and it begins to deviate from $N_x=200$ curve for strong impurity strength $W>5$.
\subsection{Transfer matrix between single-layer wavefunctions}
Once the wavefunction of the first layer, $\psi_1$, is obtained, we next propagate it to the subsequent layers to obtain a full profile of the state. This can be done by using the matrices $M_n$ as in Eq.~\eqref{tran}. Let us write down the expression of the transfer matrix for the $(k_x,E)=(0,0)$ case first from layer m to layer n
\begin{eqnarray}
T_{n\leftarrow m}&=& M_n M_m^{-1} \nonumber \\
&=& \left[ \lambda^{n-m} \ket{+}\bra{+}\,+\, \lambda^{m-n}\ket{- }\bra{-} \right] \frac{\sin (N+1-n)\theta}{\sin(N+1-m)\theta},\nonumber
\end{eqnarray}
where $\rho_{1,3} = \lambda e^{\pm i\theta}$ is used as before with $\lambda <1$, $N$ is the number of layers. We can see that the $\ket{+}$ component exponentially decays from the top surface towards the other end ($n>m$), while $\ket{-}$, if it is present, exponentially increases. Thus, it is apparent that $\ket{+}$ is a state localized to the top edge, while $\ket{-}$ is localized to the bottom edge. However, note that the transfer matrix contains an oscillating term dependent of the system size $N$, just as in the effective single-layer Hamiltonian, Eq.~\eqref{Heff}. It implies that if there is a $\ket{+}$ component in the wavefunction at exactly $(k_x, E)=(0,0)$, its oscillating part is dependent on the number of layers, which doesn't make sense in the physical picture where $N$ is much larger than the localization length of edge state. Therefore, we can say $\ket{+}$ component at $(k_x, E)=(0,0)$ must be vanishingly small as the system size is increased.
\begin{figure}[t]
\begin{center}
\includegraphics[width=90mm]{picture/vmaps_4.pdf}
\caption{The disorder-averaged LDOS at $E=0$ in wire index $n$ and momentum space $k_x$ of edge state at zero energy is plotted for increasing impurity strength. Starting from the clean edge state where the LDOS is only at $k_x=0$, the LDOS is spread out in momentum space then shifted to the second layer. The insets are showing the $k_x$-integrated LDOS as a function of layer $n$. Large enough system size $N_x=180$ is chosen such that the averaged LDOS is size-independent. 500 disorder realizations are averaged.}
\label{fig:vmap}
\end{center}
\end{figure}
The transfer matrix from the first layer to the $n$\textsuperscript{th} layer for a general $(k_x, E)$ is expressed in the following way:
\begin{align}\begin{split}
\bra{+}T_{n\leftarrow 1} \ket{+} &= \frac{1}{r_1-r_3} \left[ \rho_1^{n-1} r_1 - \rho_3^{n-1} r_3 \right] , \\
\bra{+}T_{n\leftarrow1} \ket{-} &= \frac{1}{r_1-r_3} \left[ \rho^{n-1}_1- \rho^{n-1}_3 \right] , \\
\bra{-}T_{n\leftarrow1} \ket{+} &= \frac{1}{1/r_3-1/r_1} \left[ \rho^{n-1}_1-\rho^{n-1}_3 \right] , \\
\bra{-}T_{n\leftarrow1} \ket{-} &= \frac{1}{1/r_3-1/r_1} \left[ \rho^{n-1}_1 /r_1 - \rho^{n-1}_3/ r_3 \right].
\end{split}\end{align}
with $r_i=\dirac{+}{\phi_i}/ \dirac{-}{\phi_i}$.
We apply this transfer matrix to the first layer wavefunction to obtain wavefunctions in the bulk layers. In Fig.~\ref{fig:vmap}, the disorder-averaged probability density in momentum and layer basis $P(k_x,n)=|\psi_n(k_x)|^2$ is plotted for shown impurity strength $W$. In the weak disorder regime, $W=0.1$, where impurity strength is much smaller than the energy gap, the probability density is concentrated near $k_x=0$ at zero energy. As the impurity strength increases the probability density gains width in momentum space and its weight is shifted to the second layer. While this trend is quite strong already with $W=10$, in strong impurity regime $W=40$---which is much larger than the bandwidth---the zero energy wavefunction is completely absent in the first layer, but occupies the subsequent layers in a narrow range of momentum space.
This indicates that the wavefunction has been pushed to the next layer and behaves as if the system is clean.
This behavior of the local density of states is shown for non-magnetic edge impurities, which cannot affect the transport properties of helical edge states in 2D topological insulators. Therefore, the modification of LDOS should be discussed separately from the change of transport nature, at least in 2D. For a strip geometry, the transport is studied for both non-magnetic and magnetic edge impurities in the following sections.
\begin{figure}[t]
\begin{center}
\includegraphics[width=85mm]{picture/ldoswidth-eps-converted-to.pdf}
\caption
For system size $N_x=120$, the width of LDOS in the first and the second wire in momentum space $k_x$ is plotted for as a function of disorder strength $W$ for non-magnetic impurities. In clean system limit, the edge state LDOS $P(k_x)$ is concentrated at $k_x=0$ and its width is infinitesimally small. As the disorder strength increases, the LDOS spreads in momentum space. However, for the stronger disorder, the first layer LDOS takes on all possible momenta and its width saturates, while the second layer LDOS become concentrated around $k_x=0$ again.}
\label{fig:width}
\end{center}
\end{figure}
The widths of probability density $P(k_x,n)=|\psi_n(k_x)|^2$ in the first and second layer at zero energy are plotted in Fig.~\ref{fig:width}. The width in momentum space is indicative of how disordered the edge state is due to impurities.
In the weak and strong disorder limit the wavefunction behaves like a clean system in the LDOS shape as shown in Fig.~\ref{fig:vmap}.
With strong disorder, the width of $P(k_x,n=1)$ saturates near $0.3$, although it carries little weight in that limit: $P(k_x,n=1) \ll 1$.
Meanwhile, the width of $P(k_x,n=2)$ increases and decreases again as the impurity strength is varied, peaking at around the bandwidth of the system ($W\sim8$).
\subsection{Magnetic edge impurities}
The same calculation was repeated for a system with magnetic edge impurities. We simply needed to extend the Hamiltonian to have two spin-sections and introduce random magnetic impurities, $V(x_i) = \vec{V_i} \cdot \vec{s}$, where three component random variable $\vec{V_i} = (V_{i}^{x},V_{i}^{y},V_{i}^{z})$. We found that to simulate the bulk regime for $W=20$ the system size needs to be at least $N_x=400$ as opposed to $N_x=120$ for the non-magnetic edge impurity case. In other words, the lower bound of the system size to see the thermodynamic properties is much larger and it becomes computationally challenging even for 2D system.
Figure~\ref{fig:width_mag} shows the width of the LDOS distribution $P(k_x,n)$ for $n=1,2$.
The data is qualitatively similar to the non-magnetic case, which demonstrates the universality of the result.
\begin{figure}[t]
\begin{center}
\includegraphics[width=85mm]{picture/ldoswidth_mag-eps-converted-to.pdf}
\caption
LDOS momentum-space width in the first and the second layers with {\it magnetic} impurities, at system size $N_x=120$. The overall behavior is the same with the case of non-magnetic edge impurity case (Fig.~\ref{fig:width}), but much larger systems are required to see the size-independent physical properties in bulk regime. }
\label{fig:width_mag}
\end{center}
\end{figure}
\section{Transport behavior}\label{sec:tran_beha}
The local density of states discussed in the last section can be probed by angle-resolved photoemission spectroscopy.\cite{wray2011}
The transport along the edge states, however, is provides additional information independent of the local density of states. For instance, in the case of non-magnetic impurities, edge modes can not backscatter, and their conductance remains quantized at the value of the clean system, despite the local density of states associated with it changing its support between the layers. To clarify the transport nature of the system with magnetic and non-magnetic impurities along the edge, in this section we study conductance of the systems using Landauer-B{\"u}ttiker method.
Imagine a system where two semi-infinite leads are connected to a disordered region at the ends $x=1$ and $x=N_x$. Landauer and B{\"u"ttiker \cite{land1957,land1970,butt1988} related the conductance with the transmission coefficient through the disordered region: $g = \frac{e^2}{h} T_{N1} $, where $g$ is conductance, $T_{N1}$ is a transmission coefficient through the disordered region from site 1 to site $N$. Using linear response formalism, Fisher and Lee \cite{fish1981} expressed the transmission coefficient in terms of Green's functions:
\begin{eqnarray} g = \frac{e^2}{h}\operatorname{Tr}\left[ \Gamma_L G_{N1} \Gamma_R G_{N1}^\dagger \right] \end{eqnarray}
where $\Gamma_L= i (\Sigma_L - \Sigma_L^\dagger)$, $G_{N1}$ is a Green's function from site 1 to $N$ renormalized by the presence of the leads, and $\Sigma_L$ is a self-energy of the semi-infinite left lead. Each term in this formula can be computed recursively, such that the conductance of a long system can be obtained with a reasonable computation effort. A good review of the detailed calculation can be found in Ref.~\onlinecite{meta2005}.
\subsection{Non-magnetic impurities case}\label{sec:non_impu}
Consider the 2D topological insulator system introduced earlier with non-magnetic impurities along the top edge. Because the Hamiltonian is diagonal in spin basis without magnetic impurities, we can consider transport in just one spin sector. When the chemical potential is in the energy gap, neither backscattering nor scattering into the bulk is possible. Therefore, the conductance must remain quantized even in the presence of edge impurities. This is indeed what our calculation shows. Indeed, when the disorder is non magnetic, the transport behavior does not reflect the development of the LDOS.
\subsection{Magnetic impurities case}\label{sec:magn_impu}
The two opposite-spin, counter-propagating, chiral edge modes couple as soon as magnetic edge impurities are introduced. As a result, transport through the disordered edge is suppressed, while the transport through the clean edge remains unaffected. Therefore, we expect the total conductance to rapidly approaches $e^2/h$ when introducing and increasing magnetic edge disorder.
However, in the strong-impurity quantum-dot limit, $W \gg N_x$, we found that the conductance recovers its clean system value of $2e^2/h$. In this regime the impurities are strongly bound to electrons at energies far away from the Fermi energy, and they play negligible roles in the transport at the Fermi energy. Put another way, strong disorder pushes the edge modes to the next layer where they effectively become weak scatters. This behavior is clear in Fig.~\ref{fig:spinful3_W}, which shows the conductance vs.\ the disorder strength for different system sizes. We can see that in the intermediate range of impurity strength, the conduction through the disordered top edge is significantly suppressed due to magnetic impurities, while the conductance recovers up to $2e^2/h$ value in the strong impurity limit.
We note that in the thermodynamic limit $N_x \rightarrow \infty$, the conductance is always $e^2/h$ for any disorder strength, illustrated in Fig.~\ref{fig:spinful3_dav_W}. Our calculation reflects the non-monotonic dependence of the localization length of the scattered edge mode on disorder strength, which is consistent with the expulsion of the LDOS from the disordered first layer.
\begin{figure}[h]
\includegraphics[width=85mm]{picture/cond_plot4-eps-converted-to.pdf}
\caption{Edge channel conductance vs. impurity strength when magnetic edge impurities are present. Three system sizes $N_x=20, 50, 100, 200$ are used. The edge channel conductance drops to $e^2/h$ initially. The conductance recovers to $2e^2/h$ as the system enters quantum-dot regime at strong magnetic impurity strength. This reflects the nonmonotonic behavior of the localization length of the edge mode as disorder increases. }
\label{fig:spinful3_W}
\end{figure}
\begin{figure}[h]
\includegraphics[width=85mm]{picture/cond_Nx2-eps-converted-to.pdf}
\caption{Disorder averaged conductance with increasing system size $N_x$ for three impurity strength $W$. The conductance exponentially decreases with system size as anticipated in the magnetic edge impurity case. }
\label{fig:spinful3_dav_W}
\end{figure}
\section{Summary and Conclusion}\label{sec:conc_and}
The effects of surface disorder on systems with surface states are at the focus of our work. The interest in this problem rose after an exact-diagonalization analysis showed that the response of a 3D topological insulator to surface disorder is non-monotonic: First the surface states become diffusive, and their conductance is suppressed. At a finite disorder strength, however, the surface states mean free path recovers, and they reconstitute at the disorder-free second layer of the TI \cite{schu2012}.
In this manuscript we developed a formalism that reduces solving a bulk $D$-dimensional Hamiltonian to a surface-only diagonalization problem. This, in principle, enables an exact-diagonalization analysis of much bigger bulk systems than previously possible. Roughly speaking, our method constitutes a systematic integrating out and elimination of the bulk degrees of freedom layer by layer. We show how to carry this procedure out with a technique reminiscent of transfer-matrix methods. From the resulting surface-only diagonalization problem we are able to reconstruct the wavefunctions of the bulk layers and study the effect of disorder on them. Our method is generally valid for any system composed of coupled layers but becomes particularly useful for topological systems whose surface states are fundamentally dependent on the existence of the gapped bulk.
We used our method to study 2D topological insulators with edge impurities. As the strength of edge impurities increases, we found that edge states near zero energy become more concentrated up to the moderate impurity strength (i.e., comparable to the bandwidth), and then the edge states are gradually pushed to the second layer. The ratio of the local density of states in the first two layers in Figure \ref{fig:ldosratio} shows this behavior for different system sizes. Furthermore, the width of the edge state in momentum space in the second layer reaches a maximum at moderate impurity strength and then narrows at stronger impurity strength. Despite the strong disorder, the edge states have momentum restored to being a good quantum number. This non-monotonic response to the surface disorder is equally true for magnetic and non-magnetic disorder (see Figures \ref{fig:width} and \ref{fig:width_mag}).
The transport properties, however, show a sharp contrast between the magnetic and non-magnetic cases: non-magnetic edge impurities does not affect the edge transport properties, while magnetic edge impurities immediately induce a finite localization length due to backscattering, and initially suppress the edge conductance (see Figure \ref{fig:spinful3_W}). In our numerical simulations that include magnetic impurities, however, we found that the conductance then recovers to the clean-system values when the impurity strength exceeds the system size.
This finding does not imply that the edge states completely decouple at large but finite disorder. Rather, it is an indication that the localization length of the spin-orbit locked modes exhibits a non-monotonic localization behavior as a function of disorder. In our simulations, perfect conductance will be recovered when the localization length exceeds the system width. This is precisely the regime we nicknamed the quantum-dot regime. The localization length inferred from transport calculations must be proportional to the inverse of the average width of the edge LDOS in momentum space, for momenta parallel to the edge. Therefore, our transport results are yet another manifestation of the reconstitution of the edge mode at the second layer, which is disorder free.
In the limit of infinitely strong disorder, we expect the localization length of the reconstituted edge state to diverge and translational invariance is regained.
In future work, we intend to apply our method to the 3D topological insulator with surface disorder. As we emphasize throughout, one must use large enough systems in order to explore the bulk properties in thermodynamic limit (see Figure \ref{fig:ldos_Nx} and \ref{fig:ldosratio}). This makes 3D topological systems with moderate to strong surface impurities harder to study using the ED as the cost of calculation increases. Our analytic approach will then be very useful, since it reduces the computational cost dramatically.
\section*{Acknowledgments}
It is a pleasure to acknowledge support from the Sherman-Fairchild Foundation (RM), the Packard Foundation, the Walter Burke Institue of Theoretical Physics, as well as the Institute of Quantum Information and Matter, an NSF frontier center, with the support of the Gordon and Betty Moore Foundation (KWK, GR). In addition, support from NSERC and CIFAR is gratefully acknowledged (MF).
\clearpage
|
\section{Introduction} \label{sec:intro}
\input{introduction.tex}
\section{Linear Shallow Monadic Horn Approximation} \label{sec:abstr}
We consider a standard first-order language without equality where $\Sigma$
denotes the set of function symbols.
The symbols $x,y$ denote variables, $a,b$ constants,
$f,g,h$ are functions and $s,t$ terms.
Predicates are denoted by $S,P,Q,R$, literals by $E$, clauses by $C,D$, and sets of clauses by $N,M$.
The term $t[s]_p$ denotes that the term $t$ has the subterm $s$ at position $p$.
The notion is extended to atoms, clauses, and multiple positions.
A predicate with at most one argument is called monadic.
A literal is either an atom or an atom preceded by $\neg$ and it is then respectively called positive or negative.
A term is shallow if it has at most depth one.
It is called linear if there are no duplicate variable occurrences.
A literal, where every term is shallow, is also called shallow.
A clause is a multiset of literals which we write as an implication $\Gamma \imp \Delta$ where the atoms in $\Delta$ denote the positive literals
and the atoms in $\Gamma$ the negative literals.
If $\Gamma$ is empty we omit $\imp$, e.g., we write $P(x)$ instead of $\imp P(x)$ whereas if
$\Delta$ is empty $\imp$ is always shown.
If a clause has at most one positive literal, it is a Horn clause.
If there are no variables, then terms, atoms and clauses are respectively called ground.
A substitution $\sigma$ is a mapping from variables into terms denoted by pairs $\{x \mapsto t\}$.
If for some term (literal, clause) $t$, $t\sigma$ is ground, then $\sigma$ is a grounding substitution.
A Herbrand interpretation $I$ is a - possibly infinite - set of positive ground literals and $I$ is said to satisfy a
clause $C= \Gamma \imp \Delta$, denoted by $I\sat C$, if $\Delta\sigma \cap I \neq \emptyset$ or $\Gamma\sigma \not\subseteq I$ for every grounding substitution $\sigma$.
An interpretation $I$ is called a model of $N$ if $I$ satisfies $N$, $I\sat N$, i.e.,
$I\sat C$ for every $C\in N$. Models are considered \emph{minimal} with respect to set inclusion.
A set of clauses $N$ is satisfiable, if there exists a model that satisfies $N$. Otherwise the set is unsatisfiable.
\begin{defin}[Conflicting Core]\label{conlfictCore}
A finite clause set $N^\bot$ is a conflicting core if for all grounding substitutions $\tau$ the clause set $N^\bot\tau$ is unsatisfiable.
$N^\bot$ is a conflicting core of $N$ if $N^\bot$ is a conflicting core
and for every clause $C\in N^\bot$ there exists a $C'\in N$ such that $C=C'\sigma$.
\end{defin}
\begin{defin}[Specific Instances]\label{specInst}
Let $C$ be a clause and $\sigma_1$, $\sigma_2$ be two substitutions
such that $C\sigma_1$ and $C\sigma_2$ have no common instances.
Then the \emph{specific instances} of $C$ with
respect to $\sigma_1$, $\sigma_2$ are clauses $C\tau_1,\ldots,C\tau_n$ such that
(i)~any ground instance of $C$ is an instance of some $C\tau_i$,
(ii)~there is no $C\tau_i$ such that both $C\sigma_1$ and $C\sigma_2$ are
instances of $C\tau_i$.
\end{defin}
The definition of specific instances can be extended to a single substitution $\sigma$.
In this case we require $C$ and $\sigma$ to be linear,
condition~(i) from Definition~\ref{specInst} above, $C\sigma = C\tau_1$ and no $C\tau_i$,
$i\neq 1$ has a common instance with $C\tau_1$.
Note that under the above restrictions specific instances always exist~\cite{Lassez:1987:ERT:33031.33036}.
\begin{defin}[Approximation]
Given a clause set $N$ and a relation $\Rightarrow$ on clause sets with $N \Rightarrow N'$ then
(1)~$\Rightarrow$ is called an \emph{over-approximation} if satisfiability of $N'$ implies satisfiability of $N$,
(2)~$\Rightarrow$ is called an \emph{under-approximation} if unsatisfiability of $N'$ implies unsatisfiability of $N$.
\end{defin}
Next we introduce our concrete over-approximation $\apr$ that
eventually maps a clause set $N$ to an mslH
clause set $N'$. Starting from a clause set $N$ the transformation is parameterized
by a single monadic projection predicate $T$, fresh to $N$
and for each non-monadic predicate $P$ a projection function $f_P$ fresh to $N$. The approximation
always applies to a single clause and we establish on the fly an ancestor relation between
the approximated clause(s) and the parent clause. The ancestor relation is needed for lifting and refinement.
\bigskip
\shortrules{Monadic}{$N\cup\{\Gamma \imp \Delta,P(t_1,\dots,t_n)\}$}{$N\cup\{\Gamma \imp \Delta,T(f_P(t_1,\dots,t_n))\}$}{provided $n>1$; $P(t_1,\dots,t_n)$ is the ancestor of $T(f_P(t_1,\dots,t_n))$}{MO}{10}
\bigskip
\shortrules{Horn}{$N\cup\{\Gamma \imp E_1, \dots,E_n\}$}{$N\cup\{\Gamma \imp E_i\}$}{provided $n>1$; $\Gamma \imp E_1, \dots,E_n$ is the ancestor of $\Gamma \imp E_i$}{HO}{10}
\bigskip
\shortrules{Shallow}{$N\cup\{\Gamma \imp E[s]_{p}\}$}{$N\cup\{S(x),\Gamma_1 \imp E[x]_{p}\}\cup\{ \Gamma_2 \imp S(s)\}$}{provided $s$ is a complex term, $p$ not a top position, $x$ and $S$ fresh,
and $\Gamma_1 \cup \Gamma_2 = \Gamma$; $\Gamma \imp E[s]_{p}$ is the ancestor of $S(x),\Gamma_1 \imp E[x]_{p}$ and $\Gamma_2 \imp S(s)$}{SH}{10}
\bigskip
\shortrules{Linear}{$N\cup\{\Gamma \imp E[x]_{p,q}\}$}{$N\cup\{\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q\}$}{provided $x'$ is fresh, the positions $p$, $q$ denote
two different occurrences of $x$ in $E$; $\Gamma \imp E[x]_{p,q}$ is the ancestor of $\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q$}{LI}{10}
\bigskip
For the Horn transformation, the choice of the $E_i$ is arbitrary.
In the Shallow rule, $\Gamma_1$ and $\Gamma_2$ can be arbitrarily chosen as long as they ``add up'' to $\Gamma$.
The goal, however, is to minimize the set of common variables
$\vars(\Gamma_2,s) \cap \vars(\Gamma_1,E[x]_p)$.
If this set is empty the Shallow transformation is satisfiability preserving.
In rule Linear, the duplication of $\Gamma$ is not needed if $x\not\in\vars(\Gamma)$.
\begin{defin}[$\apr$] \label{def:approx:trans}
The overall approximation $\apr$ is given by $\apr\; = \;\Rightarrow_{\text{MO}}\cup\Rightarrow_{\text{HO}}\cup\Rightarrow_{\text{SH}}\cup\Rightarrow_{\text{LI}}$
with a preference on the different rules where Monadic precede Horn precede Shallow precede Linear transformations.\\
\end{defin}
\begin{defin}
Given a non-monadic n-ary predicate $P$, projection predicate $T$, and projection function $f_P$,
define the injective function $\Proj{P}(P(t_1,\dots,t_n)) := T(f_p(t_1,\dots, t_n))$
and $\Proj{P}(Q(s_1,\dots,s_m)) := Q(s_1,\dots,s_m)$ for any atom with a predicate symbol different from $P$.
The function is extended to clauses, clause sets and interpretations.
\end{defin}
\begin{lem}[$\apr$ is sound and terminating]\label{Apr_sound}
The approximation rules are sound and terminating:
(i)~$\apr$ terminates
(ii)~the Monadic transformation is an over-approximation
(iii)~the Horn transformation is an over-approximation
(iv)~the Shallow transformation is an over-approximation
(v)~the Linear transformation is an over-approximation
\end{lem}
\begin{proof}
(i)~The transformations can be considered sequentially, because of the imposed rule preference (Definition~\ref{def:approx:trans}).
The monadic transformation strictly reduces the number of non-monadic atoms.
The Horn transformation strictly reduces the number of non-Horn clauses.
The Shallow transformation strictly reduces the multiset of term depths of the newly introduced clauses compared
to the removed ancestor clause.
The linear transformation strictly reduces the number of duplicate variables occurrences in positive literals.
Hence $\apr$ terminates.
(ii) Consider a transformation $N_k\Rightarrow^*_{\text{MO}} N_{k+j}$ that exactly removes all occurrences of atoms $P(t_1,\dots,t_n)$ and
replaces those by atoms $T(f_P(t_1,\dots,t_n))$.
Then, $N_{k+j}=\Proj{P}(N_k) $ and $N_k=\R{P}(N_{k+j}) $.
Let $I$ be a model of $N_{k+j}$ and $C \in N_{k}$.
Since $ \Proj{P}(C) \in N_{k+j}$ , $I \sat \Proj{P}(C)$ and thus, $\R{P}(I)\sat C$.
Hence, $\R{P}(I)$ is a model of $N_k$.
Therefore, the Monadic transformation is an over-approximation.
(iii) Let $N\cup\{\Gamma \imp E_1, \dots,E_n\} \Rightarrow_{\text{HO}} N\cup\{\Gamma \imp E_i\}$.
The clause $\Gamma \imp E_i$ subsumes the clause $\Gamma \imp E_1, \dots,E_n$. Therefore, for any $I$
if $I\models \Gamma \imp E_i$ then $I\models \Gamma \imp E_1, \dots,E_n$.
Therefore, the Horn transformation is an over-approximation.
(iv) Let $N_k = N\cup\{\Gamma \imp E[s]_{p}\} \Rightarrow_{\text{SH}} N_{k+1}= N\cup\{S(x),\Gamma_1 \imp E[x]_{p}\}\cup\{ \Gamma_2 \imp S(s)\}$.
Let $I$ be a model of $N_{k+1}$ and $C\in N_k$ be a ground clause.
If $C$ is an instance of a clause in $N$, then $I \models C$.
Otherwise $C = (\Gamma \imp E[s]_{p})\sigma$ for some ground substitution $\sigma$.
Then $S(s)\sigma,\Gamma_1\sigma \imp E[s]_{p}\sigma = (S(x),\Gamma_1 \imp E[x]_{p})\{x \mapsto s\}\sigma \in N_{k+1}$
and $ \Gamma_2\sigma \imp S(s)\sigma = (\Gamma_2 \imp S(s))\sigma \in N_{k+1}$.
Since $I \models N_{k+1} $, $I$ also satisfies the resolvent $\Gamma_1\sigma,\Gamma_2\sigma \imp E[s]\sigma = C$.
Hence $I \models N_k$.
Therefore, the Shallow transformation is an over-approximation.
(v) Let $N_k =N\cup\{\Gamma \imp E[x]_{p,q}\} \Rightarrow_{\text{LI}} N_{k+1}=N\cup\{\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q\}$.
Let $I$ be a model of $N_{k+1}$ and $C\in N_k$ be a ground clause.
If $C$ is an instance of a clause in $N$, then $I \models C$.
Otherwise $C = (\Gamma \imp E[x]_{p,q})\sigma$ for some ground substitution $\sigma$.
Then $(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\{x' \mapsto x\}\sigma \in N_{k+1}$ and
$I \models (\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\{x' \mapsto x\}\sigma = (\Gamma,\Gamma \imp E[x]_q)\sigma \models C $.
Hence $I \models N_k$.
Therefore, Linear transformation is an over-approximation.
\end{proof}
\begin{cor}
(i)~$\apr$ is an over-approximation.
(ii)~If $N \apr^* N'$, $P_1,\dots,P_n$ are the non-monadic predicates in $N$ and $N'$ is satisfied by model $I$, \\
then $\R{P_1}(...(\R{P_n}(I)))$ is a model of $N$.
\end{cor}
\begin{proof}
Follows from Lemma~\ref{Apr_sound} (ii)-(v).
\end{proof}
In addition to being an over-approximation, the minimal model (with respect to set inclusion)
of the eventual approximation
generated by $\apr$ preserves the skeleton term structure
of the original clause set, if it exists. The refinement introduced in Section~\ref{abstrref}
instantiates clauses. Thus it contributes to finding a model or a refutation.
\begin{defin}[Term Skeleton]\label{termSkel}
The term skeleton of term $t$ , $\skt(t)$, is defined as \\
(1) $\skt(x)=x'$, where $x'$ is a fresh variable \\
(2) $\skt(f(s_1,\dots,s_n))= f(\skt(s_1),\dots,\skt(s_n))$.
\end{defin}
\begin{lem}\label{termSkelLem2}
Let $N_k$ be a monadic clause set and $N_0$ be its approximation via $\apr$.
Let $N_0$ be satisfiable and $I$ be a minimal model for $N_0$.
If $P(s) \in I$ and $P$ is a predicate in $N_k$, then there exists a clause
$C = \Gamma \imp \Delta,P(t) \in N_k$ and a substitution $\sigma$ such that
$s = \skt(t)\sigma$ and for each variable $x$ and predicate $S$ with $C = S(x),\Gamma' \imp \Delta,P(t[x]_p)$, $S(s'') \in I$, where $s=s[s'']_p$.
\end{lem}
\begin{proof}
By induction on $k$.\\
For the base $N_k=N_0$, assume there is no $C \in N_0$ with $C\sigma= \Gamma \imp \Delta,P(s)$ and $\Gamma \subseteq I$.
Then $I\setminus \{P(s)\}$ is still a model of $N_0$ and therefore $I$ is not minimal.\\
Let $N = N_k\apr N_{k-1} \apr^* N_0$, $P(s) \in I$ and $P$ is a predicate in $N_k$ and hence also in $N_{k-1}$.
By the induction hypothesis, there exist a clause $C = \Gamma \imp \Delta,P(t) \in N_{k-1}$ and a substitution $\sigma$ such that
$s = \skt(t)\sigma$ and for each variable $x$ and predicate $S$ with $C = S(x),\Gamma' \imp \Delta,P(t[x]_p)$, $S(s'') \in I$, where $s=s[s'']_p$.
The first approximation rule application is either a Linear, a Shallow or a Horn transformation,
considered below by case analysis.
Horn Case. Let $\apr$ be a Horn transformation that replaces $\Gamma'' \imp \Delta',Q(t')$ with $\Gamma'' \imp Q(t')$.
If $C \neq \Gamma'' \imp Q(t')$, then $C \in N_k$ fulfills the claim.
Otherwise, $\Gamma'' \imp \Delta',Q(t) \in N_k$ fulfills the claim since $P=Q$ and $\Gamma' = \Gamma''$.
Linear Case. Let $\apr$ be a linear transformation that replaces $C_k = \Gamma'' \imp E[x]_{p,q}$ with $C_{k-1}=\Gamma'',\Gamma''\{x \mapsto x'\} \imp E[x']_{q}$.
If $C \neq C_{k-1}$, then $C \in N_k$ fulfills the claim.
Otherwise, $C_k=\Gamma'' \imp P(t)\{x' \mapsto x\} \in N_k$ fulfills the claim since $s = \skt(t)\sigma = \skt(t\{x' \mapsto x\})\sigma$ and $\Gamma'' \subseteq \Gamma'',\Gamma''\{x \mapsto x'\}$.
Shallow Case. Let $\apr$ be a shallow transformation that replaces $C_k = \Gamma'' \imp E[s']_{p}$ with $C_{k-1}=S(x),\Gamma_1 \imp E[x]_p$ and $C'_{k-1}=\Gamma_2 \imp S(s')$.
Since $S$ is fresh, $C \neq C'_{k-1}$.
If $C \neq C_{k-1}$, then $C \in N_k$ fulfills the claim.
Otherwise, $C = C_{k-1} = S(x),\Gamma_1 \imp P(t[x]_p)$ and hence, $s = \skt(t[x]_p)\sigma$ and $S(s'')\in I$ for $s=s[s'']_p$.
Then by the induction hypothesis, there exist a clause $C_S = \Gamma_S \imp \Delta_S,S(t_S) \in N_{k-1}$ and a substitution $\sigma_S$ such that
$s'' = \skt(t_S)\sigma_S$ and for each variable $x$ and predicate $S'$ with $C_S = S'(x),\Gamma'_S \imp \Delta_S,P(t_S[x]_q)$, $S'(s''') \in I$, where $s''=s''[s''']_q$.
By construction, $C_S = C'_{k-1}$.
Thus, $s''= \skt(s')\sigma_S$ and $s = \skt(t[x]_p)\sigma$ imply there exists a $\sigma''$ such that $s = \skt(t[s']_p)\sigma"$.
Furthermore since $\Gamma_1 \cup \Gamma_2 = \Gamma''$, if $C_k= S'(x),\Gamma''' \imp P(t[s']_{p})[x]_q$, then
either $S'(x) \in \Gamma_1$ and thus $S'(s'''') \in I$, where $s=s[s'''']_q$, or
$S'(x) \in \Gamma_2$ and thus $S'(s'''') \in I$, where $s[s'']_p=(s[s'']_p)[s'''']_q$.
Hence, $C_k \in N_k$ fulfills the claim.
\end{proof}
\begin{lem}\label{termSkelLem}
Let $N$ be a clause set and $N'$ be its approximation via $\apr$. Let $N'$ be
satisfiable and $I$ be a minimal model for $N'$.
If $P(s) \in I$ $(T(f_p(s_1,\ldots,s_n))\in I)$
and $P$ is a predicate in $N$, then there exist a clause
$\Gamma \imp \Delta,P(t) \in N$ ($\Gamma \imp \Delta,P(t_1,\ldots,t_n) \in N$)
and a substitution $\sigma$ such that
$s = \skt(t)\sigma$ ($s_i = \skt(t_i)\sigma$ for all $i$).
\end{lem}
\begin{proof}
Let $P_1,\ldots,P_n$ be the non-monadic predicates in $N$ and $N_{MO}=\Proj{P_1}(\ldots(\Proj{P_n}(N)))$. Then, $N_{MO}$ is monadic and also has $N'$ as its approximation via $\apr$.
Let $P(s) \in I$ and $P$ is a predicate in $N$. Since $P$ is monadic, $P$ is a predicate in $N_{MO}$.
Hence by Lemma ~\ref{termSkelLem2}, there exists a clause $\Gamma \imp \Delta,P(t) \in N_{MO}$
and a substitution $\sigma$ such that $s = \skt(t)\sigma$.
Then, $\R{P_1}(\ldots(\R{P_n}(\Gamma \imp \Delta,P(t)))\ldots)= \R{P_1}(\ldots(\R{P_n}(\Gamma)\ldots) \imp \R{P_1}(\ldots(\R{P_n}(\Delta)\ldots),P(t)\in N$ fulfills the claim.
Let $T(f_p(s_1,\ldots,s_n))\in I$ and $P$ is a predicate in $N$. $T$ is monadic and a predicate in $N_{MO}$.
Hence by Lemma ~\ref{termSkelLem2}, there exists a clause $\Gamma \imp \Delta,T(t) \in N_{MO}$
and a substitution $\sigma$ such that $f_p(s_1,\ldots,s_n) = \skt(t)\sigma$.
Therefore, $t=f_p(t_1,\ldots,t_n)$ with $s_i = \skt(t_i)\sigma$ for all $i$.
Then, $\R{P_1}(\ldots(\R{P_n}(\Gamma \imp \Delta,T(f_p(t_1,\ldots,t_n))))\ldots)=\R{P_1}(\ldots(\R{P_n}(\Gamma)\ldots) \imp \R{P_1}(\ldots(\R{P_n}(\Delta)\ldots),P(t_1,\ldots,t_n) \in N$ fulfills the claim.
\end{proof}
The above lemma also holds if satisfiability of $N'$ is dropped and $I$ is replaced
by the superposition partial minimal model operator~\cite{Weidenbach01handbook}.
\section{Lifting the Conflicting Core} \label{sec:lifting}
Given a monadic, linear, shallow, Horn approximation $N_k$ of $N$
and a conflicting core $N^\bot_k$ of $N_k$, using the transformations
provided in this section we attempt to lift $N^\bot_k$ to a conflicting
core $N^\bot$ of $N$.
In case of success this shows the unsatisfiability of $N$.
In case an approximation step cannot be lifted the original clause
set is refined by instantiation, explained in the next section.
Let $N_k$ be an unsatisfiable monadic, linear, shallow, Horn approximation.
Since $N_k$ belongs to a decidable first-order fragment,
we expect an appropriate decision procedure to generate a proof of unsatisfiability for $N_k$,
e.g., ordered resolution with selection \cite{Weidenbach99cade}.
A conflicting core can be straightforwardly generated out of a resolution refutation
by applying the substitutions of the proof to the used input clauses.
Starting with a resolution refutation, in order to construct the conflicting core,
we begin with the singleton set containing the pair of empty clause and the empty substitution.
Furthermore, we assume that all input clauses from $N_k$ used in the refutation are
variable disjoint.
Then we recursively choose a pair $(C,\sigma)$ from the set where $C \notin N_k$.
There exists a step in the refutation that generated this clause.
In the case of a resolution inference, there are two parent
clauses $C_1$ and $C_2$ in the refutation and two substitutions $\sigma_1$ and $\sigma_2$
such that $C$ is the resolvent of $C_1\sigma_1$ and $C_2\sigma_2$.
In the case of a factoring inference, there is one parent
clause $C'$ in the refutation and a substitution $\sigma'$
such that $C$ is the factor of $C'\sigma'$.
Replace $(C,\sigma)$ by $(C_1,\sigma_1\sigma)$ and $(C_2,\sigma_2\sigma)$ or
by $(C',\sigma'\sigma)$ respectively.
The procedure terminates in linear time in the size of the refutation.
For each pair $(C,\sigma)$, collect the clause $C\sigma$, resulting
in a conflicting core $N^\bot_k$ of $N_k$.
\begin{exmp}\label{coreex}
Let $N = \{ P(x,x');$ $P(y,a),P(z,b) \imp \}$ with signature $\Sigma = a/0, b/0$.
$N$ is unsatisfiable and a possible resolution refutation is resolving $P(b,a)$
and $P(a,b)$ with $P(b,a),P(a,b)\imp$. From this we get the conflicting core \\
$N^\bot_{ba}=\{ P(b,a);P(a,b);P(b,a),P(a,b)\imp\}$.
An alternative refutation is to resolve $P(x,x')$ and $P(y,a),P(z,b) \imp$ with
substitution $\{x\mapsto y;x' \mapsto a\}$
and then the resolvent and $P(x,x')$ with substitution $\{x\mapsto z;x' \mapsto b\}$.
From this refutation we construct the conflicting core
$N^\bot_{yz}=\{ P(y,a);$ $P(z,b);$ $P(y,a),P(z,b)\imp\}$.
\end{exmp}
Note that in Example \ref{coreex} $N^\bot_{yz}$ is more general than $N^\bot_{ba}$ since $N^\bot_{yz}\{y \mapsto b; z \mapsto a\} = N^\bot_{ba}$.
A conflicting core is minimal in that it represents the most general clauses corresponding to the refutation from that it is generated.
\subsubsection{Lifting the Monadic Transformation.}
Since the Monadic transformation is satisfiability preserving, lifting always succeeds by replacing any $T(f_P(t_1,\dots,t_n))$ atoms in the core by $P(t_1,\dots,t_n)$.
\begin{exmp}\label{LiftMonadicEx}
Let $N_0 = \{ P(x,x');$ $P(y,a),P(z,b) \imp \}$.
Then $N_k= \{ T(f_P(x,x'));$ $T(f_P(y,a)),T(f_P(z,b)) \imp \}$ is a Monadic transformation of $N_0$
and a conflicting core is $N^\bot_k= \{ T(f_P(y,a));$ $T(f_P(z,b));$ $T(f_P(y,a)),T(f_P(z,b)) \imp \}$.
Reverting the atoms in $N^\bot_k$ gives $N^\bot=\{ P(y,a);$ $P(z,b);$ $P(y,a),P(z,b)\imp\}$ a conflicting core of $N_0$.
\end{exmp}
\begin{lem}[Lifting the Monadic Transformation]\label{liftmonadic}
Let $N_k\Rightarrow^*_{\text{MO}} N_{k+l}$ be the transformation that exactly removes all occurrences of atoms $P(t_1,\dots,t_n)$ and replaces those by atoms $T(f_P(t_1,\dots,t_n))$.
If $N^\bot_{k+l}$ is a conflicting core for $N_{k+l}$ then there is a conflicting core $N^\bot_k$ of $N_k$.
\end{lem}
\begin{proof}
Since the Monadic transformation is satisfiability preserving, unsatisfiability of $N_{k+l}$ directly implies unsatisfiability of $N_{k}$ and the existence of a conflicting core of $N_{k}$.
\end{proof}
\subsubsection{Lifting the Horn Transformation.}
For a Horn transformation there are two ways for lifting.
The first, directly lifting the core, only succeeds in special cases, where the original clause and its approximation are equivalent
for the instantiations appearing in the core.
\begin{exmp}\label{LiftHornEx}
Let $N_0 = \{ P(a,b) \imp;$ $ P(x,b),P(a,y) \}$.
Then $N_k= \{ P(a,b) \imp;$ $ P(x,b) \}$ is a Horn transformation of $N_0$
and a conflicting core is $N^\bot_k= \{ P(a,b) \imp;$ $ P(a,b) \}$.
By substituting $y$ with $b$, $N^\bot_k$ lifts to $N^\bot=\{ P(a,b) \imp;$ $ P(a,b) , P(a,b) \}$ a conflicting core of $N_0$.
\end{exmp}
\begin{lem}[Lifting the Horn Transformation (direct)]\label{lifthorn1}
Let $N_k\Rightarrow_{\text{HO}} N_{k+1}$ where
$N_k = N\cup\{\Gamma \imp E_1, \dots,E_n\}$ and $N_{k+1} = N\cup\{\Gamma \imp E_i\}$.
Let $N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$.
If for all $(\Gamma \imp E_i)\sigma_j\in N^\bot_{k+1}$, $1\leq j\leq m$ there is a
substitution $\sigma'_j$ such that
$N_k^j\tau_j\models (\Gamma \imp E_1, \dots,E_n)\sigma'_j\rightarrow (\Gamma \imp E_i)\sigma_j$,
such that $N_k^j\subseteq N_k$ and
$N_k^j\tau_j \cup \{(\Gamma \imp E_1, \dots,E_n)\sigma'_j, \neg(\Gamma \imp E_i)\sigma_j\}$
is a conflicting core, then
$N^\bot_{k+1}\setminus \{(\Gamma \imp E_i)\sigma_j \mid 1\leq j\leq m\} \cup \{(\Gamma \imp E_1, \dots,E_n)\sigma'_j\mid 1\leq j\leq m\}\cup \bigcup\limits_j N_k^j\tau_j$ is
a conflicting core of $N_k$.
\end{lem}
\begin{proof}
Let $\sigma$ be a grounding substitution for $N^\bot_{k}$ and $N^\bot_{k+1}$.
Since $N_k\models (\Gamma \imp E_1, \dots,E_n)\sigma'_j\rightarrow (\Gamma \imp E_i)\sigma_j$,
$N^\bot_{k}\sigma \models N^\bot_{k}\sigma \cup \{(\Gamma \imp E_i)\sigma_j \mid 1\leq j\leq m\}\sigma \models N^\bot_{k+1}\sigma$.
Hence, $N^\bot_{k}\sigma$ is unsatisfiable because $N^\bot_{k+1}\sigma$ is unsatisfiable.
Therefore, $N^\bot_{k}$ is an conflicting core of $N_k$.
\end{proof}
Of course, the condition $N_k^j\tau_j\models (\Gamma \imp E_1, \dots,E_n)\sigma'_j\rightarrow (\Gamma \imp E_i)\sigma_j$ itself is undecidable, in general. The above lemma is meant to be a justification for
the cases where this relation can be decided, e.g, by reduction. In general, the next lemma applies.
We assume any non-Horn clauses have exactly two positive literals.
Otherwise, we would have first redefined pairs of positive literals using fresh predicates.
Further assume w.l.o.g. that Horn transformation always chooses the first positive Literal of a non-Horn clause.
The indirect method uses the information from the conflicting core to replace the non-Horn clause with a satisfiable equivalent unit clause,
which is then solved recursively. Since this unit clause is already Horn, we lifted
one Horn approximation step.
\begin{exmp}
Let $N_k = \{ P(a),Q(a); P(x)\imp\}$.
The Horn transformation $N_k=\{ P(a); P(x) \imp\}$ has a conflicting core
$N^\bot_k= \{ P(a); P(a) \imp\}$.
$N^\bot_k$ abstracts a resolution refutation with $\bot$ as the result.
If we replace $ P(a)$ with $P(a),Q(a)$ in such a refutation, the result will be $Q(a)$ instead and hence $N_k \sat Q(a)$
Since $Q(a)$ subsumes $P(a),Q(a)$, \\ $N_k$ is satisfiable if $N'_k=\{ Q(a); P(x)\imp\}$ is too.
\end{exmp}
\begin{lem}[Lifting the Horn Transformation (indirect)]\label{lifthorn2}
Let $N$ be a set of variable disjoint clauses,
$N\apr^* N_k \Rightarrow_{\text{HO}} N_{k+1}$,
$N_k = N\cup\{\Gamma \imp E_1,E_2\}$ and $N_{k+1} = N\cup\{\Gamma \imp E_1\}$ and
$N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$ where Lemma~\ref{lifthorn1} does not apply.
Let $(\Gamma \imp E_1)\sigma \in N^\bot_{k+1}$
, where $\sigma$ is a variable renaming and
$N_k^j\tau_j\not \models (\Gamma \imp E_1,E_2)\sigma'_j\rightarrow (\Gamma \imp E_1)\sigma$ for any $N_k^j\subseteq N_k$,$\tau_j$ and $\sigma'_j$.
If there exists a conflicting core $N^\bot$ of $N\cup\{E_2\}$, then a conflicting core of $N_k$ exists.
\end{lem}
\begin{proof}
From the conflicting core $N^\bot_{k+1}$, we can conclude that there exists an unsatisfiability proof of $N_{k+1}$ which derives $\bot$ and uses $(\Gamma \imp E_1)\sigma$ as the only instance of $\Gamma \imp E_1$. If we were to replace $(\Gamma \imp E_1)\sigma$ by $(\Gamma \imp E_1,E_2)\sigma$, the unsatisfiability proof's root clause would instead be $E_2\sigma$.
Hence, we know that $N_k\models N_k \cup \{E_2\sigma\}$.
Furthermore, $N_k\models N \cup \{E_2\sigma\}$ since $E_2\sigma$ subsumes $\Gamma \imp E_1,E_2$.
Let $E_2\sigma_j \in N^\bot$ for $1 \leq j \leq m$ and $N^{E_2}_k= N^\bot_{k+1} \setminus \{(\Gamma \imp E_1)\sigma\} \cup \{(\Gamma \imp E_1,E_2)\sigma\}$
Then $N^\bot \setminus \{E_2\sigma_j \mid 1 \leq j \leq m\} \bigcup\limits_j N^{E_2}_{k}\sigma_j$ is a conflict core of $N_k$.
\end{proof}
Note that $N_{k}$ now again contains the Non-Horn clause $\Gamma \imp E_1,E_2$.
Then, in a following indirect Horn lifting step $\Gamma \imp E_1,E_2$ can not necessarily be again replaced by $E_2\sigma$.
Hence, the indirect Horn lifting needs to be repeated.
\subsubsection{Lifting the Shallow Transformation.}
A Shallow transformation introduces a new predicate $S$, which is removed in the lifting step.
We take all clauses with $S$-atoms in the conflicting core and generate any
possible resolutions on $S$-atoms.
The resolvents, which don't contain $S$-atoms anymore, then replace their parent clauses in the core.
Lifting succeeds if all introduced resolvents are instances of clauses before the shallow
transformation.
\begin{exmp}\label{LiftShallowEx}
Let $N_0 = \{ P(x), Q(y) \imp R(x,f(y)); P(a);Q(b); R(a,f(b))\imp \}$.
Then $N_k = \{ S(x'),P(x) \imp R(x,x'); Q(y) \imp S(f(y)); P(a); Q(b); $ $ R(a,f(b)) \imp \}$ is a Shallow transformation of $N_0$
and a conflicting core is $N^\bot_k= S(f(b)),$ $P(a) \imp R(a,f(b));Q(b) \imp S(f(b));P(a); Q(b); R(a,f(b)) \imp$.
By replacing \\
$S(f(b)),P(a) \imp R(a,f(b))$ and $Q(b) \imp S(f(b))$ with the resolvent, $N^\bot_k$ lifts to $N^\bot =\{ P(a), Q(b) \imp R(a,f(b)); P(a);Q(b); R(a,f(b)) \imp\}$ a conflicting core of $N_0$.
\end{exmp}
\begin{lem}[Lifting the Shallow Transformation]\label{liftshallow}
Let $N_k\Rightarrow_{\text{SH}} N_{k+1}$ where $N_k = N\cup\{\Gamma \imp E[s]_{p}\}$ and $N_{k+1} = N\cup\{S(x),\Gamma_1 \imp E[x]_{p}\}\cup\{ \Gamma_2 \imp S(s)\}$.
Let $N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$.
Let $N_S$ be the set of all resolvents from clauses from $N^\bot_{k+1}$ on the $S$ literal.
If for all clauses $C_j\in N_S$, $1\leq j\leq m$ there is a substitution $\sigma_j$ such that $C_j = (\Gamma \imp E[s]_{p})\sigma_j$ then
$N^\bot_{k+1}\setminus\{C\mid C\in N^\bot_{k+1}\text{ and contains an }S\text{-atom}\}\cup\{(\Gamma \imp E[s]_{p})\sigma_j\mid 1\leq j\leq m\}$ is a conflicting core of $N_k$.
\end{lem}
\begin{proof}
Let $\sigma$ be a grounding substitution for $N^\bot_{k}$ and $N^\bot_{k+1}$ and $I$ be an interpretation.
As $N^\bot_{k+1}\sigma$ is unsatisfiable, there is a clause $D\in N^\bot_{k+1}\sigma$ such that $I\unsat D$. \\
If $D$ does not contain an $S$-atom, then $D \in N^\bot_{k}\sigma$ and hence $I\unsat N^\bot_{k}\sigma$.\\
Now assume only clauses that contain $S$-atoms are false under $I$.
By construction, any such clause is equal to either $(S(x),\Gamma_1 \imp E[x]_{p})\sigma'=C_1\sigma'$ or $(\Gamma_2 \imp S(s))\sigma'=C_2\sigma'$ for some substitution $\sigma'$.
Let $I' := \{S(s)\sigma' \mid $ $C_2\sigma' \in N^\bot_{k+1}\sigma \mathrm{ ~and~ } I\unsat C_2\sigma' \} \cup I \setminus \{ S(x)\sigma' ~\vert~ C_1\sigma' \in N^\bot_{k+1}\sigma \mathrm{ ~and~ } I\unsat C_1\sigma' \}$, i.e.,
we change the truth value for $S$-Literals such that the clauses unsatisfied under $I$ are satisfied under $I'$.\\
Since $I$ and $I'$ only differ on literals with predicate $S$ and $N^\bot_{k+1}\sigma$ is unsatisfiable, some clause $C$, containing an $S$-atom and satisfied under $I$, has to be false under $I'$.\\
Let $C=C_1\sigma_1$. Since $I\sat C$, $S(x)\sigma_1$ was added to $I'$ by some clause $D =C_2\sigma_2$, where $S(s)\sigma_2 = S(x)\sigma_1$.
Hence, $C$ and $D$ can be resolved on their $S$-literals and the resolvent $R$ is in $N^\bot_{k}\sigma$.
Since $I\unsat D$, $I'\unsat C$ and $R$ contains no $S$-atom, $I\unsat R$ and therefore $I\unsat N^\bot_{k}\sigma$.\\
For $C=C_2\sigma_2$ the proof is analogous.\\
Thus, for all interpretations $I$ and grounding substitutions $\sigma$, $I\unsat N^\bot_{k}\sigma$ and hence $N^\bot_{k}\sigma$ is a conflicting core of $N_k$.
\end{proof}
\subsubsection{Lifting the Linear Transformation.}
In order to lift a Linear transformation the remaining and the newly introduced variable
need to be instantiated the same term.
\begin{exmp}\label{liftLinEx}
Let $N_{k-1} = \{ P(x,x);$ $P(y,a),P(z,b) \imp \}$.
Then $N_k= \{ P(x,x');$ $P(y,a),P(z,b) \imp \}$ is a Linear transformation of $N_{k-1} $ and
and $N^\bot_k=\{ P(a,a);$ $P(b,b);$ $P(a,a),P(b,b)\imp\}$ is a conflicting core of $N_k$.
Since $P(a,a)$ and $P(b,b)$ are instances of $P(x,x)$ lifting succeeds and $N^\bot_k$ is also a core of $N_{k-1} $.
\end{exmp}
\begin{lem}[Lifting the Linear Transformation]\label{liftlin}
Let $N_k\Rightarrow_{\text{LI}} N_{k+1}$ where $N_k = N\cup\{\Gamma \imp E[x]_{p,q}\}$ and $N_{k+1} = N\cup\{\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q\}$.
Let $N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$.
If for all $(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\sigma_j\in N^\bot_{k+1}$, $1\leq j\leq m$ we have $x\sigma_j = x'\sigma_j$ then
$N^\bot_{k+1}\setminus\{(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\sigma_j \mid 1\leq j\leq m\}\cup\{(\Gamma \imp E[x]_{p,q})\sigma_j\mid 1\leq j\leq m\}$ is a conflicting core of $N_k$.
\end{lem}
\begin{proof}
Let $\sigma$ be a grounding substitution for $N^\bot_{k}$ and $N^\bot_{k+1}$.
As $x\sigma_j = x'\sigma_j$ for $1\leq j\leq m$, $(\Gamma \imp E[x]_{p,q})\sigma_j\sigma \sat (\Gamma,\Gamma \imp E[x]_{p,q})\sigma_j\sigma = (\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\sigma_j\sigma $.
Hence, $N^\bot_{k}\sigma \sat N^\bot_{k}\sigma \cup \{(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\sigma_j\sigma \mid 1\leq j\leq m\} \sat N^\bot_{k+1}\sigma$.
Since $N^\bot_{k+1}\sigma$ is unsatisfiable $N^\bot_{k}\sigma$ is unsatisfiable as well.
Therefore, $N^\bot_{k}$ is a conflicting core of $N_k$.
\end{proof}
\subsubsection{Lifting with Instantiation.}
By definition, if $N^\bot$ is a conflicting core of $N$, then $N^\bot\tau$ is also a conflicting core of $N$ for any $\tau$.
Example \ref{liftex} shows it is sometimes possible to instantiate a conflicting core, where no lifting lemma applies, into a core, where one does.
This then still implies a successful lifting.
\begin{exmp}\label{liftex}
Let $N_{k-1} = \{ P(x,x);$ $P(y,a),P(z,b) \imp \}$.
Then $N_k= \{ P(x,x');$ $P(y,a),P(z,b) \imp \}$ is a Linear transformation of $N_{k-1} $ and
and $N^\bot_k=\{ P(y,a);$ $P(b,b);$ $P(y,a),P(b,b)\imp\}$ is a conflicting core of $N_k$.
Since for $P(y,a)= P(x,x')\sigma$ $x\sigma=y\neq a=x'\sigma$ Lemma \ref{liftlin} is not applicable.
However, Lemma \ref{liftlin} can be applied on $N^\bot_k\{y \mapsto a; z \mapsto b\}= \{P(a,a);$ $P(b,b);$ $P(a,a),P(b,b)\imp\}$.
\end{exmp}
\section{Approximation Refinement}\label{abstrref}
In the previous section, we have presented the lifting process.
If, however, in one of the lifting steps conditions of the lemma are not met, lifting fails and
we now refine the original clause set in order to rule out the non-liftable conflicting core.
Again, since lifting fails at one of the approximation steps, we consider the different
approximation steps for refinement.
\subsubsection{Linear Approximation Refinement.}
A Linear transformation
enables further instantiations of the abstracted clause compared to the original,
that is, two variables that were the same can now be instantiated differently.
If the conflicting core of the approximation contains such instances the lifting fails.
\begin{defin}[Linear Approximation Refinement]
Let $N$ be a set of variable disjoint clauses,
$N\apr^* N_k \Rightarrow_{\text{LI}} N_{k+1}$ and
$N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$ where Lemma~\ref{liftlin} does not apply.
Let $C'\sigma=(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\sigma \in N^\bot_{k+1}$
such that $x\sigma$ and $ x'\sigma$ have no common instances.
Let $C\in N$ be the Ancestor of $C'\in N_{k+1}$.
Then the \emph{linear approximation refinement} of $N$, $C$, $x$, $x'$, $\sigma$ is the clause
set $N\setminus\{C\}\cup\{C\tau_1,\ldots,C\tau_n\}$ where the $C\tau_i$ are the specific
instances of $C$ with respect to the substitutions $\{x\mapsto x\sigma\}$ and
$\{x\mapsto x'\sigma\}$.
\end{defin}
Note that if there is no $C'\sigma$, where $x\sigma$ and $ x'\sigma$ have no common instances,
it implies that there is a substitution $\tau$
where Lemma~\ref{liftlin} applies on $N^\bot_{k+1}\tau$. Hence, $N^\bot_{k+1}\tau$ is a liftable
conflicting core.
Let $N_0\apr^* N_{k-1}= N\cup\{\Gamma \imp E[x]_{p,q}\} \Rightarrow_{\text{LI}} N_{k}= N\cup\{\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q\}$
and the core $N^\bot_k$ of $N_k$ contains the clause $C'\sigma=(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\sigma$, where $x\sigma$ and $x'\sigma$ have no common instances.
After applying Linear Approximation Refinement, there are $C\tau_i$ and $ C\tau_j$ with $i\neq j$ such that $C\tau_i$ contains all instances where $\{x \mapsto x\sigma \}$
and $C\tau_j$ contains all instances where $\{x \mapsto x'\sigma \}$.
Assume there is a $C''$ with an ancestor $C\tau$ such that $C'\sigma$ is an instance of $C''$.
This would imply that $C\tau$ has instances, where $\{x \mapsto x\sigma\} $ and $\{x \mapsto x'\sigma\} $.
Then $C\tau_i= C\tau= C\tau_j$, which is a contradiction to Definition \ref{specInst}.
\begin{exmp}\label{LinAbsEx}
Let $N_0 = \{ P(x,x);$ $P(y,a),P(z,b) \imp \}$.
Then $N_k= \{ P(x,x');$ $P(y,a),P(z,b) \imp \}$ is a Linear transformation of $N_0 $ and
and $N^\bot_k=\{ P(a,a);$ $P(a,b);$ $P(a,a),P(a,b)\imp\}$ is a conflicting core of $N_k$.\\
Due to $P(a,b)= P(x,x')\{x\mapsto a, x' \mapsto b\}$ lifting fails.
The Linear Approximation Refinement replaces $P(x,x)$ in $N_0$ with $P(a,a)$ and $P(b,b)$.
In the refined approximation $N'_k = \{ P(a,a);P(b,b);$ $P(y,a),P(z,b)\imp\}$
the violating clause $P(a,b)$ is not an instance of $N'_k$ and hence, the not-liftable conflicting core $N^\bot_k$ cannot be found again.
\end{exmp}
\subsubsection{Shallow Approximation Refinement.}
The Shallow transformation is somewhat more complex than linear transformation,
but the idea behind it is very similar to the linear case.
As mentioned before, the Shallow transformation can always be lifted
if the set of common variables
$\vars(\Gamma_2,s) \cap \vars(\Gamma_1,E[x]_p)$ is empty.
Otherwise, each such variable potentially
introduces instantiations that are not liftable.
\begin{defin}[Shallow Approximation Refinement]
Let $N$ be a set of variable disjoint clauses,
$N\apr^* N_k \Rightarrow_{\text{SH}} N_{k+1}$ and
$N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$ where Lemma~\ref{liftshallow} does not apply.
Let $C_R$ be the resolvent from the final Shallow rule application such that $C_R \neq (\Gamma \imp E[s]_{p})\sigma_R$ for any $\sigma_R$.
Let $C_1\sigma_1 \in N^\bot_{k+1}$ and $C_2\sigma_2 \in N^\bot_{k+1}$ be the parent clauses of $C_R$.
Let $y \in \text{dom}(\sigma_1) \cap \text{dom}(\sigma_2)$, where $y\sigma_1$ and $ y\sigma_2$ have no common instances.
Let $C\in N$ be the Ancestor of $C_1\in N_{k+1}$.
Then the \emph{shallow approximation refinement} of $N$, $C$, $x$, $\sigma_1$, $\sigma_2$ is the clause
set $N\setminus\{C\}\cup\{C\tau_1,\ldots,C\tau_n\}$ where the $C\tau_i$ are the specific
instances of $C$ with respect to the substitutions $\{x\mapsto x\sigma_1\}$ and
$\{x\mapsto x\sigma_2\}$.
\end{defin}
As in Linear Approximation Refinement, if for every resolvent $C_R\sigma$ $y\sigma_1$ and $ y\sigma_2$ have common instances, it implies that there is a substitution $\tau$
where Lemma~\ref{liftshallow} applies on $N^\bot_{k+1}\tau$.
After applying Shallow Approximation Refinement, there are $C\tau_i$ and $ C\tau_j$ with $i\neq j$ such that $C\tau_i$ contains all instances where $\{x \mapsto x\sigma_1\} $
and $C\tau_j$ contains all instances where $\{x \mapsto x\sigma_2\}$.
Hence, $C\tau_i$ is now the ancestor of $C_1\sigma_1$, while $C\tau_j$ is the ancestor of $C_2\sigma_2$.
Since they have different ancestors, they can no longer be resolved on their $S$-atoms which now have different predicates.
Hence $C_R$ is no longer a resolvent in the conflicting core.
\begin{exmp}
Let $N_0 = \{ P(f(x,g(x))); P(f(a,g(b))\imp\}$ with signature $\Sigma = a/0,$ $b/0,$ $g/1, f/2$.
Then $N_k= \{ S(z) \imp P(f(x,z)); S(g(y));$ $ P(f(a,g(b))\imp\}$ is a Shallow transformation of $N_0 $ and
and $N^\bot_k=\{ S(g(b)) \imp P(f(a,g(b))); S(g(b));$ $ P(f(a,g(b))\imp\}$ is a conflicting core of $N_k$.\\
The clauses $S(g(b)) \imp P(f(a,g(b)))$ and $ S(g(b))$ have the resolvent $P(f(a,g(b)))$, which is not an instance of $P(f(x,g(x)))$.
The Shallow Approximation Refinement replaces $P(f(x,g(x)))$ in $N_0$ with $P(f(a,g(a)))$, $P(f(b,g(b)))$,\\ $P(f(g(x),g(g(x))))$ and $P(f(f(x,y),g(f(x,y))))$.\\
The approximation of the refined $N_0$ is now satisfiable.
\end{exmp}
\subsubsection{Horn Approximation Refinement.}
Lifting a core of a Horn transformation fails, if the positive literals removed by
the Horn transformation are not dealt with in the approximated proof.
Since Lemma \ref{lifthorn2} only handles cases where the approximated clause appears uninstantiated in the conflicting core,
the Horn Approximation Refinement is used to ensure such a core exists.
\begin{defin}[Horn Approximation Refinement]
Let $N$ be a set of variable disjoint clauses,
$N\apr^* N_k \Rightarrow_{\text{HO}} N_{k+1}$,
$N_k = N\cup\{\Gamma \imp E_1,E_2\}$ and $N_{k+1} = N\cup\{\Gamma \imp E_1\}$ and
$N^\bot_{k+1}$ be a conflicting core of $N_{k+1}$ where Lemmas ~\ref{lifthorn1} and ~\ref{lifthorn2} do not apply.
Let $(\Gamma \imp E_1)\sigma \in N^\bot_{k+1}$
be a clause from the final Horn rule application such that $\sigma$ is not a variable renaming and
$N_k^j\tau_j\not \models (\Gamma \imp E_1,E_2)\sigma'_j\rightarrow (\Gamma \imp E_1)\sigma$ for any $N_k^j\subseteq N_k$,$\tau_j$ and $\sigma'_j$.
Let $C\in N$ be the Ancestor of $\Gamma \imp E_1\in N_{k+1}$
and $\sigma'$ a substitution such that $\sigma\sigma'$ is linear for $C$.
Then the \emph{horn approximation refinement I} of $N$, $C$, $\sigma$, $\sigma'$ is the clause
set $N\setminus\{C\}\cup\{C\sigma\sigma',C\tau_1,\ldots,C\tau_n\}$ where the $C\tau_i$ are the specific
instances of $C$ with respect to the substitutions $\sigma\sigma'$.
\end{defin}
Note that the condition for the extended version of specific instantiation to have a finite representation is not generally met by an arbitrary $\sigma$.
Therefore, $\sigma$ may need to be further instantiated or even made ground.
After the Horn Approximation Refinement, Lemma~\ref{lifthorn2} can be applied on
the clause with ancestor $C\sigma\sigma'$.
\begin{exmp}
Let $N_0 = \{ P(x),Q(x); P(a)\imp\}$ with signature $\Sigma = a/0, f/1$.
The Horn transformation $N_k=\{ P(x); P(a) \imp\}$ has a conflicting core
$N^\bot_k= \{ P(a); P(a) \imp\}$.
We pick $ \imp P(a)$ as the instance of $ P(x) \in N^\bot_k$ to use for the Horn Approximation Refinement.
The result is $N'_0=\{ P(a),Q(a);$ $ P(f(x)),Q(f(x));$ $ P(a)\imp\} $
and its approximation also has $N^\bot_k$ as a conflicting core.
However, now Lemma \ref{lifthorn2} applies.
\end{exmp}
\begin{lem}[Completeness]\label{complete}
Let $N$ be an unsatisfiable clause set and $N_k$ its approximation.
Then, there exists a conflicting core of $N_k$ that can be lifted to $N$.
\end{lem}
\begin{proof}
by induction on the number $k$ of approximation steps. The case $k=0$ is obvious.
For $k>0$, let $N\apr^* N_{k-1} \apr N_k$.
By the inductive hypothesis, there is a conflicting core $N^\bot_{k-1}$ of $N_{k-1}$ which can
be lifted to $N$.\\
The final approximation rule application is either a Linear, a Shallow, a Horn or a Monadic transformation,
considered below by case analysis.
Linear Case. Let $N\apr^* N_{k-1}= N'\cup\{\Gamma \imp E[x]_{p,q}\} \Rightarrow_{\text{LI}} N_k= N'\cup\{\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q\}$.
For every $(\Gamma \imp E[x]_{p,q})\sigma_j \in N^\bot_{k-1}$ $1\leq j\leq m$, $(\Gamma \imp E[x]_{p,q})\sigma_j \models (\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)(\{x' \mapsto x\}\sigma_j)$.
Hence $N^\bot_k=N^\bot_{k-1}\setminus\{(\Gamma \imp E[x]_{p,q})\sigma_j\mid 1\leq j\leq m\}\cup \{(\Gamma\{x \mapsto x'\},\Gamma \imp E[x']_q)\{x' \mapsto x\}\sigma_j \mid 1\leq j\leq m\}$
is a conflicting core of $N_k$.
By Lemma \ref{liftlin} $N^\bot_k$ can be lifted back to $N^\bot_{k-1}$.
Hence, the conflicting core $N^\bot_k$ can be lifted to $N$.
Shallow Case. Let $N\apr^* N_{k-1}=N'\cup\{\Gamma \imp E[s]_{p}\} \Rightarrow_{\text{SH}} N_{k}= N'\cup\{S(x),\Gamma_1 \imp E[x]_{p}\}\cup\{ \Gamma_2 \imp S(s)\}$.
We construct $N^\bot_S$ from $N^\bot_{k-1}$ by replacing every $(\Gamma \imp E[s]_{p})\sigma_j \in N^\bot_{k-1}$ $1 \leq j \leq m$ with $(S_j(x),\Gamma_1 \imp E[x]_p)\sigma_j$ and $(\Gamma_2 \imp S_j(s))\sigma_j$.
$N^\bot_S$ is a conflicting core, which by $m$ applications of Lemma \ref{liftshallow} on each $S_j$ can be lifted to $N^\bot_{k-1}$.
From $N^\bot_S$ we get $N^\bot_{k}$ by renaming every $S_j$ into $S$, which is a conflicting core of $N_k$.
The existence of $N^\bot_S$ shows that $N^\bot_{k}$ can be lifted to $N^\bot_{k-1}$.
Horn Case. W.l.o.g. let $N\apr^* N_{k-1}=N'\cup\{\Gamma \imp E_1,E_2\} \Rightarrow_{\text{HO}} N_{k}= N'\cup\{\Gamma \imp E_1\}$.
Let $C = \Gamma \imp E_1 ,E_2$ and $C' = \Gamma \imp E_1$.
If $C\sigma \in N^\bot_{k-1}$ holds for at most one $\sigma$, we construct $N^\bot_{k}$ from $N^\bot_{k-1}$ by replacing $C\sigma$ with $C'\sigma$ such that $N^\bot_{k}\subseteq N_k$.
Since $C'\sigma$ subsumes $C\sigma$, $N^\bot_{n} \sat N^\bot_{n}\cup\{C\sigma\}$.
As $N^\bot_{k}\cup\{C\sigma\}$ is a superset of $N^\bot_{k-1}$, $N^\bot_{k}$ is therefore a ground conflicting core of $N_k$.
If $C'\sigma$ and $C\sigma$ are already equivalent, $N^\bot_{k}$ can be lifted to $N^\bot_{k-1}$.
Otherwise, let $N'^\bot_{k-1}$ be $N^\bot_{k-1}$ where $C\sigma$ is instead replaced by $E_2\sigma$.
Again since $E_2\sigma$ subsumes $C\sigma$, $N'^\bot_{k-1}$ is a ground conflicting core.
As shown before, $(N'^\bot_{k-1}\setminus\{E_2\sigma\})\cup (N^\bot_{k}\setminus\{C'\sigma\})=N^\bot_{k-1}$ is a lifting from $N_k$ to $N_{k-1}$.\\
Assume $C\sigma_1 \in N^\bot_{k-1}$ and $C\sigma_2 \in N^\bot_{k-1}$ holds for $\sigma_1\neq \sigma_2$.
In this case the original clause $C$ can be specifically instantiated in such a way that $C\sigma_1$ and $C\sigma_2$ are no longer instances of the same clause,
while $N^\bot_{k-1}$ remains a conflicting core.
Hence, after finitely many such partitions eventually the first case will hold.
Monadic Case. Let $N\apr^* N_{k-j} \Rightarrow^*_{\text{MO}} N_k$ where $N_{k-j}$ has no occurrence of an atom
$T(f_P(t_1,\dots,t_n))$ and $N_{k}$ no occurrence of an atom $P(t_1,\dots,t_n)$ and
all introduced atoms in the transformation are of the form $T(f_P(s_1,\dots,s_n))$.
By the inductive hypothesis, there is a ground conflicting core $N^\bot_{k-j}$ of $N_{k-j}$ which can be lifted to $N$.
By Lemma \ref{Apr_sound}(ii) Monadic transformation preserves unsatisfiability and therefore $\Proj{P}(N^\bot_{k-j})$ is a ground conflicting core of $N_{k}$.
$\Proj{P}(N^\bot_{k-j})$ can be lifted to $\R{P}(\Proj{P}(N^\bot_{k-j}))=N^\bot_{k-j}$ a conflicting core of $N_{k-j}$.
\end{proof}
The above lemma considers static completeness, i.e., it does not tell how the conflicting
core that can eventually be lifted is found. One way is to enumerate all refutations of $N_k$
in a fair way. A straightforward fairness criterion is to enumerate the refutations by increasing
term depth of the clauses used in the refutation. Since the decision procedure on the
mslH fragment~\cite{Weidenbach99cade} generates only finitely many different
non-redundant clauses not exceeding a concrete term depth with respect to the renaming of variables, eventually the
liftable refutation will be generated.
\section{Future and Related Work} \label{sec:furewo}
\input{refinements.tex}
\subsubsection{Related Work}
In ''A theory of abstractions''~\cite{Giunchiglia:1992:TA:146945.146951} Giunchiglia and Walsh don't define an actual approximation but a general framework to classify and compare approximations, which are here called abstractions.
They informally define abstractions as ''the process of mapping a representations of a problem'' that ''helps deal with the problem in the original search space by preserving certain desirable properties`` and ''is simpler to handle``.
In their framework an abstraction is a mapping between formal systems, i.e., a triple of a language, axioms and deduction rules, which satisfy one of the following conditions:
An increasing abstraction (TI) $f$ maps theorems only to theorems, i.e., if $\alpha$ is a theorem, then $f(\alpha)$ is also a theorem,
while a decreasing abstraction (TD) maps only theorems to theorems, i.e., if $f(\alpha)$ is a theorem, then $\alpha$ was also a theorem.
Furthermore, they define dual definitions for refutations, where not theorems but formulas that make a formal system inconsistent are considered.
An increasing abstraction (NTI) then maps inconsistent formulas only to inconsistent formulas and vice versa for decreasing abstractions (NTD).
They list several examples of abstractions such as ABSTRIPS by Sacerdoti~\cite{Sacerdott:1973:PHA:1624775.1624826}, a GPS planning method by Newell and Simon~\cite{Newell:1972:HPS:1095704}, Plaisted's theory of abstractions~\cite{journals/ai/Plaisted81}, propositional abstractions exemplified by Giunchiglia~\cite{conf/ecai/GiunchigliaG88}, predicate abstractions by by Plaisted~\cite{journals/ai/Plaisted81} and Tenenberg~\cite{Tenenberg87preservingconsistency}, domain abstractions by Hobbs~\cite{Hobbs85granularity} and Iemielinski~\cite{Imielinski:1987:DAL:1625995.1626083} and ground abstractions introduced by Plaisted~\cite{journals/ai/Plaisted81}.
With respect to their notions the approximation described in this paper is an abstraction where the desirable property is the over-approximation and the decidability of the fragment makes it simpler to handle. More specifically in the context of \cite{Giunchiglia:1992:TA:146945.146951} the approximation is an NTI abstraction for refutation systems, i.e., it is an abstraction that preserves inconsistency of the original.
In Plaisted~\cite{journals/ai/Plaisted81} three classes of abstractions are defined. The first two are ordinary and weak abstractions, which share the condition that if $C$ subsumes $D$ then every abstraction of $D$ is subsumed by some abstraction of $C$. However, our approximation falls in neither class as it violates this condition via the Horn approximation. For example $Q$ subsumes $P, Q$, but the Horn
approximation $P$ of $P, Q$ is not subsumed by any approximation of $Q$.
The third class are generalization functions, which change not the problem but abstract the resolution rule of inference.
The theorem prover iProver uses the Inst-Gen~\cite{DBLP:conf/birthday/Korovin13} method,
where a first-order problem is abstracted with a SAT problem by replacing every variable by the fresh constant $\bot$.
The approximation is solved by a SAT solver and its answer is lifted to the original
by equating abstracted terms with the set they represent, e.g., if $P(\bot)$ is true in a
model returned by the SAT solver, then all instantiations of the original $P(x)$ are
considered true as well.
Inst-Gen abstracts using an under-approximation of the original clause set.
In case the lifting of the satisfying model is inconsistent,
the clash is resolved by appropriately instantiating the involved clauses, which mimics an inference step.
This is the dual of our method with the roles of satisfiability and unsatisfiability switched.
A further difference, however, is that Inst-Gen only finds finite models after approximation,
while our approximation also discovers infinite models.
For example the simple problem $\{ P(a)$, $\neg P(f(a))$, $P(x) \imp P(f(f(x)))$, $P(f(f(x))) \imp P(x)\}$ has the satisfying model where $P$ is the set of even numbers.
However, iProver's approximation can never return such a model as any $P(f^n(\bot))$ will necessarily abstract both true and false atoms and therefore instantiate new clauses infinitely.
Our method on the other hand will produce the approximation $\{P(a)$, $\neg P(f(a))$, $S(y) \imp P(f(y))$, $P(x) \imp S(f(x))$, $P(f(f(x))) \imp P(x)\}$, which is saturated after inferring $P(x) \imp P(f(f(x)))$ and $\neg S(f(a))$.
In summary, we have presented the first sound and complete calculus for
first-order logic based on an over-approximation-refinement loop. There
is no implementation so far, but the calculus will be practically useful if a problem
is close to the mslH fragment in the sense that only a few refinement loops
are needed for finding the model or a liftable refutation.
The abstraction relation is already implemented and applying it to all satisfiable non-equality
problems TPTP version~6.1 results in a success rate of 34\%, i.e., for all these problems
the approximation is not too crude and directly delivers the result.
It might be possible
to apply our idea to other decidable fragments of first-order logic.
However, then they have to support via approximation the presented lifting and refinement
principle.
Our result is also a first step towards a model-based guidance of first-order
reasoning. We proved that a model of the approximated clause set is also a model for the
original clause set. For model guidance, we need this property also for partial models. For example,
in the sense that if a clause is false with respect to a partial model operator on
the original clause set, it is also false with respect to a partial model operator on
the approximated clause set. This property does not hold for the standard superposition partial
model operator and the mslH approximation suggested in this paper. It is subject to future research.
\bibliographystyle{plain}
\section{Skeleton and Partial Minimal Model Construction}
As mentioned before, Lemma \ref{termSkelLem} also holds if satisfiability of $N'$ is dropped and $I$ is replaced
by the superposition partial minimal model operator~\cite{Weidenbach01handbook}.
\begin{defin}[Partial Minimal Model Construction]
Given the set of ground clauses $N_g$ of $N$ and an ordering $\prec$ we construct an interpretation $\I{N}$ for $N$, called a partial model, inductively as follows:
\begin{align*}
\I{C} &:= \bigcup_{ D \in N_g,D\prec C} \delta_D\\
\delta_D &:= \left\{
\begin{array}{l l}
\{P\} & \quad \text{if $D=D'\vee P$, P strictly maximal and $\I{D}\unsat D$}\\
\emptyset & \quad \text{otherwise}
\end{array} \right. \\
\I{N} &:= \bigcup_{C \in N_g} \delta_C
\end{align*}
Clauses $C$ with $\delta_C \neq \emptyset$ are called productive.
\end{defin}
Note that this construction doesn't terminate since the ground clause set of $N$ is generally infinite.
\begin{lem}\label{termSkelLem3}
Let $N_k$ be a monadic clause set and $N_0$ be its approximation via $\apr$.
If $P(s) \in \I{N_0}$ and $P$ is a predicate in $N_k$, then there exists a clause
$C = \Gamma \imp \Delta,P(t) \in N_k$ and a substitution $\sigma$ such that
$s = \skt(t)\sigma$ and for each variable $x$ and predicate $S$ with $C = S(x),\Gamma' \imp \Delta,P(t[x]_p)$, $S(s'') \in \I{N_0}$, where $s=s[s'']_p$.
\end{lem}
\begin{proof}
By induction on $k$.\\
The base $N_k=N_0$ holds by definition of the model operator $\I{}$. \\
Let $N = N_k\apr N_{k-1} \apr^* N_0$, $P(s) \in \I{N_0}$ and $P$ is a predicate in $N_k$ and hence also in $N_{k-1}$.
By the induction hypothesis, there exist a clause $C = \Gamma \imp \Delta,P(t) \in N_{k-1}$ and a substitution $\sigma$ such that
$s = \skt(t)\sigma$ and for each variable $x$ and predicate $S$ with $C = S(x),\Gamma' \imp \Delta,P(t[x]_p)$, $S(s'') \in \I{N_0}$, where $s=s[s'']_p$.
Let $\apr$ be a Horn transformation that replaces $\Gamma'' \imp \Delta',Q(t')$ with $\Gamma'' \imp Q(t')$.
If $C \neq \Gamma'' \imp Q(t')$, then $C \in N_k$ fulfills the claim.
Otherwise, $\Gamma'' \imp \Delta',Q(t) \in N_k$ fulfills the claim since $P=Q$ and $\Gamma' = \Gamma''$.
Let $\apr$ be a linear transformation that replaces $C_k = \Gamma'' \imp E[x]_{p,q}$ with $C_{k-1}=\Gamma'',\Gamma''\{x \mapsto x'\} \imp E[x']_{q}$.
If $C \neq C_{k-1}$, then $C \in N_k$ fulfills the claim.
Otherwise, $C_k=\Gamma'' \imp P(t)\{x' \mapsto x\} \in N_k$ fulfills the claim since $s = \skt(t)\sigma = \skt(t\{x' \mapsto x\})\sigma$ and $\Gamma'' \subseteq \Gamma'',\Gamma''\{x \mapsto x'\}$.
Let $\apr$ be a shallow transformation that replaces $C_k = \Gamma'' \imp E[s']_{p}$ with $C_{k-1}=S(x),\Gamma_1 \imp E[x]_p$ and $C'_{k-1}=\Gamma_2 \imp S(s')$.
Since $S$ is fresh, $C \neq C'_{k-1}$.
If $C \neq C_{k-1}$, then $C \in N_k$ fulfills the claim.
Otherwise, $C = C_{k-1} = S(x),\Gamma_1 \imp P(t[x]_p)$ and hence, $s = \skt(t[x]_p)\sigma$ and $S(s'')\in \I{N_0}$ for $s=s[s'']_p$.
Then by the induction hypothesis, there exist a clause $C_S = \Gamma_S \imp \Delta_S,S(t_S) \in N_{k-1}$ and a substitution $\sigma_S$ such that
$s'' = \skt(t_S)\sigma_S$ and for each variable $x$ and predicate $S'$ with $C_S = S'(x),\Gamma'_S \imp \Delta_S,P(t_S[x]_q)$, $S'(s''') \in \I{N_0}$, where $s''=s''[s''']_q$.
By construction, $C_S = C'_{k-1}$.
Thus, $s''= \skt(s')\sigma_S$ and $s = \skt(t[x]_p)\sigma$ imply there exists a $\sigma''$ such that $s = \skt(t[s']_p)\sigma"$.
Furthermore since $\Gamma_1 \cup \Gamma_2 = \Gamma''$, if $C_k= S'(x),\Gamma''' \imp P(t[s']_{p})[x]_q$, then
either $S'(x) \in \Gamma_1$ and thus $S'(s'''') \in \I{N_0}$, where $s=s[s'''']_q$, or
$S'(x) \in \Gamma_2$ and thus $S'(s'''') \in \I{N_0}$, where $s[s'']_p=(s[s'']_p)[s'''']_q$.
Hence, $C_k \in N_k$ fulfills the claim.
\end{proof}
\begin{lem}\label{termSkelLem4}
Let $N$ be a clause set and $N'$ be its approximation via $\apr$.
If $P(s) \in \I{N'}$ $(T(f_p(s_1,\ldots,s_n))\in \I{N'})$
and $P$ is a predicate in $N$, then there exist a clause
$\Gamma \imp \Delta,P(t) \in N$ ($\Gamma \imp \Delta,P(t_1,\ldots,t_n) \in N$)
and a substitution $\sigma$ such that
$s = \skt(t)\sigma$ ($s_i = \skt(t_i)\sigma$ for all $i$).
\end{lem}
\begin{proof}
Let $P_1,\ldots,P_n$ be the non-monadic predicates in $N$ and $N_{MO}=$ $\Proj{P_1}(\ldots(\Proj{P_n}(N)))$.
Then, $N_{MO}$ is monadic and also has $N'$ as its approximation via $\apr$.
Let $P(s) \in \I{N'}$ and $P$ is a predicate in $N$. Since $P$ is monadic, $P$ is a predicate in $N_{MO}$.
Hence by Lemma ~\ref{termSkelLem3}, there exists a clause $\Gamma \imp \Delta,P(t) \in N_{MO}$
and a substitution $\sigma$ such that $s = \skt(t)\sigma$.
Then, $\R{P_1}(\ldots(\R{P_n}(\Gamma \imp \Delta,P(t)))\ldots)= \R{P_1}(\ldots(\R{P_n}(\Gamma)\ldots) \imp \R{P_1}(\ldots(\R{P_n}(\Delta)\ldots),P(t)\in N$ fulfills the claim.
Let $T(f_p(s_1,\ldots,s_n))\in \I{N'}$ and $P$ is a predicate in $N$. $T$ is monadic and a predicate in $N_{MO}$.
Hence by Lemma ~\ref{termSkelLem3}, there exists a clause $\Gamma \imp \Delta,T(t) \in N_{MO}$
and a substitution $\sigma$ such that $f_p(s_1,\ldots,s_n) = \skt(t)\sigma$.
Therefore, $t=f_p(t_1,\ldots,t_n)$ with $s_i = \skt(t_i)\sigma$ for all $i$.
Then, $\R{P_1}(\ldots(\R{P_n}(\Gamma \imp \Delta,T(f_p(t_1,\ldots,t_n))))\ldots)=\R{P_1}(\ldots(\R{P_n}(\Gamma)\ldots) \imp \R{P_1}(\ldots(\R{P_n}(\Delta)\ldots),P(t_1,\ldots,t_n) \in N$ fulfills the claim.
\end{proof}
|
\section{Introduction}
\mbox{\v{C}erenkov}{} counters have been used in high energy electron scattering
experiments as part of detector stacks within shielded spectrometers,
especially at low duty cycle beam facilities where the instantaneous beam
current was at the milliamp level. With the advent of continuous electron beams
with very high duty cycle it became feasible to use \mbox{\v{C}erenkov}{} counters in an
open environment for the same average beam current. In the SANE experiment data
were collected in Hall C at Jefferson Lab using a threshold gas \mbox{\v{C}erenkov}{}
counter, as an integral part of a new detector package known as the Big
Electron Telescope Array (BETA), which identified scattered electrons, measured
their scattered angle and their energy. BETA, shown in figure~\ref{fig:BETA0},
consisted of a forward tracker, a gas \mbox{\v{C}erenkov}{} detector which is the focus
of this report, a Lucite hodoscope and a lead-glass calorimeter.
Using a polarized electron beam and a polarized ammonia target,
The SANE collaboration~\cite{SANEproposal} set out to measure two observables,
known as double spin asymmetries, $A_{\parallel}$ and $A_{\perp}$. These are
measured with two different configurations, parallel or perpendicular, of the
spin directions of the beam and the target. The electron beam's helicity was
flipped while the spin direction of the target has been fixed. To detect the
scattered electrons BETA was positioned at a central scattering angle of $40$
degrees, covering roughly $\pm5$\mbox{$^\circ$}{} in scattering angle, and from about
$700$ MeV up to about $2.5$ GeV in energy, thus covering a kinematic range of
Bjorken $x$ and four momentum transfer $Q^2$ corresponding to $0.3 < x < 0.8$
and $2.5$~GeV$^2 < Q^2 < 6.5$~GeV$^2$.
The gas \mbox{\v{C}erenkov}{} counter's role was to identify inclusively scattered
electrons and to reject pions during the experiment, thus providing an
efficient trigger for clean electrons in a high background environment.
A description of the SANE apparatus follows in section \ref{sec:BETA}. In
addition to a mechanical overview, section \ref{sec:Design} discusses the
design choices and their motivation. Section~\ref{sec:construction} presents,
in detail, the construction of the SANE \mbox{\v{C}erenkov}{} counter. This is followed
by a description of the detector's calibrations and initial commissioning.
Electronics and calibrations are the focus of section~\ref{sec:PMTs}. Due to
differences in background rates, magnetic field, and particle trajectories, the
performance during each target field orientation is separately discussed in
section~\ref{sec:Performance}. Our conclusions are described in
section~\ref{sec:conclusion}.
\section{SANE Apparatus}
\label{sec:BETA}
To reach the statistical precision required by SANE in a limited amount of
incident beam time BETA, a device unique in its open configuration, was built
to provide the required angular and momentum acceptance for the experiment.
\begin{center}
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{beta_diagram-figure13.pdf}
\caption{ BETA detectors and polarized target.\label{fig:BETA0}}
\end{figure}
\end{center}
\subsection{The BETA Detector Package}
BETA comprises four detectors, a forward tracker placed as close to the target
as possible, followed by a \mbox{\v{C}erenkov}{} counter, a Lucite hodoscope and a large
electromagnetic calorimeter dubbed BigCal. As can be seen in
Figure~\ref{fig:BETA0}, the front window of the \mbox{\v{C}erenkov}{} counter was
positioned just behind the forward tracker while BigCal was set at roughly
$3.5$m from the target. The dimensions and positions of the detectors and
target are shown in Figure~\ref{fig:BETADimensions}.
BigCal provided position and energy measurements of electrons and photons. It
consisted of 1792 lead glass blocks and was divided into two sections. The
lower section makes use of $3.8$~cm$\times3.8$~cm$\times45$~cm blocks arranged
in a $32\times32$ grid, while the upper section contains
$4$~cm$\times4$~cm$\times40$~cm blocks arranged in a $30\times24$ grid.
The Lucite hodoscope includes 28 curved Lucite bars with light guides mounted
to edges cut at 45\mbox{$^\circ$}{}. Stacked vertically, each bar was 6 cm tall and 3.5
cm thick and provided a vertical position measurement. Photomultipliers (PMTs)
were connected at both ends of the bar, providing an additional position
determination in the horizontal direction by taking the time difference between
two discriminated PMT signals.
The forward tracker used wavelength shifting fibers glued to BC-408 plastic
scintillator to detect the scattered particles as close to the target as
possible. Two layers of $3$~mm$\times3$~mm$\times22$~cm scintillators, stacked
vertically and offset by $1.5$ mm along with a layer of
$3$~mm$\times3$~mm$\times40$~cm scintillators piled horizontally, provided a
position measurement with a resolution sufficient to distinguish between
electron and positron trajectories of momenta, below $1$~GeV/c.
\subsection{Polarized Target }
Used in previous deep-inelastic scattering experiments that measured also
double spin asymmetries at
SLAC~\cite{Averett:1999nz,Abe:1998wq,Anthony:2002hy,McNulty:2002yf} and at
Jefferson Lab~\cite{Wesselmann:2006mw}, the polarized target system used a 5.1
T magnet to polarize ammonia through the mechanism of dynamic nuclear
polarization. In addition to polarizing the target material, the presence of
this magnetic field reduces background radiation by trapping charged particles
with a momentum less than about $180$~MeV. As will be discussed in
section~\ref{sec:transverse}, magnetic trapping also increased background
during transverse polarization running. Due to depolarization effects, the
nominal production beam current was kept roughly at about $100$~nA which in
turns limited the experimental luminosity to about $\sim$1.7$\times$10$^{35}$
cm$^{-2}$ s$^{-1}$ when using the $3$~cm thick ammonia target.
The experiment required two target configurations, (anti-)parallel, with the
magnetic field pointed along the incoming electron beam direction, and
transverse, where mechanical constraints limited the target angle to be 80
degrees with respect to the beam direction. In addition to rotating the magnet,
additional upstream chicane magnets were used to compensate for the beam
deflection during transverse running. Deflection of background particles in the
downstream target field required a non-standard beamline incorporating a helium
gas bag and extended exit beam pipe.
\begin{center}
\begin{figure}[h]
\centering
\includegraphics[trim = 1mm 12mm 10mm 1mm, clip, width=\textwidth]{beta_diagram-figure10.pdf}
\caption{BETA dimensions with side view (upper figure) and a top view (lower
figure). Shown from left to right are the calorimeter, hodoscope, \mbox{\v{C}erenkov}{} counter, forward tracker and polarized target.
\label{fig:BETADimensions} }
\end{figure}
\end{center}
\section{Design of the \mbox{\v{C}erenkov}{} Counter}
\label{sec:Design}
The primary requirements for the SANE gas \mbox{\v{C}erenkov}{} counter were to identify
electrons with high efficiency (greater than $90$\%) while maintaining a pion
rejection factor better than 1000:1 with software cuts. The open configuration
of the experiment required a high $\pi^\pm$ momentum threshold to reject much
of the low energy background. The \mbox{\v{C}erenkov}{} counter needed to cover a rather
large solid angle of about 200 msr and thus was placed
as close as possible to the target.
Charged pions needed to be rejected for momenta up to $5$~GeV. Pions above
this momentum threshold should be extremely rare with a 4.7 GeV or 5.9 GeV
incident electron beam and would be removed in any case with our software
electron selection cuts. Looking at the thresholds for commonly used gases
shown in Table~\ref{tab:certhresholds}, N$_2$ gas was selected as it best meets
this rejection requirement.
\begin{table}[h]
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
\emph{Charged particle}& N$_2$ & CO$_2$ & C$_4$F$_{10}$ \\
\hline
$e^+$, $e^-$ & 21 MeV & 17 MeV & 7.8 MeV \\
$\mu^+$, $\mu^-$ & 4.33 GeV & 3.5 GeV & 1.6 GeV \\
$\pi^+$, $\pi^-$ & 5.75 GeV & 4.6 GeV & 2.1 GeV \\
\hline
\end{tabular}
\caption{\mbox{\v{C}erenkov}{} thresholds for charge particles in a N$_2$ gas at
atmospheric pressure compared with other common gases.
\label{tab:certhresholds}}
\end{center}
\end{table}
Operating at atmospheric pressure helped simplify the mechanical design and
minimize window thicknesses. The choice of a N$_2$ gas \mbox{\v{C}erenkov}{} for radiator
was a trade-off between \mbox{\v{C}erenkov}{} photon-yield, material budget, undesired
scintillation, and (ultraviolet) transparency. Although N$_2$ gas is known to
scintillate, the low density and optical design minimized the impact of the
isotropic scintillation background (Section~\ref{sec:Optics}). At $20$\mbox{$^\circ$}
C, the index of refraction of N$_2$ gas is $n=1.000279$, yielding a threshold
velocity, $\beta_{0}$ , for \mbox{\v{C}erenkov}{} light emission by charged particles of
\begin{displaymath}
\beta_{0} = \frac{1}{n} = 0.999721.
\end{displaymath}
Furthermore, N$_2$ is inexpensive in comparison to other gasses making a gas
recovery system unnecessary.
The leading mechanical design constraint was the optics design for focusing the
\mbox{\v{C}erenkov}{} light onto PMTs. Additional constraints on the mechanical design
included a high rate of background at small angles, access to any mirrors and
PMTs, and operation in the fringe magnetic field from the target.
A symmetric mirror-PMT design could not be considered as a viable option given
the high background rates and long experimental run times in the area closer to
the beam line. In order to allow for easy maintenance of the PMTs and to
minimize worker radiation exposure\cite{ALARA} all PMTs were set on the large
scattering angle side of the detector. Furthermore, with the PMTs on one side
(as seen in figure \ref{fig:BETADimensions}), background shielding was
consolidated into two large walls.
This asymmetric design called for elliptical mirrors on the far side of the
\mbox{\v{C}erenkov}{} tank instead of spherical mirrors. Although complicating mirror
procurement, these mirrors provided a focused light profile necessary to
operate with three-inch diameter PMTs. Five-inch diameter tubes were initially considered,
however their increased size made it very difficult to shield magnetically, and
shield from background radiation. The polarized target's $5.1$ Tesla
superconducting magnet sat just over one meter away necessitating a significant
amount of mu-metal shielding which housed the PMTs. The mu-metal was rolled
into $3$ mm thick cylinders to mitigate the effects of a $100-200$ Gauss
magnetic field around the three-inch PMTs.
\section{\mbox{\v{C}erenkov}{} Construction}
\label{sec:construction}
\subsection{Mirror Optics}
\label{sec:Optics}
After an extensive detailed optical ray-trace analysis which included among other things
the effect of the target magnetic field on the scattered electron trajectories,
the shapes of the mirrors were determined\cite{BETAMirrorDesign}. The
optimized result called for eight, roughly 40 cm by 40 cm glass mirrors
arranged in two overlapping columns of four mirrors to cover the rather large
acceptance shown in Fig~\ref{fig:BETADimensions}. Four spherical mirrors cover
one column at large scattering angles and four elliptical mirrors cover another
column at small scattering angles. For machining purposes, the elliptical
mirror was approximated with a spindle torus without any effect on the
performance \cite{MirrorBlanks}. The dimensions of the mirrors are listed in
Table~\ref{tab:mirrorDimensions}. Light from each mirror is focussed onto
individual 3-inch quartz window photomultipliers (Photonis XP4318B).
\begin{table}[h]
\begin{center}
\begin{tabular}{|l|l|}
\hline
Spherical Mirror & \\
\hline
radius of curvature & $92.0$ cm \\
vertical size & $36.5$ cm \\
horizontal size & $35.5$ cm \\
\hline \hline
Toroidal Mirror & \\
\hline
minor circle (`tube') radius & $85.8$ cm \\
circle of revolution (`donut') radius & $25.8$ cm \\
vertical size & $36.5$ cm \\
horizontal size & $43.0$ cm \\
\hline
x semiaxis (ellip. equiv.)& $97.0$ cm \\
y and z semiaxes (ellip. equiv.)& $86.0$ cm \\
\hline
\end{tabular}
\caption{Dimensions for the spherical and toroidal mirrors. Also the torus-approximated elliptical surface's semiaxes for a mirror in the x-y plane are given. }
\label{tab:mirrorDimensions}
\end{center}
\end{table}
\begin{figure}[h]
\centerline{\includegraphics[width=\textwidth,scale=1.0,angle=0.0]{cherenkov_mirror_reflectivity.pdf}}
\caption{\label{fig:reflectivity}Mirror reflectivity as measured at CERN \cite{mirrorcoating}. }
\end{figure}
The glass mirrors were coated with aluminum and MgF$_2$ for maximum
reflectivity into the far UV in order to compliment the sensitivity of the
photomultiplier tube at these wavelengths. The reflectivity for each mirror was
measured at CERN\cite{mirrorcoating} and shown in
Figure~\ref{fig:reflectivity}.
The electrons of interest have momentum above 0.7~GeV/c and are deflected
by the target field less than a few degrees. Thus, to a good
approximation, the mirrors have been designed for point-to-point focusing
from the target cell to the photomultiplier photocathodes. This
permits the two towers of mirrors to be optimally aligned with a
small, bright light bulb located at the same target-mirror
distance. This geometry also permits good rejection of stray light
from scintillation and low energy $\delta$ rays (which are
preferentially emitted at angles several times larger than the
\mbox{\v{C}erenkov}{} cone).
\subsection{Mirror Alignment}
\begin{figure}[h]
\centering
\includegraphics[width=0.75\textwidth]{mirror_holder-figure0.pdf}
\caption{A cross section view of the mirror holder u-channel. The channel was machined to match the mirror curvature. }
\label{fig:mirrorHolder}
\end{figure}
Mirror alignment was completed before the detector was placed into its final
position. A curved u-channel assembly (Figure~\ref{fig:mirrorHolder}) held the mirror
and connected to the frame-mounted holding arms that adjusted the mirror positions.
The mirrors were aligned with the front window removed and using an
attachment mounted to the very front of the detector, which held a small
incandescent light bulb at the location of the target relative to the detector.
Figures \ref{fig:mirroralign}a and \ref{fig:mirroralign}b show the alignment
procedure using the front attachment which illuminated half of the detector at
a time. A cap placed on the photo-multipliers protected them from light while
the detector was open and during mirror alignment. These caps also had
concentric circle targets on the outside to aid in the visual mirror alignment.
With the attachment in place, the light bulb along with other tooling balls
were used to survey the detector, allowing for very accurate final placement
relative to the target as shown in Figure \ref{fig:mirroralign}c.
\begin{figure}[h]
\centering
\subfloat[Align top
mirrors.]{\includegraphics[width=0.3\textwidth]{mirror_alignment-figure0} }
\subfloat[Align bottom mirrors.]{\includegraphics[width=0.3\textwidth]{mirror_alignment-figure1} }
\subfloat[Completed mirror alignment.]{\includegraphics[width=0.35\textwidth]{mirror_alignment-figure2} }
\caption{\mbox{\v{C}erenkov}{} mirror alignment procedure.}
\label{fig:mirroralign}
\end{figure}
In order to calibrate and observe changes in the detector a LED/mirror system,
shown in Figure~\ref{fig:ledSystem}, was installed inside the \mbox{\v{C}erenkov}{}. An
actuator rotated a thin mirror just inside the front window to reflect light
from a LED, which also rotated to mirror the nominal target position.
When not in use the mirror and LED are rotated flush against the inside of the \mbox{\v{C}erenkov}{} snout.
This LED system was used to check for position shifts after the \mbox{\v{C}erenkov}{} was
moved into its final position near the target.
\begin{figure}[h]
\centering
\includegraphics[ trim = 40mm 65mm 5mm 41mm, clip,width=1\textwidth]{led_system-figure0}
\caption{Plan view of the \mbox{\v{C}erenkov}{} LED/mirror monitoring system.}
\label{fig:ledSystem}
\end{figure}
\subsection{Photo-multipliers}
\label{sec:PMTs}
Three inch diameter Photonis XP4318B quartz window photomultiplier tubes were
placed inside a $3$~mm thick mu-metal shield with its photocathode recessed 2
inches from the end of the cylinder as shown in Figure~\ref{fig:pmtholder}.
Quartz windows provided a complementary UV transparency needed to match the
reflectivity of the mirrors. Shown as the dashed-dotted curve in
Figure~\ref{fig:cerspectrum}, the relative efficiency of the PMT is a
combination of the photo-cathode quantum efficiency, the collection efficiency
and quartz window transparency. The \mbox{\v{C}erenkov}{} photon spectrum is given by
the expression
\begin{equation}
\frac{d^2N}{d\lambda dx} = 2 \pi \alpha (1-\frac{1}{ n^2 \beta^2})\frac{1}{\lambda^2}
\end{equation}
where $\alpha$ is the fine structure constant, $\beta$ is the velocity of the
particle and $\lambda$ is the photon wavelength. The number of photons, N, is
simply proportional to the gas length, L, which we take on average to be
$1.3$~meters.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{PMT_holder-figure0.pdf}
\caption{ A diagram (not to scale) of the $\mu$-metal PMT holder. \label{fig:pmtholder} }
\end{figure}
Neglecting the N$_2$ gas absorption, each scattered electron causes roughly 185
photons to fall onto the mirrors and about 155 of these reflect towards the
photo-multiplier tube. From such a light pulse about 20 photo-electrons should
be counted. This result can be calculated by integrating the solid curve in
Figure~\ref{fig:cerspectrum} which is given by
\begin{equation}\label{eq:cerSpectrum}
\frac{dN}{d\lambda} = 2 \pi \alpha (1-\frac{1}{ n^2
\beta^2})\frac{\mathrm{L}}{\lambda^2} \eta(\lambda) R(\lambda)
\end{equation}
where $R$ is the reflectivity and $\eta$ is the relative efficiency of the PMT.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{dndL_sensitivity.pdf}
\caption{The mirror reflectivity (top, dashed), effective PMT quantum
efficiency (top, solid line), the \mbox{\v{C}erenkov}{} spectrum (bottom, dashed), and
the photo-electron yield integrand, Eq.~\ref{eq:cerSpectrum} (bottom, solid
line). \label{fig:cerspectrum}}
\end{figure}
\subsection{Gas System and Tank Construction}
\label{sec:Gas_System}
As previously mentioned, N$_2$ was selected as the \mbox{\v{C}erenkov}{} radiator gas due
to the very high energy threshold for \mbox{\v{C}erenkov}{} light production from
particles heavier than an electron. Although air is mostly N$_2$ gas, it
contains moisture and other gases (such as oxygen) which attenuate and distort
propagating photons.
Using a dry N$_2$ gas source, a controller monitored and regulated the gas
pressure inside the tank over the course of the experiment. A pressure
transducer measured the pressure relative to the current atmospheric pressure
and the controller opened and closed the appropriate solenoid valves to
maintain a differential pressure of a few torr. This slight constant
overpressure reduced the possible contamination from the outside gases.
Based on initial measurements, the average delivery flow rate of gas while
maintaining a differential pressure of 10 torr is less than 0.1 Standard Cubic
Foot per Hour (SCFH). For an ideal hermetic and perfectly sealed \mbox{\v{C}erenkov}{}
tank this rate would be zero, and therefore this rate is also referred to as a
\textit{leak rate}. This is the rate that gas will enter the \mbox{\v{C}erenkov}{} tank
when the manometer calls for more gas and automatically opens the fill solenoid
valve to maintain a set point pressure.
Flushing the tank was important once it was sealed to purge all atmospheric gas
present. The controller was placed in flush mode and the relative humidity of
the venting gas monitored. A relative humidity of less than a few percent was
achieved in about 2 hours as shown in figure~\ref{flushtest3}.
Desiccant was placed at the bottom of the detector to help remove any remaining moisture from the gas.
\begin{figure}[h]
\centering
\includegraphics[width=1\textwidth]{flushtest3.pdf}
\caption{Relative humidity while flushing at a differential pressure of about $10$~torr and a gas flow rate of $1.1$~SCFM. \label{flushtest3}}
\end{figure}
The tank's frame was constructed of welded and leak-checked two inch square
aluminum tubing. Attached to the front of the frame was the snout which
narrows down to the front window frame holder. Attached to this frame holder
was a thin piece of aluminum which is completely opaque, painted flat black,
and framed by an aluminum bracket. The side panels of the main tank's frame
were detachable, as was the rear window, a 1/16 inch aluminum sheet held by a
large bracket. Neoprene gaskets and O-rings were used throughout when
attaching or sealing any panels or flanges.
Two large lead walls were constructed to shield the PMTs from background. The
first sat between the target and \mbox{\v{C}erenkov}{} PMTs to shield them from
scattering originating at the target. A second wall on the opposite side
shielded the PMTs from secondary sources of scattering located around the beam
dump.
\section{Computer Simulations}
\label{sec:simulation}
A GEANT4 \cite{Agostinelli:2002hh} simulation was developed to simulate the
full BETA detector package called BETAG4\cite{BETAG4}. In addition to the BETA
detectors, various aspects of the experiment and apparatus were implemented
such as the target geometry, materials, magnetic field and emulation of the
various triggers and scalers. Beyond providing insight into the behavior of the
\mbox{\v{C}erenkov}{} counter alone, the simulation was used to explore various
correlations with other detectors, observe the effects of extra material
thicknesses, and study the event reconstruction from the calorimeter back to
the target for each target polarization direction.
One important study undertaken using the simulation was to investigate the
effect of the extra material in front of the \mbox{\v{C}erenkov}{} counter. This extra
material thickness is almost entirely due to the forward tracker (which roughly
totals to the same thickness in radiation lengths as the ammonia target).
Beyond the small energy loss in this material, the primary problem comes from
photons, which produce electron-positron pairs in the extra material. These
pairs can produce a track that is misidentified as a scattered electron by the
\mbox{\v{C}erenkov}{} counter. To understand the effects of these background events they
are classified by the location of their vertex and fall in two categories.
\begin{figure}[h]
\centering
\includegraphics[trim = 4mm 4mm 4mm 14mm, clip,width=\textwidth]{simulated_cherenkov2_900.pdf}
\caption{Simulated \mbox{\v{C}erenkov}{} counter ADC spectrum for isotropic photons on
BETA with energies uniformly sampled between $0.5$~GeV and $4$~GeV. The ADC
values are normalized by the average ADC single for a single electron track.
\emph{Intra-target} pairs produced within the target material in the central
region of the target vacuum chamber are exposed to a very strong magnetic
field. This field separates the electron and positron that at least one of the
particles is ejected from BETA's acceptance. \emph{Intra-target} pairs dominate
the first peak located around 1 track and \emph{extra-target} pairs are
responsible for the second (double track) peak.
\label{fig:cerSimGamma}}
\end{figure}
Photons produce pairs within the target material in the central region of the
target vacuum chamber, where the target magnetic field strength is quite large
compared to the forward tracker location, just outside of the vacuum chamber.
For these \emph{intra-target} pairs, the strong magnetic field significantly
deflects (in opposite directions) the electron and positron trajectories, so
much so, that even the high energy pairs almost never produce two tracks in
BETA.
However, \emph{extra-target} pairs produced in the forward tracker material
generate nearly undeflected track into BETA, and more importantly, the pair
produces \emph{twice} the amount of \mbox{\v{C}erenkov}{} photons as a good scattered
electron track. This is clearly demonstrated in the simulation results of the
ADC spectrum shown in Figure~\ref{fig:cerSimGamma}. We will return to the
discussion of the pair symmetric background in
section~\ref{sec:PairBackground}.
\section{Calibration and Commissioning}
The PMTs were calibrated to about 100 ADC channels per photo-electron.
Following a technique of photo-electron counting described in
\cite{Dossi:1998zn}, the charge response of each PMT was modeled as the sum of
a pedestal Gaussian, single electron response function,
$\text{SER}(\mathrm{x})$, and a multiple photoelectron response
$M(\mathrm{x})$. The pedestal provides the so-called ``noise'' function which
is convoluted with the single electron response. To account for the non-ideal
first dynode response and dynode noise, an exponential function is added to the
ideal single (photo-) electron Gaussian function. This ideal single electron
response function is defined as \begin{equation}
\text{SER}_0(\mathrm{x}) =
\begin{cases}
\displaystyle \frac{p_E}{A}~\mathrm{e}^{-\frac{\mathrm{x}-\mathrm{x}_p}{A}}
+ \frac{1-p_E}{g_N
\sqrt{2\pi\sigma_0}}~\mathrm{e}^{-\left(\frac{\mathrm{x}-\mathrm{x}_0-\mathrm{x}_p}{\sqrt{2}\sigma_0}\right)^2
}
& \mathrm{x} > \mathrm{x}_p \\
0 & \mathrm{x} < \mathrm{x}_p
\end{cases}
\end{equation}
where $\mathrm{x}_p$ is the pedestal peak position,
$\sigma_p$ is the pedestal width,
$p_E$ is the fraction of events which fall under the exponential,
$A$ is the decay constant of exponential,
$\mathrm{x}_0$ is the single photo-electron peak position, and
$\sigma_0$ is the width of the single photo-electron peak.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{OnePE_calibration_2_71805_1.pdf}
\caption{Single photo-electron response fit result (solid) using
equation~\ref{eq:PMTfit}, neglecting the $M(\mathrm{x})$ term. Also shown are
the contributions from the pedestal (dotted), exponential (dashed), and Gaussian
(solid) part of $SER(\mathrm{x})$. \label{fig:onePEsingle}}
\end{figure}
In order to take into account non-ideal issues such as electrical noise or ADC resolution, a noise function, defined as
\begin{equation} \label{eq:PMTNoise}
\displaystyle \text{Noise}(\mathrm{x}) = \frac{1}{\sqrt{2\pi}\sigma_p}
\mathrm{e}^{-\left(\frac{\mathrm{x}-\mathrm{x}_p}{\sqrt{2}\sigma_p}\right)^2},
\end{equation}
is convoluted with the ideal single electron response to yield a realistic
single (photo-)electron response function.
\begin{equation} \label{eq:PMTSER}
\displaystyle\text{SER}(\mathrm{x}) = (\text{Noise}\otimes\text{SER}_0)(\mathrm{x})
\end{equation}
This function used in the model of the full ADC spectrum,
\begin{equation} \label{eq:PMTfit}
\displaystyle f(\mathrm{x}) = N_0( P(0) \text{Noise}(\mathrm{x}) + P(1) \text{SER}(\mathrm{x}) + M(\mathrm{x}))
\end{equation}
where $P(0)$ and $P(1)$ are the probabilities for 0 and 1 photo-electron response, and $N_0$ is a normalization.
The multiple photo-electron response, $M(\mathrm{x})$ is
\begin{equation} \label{eq:Mx}
\displaystyle M(\mathrm{x}) = \sum_{n=2}^{N_{M}}
\frac{P(n;\mu)}{\sqrt{2n\pi}\sigma_1} \mathrm{e}^{-\frac{(\mathrm{x} - n
\mathrm{x}_1
- \mathrm{x}_p)}{2 n \sigma_1^2}}
\end{equation}
where
$\mu$ is the average number of photo-electrons,
$\mathrm{x}_1$ ($\simeq\mathrm{x}_0$) is the average number of channels per photo-electron,
$\sigma_1$ is the average width of the (Gaussian) PMT response to a single photo-electron,
$N_{M}$ is the cut off in the sum which should be much greater than $\mu$,
and
$P(n;\mu)$ is the Poisson probability distribution for $n$ photo-electrons with mean value $\mu$.
In order to determine the location of the single photo-electron peak, the LED
pulser was set very low, $\mu < 0.01$. The pedestal amplitude was roughly
determined by counting the number of events within the pedestal, $N_p$, and
then taking the ratio to yield the zero photo-electron probability
$P(0)=N_p/N_{Tot}$, where $N_{Tot}$ is the total number of LED triggers.
Since $\mu$ is much less than unity ($\mu \ll 1$), the contribution of multiple
photo-electrons in Eq.~\ref{eq:PMTfit} can be neglected when fitting the LED
data. An example of such a fit is shown in figure~\ref{fig:onePEsingle}.
Fitting yields values for the parameters $\mathrm{x}_p$, $\sigma_p$,
$\mathrm{x}_0$, and $\sigma_0$ for each PMT.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{LED_fit5_0_1_2.pdf}
\caption{A fit to the photomultiplier ADC spectrum with an LED amplitude of a
few photo-electrons. It shows the various terms including the pedestal
(dotted), exponential (dashed), single photo-electron peak (solid), and the
first few terms in the multiple photo-electron response (solid).
\label{fig:LEDfit}}
\end{figure}
Next, in order to calculate the contribution from multiple photoelectrons the
parameters in equation~\ref{eq:Mx} are needed. To first approximation
$\mathrm{x}_1 \simeq \mathrm{x}_0$ and $\sigma_1 \simeq \sigma_0$, however, for
larger signals this approximation may not hold. In order to check this, the LED
pulser amplitude is turned up to $\mu\simeq5$. Subsequently using the fixed parameters
$\mathrm{x}_p$, $\sigma_p$, $\mathrm{x}_0$, and $\sigma_0$, a spectrum fit is
performed.
From this fit the values of $\mu_{1}$ and $\sigma_{1}$ are determined as shown in Figure~\ref{fig:LEDfit}.
\begin{figure}[h]
\centering
\includegraphics[trim = 0mm 0mm 0mm 10mm,clip,width=0.8\textwidth]{mirror_edges.pdf}
\caption{The mirror edges are shown in the calorimeter cluster positions when
a TDC hit is required on two adjacent mirrors for a cluster. Also note how the
size of the \mbox{\v{C}erenkov}{} cone can be estimated from the distribution of events
across the mirror edges.\label{fig:edges}}
\end{figure}
For some events the \mbox{\v{C}erenkov}{} cone is split between multiple mirrors, thus
locating the mirror edges is important for knowing how to form an appropriate
ADC sum. Plotting the calorimeter's cluster positions while requiring a TDC
for at least two a adjacent mirrors produces a projection of the mirror edges
as shown in Figure~\ref{fig:edges}. When a cluster falls within the region of an
edge, the corresponding mirror's ADCs are added to form the \mbox{\v{C}erenkov}{} sum
associated with the cluster.
\section{Performance }
\label{sec:Performance}
\subsection{Data Acquisition and Analysis}
\label{sec:DAQ}
The eight analog \mbox{\v{C}erenkov}{} signals were summed using a Lecroy 428F Quad
Linear Fan-In/Fan-Out NIM module. Using a single discriminator, the analog
sum, in coincidence with the calorimeter, formed the primary trigger used
during the SANE experiment. The performance of the SANE Gas \mbox{\v{C}erenkov}{} is
reported for two configurations which are characterized by the polarized
target's magnetic field orientation.
The SANE experiment used a 5.1 Tesla superconducting magnet with its field
aligned anti-parallel or transverse ($80^{\circ}$) to the beam direction. The
\mbox{\v{C}erenkov}{} detector was exposed to very different fringe fields and background
rates (Table~\ref{tab:cerRates}) with each target field configuration. This
difference in rates is primarily due to lower energy forward scattered
particles circling along the polarization direction of the magnet, which in the
parallel case travels along the beam direction towards the beam dump. However,
for the transverse polarization, these low energy particles are directed
towards the detector thus increasing the background compared to the parallel
orientation. The individual PMT rates were roughly proportional to the solid
angle covered by the corresponding mirror.
\begin{table}[h]
\begin{center}
\begin{tabular}{|l|r|r|}
\hline
Channel & Parallel & Perpendicular \\
\hline
1 & 33 kHz & 1532 kHz \\
2 & 72 kHz & 685 kHz\\
3 & 41 kHz & 960 kHz\\
4 & 74 kHz & 473 kHz\\
5 & 43 kHz& 587 kHz\\
6 & 86 kHz & 516 kHz\\
7 & 53 kHz & 1023 kHz\\
8 & 60 kHz & 338 kHz\\
\hline
Lower 4 sum & 103 kHz& 928 kHz\\
Middle 4 sum & 131 kHz & 1059 kHz\\
Upper 4 sum & 114 kHz & 1125 kHz \\
\hline
Total sum & 206 kHz & 1795 kHz \\
\hline
\end{tabular}
\caption{Rates during normal operations for the two configurations. \label{tab:cerrates}}
\label{tab:cerRates}
\end{center}
\end{table}
The detector commissioning occurred with the target polarization in the
transverse orientation due to target magnet problems. Therefore, we begin by
reporting the performance during the first part of the experiment.
\subsection{Transverse Field Orientation}
\label{sec:transverse}
The first \mbox{\v{C}erenkov}{} counter data were taken with the target in transverse
field orientation. A few unanticipated problems arose during initial
commissioning of the detector.
The spherical mirror PMTs had a slightly larger than anticipated longitudinal
magnetic field component which degraded the efficiency of the PMTs on that
side. Their performance degradation was monitored using the LED pulsing
system. To help mitigate this inefficiency, a large 3/4-inch iron plate was
quickly mounted on the front of the lead shielding wall to act as a magnetic
yoke. The performance improved markedly after the plate was installed.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{72795_cherenkov_sums.pdf}
\caption{ADC spectra during transverse field configuration are shown: the
analog sum of all eight mirrors (no hatches), cluster correlated ADC sum
(single hatches), and cluster correlated ADC sum with a \mbox{\v{C}erenkov}{} TDC cut.
\label{fig:PerpADCSums}}
\end{figure}
The transverse field production runs are characterized by high background
rates on all detectors, including the \mbox{\v{C}erenkov}{} counter. Therefore, clean
event selection required a correlated calorimeter cluster and hodoscope hit.
The necessity of these cuts can be seen in Figure~\ref{fig:PerpADCSums}.
Comparing the analog sum, which was used in the trigger, to the software sum
of the cluster correlated mirrors, there remained a substantial background.
With the addition of a \mbox{\v{C}erenkov}{} timing cut, the background is dramatically
reduced, indicating a background composed of uncorrelated, low energy electrons
or positrons, consistent with the description of nearly trapped particles in
the target's transverse magnetic field.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{72795_Cherenkov8_log.pdf}
\caption{\mbox{\v{C}erenkov}{} counter ADC spectrum from a transverse field
configuration run for all the toroidal mirrors (top) and spherical mirrors
(bottom).\label{fig:PerpClean}}
\end{figure}
\subsection{Anti-parallel Field Orientation}
\begin{figure}[h]
\centering
\includegraphics[width=0.8\textwidth]{72999_cherenkov_sums.pdf}
\caption{ The histograms above show the calibrated ADC spectra for an
anti-parallel configuration. The ADC spectrum of the analog sum of all eight
mirrors (no hatches), cluster correlated ADC sum (single hatches), and cluster
correlated ADC sum with a \mbox{\v{C}erenkov}{} TDC cut. \label{fig:ParaADCSums}}
\end{figure}
The background rates were significantly lower during parallel target running.
Unlike the transverse configuration the software sums have much less
background. Even without the aid of BETA detector event selection, the
\mbox{\v{C}erenkov}{} spectrum appeared relatively free of background as shown in
Figure~\ref{fig:ParaADCSums}.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\textwidth]{Cherenkov8_log.pdf}
\caption{Same as Figure~\ref{fig:PerpClean} for a anti-parallel field
run.\label{fig:ParaClean} }
\end{figure}
\subsection{Pair Symmetric Background Identification}
\label{sec:PairBackground}
It can be seen from the ADC spectra of Figures~\ref{fig:PerpClean} and
\ref{fig:multiFit6} that there was a significant contribution from tracks that
have twice the amount of light than a single electron track produces.
Most background events come from electron positron pair-production which can
originate from one of two locations as previously discussed
(Section~\ref{sec:simulation}).
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{performanc_by_mirror5_6.pdf}
\caption{A typical ADC spectrum showing the one and two track peaks and the
convolution fit. \label{fig:multiFit6}}
\end{figure}
Simulation has shown that \emph{intra-target} pairs produced at the target are
very unlikely to produce a double hit due to the strong (and opposite)
deflections through the target magnetic field. This background cannot be
removed and must be accounted for in the data analysis.
Only \emph{extra-target} pairs remain close enough to each other to appear as a
single cluster at the calorimeter. Fortunately, placing an ADC window cut on
the \mbox{\v{C}erenkov}{} ADC spectrum alone removes a significant amount of these pair
events and allows for an accurate estimation of the cuts efficiency. The ADC
window cut consists of the usual lower ADC limit and an upper limit located
between the single and double track peaks. The precise location of the upper
limit is a compromise between statistics and (extra-target) pair background
rejection. For a systematic error limited measurement, pushing this limit
closer to the single track eliminates a large fraction of the background, and
by fitting the ADC spectrum as shown in Figure~\ref{fig:multiFit6}, the
background contamination can be calculated by integrating the double track
peak.
\section{Conclusion }
\label{sec:conclusion}
The SANE gas \mbox{\v{C}erenkov}{} detector operated successfully in a high luminosity
environment with an open configuration and covered a large solid angle and
momentum acceptance. With a large photo-electron yield, it identified
electrons with high efficiency. This large photo-electron yield provided enough
ADC separation of double track events, allowing for novel event selection with
the removal of (extra-target) pair symmetric background events. Although the
\mbox{\v{C}erenkov}{} counter operated efficiently under both target configurations, the
transverse field orientation proved most challenging due to higher background
rates.
\section{Acknowledgements}
This work is supported by DOE grant DE-FG02-94ER4084
|
\section{Introduction}
The analysis of complex systems \citep{friedrich2011approaching,kwapien2012physical,crutchfield2012between} is recently gaining traction with the development of new computational and theoretical tools, specifically designed to tackle the problem of disentangling the complex interactions among different parts of a process.
A common approach for this purpose is the study of scale-invariant properties \cite{stanley1999scaling}, i.e., those features that manifest in a similar manner at all magnification levels of the system. The scale-invariance property is a peculiar feature of fractal structures, and it is quantifiable by evaluating a suitable collection of scaling exponents that compactly describe how the system evolves when observed at different scales \cite{harte2010multifractals}.
In particular, the scaling of the correlation properties of a stochastic process could provide valuable information on the underlying dynamical system, especially in cases in which the observed system denotes long-term memory and nonlinear behaviors \cite{kantelhardt2009fractal,drozdz2009quantitative}.
Long-term memory properties of a stochastic process are firstly described by power-law like decreasing of the autocorrelation calculated for suitable observables.
This feature can be characterized by the value of the so-called Hurst exponent $H$ \cite{serinaldi2010use}, which assumes values in $[0, 1]$.
The value of $H$ is 0.5 when the process corresponds to uncorrelated noise, whereas if the process is persistent (correlated) or antipersistent (anticorrelated) then $H$ is greater than or less than 0.5, respectively.
However, conventional methods employed to analyze the long-term correlations of a time series (e.g., spectral analysis and Hurst analysis) are known to be misleading when data are non-stationary and contains deterministic trends \cite{PhysRevE.65.041107,PhysRevE.64.011114}.
Therefore, in many experiments it is important to distinguish intrinsic fluctuations characterizing the process from pure trending behaviors.
A well-established approach employed to this end is the Detrended Fluctuation Analysis (DFA) \cite{shao2012comparing}, which has been successively generalized in the so-called Multifractal Detrended Fluctuation Analysis (MF-DFA) \cite{kantelhardt2002multifractal,PhysRevE.74.016103,kantelhardt2009fractal}. MF-DFA accounts for the possibility to characterize different degrees of correlation by considering a collection of scaling exponents.
Such scaling exponents provide a way to perform classical multifractal analyses on time series.
Notably, it is possible to calculate the so-called multifractal spetrum, which describes all relevant multifractal characteristics of the system under analysis.
MF-DFA technique has been used in a multitude of contexts, such as breast cancer imaging \cite{gerasimova2013multifractal}, river flow modeling \cite{movahed2008fractal}, brain signals \cite{fetterhoff2014multifractal}, solar filaments and sunspots \cite{hu2009multifractal,Wu20151}, financial returns \cite{zhou2009components,meng2012effects}, high-energy physics \cite{mali2015multifractal}, earth's gravity \cite{Telesca2015426}, protein contact networks \cite{mixbionets2}, and protein molecular dynamics \cite{zhou2014fractal}.
Detrending is a very crucial aspect in multifractal analysis of time series \cite{bashan2008comparison,ludescher2011spurious,PhysRevE.71.011104}.
In fact, trends are usually a source of spurious multifractal signatures observed in the analyzed series.
Conventional MF-DFA is able to deal with trends compatible with (higher-order) polynomials.
Nonetheless, trending could appear also as a consequence of periodicities, such as those induced by seasonality for long time series.
To this end, \citet{chianca2005fourier} proposed to perform a preprocessing stage involving a Fourier analysis in order to identify and then remove specific frequencies with large amplitude.
Such a filtering technique demonstrated to provide good results in several different contexts \cite{nagarajan2005minimizing,movahed2008fractal,zhao2011multifractal}, while however criticisms and limitations of the method have been highlighted in the context of sunspot time series \cite{hu2009multifractal}.
In this paper, we study long-term correlations and multifractal properties elaborated from time series of three-phase current signals coming from an electric arc furnace (EAF) \cite{sadeghian1999nonlinear,sadeghian2011dynamic,sadeghian1999application}.
EAF modeling and analysis of related data is an active area of research \cite{1413319,6690260,janabi2009adaptive,samet2014dynamic,samet2012wide,samet2014maximizing}. Previous studies focused mostly on system identification and prediction problems with the aim of power optimization and control.
However, to our knowledge there is no track in the literature of studies employing MF-DFA to characterize EAF signals.
An EAF is a plant utilized in industrial environments for the production of steel by a melting process of metallic scraps.
EAF consists of a cylindrical melting pot covered by a removable vault, inside of which are placed one or more electrodes.
The scraps to be molten (metallic charge) are deposited on the bottom of the melting pot and, by applying AC or DC tension to the electrodes, an electric arc is generated between the extremities of such electrodes and the scraps. The current traversing the metallic charge raises its temperature until the fusion point is reached, initiating the melting process.
The V--I characteristic of the EAF is the result of a nonlinear process with memory \cite{sadeghian2011dynamic}, since the electrical current value is determined as a function of both the voltage and the internal state, which in turn is determined by the past states of the system.
Such a process gives rise to an electrical phenomenon called \textit{hysteresis} \cite{chua1987linear}.
Toward removing all implicit periodicities proper of an alternating current setting, the EAF time series are preprocessed via a Fourier-based analysis in order to remove the first few complex coefficients.
By analyzing the filtered time series we detected long-term, positive correlations. This outcome is in agreement with the typical V--I characteristic of the EAF, which shows a clear hysteretic behavior, as in fact the equivalent dipole of the arc is a dynamic and nonlinear one \cite{chua1987linear}.
The multifractal signature is consistent among the three-phases and it is strong enough in the filtered time series to be effectively classified as multifractal.
Finally, we show that shuffling the series reveals that the multifractal signature is entirely due to the long-term correlations.
The paper is structured as follows.
Section \ref{sec:eaf} provides an introduction to the EAF context.
In Section \ref{sec:results} we discuss our results regarding long-term correlations and multifractal properties of the analyzed three-phase current signals.
Section \ref{sec:conclusions} provides the concluding remarks and future directions.
Finally, in Appendix \ref{sec:mfdfa} we introduce the MF-DFA technique.
\section{Electric arc furnace and the V--I characteristic}
\label{sec:eaf}
The analyzed EAF \cite{sadeghian2011dynamic} is an alternating current furnace powered by a three-phase electrical supply.
The plant contains three electrodes placed on the three vertices of a triangle, supported by vertical guides that allow controlling their distance with respect to the underlying metallic charge.
When the applied voltage is sufficiently high to break the air resistance, an electric arc is struck from the extremity of each electrode to the metallic charge.
In an ideal setting, the electric arc is created each time the absolute value of the alternating voltage exceeds a threshold and, consequently, the current assumes an almost sinusoidal waveform.
However, in real-world situations, the V--I characteristic is subject to many sources of noise, such as arc instabilities, non-monochromaticity of the input signal, and metallic charge cave-ins that hinder the analysis of the measured data.
The latter effect, for example, might cause sudden drops in the current magnitude.
The V--I characteristic of the system forms a closed trajectory in the tension--current phase space (see Fig. \ref{fig:hysteresis}). This behavior denotes a branching process in the joint evolution of those two quantities and it is the distinctive feature of a system subject to hysteresis \cite{chua1987linear}.
This implies that the system possesses memory of the previous states.
As we will show in this study, the hysteresis can be investigated through the analysis of the long-term correlations of the related EAF current time series.
\begin{figure}[h]
\centering
\includegraphics[viewport=0 0 654 456,scale=0.30,keepaspectratio]{./hysteresis}
\caption{Hysteresis of the V--I characteristic of EAF. Image courtesy of \citet{sadeghian2011dynamic}.}
\label{fig:hysteresis}
\end{figure}
\section{Results and discussion}
\label{sec:results}
Our data consists in three current time series, each related to one of the three phases of the EAF under analysis \cite{sadeghian2011dynamic}.
In the following, we will denote the three phases as A, B, and C.
Each time series is composed 1150000 samples, having a sampling frequency of 1920 sample per second. The inter-sample time is hence $520.833\times 10^{-6}$ seconds, which gives a total of roughly $10$ minutes of recording. All data are standardized prior to processing.
Fig. \ref{fig:correlations} shows the autocorrelation \ref{fig:IA__autocorr} and power spectrum \ref{fig:IA__power} for phase A.
Autocorrelation is in agreement with a typical sinusoidal waveform of a current signal (60 Hz).
The power spectrum offers an insight on the fact that such a signal contains many harmonics of significant power, which correspond to strong periodicities in the series.
This is fairly expected, due to the nature of the physical system under analysis.
However, in order to avoid the detection of spurious multifractal signatures, it is necessary to eliminate such periodicities.
The interest in the analysis of such EAF signals lies hence in the non-linearities that are super-imposed during the melting process on the baseline sinusoidal signal.
\begin{figure}[ht!]
\centering
\subfigure[Autocorrelation function.]{
\includegraphics[viewport=0 0 349 243,scale=0.6,keepaspectratio=true]{./IA__ACF}
\label{fig:IA__autocorr}}
~
\subfigure[Power spectrum.]{
\includegraphics[viewport=0 0 343 246,scale=0.6,keepaspectratio=true]{./IA__power}
\label{fig:IA__power}}
\caption{Autocorrelation \ref{fig:IA__autocorr} and power spectrum \ref{fig:IA__power} of current time series of phase A. It is possible to recognize the presence of several harmonics with significant power.}
\label{fig:correlations}
\end{figure}
In Fig. \ref{fig:fluctuations_crossover} we show the scaling of the fluctuations (\ref{eq:Fqshq}) for the three original signals.
A cross-over is visible in Fig. \ref{fig:fluctuations_crossover} at a scale of roughly 32--40 time instants, which corresponds (considering the aforementioned inter-sample time) to a regime of 48--60 Hz. The cross-over is consistent among the three phases, as in fact in the figure the three fluctuation functions overlap almost perfectly.
Therefore, fluctuations, when calculated on the original signals, immediately pick up the main sinusoid of the signal due to the frequency of the electrical current.
In order to remove traces of all periods in the three current signals, we applied Fourier detrending \cite{nagarajan2005minimizing,chianca2005fourier,movahed2008fractal,zhao2011multifractal}. The procedure consists in transforming the signal in the Fourier space with the fast Fourier transform. Complex-valued Fourier coefficients are then order according to their amplitude (from larger to smaller).
In order to obtain a suitable signal to be processed with MF-DFA, we have zeroed the first 3000 coefficients (corresponding to roughly 1.5 seconds) and then antitransformed the signal back to the time domain.
The result of the filtering can be appreciated in Fig. \ref{fig:signals}. In the figure, we show an excerpt of the original signals together with the identified trends (left panels), while in the right panels we report the corresponding filtered versions. It is worth noting that, as expected, the identified trend is not compatible with a single sinusoid, as in fact the trend in the signal is a combination of the main sinusoid with several harmonics of different power (see Fig. \ref{fig:IA__power}).
As we will demonstrate in the following, such filtered signals contain long-term, positive correlations and a fair multifractal signature.
\begin{figure}
\centering
\includegraphics[viewport=0 0 344 244,scale=0.6,keepaspectratio=true]{./crossover}
\caption{(colors online) Scaling of fluctuations of original (standardized) time series. Fluctuations, calculated with a linear fitting and $q=2$, denote a cross-over compatible with the 60 Hz of the electrical current.}
\label{fig:fluctuations_crossover}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 360 246,scale=1,keepaspectratio=true]{./500samples}
\caption{(colors online) First 500 samples of the time series (left panels, black lines indicate the original signal while in red we show the trend) and related filtered version (right panels, blue lines). Phase A, B, and C shown from top to bottom.}
\label{fig:signals}
\end{figure}
Filtered signals are then analyzed with the MF-DFA procedure, which is described in Appendix \ref{sec:mfdfa}.
MF-DFA has been executed by using scales from 16 up to 16384, $q$ ranging in [-5, 5], and with a linear fitting for the local detrending; same outcome has been obtained with a quadratic fitting. Results are reported in Fig. \ref{fig:mfdfa}.
Scaling of fluctuations, shown in Fig. \ref{fig:fluctuations}, provides a confirmation that the filtering stage removed all relevant periodicities in the original signals. In fact, log-log plot of the filtered signals, for all three phases, is now largely consistent with a global power-law scaling.
Let us now focus on Fig. \ref{fig:hurst}, which shows the generalized Hurst exponents (\ref{eq:Fqshq}).
All three signals denote long-term, positive correlations: $H(2)\simeq 0.75$ for phase A, $H(2)\simeq 0.78$ for phase B and C.
In Fig. \ref{fig:localslopes} we show the local slopes of the fluctuation functions for $q=2$. Local slopes of the three signals fluctuate around the estimated Hurst exponents (the global slope of the entire fluctuation function). As typically found in synthetic time series \cite{shao2012comparing}, higher deviations are observed when moving to larger scales.
The discovery of long-term, positive correlations in all three (filtered) electrical current signals is well-justified and it is in agreement with the typical V--I characteristic of EAF (Sec. \ref{sec:eaf}).
In fact, hysteresis is a phenomenon that is observed when the output of a (nonlinear) system does not depend only on its current input, but also on past system states.
Such a memory effect is imprinted in the EAF time series, as highlighted by the generalized Hurst exponents shown in Fig. \ref{fig:hurst}.
Another important fact in multifractal analysis is the distinction between monofractal and multifractal scaling: if the scaling of a system can be fully characterized by a single exponent (Hurst exponent in our case) then it is monofractal, otherwise it is called multifractal.
The analysis of the multifractal spectrum (\ref{eq:mutifractal_spectrum}) shown in Fig. \ref{fig:mfspectra} provides the answer to such a question.
All three signals denote a fair multifractal scaling: the width (\ref{eq:width}) of the spectrum is $\Delta\alpha \simeq 0.30$ for phase A, $\Delta\alpha \simeq 0.21$ for phase B, and $\Delta\alpha\simeq 0.31$ for phase C.
It is worth mentioning for phase A and C a small but yet visible tendency of being insensitive to larger fluctuations, which can be deduced from the fact that the spectra are slightly right sided. The converse holds for phase B, which instead appear as slightly left sided -- as in fact the corresponding Hurst exponents tends to be flat for negative $q$.
As discussed by \citet{drozdz2015detecting}, right sided spectra are a less usual in experimental time series.
The authors \cite{drozdz2015detecting} offer a mechanism for reproducing such an effect by constructing a synthetic time series starting from a binomial cascade and successively super-imposing a white noise signal considering a suitable threshold for editing the original cascade.
However, we believe that in our case such an asymmetry is likely to be a byproduct of the preliminary filtering stage performed via Fourier coefficient truncation.
Our results indicate that all three (filtered) time series are compatible with signals having both long-term, positive correlations and a fair amount of multifractal features.
This suggests that the non-linearities and memory effects of the melting process are super-imposed on the original current signals as noise.
Such a noise retains relevant information regarding the physical process, which we suitably detected by means of MF-DFA.
Multifractality could be observed as a byproduct of the long-range correlations and/or it could emerge as a consequence of an underlying broad probability distribution \cite{kantelhardt2009fractal}. Shuffling the time series provides hence a reliable proof to disambiguate the origins of the observed multifractality. In fact, if the multifractal behavior is entirely due to the long-term correlations, then the shuffled signals would not present any multifractality at all; shuffling the data destroys all long-term correlations. This can be quantitatively verified by observing a relevant shrinking of the multifractal spectrum width.
On the other hand, if multifractality is (also) due to a broad underlying density, then shuffling would have only a visible but yet limited impact on the width of the resulting multifractal spectrum.
As shown in the inset of Fig. \ref{fig:mfspectra}, the shuffled signals do not contain any sign of multifractal signature.
This leads us to claim that the observed multifractal properties of the current time series are entirely due to the long-term correlations -- the memory effects due to the hysteresis. We argue that a broad probability density function, in our case, might not be possible, due to the EAF physical constraints. In fact, EAF operates in an industrial setting that is suitably controlled in order to achieve specific operating conditions.
To summarize, in Table \ref{tab:values} we report all relevant coefficients determined via MF-DFA.
\begin{figure}[ht!]
\centering
\subfigure[Scaling of fluctuations.]{
\includegraphics[viewport=0 0 348 244,scale=0.6,keepaspectratio=true]{./fluctuations}
\label{fig:fluctuations}}
~
\subfigure[Generalized Hurst exponent.]{
\includegraphics[viewport=0 0 345 244,scale=0.6,keepaspectratio=true]{./hurst}
\label{fig:hurst}}
\subfigure[Local slopes of fluctuations for $q=2$.]{
\includegraphics[viewport=0 0 348 242,scale=0.6,keepaspectratio=true]{./localslopes}
\label{fig:localslopes}}
~
\subfigure[Multifractal spectrum.]{
\includegraphics[viewport=0 0 343 242,scale=0.6,keepaspectratio=true]{./mfspectra}
\label{fig:mfspectra}}
\caption{Scaling of fluctuations \ref{fig:fluctuations}, generalized Hurst exponents \ref{fig:hurst}, local slopes \ref{fig:localslopes} of the fluctuation functions for $q=2$, and multifractal spectra \ref{fig:mfspectra} of the filtered time series. In Fig. \ref{fig:fluctuations}, fluctuation functions are shown for all considered $q$ and vertically shifted to improve visualization. Horizontal lines in Fig. \ref{fig:localslopes} denote the respective estimated Hurst exponents of each phase. The inset in Fig. \ref{fig:mfspectra} shows the spectra obtained after shuffling the time series.}
\label{fig:mfdfa}
\end{figure}
\begin{table}[thp!]
\begin{center}
\caption{Summary of all relevant exponents and coefficients found with MF-DFA.}
\begin{tabular}{|c|c|c|c|}
\hline
& \textbf{phase A} & \textbf{phase B} & \textbf{phase C} \\
\hline
$H(2)$ & 0.75 & 0.78 & 0.78 \\
$\tau(2)$ & 0.51 & 0.56 & 0.56 \\
$\Delta\alpha$ & 0.30 & 0.21 & 0.31 \\
$\Delta\alpha$ (shuffled) & 0.04 & 0.06 & 0.05 \\
\hline
\end{tabular}
\end{center}
\label{tab:values}
\end{table}
\section{Conclusions and future directions}
\label{sec:conclusions}
In this paper, we studied the long-term correlations and multifractal properties elaborated from time series of three-phase electric current measured from an electric arc furnace.
Electric arc furnaces are industrial plants used for the production of steel by a melting process of metallic scraps.
The V--I characteristic of the EAF under analysis is the result of a nonlinear process having memory.
In fact, the value of the electrical current is determined as a function of both voltage and internal state, which in turn is determined by considering the history of the states of the system. Such a process gives rise to an electrical phenomenon called hysteresis.
We initially filtered the original signals via a Fourier based detrending, which we have performed in order to remove the implicit sinusoidal trends due to the alternate current.
Our analysis showed that the memory effects of EAF process are suitably encoded into the filtered time series in the form of long-terms, positive correlations, which are quantified by the high values found for the Hurst exponent.
Successively, we have shown the presence of a fair multifractal signature in all three time series, confirming thus the complexity of such noisy signals.
Finally, shuffling of the (filtered) time series demonstrated that the multifractal signature was entirely due to the long-term correlations of the signals.
Future research works will focus on the study of the long-term and multifractal properties of the cross-correlations between current and voltage time series related to the electric arc furnace.
|
\section{Introduction}
Although the LHC has unearthed the standard model (SM) Higgs boson~\cite{Aad:2012tfa,Chatrchyan:2012ufa,Giardino:2013bma}, so far, it has failed to identify any sign of new physics beyond the SM. New physics at the TeV scale is invoked in the form of supersymmetry, extra dimensions {\em etc.\ } to address the issue of the hierarchy problem. Absence of such new physics can imply that the solution of the hierarchy problem lies somewhere else. Hence, it is likely that the SM could be valid up to the Planck scale ($M_{\rm Pl}$), above which gravity is expected to dominate over other forces.
However, we know that the presence of dark matter (DM) and nonzero neutrino masses indicate physics beyond the SM. Hence, it is worthwhile to explore a scenario where the SM, augmented by a low energy DM model, is valid up to $M_{\rm Pl}$. Earlier we had explored such a scenario where a gauge singlet scalar is added to the SM to pose as a weakly interacting massive particle (WIMP) DM~\cite{Khan:2014kba}. In the present work, we extend the SM by a scalar doublet protected by a $\mathbb{Z}_2$ symmetry. The model is popularly known in the literature as the inert doublet~(ID) model, first proposed by Deshpande and Ma~\cite{Deshpande:1977rw}.
Given the experimental measured values for $M_t$, $M_h$, and $\alpha_s$, the electroweak (EW) vacuum is reported to reside in a metastable state within the framework of the SM. This has been verified in the literature~\cite{Bezrukov:2012sa, Degrassi:2012ry, Masina:2012tz, Buttazzo:2013uya, Spencer-Smith:2014woa} only very recently using state of the art next-to-next-to leading order (NNLO) loop corrections contributing to the Higgs effective potential. In Ref.~\cite{Khan:2014kba}, this analysis was extended to the case of a singlet scalar DM model. The stability of the EW vacuum was shown to depend on new physics parameters.
In this paper, we extend such an analysis to the ID model. We assume that ID DM is the only DM particle which saturates the entire DM relic density. In this context, we review the constraints on the parameters of the ID model.
A detailed study on the ID parameter space was recently performed in Refs.~\cite{Dolle:2009fn, Gustafsson:2012aj} indicating bounds from EW stability, perturbativity, collider study, electroweak precision tests (EWPT), etc., considering DM annihilation to two-body final states only. In Refs.~\cite{Honorez:2010re, Gustafsson:2012aj} DM relic density was calculated including three-body final states as well. An updated detailed analysis on the ID parameter space was performed in Refs.~\cite{Goudelis:2013uca, Arhrib:2013ela} reflecting the impact of the $M_h\approx 126$~GeV Higgs boson discovery at the LHC. The authors of Refs.~\cite{Gustafsson:2012aj, Goudelis:2013uca} looked at renormalization group evolution (RGE) of model parameters to check the validity of the ID model at higher energies. The constraints from the measurements of the diphoton decay channel of the Higgs boson on the ID parameter space was also discussed. High scale validity of this model in the presence of a right-handed neutrino has been looked at in Ref.~\cite{Chakrabarty:2015yia}. The influence of ID in the Higgs effective potential was studied in Ref.~\cite{Hambye:2007vf} to explore the possibility of electroweak symmetry breaking {\it \`a la} Coleman-Weinberg~\cite{Coleman:1973jx}. High scale validity of the two Higgs doublet model has been explored in Refs.~\cite{Chakrabarty:2014aya, Das:2015mwa} with broken $\mathbb{Z}_2$, symmetry considering tree level potential only.
In this paper we improve earlier studies on ID model parameter space by including radiative corrections to the scalar potential to explore the stability of the electroweak vacuum. In particular, we permit the Higgs quartic coupling to assume small negative values to render a metastable EW vacuum. As our analysis improves the scalar potential considering radiative corrections, the tree level stability constraints used in earlier analyses need to be reviewed. In addition, as we demand that the theory be valid up to $M_{\rm Pl}$, the model parameters at the EW scale are constrained by the requirement that they satisfy all the bounds up to the $M_{\rm Pl}$. If in the near future the direct detection experiments and colliders confirm the ID model as the dark matter model realized in nature, then our study will help to estimate the lifetime of the electroweak vacuum, especially if it still remains in the metastable state. As metastability of EW vacuum is considered to be an indication of new physics, if the vacuum is metastable even after consideration of the ID model, then that could indicate the presence of additional new physics before $M_{\rm Pl}$.
The paper is organized as follows. Section~\ref{sec:IDM} starts with an introduction of the ID model, followed by a discussion of the effective scalar potential and renormalization group (RG) running of parameters of the model. Constraints on the model are discussed in Sec.~\ref{sec:constraints}. Tunneling probability of the metastable vacuum and the restrictions on the parameters to avoid a potential unbounded from below, are mentioned in Sec.~\ref{sec:metastability}. A detailed study of the parameter space identifying regions of EW vacuum stability and metastability is performed in Sec.~\ref{sec:phasediag} with the help of various phase diagrams. Section~\ref{sec:veltman} contains a short discussion on Veltman's conditions in the context of the ID model. We finally conclude in Sec.~\ref{sec:conclusion}.
\section{Inert Doublet Model}
\label{sec:IDM}
In this model, the standard model is extended by adding an extra $SU(2)$ doublet scalar, odd under an additional discrete $\mathbb{Z}_2$ symmetry. Under this symmetry, all standard model fields are even. The $\mathbb{Z}_2$ symmetry prohibits the inert doublet to acquire a vacuum expectation value.
The scalar potential at the tree level is given by
\begin{eqnarray}
V(\Phi_1,\Phi_2) &=& \mu_1^2 |\Phi_1|^2 + \lambda_1 |\Phi_1|^4+ \mu_2^2 |\Phi_2|^2 + \lambda_2 |\Phi_2|^4\nonumber\\
&&+ \lambda_3 |\Phi_1|^2 |\Phi_2|^2
+ \lambda_4 |\Phi_1^\dagger \Phi_2|^2 + \frac{\lambda_5}{2} \left[ (\Phi_1^\dagger \Phi_2)^2 + {\rm H.c.}\right] \, ,
\label{Scalarpot}
\end{eqnarray}
where the SM Higgs doublet $\Phi_1$ and the inert doublet $\Phi_2$ are given by
\begin{equation}
\Phi_1 ~=~ \left( \begin{array}{c} G^+ \\ \frac{1}{\sqrt{2}}\left(v+h+i G^0\right) \end{array} \right),
\qquad
\Phi_2 ~=~ \left( \begin{array}{c} H^+\\ \frac{1}{\sqrt{2}}\left(H+i A\right) \end{array} \right) \, \nonumber.
\end{equation}
$G^\pm$ and $G^0$ are Goldstone bosons and $h$ is the SM Higgs.
$\Phi_2$ contains a $CP$ even neutral scalar $H$, a $CP$ odd neutral scalar $A$, and a pair of charged scalar fields $H^\pm$. The $\mathbb{Z}_2$ symmetry prohibits these particles to decay entirely to SM particles. The lightest of $H$ and $A$ can then serve as a DM candidate.
After electroweak symmetry breaking, the scalar potential is given by
\begin{eqnarray}
V(h, H,A,H^\pm) &=& \frac{1}{4} \left[ 2 \mu_1^2 (h+v)^2 + \lambda_1 (h+v)^4 +2 \mu_2^2 (A^2+H^2+2 H^+ H^-) \right. \nonumber \\
&& \left. + \lambda_2 (A^2 + H^2 + 2 H^+ H^-)^2 \right] \nonumber \\
&& + \frac{1}{2} (h+v)^2 \left[ \lambda_3 H^+ H^-
+ \lambda_S A^2
+ \lambda_L H^2 \right] \label{Scalarpot2}
\end{eqnarray}
where
\begin{eqnarray}
\lambda_{L,S}&=&\frac{1}{2}\left(\lambda_3+\lambda_4\pm\lambda_5\right) \, .
\end{eqnarray}
Masses of these scalars are given by
\begin{align}
M_{h}^2 &= \mu_1^2 + 3 \lambda_1 v^2,\nonumber \\
M_{H}^2 &= \mu_2^2 + \lambda_L v^2, \nonumber \\
M_{A}^2 &= \mu_2^2 + \lambda_S v^2,\nonumber \\
M_{H^\pm}^2 &= \mu_2^2 + \frac{1}{2} \lambda_3 v^2 \,\nonumber .
\end{align}
For $\lambda_4-\lambda_5<0$ and $\lambda_5>0$~($\lambda_4+\lambda_5<0$ and $\lambda_5<0$),
$A$~($H$) is the lightest $\mathbb{Z}_2$ odd particle (LOP). In this work, we take $A$ as the LOP and hence, as a viable DM candidate. Choice of $H$ as LOP will lead to similar results.
As we will explain later, for large DM mass, $M_A\gg M_Z$, appropriate relic density of DM is obtained if $M_A$, $M_H$, and $M_{H^\pm}$ are nearly degenerate. Hence, in anticipation, we define
\begin{eqnarray}
\Delta M_H&=& M_H-M_A, \nonumber\\
\Delta M_{H^\pm}&=&M_{H^\pm}-M_A \nonumber\, .
\end{eqnarray}
so that the new independent parameters for the ID model become $\{M_A, \Delta M_H, \Delta M_{H^\pm}, \lambda_2, \lambda_S\}$. Here we have chosen $\lambda_S$ as we treat $A$ as the DM particle.
The one-loop effective potential for $h$ in the $\overline{\mbox{\sc ms}}$ scheme and the Landau gauge is given by
\begin{equation}
V_1^{{\rm SM}+{\rm ID}}(h)= V_1^{\rm SM}(h) + V_1^{\rm ID}(h)
\end{equation}
where
\begin{equation}
V_1^{\rm SM}(h)=\sum_{i=1}^5 \frac{n_i}{64 \pi^2} M_i^4(h) \left[ \ln\frac{M_i^2(h)}{\mu^2(t)}-c_i\right] \, .
\end{equation}
$n_i$ is the number of degrees of freedom. For scalars and gauge bosons, $n_i$ is positive, whereas for fermions it is negative. Here $c_{h,G,f}=3/2$, $c_{W,Z}=5/6$, and the running energy scale $\mu$ is expressed in terms of a dimensionless parameter $t$ as $\mu(t)=M_Z \exp(t)$. $M_i$ is given as
\begin{equation}
M_i^2(h)= \kappa_i(t)\, h^2(t)-\kappa_i^{\prime}(t) \, .
\nonumber \end{equation}
$n_i$, $\kappa_i$ and $\kappa_i^{\prime}$ can be found in Eq.~(4) in Ref.~\cite{Casas:1994qy} (see also Refs.~\cite{Altarelli:1994rb, Casas:1994us, Casas:1996aq, Quiros:1997vk}).
The additional contribution to the one-loop effective potential due to the inert doublet is given by~\cite{Hambye:2007vf}
\begin{equation}
V_1^{\rm ID}(h)= \sum_{j=H,A,H^+,H^-} \frac{1}{64 \pi^2} M_j^4(h) \left[ \ln\left(\frac{M_j^2(h)}{\mu^2(t)} \right)- \frac{3}{2} \right]
\end{equation}
where
\begin{equation}
M_j^2(h)=\frac{1}{2} \,\lambda_{j}(t) \, h^2(t)+\mu_2^2(t)
\end{equation}
with $\lambda_{A}(t)=2 \lambda_{S}(t)$, $\lambda_{H}(t)=2 \lambda_{L}(t)$, and $\lambda_{H^\pm}(t)= \lambda_{3}(t)$.
In the present work, in the Higgs effective potential, SM contributions are taken at the two-loop level~\cite{Degrassi:2012ry, Buttazzo:2013uya,Ford:1992pn,Martin:2001vx}, whereas the ID scalar contributions are considered at one loop only.
For $h \gg v$, the Higgs effective potential can be approximated as
\begin{equation}
V_{\rm eff}^{{\rm SM}+{\rm ID}}(h) \simeq \lambda_{\rm 1,eff}(h) \frac{h^4}{4}\, ,
\label{efflam}\end{equation}
with
\begin{equation}
\lambda_{\rm 1,eff}(h) = \lambda_{\rm 1,eff}^{\rm SM}(h) +\lambda_{\rm 1,eff}^{\rm ID}(h)\, ,
\end{equation}
where~\cite{Buttazzo:2013uya}
\begin{eqnarray}
\lambda_{\rm 1,eff}^{\rm SM}(h) &=& e^{4\Gamma(h)}
\left[ \lambda_1(\mu=h) + \lambda_{\rm 1,eff}^{(1)}(\mu=h) + \lambda_{\rm 1,eff}^{(2)}(\mu=h)\right] \nonumber\\
\lambda_{\rm 1,eff}^{\rm ID}(h)&=&\sum_{j=L,S,3} e^{4\Gamma(h)} \left[\frac{\delta_j \lambda_j^2}{64 \pi^2} \left(\ln\left(\delta_j\lambda_j\right)-\frac{3}{2}\right ) \right]\, . \nonumber
\end{eqnarray}
Here $\delta_j=1$ when $j=L,S$; $\delta_j=\frac{1}{2}$ for $j=3$; and
\begin{equation}
\Gamma(h)=\int_{M_t}^{h} \gamma(\mu)\,d\,\ln\mu \, .\nonumber
\end{equation}
Anomalous dimension $\gamma(\mu)$ of the Higgs field takes care of its wave function renormalization.
As quartic scalar interactions do not contribute to wave function renormalization at the one-loop level, ID does not alter $\gamma(\mu)$ of the SM. The expressions for the one- and two-loop quantum corrections $\lambda_{\rm 1,eff}^{(1,2)}$ in the SM can be found in Ref.~\cite{Buttazzo:2013uya}. All running coupling constants are evaluated at $\mu=h$. This choice guarantees that the perturbative expansion of the potential is more reliable~\cite{Ford:1992mv}, which can be understood as follows. The perturbative expansion of the potential at high field values of $h$ consists of terms of the form $\epsilon\,\ln(h/\mu)$, where $\epsilon$ is some dimensionless coupling. Clearly the perturbation series works better if we choose the renormalization scale $\mu \sim h$ and ensure that $\epsilon$ is small as well. Hence, our choice of $\mu=h$, along with the requirement that all the couplings remain within the perturbative domain, ensures faster convergence of the perturbation series of the potential.
To compute the RG evolution of all the couplings, we first calculate all the SM couplings at $M_t$, taking care of the threshold corrections as in Refs.~\cite{Sirlin:1985ux,Bezrukov:2012sa,Degrassi:2012ry,Khan:2014kba}. Then we evolve them up to $M_{\rm Pl}$ using our own computer codes incorporating the RG equations~\cite{Chetyrkin:2012rz,Zoller:2012cv,Chetyrkin:2013wya,Zoller:2013mra}. The corresponding $\beta$-functions include three-loop SM effects and two-loop ID contributions (one-loop RGEs are presented in the Appendix for completeness). If DM mass $M_A$ is larger than $M_t$, then the ID starts to contribute after the energy scale $M_A$. For $M_A<M_t$, the contributions of ID to the $\beta$-functions are rather negligible for the running from $M_A$ to $M_t$, as is evident from the expressions.
To help the reader in reproducing our results, we provide in Table~\ref{table1} a specific set of values of $\lambda_i$ at $M_t=173.1$~GeV and at $M_{\rm Pl} = 1.2 \times 10^{19}$~GeV for $M_h=125.7$~GeV and $\alpha_s\left(M_Z\right)=0.1184$. In Fig.~\ref{fig:SMInert} we explicitly show running of the scalar couplings $(\lambda_i)$ for this set of parameters. We see that for this specific choice of parameters, $\lambda_1$ assumes a small negative value\footnote{$\lambda_1$ and $\lambda_{\rm 1,eff}$, being almost equal at $M_{\rm Pl}$, are interchangeable at that scale.} leading to a metastable EW vacuum as discussed in the following sections. This set is chosen to reproduce the DM relic density in the right ballpark.
\begin{table}[h!]
\begin{center}
\begin{tabular}{| c | c | c | c | c | c | c | c |}
\hline
&$\lambda_S$ & $\lambda_L$ & $\lambda_2$ & $\lambda_3$ & $\lambda_4$ & $\lambda_5$ & $\lambda_1$\\
\hline
~$M_{t}$ ~&~ 0.001 ~&~ 0.039 ~&~ 0.10 ~&~ 0.0399 ~&~ 0.00003 ~&~ 0.038 ~&~ 0.127~\\
\hline
$M_{\rm Pl}$ & 0.046 & 0.082 & 0.127 & 0.090 & 0.038 & 0.036 & $-0.009$\\
\hline
\end{tabular}
\caption{A set of values of all ID model coupling constants at $M_t$ and $M_{\rm Pl}$ for $M_{A}=573$ GeV, $\Delta M_{H^\pm}=1$ GeV, $\Delta M_{H}=2$ GeV, and $\lambda_S\left(M_Z\right)=0.001$.}
\label{table1}
\end{center}
\end{table}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=3.7in,height=3.1in, angle=0]{couplingvsenergy}
\caption{\label{fig:SMInert} \text{RG evolution of the couplings $\lambda_i\left(i=1,..,5\right), \lambda_L, \lambda_S$ for the set of parameters in Table~\ref{table1}. } }
\end{center}
\end{figure}
\newpage
\section{Constraints on ID model: A brief review}
\label{sec:constraints}
ID model parameter space is constrained from theoretical considerations like absolute vacuum stability, perturbativity and unitarity of the scattering matrix. Electroweak precision measurements and direct search limits at LEP put severe restrictions on the model. The recent measurements of Higgs decay width at the LHC put additional constraints. The requirement that the ID DM saturates the DM relic density all alone restricts the allowed parameter space considerably. Although these bounds are already discussed in the literature~\cite{LopezHonorez:2006gr}, here we apply these bounds with the requirement that the model needs to be valid till $M_{\rm Pl}$.
\subsection{Vacuum stability bounds}
The tree level scalar potential potential $V(\Phi_1,\Phi_2)$ is stable and bounded from below if~\cite{Deshpande:1977rw}
\begin{equation}
\lambda_{1,2}(\Lambda) \geq 0, \quad \lambda_{3}(\Lambda) \geq -2 \sqrt{ \lambda_{1}(\Lambda)\lambda_{2}(\Lambda)}, \quad \lambda_{L,S}(\Lambda) \geq - \sqrt{ \lambda_{1}(\Lambda)\lambda_{2}(\Lambda)} \label{stabilitybound}
\end{equation}
where the coupling constants are evaluated at a scale $\Lambda$ using RG equations. However, these conditions become nonfunctional if $\lambda_1$ becomes negative at some energy scale to render the EW vacuum metastable. Under such circumstances we need to handle metastability constraints on the potential differently, which we pursue in the next section.
\subsection{Perturbativity bounds}
The radiatively improved scalar potential remains perturbative by requiring that all quartic couplings of $V(\Phi_1,\Phi_2)$ satisfy
\begin{equation}
\mid \lambda_{1,2,3,4,5}\mid \leq 4 \pi
\end{equation}
\subsection{Unitarity bounds}
Unitarity bounds on $\lambda_i$ are obtained considering scalar-scalar, gauge boson-gauge boson, and scalar-gauge boson scatterings~\cite{Lee:1977eg}. The constraints come from the eigenvalues of the corresponding S-matrix~\cite{Arhrib:2012ia}:
\begin{eqnarray}
\vert \lambda_3 \pm \lambda_4\vert\leq 8\pi , ~~~~~\vert\lambda_3 \pm \lambda_5\vert\leq 8\pi\nonumber\\
\vert \lambda_3+ 2 \lambda_4 \pm 3\lambda_5\vert\leq 8 \pi\nonumber\\
\vert-\lambda_1 - \lambda_2 \pm \sqrt{(\lambda_1 - \lambda_2)^2 + \lambda_4^2}\vert\leq 8 \pi
\\
\vert-3\lambda_1 - 3\lambda_2 \pm \sqrt{9(\lambda_1 - \lambda_2)^2 + (2\lambda_3 + \lambda_4)^2}\vert\leq 8 \pi
\nonumber\\
\vert -\lambda_1 - \lambda_2 \pm \sqrt{(\lambda_1 - \lambda_2)^2 + \lambda_5^2}\vert\leq 8 \pi\nonumber
\label{unitary}
\end{eqnarray}
\subsection{Bounds from electroweak precision experiments}
Bounds ensuing from electroweak precision experiments are imposed on new physics models {\it via} Peskin-Takeuchi~\cite{Peskin:1991sw} $S, T, U$ parameters. The additional contributions from ID are given by~\cite{Barbieri:2006dq,Arhrib:2012ia}
\begin{equation}
\Delta S = \frac{1}{2\pi}\Bigg[ \frac{1}{6}\ln\left(\frac{M^2_{H}}{M^2_{H^\pm}}\right) -
\frac{5}{36} + \frac{M^2_{H} M^2_{A}}{3(M^2_{A}-M^2_{H})^2} +
\frac{M^4_{A} (M^2_{A}-3M^2_{H})}{6(M^2_{A}-M^2_{H})^3} \ln \left(\frac{M^2_{A}}{M^2_{H}}\right)\Bigg]
\label{Sparam}
\end{equation}
and
\begin{eqnarray}
\Delta T &=& \frac{1}{32\pi^2 \alpha v^2}\Bigg[ F\left(M^2_{H^\pm}, M^2_{A}\right)
+ F\left(M^2_{H^\pm}, M^2_{H}\right) - F\left(M^2_{A}, M^2_{H}\right)\Bigg]
\label{Tparam}
\end{eqnarray}
where the function $F$ is given by
\begin{equation}
F(x,y) = \left\{ \begin{array}{lr}
\frac{x+y}{2} - \frac{xy}{x-y}\ln\left(\frac{x}{y}\right), & x\neq y \\
0, & \, x=y
\end{array} \right. \, .
\end{equation}
We use the NNLO global electroweak fit results obtained by the Gfitter group~\cite{Baak:2014ora},
\begin{equation}
\Delta S = 0.06 \pm 0.09, ~ ~ \Delta T = 0.1 \pm 0.07
\label{STU1}
\end{equation}
with a correlation coefficient of $+0.91$, fixing $\Delta U$ to zero. We use this fit result as the contribution of the scalars in the ID model to $\Delta U$ is rather negligible.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=2.8in,height=2.8in, angle=0]{STUCheckplotnew11}
\caption{\label{fig:STUcheck} \textit{\rm{Allowed parameter space in $\Delta M_{H^\pm}-\Delta M_{H}$ plane for $M_A=70$~GeV and $\lambda_S=0.007$. Constraints from $S$ and $T$ parameters are shown by solid black, green and red lines. The blue band corresponds the 3$\sigma$ variation in $\Omega h^2=0.1198\pm0.0026$~\cite{Ade:2013zuv}. On the brown region the unitarity bound is violated on or before $M_{\rm Pl}$. The cross-hatched region is excluded from LEP\,II data. }} }
\end{center}
\end{figure}
To assess the implications of $S$ and $T$ constraints on the ID model, in Fig.~\ref{fig:STUcheck}, in the $\Delta M_{H^\pm}-\Delta M_{H}$ plane, we display various constraints for $M_A=70$~GeV and $\lambda_S=0.007$. This is the maximum value of $\lambda_S$ for the given DM mass, allowed by LUX~\cite{Akerib:2013tjd} direct detection data at 1$\sigma$. The blue region allowed by relic density constraints shifts upwards for smaller $\lambda_S$. Constraints on $\Delta S$ and $\Delta T$, as mentioned in Eq.~(\ref{STU1}), are marked as black, green, and red solid lines. We see that the $1\sigma$ bound on $\Delta T$ is the most stringent one. The black line corresponding to the lower limit on $\Delta S$ at $1\sigma$ can also be interesting. But the LEP\,II bounds, represented by the cross-hatched bar, take away a considerable part. The line representing the $\Delta S$ upper limit is beyond the region considered and lies towards the bottom-right corner of the plot. In this plot, increasing $M_H$ enhances $S$, but $T$ gets reduced. In addition, if we demand unitarity constraints to be respected up to $M_{\rm Pl}$, assuming no other new physics shows up in between, the parameter space gets severely restricted. In this plot, a small window is allowed by $\Delta T$ only at $2\sigma$, which satisfies DM relic density constraints. As mentioned earlier, this window gets further reduced with smaller $\lambda_S$ (the relic density band takes an ``L'' shape as shown in Fig.~\ref{fig:STU2}).
\subsection{Direct search limits from LEP}
The decays $Z\rightarrow AH$, $Z\rightarrow H^+ H^-$, $W^\pm\rightarrow A H^\pm$, and $W^\pm\rightarrow H H^\pm$ are restricted from $Z$ and $W^\pm$ decay widths at LEP. It implies $M_{A} + M_{H} \geq M_{Z}$, $2M_{H^\pm}\geq M_{Z}$, and $M_{H^\pm} + M_{H,A} \geq M_{W}$. More constraints on the ID model can be extracted from chargino~\cite{Pierce:2007ut} and neutralino~\cite{Lundstrom:2008ai} search results at LEP\,II: The charged Higgs mass $M_{H^\pm}\geq 70$~GeV. The bound on $M_A$ is rather involved: If $M_A<80$ GeV, then $\Delta M_H$ should be less than $\sim 8$ GeV, or else $M_H$ should be greater than $\sim 110$~GeV ({see Fig.~\ref{fig:STU1}}).
\subsection{Bounds from LHC}
In the ID model, Higgs to diphoton signal strength $\mu_{\gamma\gamma}$ is defined as
\begin{equation}
\mu_{\gamma\gamma} = \frac{\sigma(gg\rightarrow h\rightarrow\gamma\gamma)}{\sigma(gg\rightarrow h\rightarrow\gamma\gamma)_{\rm SM}}\approx \frac{Br(h \rightarrow {\gamma\gamma})_{\rm ID}}{Br(h \rightarrow {\gamma\gamma})_{\rm SM}}
\end{equation}
using the narrow width approximation for the production cross section of $\sigma(gg\rightarrow h\rightarrow \gamma\gamma)$ and the fact that $\sigma(gg\rightarrow h)$ in both the SM and ID are the same.
Now if the ID particles have masses less than $M_h/2$, $h\rightarrow \rm ID,ID$ decays are allowed. In that case,
\begin{equation}
\mu_{\gamma\gamma}= \frac{\Gamma(h\rightarrow \gamma\gamma)_{\rm ID}} {\Gamma(h\rightarrow \gamma\gamma)_{\rm SM}} ~ \frac{\Gamma_{\rm tot}(h\rightarrow \rm SM,SM)}{\Gamma_{\rm tot}(h\rightarrow \rm SM,SM)+\Gamma_{\rm tot}(h\rightarrow \rm ID,ID)}\, ,
\label{mugagalow}
\end{equation}
where~\cite{Cao:2007rm}
\begin{equation}
\Gamma\left(h\rightarrow \rm ID,ID\right)= \frac{ v^2}{16\pi M_h }\lambda_{\rm ID}^2 \left(1-\frac{4 M_{\rm ID}^2}{M_{h}^2}\right)^{1/2}\, ,
\label{decaywidth}
\end{equation}
where for ${\rm ID}=A,H,H^\pm$, $\lambda_{\rm ID}=\lambda_S, \lambda_L, \sqrt{2}\,\lambda_3$.
In this case, the ID particles are heavier than $M_h/2$,
\begin{equation}
\mu_{\gamma\gamma}= \frac{\Gamma(h\rightarrow \gamma\gamma)_{\rm ID}}{\Gamma(h\rightarrow \gamma\gamma)_{\rm SM}}\, .
\label{mugagahigh}
\end{equation}
In the ID model, the $H^\pm$ gives additional contributions at one loop.
The analytical expression is given by~\cite{Djouadi:2005gj, Swiezewska:2012eh,Krawczyk:2013jta}
\begin{equation}
\Gamma(h\rightarrow \gamma\gamma)_{\rm ID}={\alpha^2 m_h^3\over 256\pi^3
v^2}\left|\sum_{f}N^c_fQ_f^2y_f
F_{1/2}(\tau_f)+ y_WF_1(\tau_W)
+Q_{H^{\pm}}^2{v\mu_{{hH^+H^-}}\over
2m_{H^{\pm}}^2}F_0(\tau_{H^{\pm}})\right|^2\
\label{hgaga}
\end{equation}
where $\tau_i=m_h^2/4m_i^2$. $Q_{f}$, $Q_{H^{\pm}}$ denote electric charges of corresponding particles. $N_f^c$ is the color factor. $y_f$ and
$y_W$ denote Higgs couplings to $f\bar{f}$ and $WW$. $\mu_{hH^+H^-}= \lambda_3 v$ stands for the coupling constant of the $hH^+H^-$ vertex. The loop functions $F_{(0,\,1/2,\,1)}$ are defined as
\begin{align}
F_{0}(\tau)&=-[\tau-f(\tau)]\tau^{-2}\, ,\nonumber\\
F_{1/2}(\tau)&=2[\tau+(\tau-1)f(\tau)]\tau^{-2}\, ,\nonumber\\
F_{1}(\tau)&=-[2\tau^2+3\tau+3(2\tau-1)f(\tau)]\tau^{-2}\, , \nonumber
\label{loopfn}
\end{align}
where
\begin{equation}
f(\tau)= \bigg\{\begin{array}{ll}
(\sin^{-1}\sqrt{\tau})^2\,,\hspace{60pt}& \tau\leq 1\\
-{1\over4}[\ln{1+\sqrt{1-\tau^{-1}}\over1-\sqrt{1-\tau^{-1}}}-i\pi]^2\,,\quad
&\tau>1
\label{ftau}
\end{array}\;,\;
\end{equation}
From the diphoton decay channel of the Higgs at the LHC, the measured values are $\mu_{\gamma\gamma}$=$1.17\pm0.27$ from ATLAS~\cite{Aad:2014eha} and $\mu_{\gamma\gamma}$=$1.14^{+0.26}_{-0.23}$ from CMS~\cite{Khachatryan:2014ira}.
One can see that a positive $\lambda_3$ leads to a destructive interference between SM and ID contributions in En.~(\ref{hgaga}) and {\it vice versa}. Hence, for ID particles heavier than $M_h/2$, $\mu_{\gamma\gamma}<1$ ($\mu_{\gamma\gamma}>1$) when $\lambda_3$ is positive (negative). However, if these ID particles happen to be lighter than $M_h/2$, they might contribute to the invisible decay of the Higgs boson. Using the global fit result~\cite{Belanger:2013xza} that such an invisible branching ratio is less than $\sim 20$\%, in Eq.~(\ref{mugagalow}), the second ratio provides a suppression of $\sim 0.8$ -- $1$.
Now can we work with a negative $\lambda_3$ in the ID model? We will discuss this at the end of this section. For the benchmark points used in this paper, we have worked only with positive values of $\lambda_3$, allowed at 1$\sigma$ by both CMS and ATLAS experiments. It is important to note that the constraints from $\mu_{\gamma\gamma}$ pass through the sign of $\lambda_3$ in our vacuum stability considerations.
\subsection{Constraints from dark matter relic density and direct search limits}
The ID dark matter candidate $A$ can self-annihilate into SM fermions. Once the DM mass is greater than the $W$-mass, so that the DM can annihilate into a pair of $W$ bosons, the cross section increases significantly, thereby reducing DM relic density. Hence, it becomes difficult to saturate $\Omega h^2$ after $\sim 75$~GeV with positive $\lambda_S$, although for $M_A<75$~GeV, both signs of $\lambda_S$ can be allowed to arrive at the right DM relic density $\Omega h^2$.
The role of the sign of $\lambda_S$ can be understood from the contributing diagrams to the $AA\rightarrow W^+ W^-$ annihilation processes. Four diagrams contribute: the $AAW^+ W^-$ vertex driven point interaction diagram (henceforth referred to as the $p$-channel diagram), $H^+$-mediated $t$- and $u$-channel diagrams, and the $h$-mediated $s$-channel diagram. For $AA\rightarrow ZZ$ annihilation, the $t$- and $u$-channel diagrams are mediated by $H$. A negative $\lambda_S$ induces a destructive interference between the $s$-channel diagram with the rest, thereby suppressing $AA\rightarrow W^+ W^-, ZZ$ processes. For DM masses of $75-100$~GeV, this can be used to get the appropriate $\Omega h^2$~\cite{Gustafsson:2012aj, LopezHonorez:2010tb}. To avoid large contributions from $t$- and $u$-channel diagrams and coannihilation diagrams, the $M_H$ and $M_{H^\pm}$ can be pushed to be rather large $\gtrsim 500$~GeV. However, to partially compensate the remaining $p$-channel diagram by the $s$-channel one, $\lambda_S$ assumes a large negative value $\sim -0.1$, which is ruled out by the DM direct detection experiments. That is why in the ID model, DM can be realized below 75 GeV, a regime we designate as the ``low'' DM mass region.
At ``high'' DM mass $M_A \gtrsim 500$~GeV, one can get the right $\Omega h^2$ due to a partial cancellation between different diagrams contributing to the $AA\rightarrow W^+ W^-$ and $AA\rightarrow ZZ$ annihilation processes. For example, in $AA\rightarrow W^+ W^-$, the $p$-channel diagram tends to cancel with the $H^+$-mediated $t$- and $u$-channel diagrams~\cite{Hambye:2009pw, Goudelis:2013uca} in the limit $M_A\gg M_W$, and the sum of amplitudes of these diagrams in this limit is proportional to $M_{H^\pm}^2-M_A^2$. Hence, at high $M_A$, a partial cancellation between these diagrams is expected for nearly degenerate $M_{H^\pm}$ and $M_A$. Similarly, for $AA\rightarrow ZZ$, a cancellation is possible when the masses $M_{H}$ and $M_A$ are close by. For $M_A \gtrsim 500$~GeV, keeping the mass differences of $M_{H^\pm}$ and $M_{H}$ with $M_A$ within $8$~GeV, such cancellations help reproduce the correct $\Omega h^2$. It is nevertheless worth mentioning that such nearly degenerate masses will lead to coannihilation of these $\mathbb{Z}_2$ odd ID scalars~\cite{griest} to SM particles. Despite such near degeneracy, both $H$ and $H^+$, being charged under the same $\mathbb{Z}_2$ as $A$, decay promptly to the LOP $A$, so that they do not become relics. We use {\tt FeynRules}~\cite{Alloul:2013bka} along with {\tt micrOMEGAs}~\cite{Belanger:2010gh, Belanger:2013oya} to compute the relic density of $A$.
DM direct detection experiments involve the $h$-mediated $t$-channel process $AN\rightarrow AN$ with a cross section proportional to $\lambda_S^2/M_A^2$ in the limit $M_A\gg M_N$:
\begin{equation}
\sigma_{A,N} = \frac{m_r^2}{\pi} f^2 m_N^2\left(\frac{\lambda_S}{M_{A} M_h^2}\right)^2 \label{directcs}
\end{equation}
where $f\approx0.3$ is the form factor of the nucleus. $m_r$ represents the reduced mass of the nucleus and the scattered dark matter particle.
Thus, $\lambda_S$ is constrained from nonobservation of DM signals at XENON\,100~\cite{Aprile:2011hi,Aprile:2012nq} and LUX~\cite{Akerib:2013tjd}. For $M_{A}= 70$~GeV, the ensuing bound from LUX~\cite{Akerib:2013tjd,dmtools} data at 1$\sigma$ is $|\lambda_S|< 0.007$.
The constraint on $\lambda_S$ from DM direct detection experiments gets diluted with $M_A$ [see Eq.~(\ref{directcs})]. Hence, for low DM mass, direct detection bounds are more effective. At high mass, the relic density constraints are likely to supersede these bounds. For example, for $M_{DM}= 573$~GeV, the upper limit on $|\lambda_S|$ is 0.138 from LUX. However, to satisfy the relic density constraints from the combined data of WMAP and Planck within 3$\sigma$, $\lambda_S$ can be as large as $0.07$ only.
Within the framework of the ID model, it is possible to explain the observations in various indirect DM detection experiments~\cite{Arhrib:2013ela, Modak:2015uda} for some regions of the parameter space. In this paper, however, we do not delve into such details as such estimations involve proper understanding of the astrophysical backgrounds and an assumption of the DM halo profile which contain some arbitrariness. For a review of constraints on the ID model from astrophysical considerations see, for example, Ref.~\cite{Gustafsson:2010zz}.
\vskip 30pt
\noindent{\bf\underline {Sign of $\lambda_3$}}\\
Whether $\lambda_3$ can be taken as positive or negative depends on the following:
\begin{itemize}
\item
If the ID model is not the answer to the DM puzzle, so that both relic density and direct detection constraints can be evaded, no restriction exists on the possible sign of $\lambda_3(M_Z)$. Otherwise, the following two cases need be considered:
\begin{itemize}
\item
A negative $\lambda_3(M_Z)$ implies
\begin{equation}
\lambda_S(M_Z) < -\frac{1}{v^2} (M_{H^\pm}^2-M_A^2) \nonumber\, .
\end{equation}
As we are considering $A$ as the DM candidate, so that $M_A<M_{H^\pm}$, $\lambda_S(M_Z)$ is always negative when $\lambda_3(M_Z)<0$ .
For low DM mass, the splitting $(M_{H^\pm}-M_A)\gtrsim 10$~GeV, as otherwise DM coannihilation processes cause an inappreciable depletion in $\Omega h^2$. For $M_A=70$~GeV, this implies a lower bound $\lambda_S(M_Z)\lesssim -0.025$, which violates the DM direct detection bound $|\lambda_S|< 0.007$. Hence, for low DM mass, a negative $\lambda_3(M_Z)$ is not feasible.
\item
For high mass DM, the right relic density can be obtained when the splitting $(M_{H^\pm}-M_A)\sim$ a few GeV or less. The above logic then implies that a negative $\lambda_3$ does not put any severe restriction on $\lambda_S$ to contradict DM direct detection bounds as earlier. Hence, for a high DM mass, $\lambda_3(M_Z)$ can assume both the signs. Moreover, due to propagator suppression for large $M_{H^\pm}$ in the $h\gamma\gamma$ vertex, the ID contribution to $\mu_{\gamma\gamma}$ is negligibly small and hence, the sign of $\lambda_3(M_Z)$ is not constrained by measurements on $\mu_{\gamma\gamma}$ as well.
\end{itemize}
\item
If at any scale, $\lambda_3$ is negative while $\lambda_1>0$, then the bound~(\ref{stabilitybound}) must be respected.
\item
If at some scale, $\lambda_1<0$, then a negative $\lambda_3$ makes the potential unbounded from below, as mentioned in the following section. This means one can start with a negative $\lambda_3(M_Z)$, but with RG evolution when $\lambda_1$ turns negative, $\lambda_3$ evaluated at that scale must be positive. Such parameter space does exist. Here, we note a significant deviation of our analysis from earlier analyses which did not allow a negative $\lambda_1$. For example, in Ref.~\cite{Goudelis:2013uca} a negative $\lambda_3(M_Z)$ was not allowed from stability of the Higgs potential if the theory has to be valid up to $10^{16}$~GeV together with relic density considerations.
\end{itemize}
\section{Tunneling Probability and Metastability}
\label{sec:metastability}
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{Inert_Tunnelingplots573}}
\hskip 15pt
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{tunnelinglamSplots}}
\caption{\label{fig:Tun} \textit{ \rm{(a) Tunneling probability ${\cal P}_0$ dependence on $M_t$. The left band (between dashed lines) corresponds to {\rm SM}. The right one (between dotted lines) is for $ID$ model for DM mass $M_A=573$~GeV. Constraints from WMAP and Planck measured relic density, as well as XENON\,100 and LUX DM direct detection null results are respected for these specific choice of parameters. Light-green band stands for $M_t$ at $\pm 1\sigma$. (b) ${\cal P}_0$ is plotted against Higgs dark matter coupling $\lambda_S(M_Z)$ for different values of $\lambda_2(M_Z)$.}}}
\end{center}
\end{figure}
In the standard model, the present measurements on $M_h$ and $M_t$ indicate that the electroweak vacuum, in which the Universe is at present residing, may be a false one. From this metastable vacuum, it might tunnel into a deeper true vacuum, residing close to $M_{\rm Pl}$.
The decay probability of the EW vacuum to the true vacuum at the present epoch can be expressed as~\cite{Coleman:1977py, Isidori:2001bm, Buttazzo:2013uya}
\begin{equation}
{\cal P}_0=0.15 \frac{\Lambda_B^4}{H^4} e^{-S(\Lambda_B)}
\label{prob}\\
\end{equation}
where $H$ is the Hubble constant and the action is given by
\begin{equation}
S(\Lambda_B)=\frac{8\pi^2}{3|\lambda_1(\Lambda_B)|}\, .
\label{action}\\
\end{equation}
$S(\Lambda_B)$ is called the action of bounce of size $R=\Lambda_B^{-1}$. The value of $R$ for which $S(\Lambda_B)$ is minimum gives the dominant contribution to the tunneling probability ${\cal P}_0$. It occurs when $\lambda_1(\Lambda_B)$ is minimum, {\em i.e.,\ } $\beta_{\lambda_1}(\Lambda_B)=0$. Henceforth, $\Lambda_B$ denotes the scale where $\lambda_1$ is minimum. Here we neglect the loop correction to the action, as in Ref.~\cite{Isidori:2001bm} it had been argued that setting the running scale to $R^{-1}$ significantly restricts the size of such corrections. In the numerical evaluation we have used $\lambda_{1,\rm{eff}}$ in place of $\lambda_{1}$ in Eq.~(\ref{action}). We also neglect gravitational corrections~\cite{Coleman:1980aw,Isidori:2007vm} to the action as in Ref.~\cite{Khan:2014kba}.
In Ref.~\cite{Isidori:2001bm} it was pointed out that thermal corrections are important at very high temperatures.
Finite temperature effects to EW vacuum stability in the context of SM have been calculated recently in a preprint~\cite{Rose:2015lna}\footnote {See Ref.~\cite{Espinosa:1995se} for an earlier work.}. It had been claimed that the parameter space corresponding to EW metastability shrinks considerably. In this paper we work with zero temperature field theory only.
The presence of the inert doublet induces additional contributions to $\beta_{\lambda_1}$ [see Eq.~(\ref{betal_1})]. As a result, which is generic for all scalars, $\lambda_1$ receives a positive contribution compared to the SM, which pushes a metastable vacuum towards stability, implying a lower ${\cal P}_0$.
Electroweak metastability in the ID model has been explored earlier in the literature, albeit in a different context~\cite{Barroso:2013awa,Barroso:2012mj,Barroso:2013kqa,Gil:2012ya}. If $H^+$ gets a VEV, there could exist another charge-violating minimum. If instead, $A$ receives a VEV, another $CP$-violating minimum could pop up. But these vacua always lie higher than the usual EW vacuum. If $\mathbb{Z}_2$ is broken by introducing additional soft terms in the Lagrangian, then the new $\mathbb{Z}_2$-violating minimum can be lower than the usual $\mathbb{Z}_2$-preserving EW minimum. As in our present work, $\mathbb{Z}_2$ is an exact symmetry of the scalar potential, so such cases need not be considered. However, as mentioned earlier, if at some scale before $M_{\rm Pl}$, the sign of $\lambda_1$ becomes negative, there might exist a deeper minimum which is charge-, $CP$- and $\mathbb{Z}_2$-preserving and lying in the SM Higgs $h$ direction.
Whether the EW vacuum is metastable or unstable, depends on the minimum value of $\lambda_1$ before $M_{\rm Pl}$, which can be understood as follows. For EW vacuum metastability, the decay lifetime should be greater than the lifetime of the Universe, implying ${\cal P}_0 < 1$. This implies~\cite{Isidori:2001bm, Khan:2014kba}
\begin{equation}
\lambda_{\rm 1, eff}(\Lambda_B) > \lambda_{\rm 1,min}(\Lambda_B)=\frac{-0.06488}{1-0.00986 \ln\left( {v}/{\Lambda_B} \right)}\,.
\label{lammin}
\end{equation}
Hence we can now reframe the vacuum stability constraints on the ID model, when $\lambda_{\rm 1, eff}$ runs into negative values, implying metastability of the EW vacuum. We remind the reader that in the ID model, instability of the EW vacuum cannot be realized as addition of the scalars only improves the stability of the vacuum.
\begin{itemize}
\item
If $0>\lambda_{\rm 1, eff}(\Lambda_B)>\lambda_{\rm 1,min}(\Lambda_B)$, then the vacuum is metastable.
\item
If $\lambda_{\rm 1, eff}(\Lambda_B)<\lambda_{\rm 1,min}(\Lambda_B)$, then the vacuum is unstable.
\item
If $\lambda_2<0$, then the potential is unbounded from below along the $H, A$ and $H^\pm$ direction.
\item
If $\lambda_3(\Lambda_{\rm I})<0$, the potential is unbounded from below along a direction in between $H^\pm$ and $h$.
\item
If $\lambda_L(\Lambda_{\rm I})<0$, the potential is unbounded from below along a direction in between $H$ and $h$.
\item
If $\lambda_S(\Lambda_{\rm I})<0$, the potential is unbounded from below along a direction in between $A$ and $h$.
\end{itemize}
In the above, $\Lambda_{\rm I}$ represents any energy scale for which $\lambda_{\rm 1, eff}$ is negative and the conditions for unboundedness of the potential follow from Eq.~(\ref{Scalarpot2}). At this point note the significant deviations we are making in the allowed parameter space compared to the usual vacuum stability conditions: According to Eq.~(\ref{stabilitybound}), $\lambda_{3,L,S}$ can take slightly negative values. But at a scale where $\lambda_{\rm 1,eff}$ is negative, with the new conditions $\lambda_{3,L,S}$ have to be positive.
The tunneling probability ${\cal P}_0$ is computed by putting the minimum value of $\lambda_{\rm 1, eff}$ in Eq.~(\ref{action}) to minimize $S(\Lambda_B)$. In Fig.~\ref{fig:Tun}$\rm{ \left(a\right)}$, we have plotted ${\cal P}_0$ as a function of $M_t$. The right band corresponds to the tunneling probability for our benchmark point as in Table~\ref{table1}. For comparison, we plot ${\cal P}_0$ for SM as the left band in Fig.~\ref{fig:Tun}$\rm{(a)}$. 1$\sigma$ error bands in $\alpha_s$ and $M_h$ are also shown. The error due to $\alpha_s$ is more significant than the same due to $M_h$. As expected, for a given $M_t$, the presence of ID lowers tunneling probability. This is also reflected in Fig.~\ref{fig:Tun}$\rm {\left(b\right)}$, where we plot ${\cal P}_0$ as a function of $\lambda_S(M_Z)$ for different choices of $\lambda_2(M_Z)$, assuming $M_h=125.7$~GeV, $M_t=173.1$~GeV, and $\alpha_s=0.1184$. Here DM mass $M_A$ is also varied with $\lambda_S$ to get $\Omega h^2=0.1198$. For a given $\lambda_S(M_Z)$, the higher the value of $\lambda_2(M_Z)$, the smaller ${\cal P}_0$ gets, leading to a more stable EW vacuum.
\section{Phase diagrams}
\label{sec:phasediag}
The stability of EW vacuum depends on the value of parameters at low scale, chosen to be $M_Z$. In order to show the explicit dependence of EW stability on various parameters, it is customary to present phase diagrams in various parameter spaces.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=2.8in,height=2.8in, angle=0]{MHvsMDMcontourplot}
\caption{\label{fig:STU1} \textit{\rm {Constraints in $M_{H}-M_{A}$ plane. The cross-hatched region is excluded from LEP~{\rm \cite{Lundstrom:2008ai}}. Choosing $M_{H^\pm}=120$~GeV and $\lambda_S(M_Z)=0.001$, relic density constraint is satisfied at 3$\sigma$ on the blue band. The green (yellow) region corresponds to EW vacuum stability (metastability). The solid brown line correspond to $M_H=M_A$. The grey area on the left to it is of no interest to us as we have chosen $M_H>M_A$. The dashed brown line shows the LEP\,I limit. On the brown region, unitarity constraints are violated before $M_{\rm Pl}$. } }}
\end{center}
\end{figure}
In Fig.~\ref{fig:STU1} we show the LEP constraints in the $M_{H}- M_A$ plane as in Ref.~\cite{Lundstrom:2008ai}. We update this plot identifying regions of EW stability and metastability. As we are considering a scenario where the ID model is valid till $M_{\rm Pl}$, there are further limits from unitarity. The relic density constraint imposed by WMAP and Planck combined data is represented by the thin blue band. The choice of $\lambda_2(M_Z)$ does not have any impact on relic density calculations, but affects EW stability as expected. In this plot, for higher values of $\lambda_2(M_Z)$, the region corresponding to EW metastability will be smaller. The chosen parameters satisfy the LUX direct detection bound.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=2.8in,height=2.8in, angle=0]{metastableinertSTUplot70unitary}
\caption{\label{fig:STU2} \textit{\rm {Phase diagram in $\Delta M_{H}-\Delta M_{H^\pm}$ plane for $M_A=70$~GeV. The green and yellow regions correspond to EW vacuum stability and metastability respectively. The cross-hatched band is excluded from LEP. The brown region suffers from unitarity violation before $M_{\rm Pl}$. The blue band reflects relic density constraint at 3$\sigma$.
} }}
\end{center}
\end{figure}
As small splitting among $M_A$, $M_H$, and $M_{H^\pm}$ leads to some cancellations among diagrams contributing to DM annihilation, $\Delta M_{H^\pm}$ and $\Delta M_{H}$ are often used as free parameters in the ID model. In Fig.~\ref{fig:STU2}, we present constraints on this parameter space for $M_A=70$~GeV. As before, the brown region corresponds to unitarity violation before $M_{\rm Pl}$. For small $\Delta M_{H^\pm}$ and $\Delta M_{H}$, $\lambda_{3,4,5}$ are required to be small, which leads to little deviation from SM metastability. The metastable region is shown by the yellow patch, which shrinks for larger $\lambda_2$. The blue band reflects the relic density constraint for $\lambda_S(M_Z)=0.001$. For such small $\lambda_S(M_Z)$, the $h$-mediated $s$-channel diagram in $AA\rightarrow WW$ or $AA\rightarrow ZZ$ contributes very little. $H^+$- or $H$-mediated $t$- and $u$-channel diagrams are also less important than the quartic vertex driven diagram due to propagator suppression. This explains the ``L'' shape of the blue band. For higher values of $\lambda_S(M_Z)$, the shape of the band changes and ultimately leads to a closed contour. It appears that due to LEP constraints, EW vacuum metastability is almost ruled out. Although the LEP constraint permits $\Delta M_H<8$~GeV, allowing a narrow strip towards the left, the relic density constraints cannot be satisfied on this strip as it leads to an increased rate of DM coannihilation processes, leading to a dip in $\Omega h^2$. But as we will see later, if $M_t$ and $\alpha_s$ are allowed to deviate from their respective central values, for some region in this parameter space, it is possible to realize a metastable EW vacuum.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{LamsMDMcontourplot}}
\hskip 15pt
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{LamsMDMcontourplothigh}}
\caption{\label{fig:MDM_LamS} \textit{\rm {Phase diagram in $\lambda_S(M_Z)-M_A$ plane for $\lambda_2(M_Z)=0.1$. Panel (a) stands for `low' DM mass. The blue band corresponds the 3$\sigma$ variation in $\Omega h^2$ when $\Delta M_{H^\pm}=40$~GeV and $\Delta M_H=40$~GeV. LEP direct search constraints are represented by the cross-hatched band at the bottom. Entire green region imply EW vacuum stability. Panel (b) stands for `high' DM masses. The relic density band (blue) now correspond to $\Delta M_{H^\pm}=1$~GeV and $\Delta M_H=2$~GeV. The corresponding stable and metastable phases for EW vacuum are represented by green and yellow patches respectively. The relic density band (red) corresponds to $\Delta M_{H^\pm}=5$~GeV and $\Delta M_H=2$~GeV. For this, the boundary separating the EW phases is denoted by the red dashed line.} }}
\end{center}
\end{figure}
To delineate the role of $M_H$ in EW vacuum stability, in Figs.~\ref{fig:STU1} and \ref{fig:STU2}, $\lambda_2(M_Z)$ was chosen to be small. Now to demonstrate the effect of $\lambda_S$, we will now present in Fig.~\ref{fig:MDM_LamS} phase diagrams in the $\lambda_S(M_Z) - M_A$ plane. Panel (a) deals with low DM masses. For $\Delta M_{H^\pm}=40$~GeV and $\Delta M_H=40$~GeV, part of the allowed relic density band (blue) is allowed from LEP constraints (cross-hatched band). The entire parameter space corresponds to EW vacuum stability. Choosing small $\Delta M_{H^\pm}$ and $\Delta M_H$, which imply small values of $\lambda_{3,4,5}$, can lead to metastability. But those regions are excluded by LEP. Again, metastability can creep in if $M_t$ and $\alpha_s$ are allowed to deviate from their central values.
In Fig.~\ref{fig:MDM_LamS}(b), we study the same parameter space for high DM masses. As mentioned before, to obtain the correct relic density, smaller mass splitting among various ID scalars needs to be chosen. For $\Delta M_{H^\pm}=1$~GeV and $\Delta M_H=2$~GeV, the 3$\sigma$ relic density constraint is shown as the blue band. The blue dashed line demarcates the boundary between stable (green) and metastable (yellow) phases of EW vacuum. The choice of small values of $\Delta M_{H^\pm}$ and $\Delta M_H$, in turn, leads to a large region pertaining to EW metastability.
To illustrate the sensitivity to the mass splitting, in Fig.~\ref{fig:MDM_LamS}(b), we present another relic density band (red) when $\Delta M_{H^\pm}=5$~GeV and $\Delta M_H=2$~GeV. The corresponding boundary between the phases is denoted by the red dashed line. The region on the right implies EW stability (the green and yellow regions do not apply to this case). As for high DM masses, EW metastability can be attained for a sizable amount of the parameter space; $\lambda_2(M_Z)$ need not be chosen to be very small to maximize the metastable region for the sake of demonstration.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{MhMtcpoddplot70}}
\hskip 15pt
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{MhMtcpoddplot573}}
\caption{\label{fig:Mt_Mh} \textit{\rm {Phase diagrams in $M_h - M_t$ plane. Panels (a) and (b) stand for `low' and `high' DM masses respectively. Regions of
absolute stability~(green), metastability~(yellow), instability~(red) of the EW vacuum are also marked. The grey zones represent error ellipses at $1$, $2$ and $3\sigma$. The three boundary lines (dotted, solid and dotted red) correspond to $\alpha_s(M_Z)=0.1184 \pm 0.0007$. Details of benchmark points are available in the text. } }}
\end{center}
\end{figure}
The fact that for SM the EW vacuum stability is ruled out at $\sim 3\sigma$, is demonstrated by a phase space diagram in the $M_t-M_h$ plane~\cite{Degrassi:2012ry, Buttazzo:2013uya}. In Ref.~\cite{Khan:2014kba}, similar diagrams were presented for a singlet scalar extended SM. To demonstrate the impact of ID scalars to uplift the EW vacuum metastability, we present phase diagrams in the $M_t-M_h$ plane for two sets of benchmark points in Fig.~\ref{fig:Mt_Mh}. Panel (a) is drawn for $M_A = 70$~GeV, $\Delta M_{H^\pm}=11.8$~GeV, $\Delta M_H=45$~GeV, $\lambda_S(M_Z)=0.001$, and $\lambda_2 (M_Z) = 0.1$. For panel (b) the set of parameters in Table~\ref{table1} is being used.
Both sets of parameters are chosen so that they respect the WMAP and Planck combined results on DM relic density and the direct detection bounds from XENON\,100 and LUX. As in Ref.~\cite{Khan:2014kba}, the line demarcating the boundary between stable and metastable phases of EW vacuum is obtained by demanding that the two vacua be at the same depth, implying $\lambda_1(\Lambda_B)=\beta_{\lambda_1}(\Lambda_B)=0$. The line separating the metastable phase from the unstable one is drawn using the conditions $\beta_{\lambda_1}(\Lambda_B)=0$ and $\lambda_1(\Lambda_B)=\lambda_{1, \rm min}(\Lambda_B)$, as in Eq.~(\ref{lammin}). The variations due to uncertainty in the measurement of $\alpha_s$ are marked as dotted red lines. In each panel, the dot representing central values for $M_h$ and $M_t$ is encircled by 1$\sigma$, 2$\sigma$, and 3$\sigma$ ellipses representing errors in their measurements. According to Fig.~\ref{fig:Mt_Mh}(a), EW vacuum stability is allowed at 1.5$\sigma$, whereas in Fig.~\ref{fig:Mt_Mh}(b), it is excluded at 2.1$\sigma$, indicated by blue-dashed ellipses.
\begin{figure}[h!]
\begin{center}
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{AlphavsmtplotsCPodd70}}
\hskip 15pt
\subfigure[]{
\includegraphics[width=2.8in,height=2.8in, angle=0]{AlphavsmtplotsCPodd573}}
\caption{\label{fig:Alpha_Mt} \textit{\rm {Phase diagrams in $M_t-\alpha_s(M_Z)$ plane for the same sets of benchmark points as in Fig.~\ref{fig:Mt_Mh}. Notations used are also the same as in Fig.~\ref{fig:Mt_Mh}. } }}
\end{center}
\end{figure}
As in the literature SM EW phase diagrams are also presented in the $\alpha_s(M_Z)-M_t$ plane~\cite{Bezrukov:2012sa, EliasMiro:2011aa}, we do the same in the ID model as well. In Fig.~\ref{fig:Alpha_Mt}, we use the same sets of benchmark parameters as in Fig.~\ref{fig:Mt_Mh}. As a consistency check, one can note that the EW vacuum is allowed or ruled out at the same confidence levels.
\begin{figure}[h!]
\begin{center}{
\includegraphics[width=2.8in,height=2.8in, angle=0]{lamsconfdncplot44}}
\caption{\label{fig:confidence} \textit{\rm {Dependence of confidence level at which EW vacuum stability is excluded (one-sided) or allowed on $\lambda_S(M_Z)$ and $\lambda_2(M_Z)$. Regions of absolute stability (green) and metastability (yellow) of EW vacuum are shown for $\lambda_2(M_Z)=0.1$. The positive slope of the line corresponds to the stable electroweak vacuum and negative slope corresponds to the metastability.
} }}
\end{center}
\end{figure}
To study the impact of nonzero ID couplings, however, it is instructive to study the change in the confidence level ($\sigma$) at which EW stability is modified with respect to these couplings. As in Ref.~\cite{Khan:2014kba}, we plot in Fig.~\ref{fig:confidence} ~$\sigma$ against $\lambda_S(M_Z)$ for different values of $\lambda_2(M_Z)$. We vary $M_A$ along with $\lambda_S(M_Z)$ to keep DM relic density fixed at $\Omega h^2=0.1198$ throughout the plot. Note that changing $\lambda_2(M_Z)$ does not alter $\Omega h^2$. The masses of other ID particles are determined using $\Delta M_{H^\pm}=1$~GeV and $\Delta M_H=2$~GeV. The parameter space considered does not yield too large DM-nucleon cross section, inconsistent with XENON\,100 and LUX DM direct detection null results. For a specific value of $\lambda_2(M_Z)=0.1$, with the increase of $\lambda_S(M_Z)$, the confidence level at which EW is metastable (yellow region) gets reduced and becomes zero at $\lambda_S(M_Z)\simeq 0.04$. After this, EW vacuum enters in the stable phase (green). With further increases in $\lambda_S(M_Z)$, the confidence level at which EW is stable keeps increasing. To illustrate the role of $\lambda_2(M_Z)$, we use two other values in the same plot. The value of $\lambda_S(M_Z)$ at which the EW vacuum enters in the stable phase increases with decreases in $\lambda_2(M_Z)$, as expected. The yellow and green marked regions are not applicable when $\lambda_2(M_Z)=0.05, 0.15$.
\section{Veltman's conditions}
\label{sec:veltman}
As we have extended the validity of the ID model till $M_{\rm Pl}$, it is interesting to explore whether Veltman's condition (VC) can be satisfied in this model at any scale on or before $M_{\rm Pl}$. It is particularly interesting as similar studies for SM have been carried out in Ref.~\cite{Degrassi:2012ry}. It has been shown that if one imposes VC in the SM at $M_{\rm Pl}$, then the top mass measurement $M_t=173.1\pm 0.6$~GeV implies $M_h\approx 135\pm2.5$~GeV, which is excluded at more than 3$\sigma$.
Veltman's condition implies that the quadratic divergences in the radiative corrections to the Higgs mass can be handled if the coefficient multiplying the divergence somehow vanishes~\cite{Veltman:1980mj, Hamada:2012bp}. VC includes the contributions from the infrared degrees of freedom of the theory and does not carry any special information about the ultraviolet divergences. In SM, it suggests the combination
\begin{equation}
6 \lambda_1+\frac{9}{4} g_2^2+\frac{3}{4} g_1^2 -12 y_t^2 = 0\nonumber\, .
\end{equation}
Due to the large negative contribution from the term containing the top Yukawa coupling, it is not possible to satisfy VC till $M_{\rm Pl}$ given the experimental measurements of $M_t$ and $M_h$ within the context of SM.
In the ID model, as we add more scalars a possibility opens up to satisfy VC, as their contributions can offset the large negative contribution from the top quark.
The above VC for the SM associated with $\mu_1$ is promoted in the ID model to~\cite{Barbieri:2006dq, Chakraborty:2014oma}
\begin{equation}
6 \lambda_1+ 2\lambda_3+\lambda_4+\frac{9}{4} g_2^2+\frac{3}{4} g_1^2 -12 y_t^2 = 0.
\end{equation}
If $2\lambda_3+\lambda_4$ is positive, then we have checked with our RG improved coupling constants, and it is possible to satisfy the above VC at a scale before $M_{\rm Pl}$.
However, in the ID model $\mu_2$ also receives quadratically divergent radiative corrections. The corresponding VC reads as
\begin{equation}
6 \lambda_2+ 2\lambda_3+\lambda_4+\frac{9}{4} g_2^2+\frac{3}{4} g_1^2 = 0\, .
\end{equation}
Note that it lacks the Yukawa contribution as the unbroken $\mathbb{Z}_2$ forbids fermionic interactions of the inert doublet $\Phi_2$. As $2\lambda_3+\lambda_4$ is already positive, this VC can be satisfied if $\lambda_2$ is negative. But a negative $\lambda_2$ renders the potential unbounded from below as evident from our earlier discussions. Note that amongst our RG improved coupling constants $\lambda_1$ can be driven to negative values at high scales. But this makes the required cancellations for VCs even worse. Hence, it is not possible to satisfy Veltman's conditions in a scenario where only the ID model reigns the entire energy regime up to the $M_{\rm Pl}$.
\section{Summary and Conclusion}
\label{sec:conclusion}
If the standard model is valid up to the Planck scale, the present measurements on the masses of the top quark and Higgs indicate the presence of a deeper minimum of the scalar potential at a very high energy scale, threatening the stability of the present electroweak vacuum. State of the art NNLO calculations performed to evaluate the probability that the present EW vacuum will tunnel into the deeper vacuum lying close to $M_{\rm Pl}$ suggest that the present EW vacuum is metastable at $\sim 3\sigma$. The lack of stability might be the artifact of incompleteness of the SM.
Although the LHC has yet to find any signal suggesting existence of any new physics beyond the standard model of particle physics, other experimental evidence points towards the existence of dark matter, which so far could have escaped detection in colliders and DM direct detection experiments. Hence, it is important to look into the problem of EW stability in a scenario which addresses the issue of DM as well. In particular, we extend SM by adding an inert scalar doublet, offering a viable DM candidate and assume that this model is valid up to $M_{\rm Pl}$.
In this paper, our intentions are twofold. First, in such a scenario we have consolidated the bounds imposed on the ID model. As we are demanding validity of the model up to $M_{\rm Pl}$, the RG evolution of the couplings can disturb the unitarity of the S-matrix governing various scattering processes, which in turn imposes stringent limits on the parameter space at the EW scale. In this light, we present a consolidated discussion updating the existing bounds on the ID model.
The other and the main goal of this paper is to check the stability of EW vacuum in the ID model. If the ID DM happens to be the only DM particle, which saturates the observed DM relic density, can the ID model modify the stability of the EW vacuum? Note that rather than considering new physics effects close to the Planck scale, here new physics is added at the EW scale only.
It is well known that addition of a scalar can improve stability of the EW vacuum. But if we are to solve both the DM and EW vacuum stability problems in the context of the ID model, it is important to study the parameter space which allows us to do so. As ID introduces a few new parameters and fields, the study of the parameter space is quite involved when we consider radiatively improved scalar potentials containing SM NNLO corrections and two-loop ID contributions. Inclusion of these NNLO corrections is mandatory to reproduce the correct confidence level at which EW vacuum is metastable in the SM. In case of ID, if one works with one-loop ID contributions in the RGEs instead of two-loop ones, the changes in the plots presented in the paper are limited to $5$\% only. However, as a small change in the action $S$ is amplified exponentially in the vacuum decay lifetime, two-loop effects have also been taken care of. For the benchmark point used in Table~\ref{table1}, ${\cal P}_0$ using one-loop ID RGEs turns out to be $5.8 \times 10^{-1271}$, whereas for two-loop ID RGEs, ${\cal P}_0$ changes to $1.4 \times 10^{-1310}$. We neglect other loop corrections to bounce action, as a $5$\%--$10$\% correction can lead to an appreciable change in the lifetime, but as demonstrated above with the example of inclusion of two-loop ID corrections, the phase diagrams essentially remain unchanged (see Ref.~\cite{Branchina:2014rva} for a related example). We note in passing that unlike Ref.~\cite{Goudelis:2013uca}, our analysis is valid up to $M_{\rm Pl}$. Requiring the potential bounded from below at all scales below $M_{\rm Pl}$ puts severe constraints on the model.
As far as the allowed parameter space is concerned, we focus on metastability of EW vacuum which is realized when $\lambda_1$ is negative. In this context, our analysis is much different from earlier analyses in the literature, for example, as in Ref.~\cite{Goudelis:2013uca}. When $\lambda_1$ is negative, the vacuum stability conditions (\ref{stabilitybound}) are no longer useful and the parameter space analysis in Ref.~\cite{Goudelis:2013uca} is not valid. For example, when $\lambda_1$ is positive, $\lambda_{3,L,S}$ can assume small negative values as positivity of the $\lambda_1 h^4$ term ensures the potential bounded from below. However as we point out, one can no longer afford a negative $\lambda_{3,L,S}$ at a scale where $\lambda_1$ is negative. If we demand in addition that the ID model saturates DM relic density, then
even when $\lambda_1$ is positive, $\lambda_{3}$ cannot be negative. But as mentioned earlier, for negative $\lambda_{1}$, although $\lambda_{3}$ has to be positive at that scale, its value at $M_Z$ can still be negative. Thus, one of the main motives for this work is to find out the allowed parameter space in the ID DM model in case the EW vacuum is metastable.
We see that for DM masses of 70~GeV, the allowed parameter space corresponds to absolute stability unless we allow some deviation of $M_t$ and $\alpha_s$ from their measured central values. For higher DM masses more than 500~GeV, it is possible to realize a metastable EW vacuum for a large parameter space. This vacuum will have a longer lifespan than the SM one as the addition of scalars improves the stability of the EW vacuum.
\vskip 20pt
\noindent{\bf Acknowledgements:}\\
The work of N.K. is supported by a fellowship from UGC. S.R. is indebted to Sourov Roy and Dilip Ghosh for useful discussions.
|
\section{Introduction}
This paper proposes the recipe for the definition of a simplified system architecture that aims at exploiting a collaborative exchange of information between peers belonging to a highly distributed infrastructure in order to build a computing/network approach able to link similar users in order to foster the process of information percolation within the nodes of the network.
More in detail, we push further the idea of realising collaborative recommender mechanisms, by means of solutions fostering interest clustering, that are obtained by means of interactions happening among users.
Our approach couples epidemic-based P2P overlay networks to ease the gathering of users with similar interests and then use the connections established to let peer exchange recommendations one each others.
Our goal is twofold. On one hand we aim at building an adaptive system supporting the recognition of communities of users\' interests in a decentralized, distributed way. The approaches that have been proposed so far in the area of P2P computing (and the epidemic ones, in particular) are able to manage a very large amount of peers and to deal gracefully with churn, whereas centralized systems require expensive and, often, very complex techniques to ensure continuous operation under node and link failures.
The service is implemented by means of the collaborations established between computing nodes, without needing any centralized authority devoted to store all the profiles and the ratings of users as well as to provide centralized-controlled suggestions.
On the other hand, our goal is to exploit such communities not limiting our aim to the knowledge sharing about interesting items within them, but also to address some of the traditional problems affecting recommender systems. In particular, the ability to recommend new, almost unknown, items.
The system we are sketching, assumes that each neighbor of a computing node (peer) $P$ pushes recommendations to it focused on the items that might be of potential interest for $P$. It is worth to notice that this decision is taken locally, when a neighbor selects or becomes aware from its links of the existence of a new item, whose characteristics, are someway related with one (or more) of its communities.
Then, it can then recommend such item both to $P$ and to its other neighbors, of all the related communities, as well. This approach would allow a more efficient and rapid percolation of the information within the overlay network.
The remain of this paper is organized as follows: in Sec.~\ref{rel} we shortly present the scientific literature about the subject of this paper; in Sec.~\ref{arch} we describe the architecture of our proposed system. Finally, in Sec.~\ref{concl}, conclusions are given and potential further exploitations of this work are proposed.
\section{Related Work}\label{rel}
The correlations of interests amongst a group of distributed users has been leveraged in a variety of contexts and for designing or enhancing various distributed systems~\cite{carlini2012reducing, carlini2011evaluating}.
For peer-to-peer file sharing systems that include file search facilities (e.g., Gnutella, eMule, etc.), an approach to increase recall and precision of the search is to group users based on their past search history or based on their current cache content ~\cite{Fraigniaud2005Combining-the-use-of,Handurukande2006Peer-Sharing-Behavio}.
Another potential use of interest clustering is to form groups of peers that are likely to be interested in the same content in the future, hence forming groups of subscribers in a content-based information diffusion system~\cite{carlini2009service,baraglia2013peer,dazzi2009peer,baraglia2011group,mordacchini2009challenges,dazzi2011experiences,carlini2014toward,mordacchini2013towards,marzolla2006p2p,gennaro2007mroute}.
Moreover, interest correlation can be used to help bootstrapping and self-organization of dissemination structures such as network-delay-aware trees for RSS dissemination~\cite{Patel2009Rappel:-Exploiting-i}. The correlation between the users\' past and present accesses has been used for user-centric ranking.
In order to improve the customisation of search results, the most probable expectations of users are determined using their search log stored on a centralized server~\cite{tan_mining_2006,teevan_personalizing_2005}. However, the correlation between users is not leveraged to improve the quality of result personalization, hence making the approach viable only for users with sufficiently long search logs. An alternative class of clustering search engines uses semantic information in order to cluster results according to the general domain they belong in (and not as in our approach to cluster users based on their interests). This can be seen as a centralized, user-agnostic approach to improve user experience. The clustering amongst data elements is derived from their vocabulary. It presents the user with results along different interest domains and can help the user to disambiguate these results from a query that may cover several domains, e.g., the query word apple can relate to both food/fruits and computers domains. Examples of such systems are EigenCluster~\cite{Cheng2006A-divide-and-merge-m}, or TermRank~\cite{Gelgi2007Term-Ranking-for-Clu}. Nonetheless, these systems simply modify the presentation of results so that the user decides herself in which domain the interesting results may fall these results are not in any way automatically tailored to her expectations. They do not also consider the clustering of
interest amongst users, but only the clustering in content amongst the data.
Other approaches cluster users on the basis of similarity between their semantics profile. Approaches of this kind of systems includes GridVine~\cite{GridVine}, the semantic overlay networks~\cite{SON} and p2pDating~\cite{p2pDating}.
They build a semantic P2P overlay infrastructure that relies on a logical layer storing data.
They make use of heterogeneous but semantically related information sources whereas our approach does not rely on any kind of semantic interpretation. It, in principle, enables a broader exploitation of more heterogeneous data sources.
Related with our proposal is Tribler~\cite{tribler}, a P2P television recommender system. In contrast with our approach, neighbor lists can be directly filled in by the user herself using an interface. No topology or affinity property is considered. We propose a gossip system that construct and maintain in rest groups of dynamic users based on their past activities, without needing their direct intervention.
\section{Proposed approach}\label{arch}
This section introduces the main pillars that would be needed to support the construction and the exploitation of an overlay network made of peers that share common interests.
Figure~\ref{fig:communities} sketches the architecture of an overlay network organised accordingly.
As can be observed in figure, the links between peers are established when they are characterised by a common interest. This information is derived recognising the accesses performed by peers to the same content in the past.
The surrounding idea is that they are considered interested to share similar interest if can potentially show interests for the same content in the future. Thus, peers collaboratively exchange useful recommendations among themselves.
\begin{figure}[ht!]
\begin{minipage}[t]{20em}
\begin{center}
\includegraphics[width=0.5\columnwidth]{communities}
\caption[Interest Communities]{Interest Communities}\label{fig:communities}
\end{center}
\end{minipage}
\hfill
\begin{minipage}[t]{20em}
\begin{center}
\includegraphics[width=0.5\columnwidth]{communities2}
\caption[Interest Overlays]{Interest Overlays}\label{fig:interests}
\end{center}
\end{minipage}
\end{figure}
The protocol, to group similar users in communities, adopts a clustering algorithm. As a first step each peer determines, independently, the peers to link with.
These one-to-one connections are established on the basis of an interest-based degree, that is measured amongst the peer it encounters. Every time it becomes aware of a new peer, it can, in turn, learn of the existence of new potential neighbors and possibly communicate with them. Finally, it can also be aware of other, potentially better neighbors.
The idea is that this process is stabilized when each node composing the neighborhood of a peer can be considered as the representative of a community of a shared interest. An important side-effect of this vision is that a peer is characterized by multiple interests, with different ``entry-points'' for each interest.
The process is conducted separately for each of the interests of a peer. Consequently, the connections are established and maintained separately for every distinct interest. At the end will be created a set of virtual different overlays, where each peer participates in as many groups is required to cover its interests.
The resulting scenario situation is depicted in Fig.\ref{fig:interests}. In order to obtain such organisation, each peer initiate a different stream of messages, one for each of its interests.
\subsection{Profiles}
Profiles of peers need to be modeled according to the users' interests. A possible approach would be based on recently accessed resources, purchased items, visited pages, etc.
Such information, once gathered, has to be considered in a proper way. Basically, it consists in the basis over which the overlay network will be organised.
Generally speaking, let $\Im$ be the set of items belongings to the whole set of profiles of users and let $\Im_p \subseteq \Im$ be the subset of items belonging to a specific peer $p$. We consider that the profile $\pi$ of $p$ can be defined as
\begin{eqnarray*}
\pi_p = \{ (i, C(i), R(i)) | i \in \Im_p \}
\end{eqnarray*}
where $i$ is an item belongings to the set of $\Im_p$, $C(i)$ is the content associated with $i$ and $R(i)$ is the rating given by $p$ to $i$. The peer $p$ has also associated a set $I^p = \{I^{p}_1,\ldots,I^{p}_k\}$ of interests.
Each of the items in $\pi_p$ may be associated with an interest $I^{p}_j$. We can then represent $\pi_p$ in the following way:
\begin{eqnarray*}
\pi_p = \displaystyle\bigcup_{j=1,...,k} \pi_{p}(I^{p}_j)
\end{eqnarray*}
where $\pi_{p}(I^{p}_j) $ is the set of items related to the interest $I^{p}_j$.
For realising this association, we introduce a function $\gamma$ that given an item belonging to $\Im$ decides the interest it should be associated with. More formally:
\begin{eqnarray*}\gamma_p(i) = I^p_j~~with~i \in \Im_p\end{eqnarray*}
It is worth to note that the set $I^p$ is specific for each distinct peer $p$.
In fact, we do not assume any globally known labeling, classification or partitioning of the objects in $\Im$. Each peer performs its own subdivision of $\Im_p$ in the interests of $I_p$.
It can then compare its objects divided per interest with the sets of the other peers it will contact. Given two peers $p_1$ and $p_2$, $p_1$ would consider its local interest $I^{p_1}_s$ similar to the interest $I^{p_2}_t$ if it would contain the most similar set of items among the other sets in $I^{p_2}$ with respect to the items in $I^{p_1}_s$.
As a consequence of having a solution to describe each user interests coded in the peer profiles, it is important to pay attention on adopting a proper similarity function $sim: \Pi^2 \rightarrow \mathbb{R}$ to compare profiles, where $\Pi$ is the set of all possible profiles.
This is a key aspect, since this function specify the relationships between peers according to their interests.
If each distinct interest is determined by different type of features, different similarity measures could be used to evaluate peers proximities with respect to each interest.
Several measures can be adopted to this end. As an example, a typical approach is to use a metric that takes into account the size of each profile, such as the Jaccard similarity, which has proven to be an effective similarity measure~\cite {Fraigniaud2005Combining-the-use-of,Patel2009Rappel:-Exploiting-i}. Given two peers $p_1$ and $p_2$ and two interests $I^{p_1}_s$ and $I^{p_2}_t$, the similarity can be computed as
\begin{eqnarray*}
sim(p_1,p_2) = \frac{|\pi_{p_1}(I^{p_1}_s) \cap \pi_{p_2}(I^{p_2}_t)|}{|\pi_{p_1} (I^{p_1}_s)\cup \pi_{p_2}(I^{p_2}_t)|}
\end{eqnarray*}
\subsection{Setup of Interest Communities}
One of the base assumptions of our envisioned system is that every peer is able to compute its interest-based distance to any other peer in the network. This measure allows it drive is ability to \emph{group} with other peers that have close-by interests, in order to form the basis for \emph{interests communities}.
This process is conducted automagically in a self-organizing and completely decentralized way, using a epidemic communication.
Each peer knows a set of other peers, namely its neighbors, and tries periodically to choose new neighbors that are closer to its interest than the previous ones.
In our envisioned system, this is simply obtained by discovering new peers from some other peer, then retrieving their profiles. Finally, choosing the $C$ nearest neighbors in the union of present and potential neighbors.
When a peer $p$ joins the network, it becomes in contact with one or more peers already belonging to in the interest-proximity network overlay.
They use the profile similarity function to compute how similar they are. They consider each interest in the $I$ of $\pi_p$ separately and they compare it against their own.
Furthermore, the peers contacted by $p$ use the same similarity function to determine which are, among their neighbors, the most similar to $p$. Once determined, the join request of $p$ are routed toward them.
All the peers that receive that request will react using the same protocol described above. All the interactions are shown in Algorithms 1 and 2.
This approach will lead $p$ to become aware of the existence of the most similar peers in the network overlay and allow it to connect with them. In doing this process, the involved peers can only use their local knowledge to compare their respective profiles.
\begin{table}[tb]
\begin{tabular}{l|l}
\begin{minipage}{2.35in}
{\scriptsize\textbf{Algorithm 1}}
\begin{algorithmic}
{\scriptsize
\STATE Let $CR(P')$ be a connection request from another peer $P'$
\STATE Let $NewPeers = \emptyset$
\IF{$Sim(P,P^{'}) \geq \displaystyle\min_{P_i \in N(P)}Sim(P,P_i)$}
\STATE Accept $CR(P')$
\FORALL{$P_i \in N(P)$}
\IF{$Sim(P_i,P') \geq \theta$}
\STATE add $P_i$ to $NewPeers$
\ENDIF
\ENDFOR
\STATE add $P'$ to $N(P)$
\STATE send $NewPeers$ to $P'$
\ELSE
\STATE refuse $CR(P')$
\ENDIF
}
\end{algorithmic}\label{alg:1}
\end{minipage}
&
\begin{minipage}{2.35in}
{\scriptsize\textbf{Algorithm 2} }
\begin{algorithmic}
{\scriptsize
\STATE Let $N(P)$ be the set of $P'$s actual neighbors
\FORALL{$P_i \in N(P)$}
\STATE Get from $P_i$ a set $NewPeers$ from its neighborhood
\FORALL{$P^{'} \in NewPeers$}
\IF{$P^{'} \notin N(P)$}
\STATE connect with $P^{'}$
\IF{$Sim(P,P^{'}) \geq \displaystyle\min_{P_j \in N(P)}Sim(P,P_j)$}
\STATE add $P^{'}$ to $N(P)$
\ENDIF
\ENDIF
\ENDFOR
\ENDFOR
}
\end{algorithmic}
\end{minipage} \label{alg:2}
\\
\\ \hline
\\
\begin{minipage}{2.35in}
{\scriptsize\textbf{Algorithm 3}}
\begin{algorithmic}
{\scriptsize
\STATE Let $N_p(I_j)$ be the set of $P'$s neighbors for the interest $I_j$
\STATE Receive a recommendation request from $p' \in N_P(I_j)$
\FORALL{$i \in \pi_p(I_j)$}
\IF{$Sim(p',i) \geq \theta$}
\STATE recommend $i$ to $p'$
\ENDIF
\ENDFOR
}
\end{algorithmic} \label{alg:3}
\end{minipage}
&
\begin{minipage}{2.35in}
{\scriptsize\textbf{Algorithm 4}}
\begin{algorithmic}
{\scriptsize
\STATE Know about a new item $h$
\STATE Let $I_j$ be the interest $h$ is related to
\STATE Let $N_P(I_j)$ be the neighborhood of peers interested in $I_j$
\FORALL{$p' \in N_P(I_j)$}
\IF{$Sim(p',h) \geq \theta$}
\STATE recommend $h$ to $p'$
\ENDIF
\ENDFOR
}
\end{algorithmic} \label{alg:4}
\end{minipage}
\end{tabular}
\\ \\
\caption{Active and passive threads and pull and push recommender algorithms}
\vspace{-15pt}
\end{table}
Once the process is stabilized, $p$ can consider its neighbors as the representatives of a personal community of ``friends'' from which request and to which forward recommendations. Thus, the gossip protocol provide the basis for classical recommender systems in forming the set of similar users.
This is done distributively and adaptively and the epidemic protocol ensure a robust and constant maintenance over time. Recommendations can then be requested by $p$ to its neighborhood and it can forward the newly items it discovered to its neighbors using Algorithms 3 and 4.
\section{Conclusion}\label{concl}
The focus of this paper is on giving a simple recipe for addressing the problem of clustering users in a purely decentralized way to foster information exchange. This is a particularly useful brick for enabling self-emerging and automated creation of communities of nodes representing users, which share common interests. In this paper we sketched the overall architecture of a epidemic-based distributed system exploiting a collaboratively built recommender system. The solution sketched in this work is simpler than most part of the existing solutions. This is inline with our goal: keep the solution as simple as possible but still providing a solution that exploit collaborative filtering is able to provide recommendations that are tailored and offer an acceptable degree of serendipity.
\bibliographystyle{unsrt}
|
\section{Introduction}
Because massive amounts of structured and unstructured data continue to accumulate, the importance of effective big data analysis is rapidly increasing.
One well-known big data analysis tool is Boltzmann machine learning. This technique is physics-friendly, because it is a form of probability density defined by the Hamiltonian of the Ising model \cite{Ackley1985}.
We assume that the generative model has a bias on each variable, the magnetic field, and the pair-wise interactions between the different variables (i.e., the interaction between adjacent spins).
Boltzmann machine learning has proven effective, and has stimulated increasing interest in deep learning \cite{Hinton2006,Hinton2006sci,RG2014,Ohzeki2015}.
Deep learning typically needs large volumes of data for its implementation.
Currently, this demand is often satisfied because we are in the so-called big data era; however, we require hard computation as a return.
Thus, the study of Boltzmann machine learning may involve constructing a good approximation \cite{Sessak2009,Cocco2011,Cocco2012,Ricci2012,Yasuda2013,Raymond2013,Ohzeki2013}.
Otherwise, we require a novel method to achieve efficient learning, even from a small amount of data.
Effective big data analysis can produce a substantial amount of valuable information.
An objective of this analysis is to elucidate a small number of relevant quantities to describe the acquired data, a process known as variable selection.
The goal of data-driven science is to capture an essential portion of the generative model and to identify the characteristics that describe its origin.
In order to achieve this goal, sparseness may be imposed on the bias or pair-wise interactions of the generative model.
One successful approach is to employ the regularization of the $L_1$ norm of the bias and pair-wise interactions.
However, because of the $L_1$ norm's lack of differentiability, the application of the simple gradient method is not straightforward.
A different method employs a greedy algorithm, which seeks a small number of non-zero components satisfying some criteria.
Under some conditions, greedy algorithms can overcome the $L_1$ regularization \cite{Aurelien2014,Yamanaka2015}.
However, greedy algorithms depend on the properties of the parameters to be estimated; moreover, $L_1$ regularization cannot be discarded, because it has a wide range of applications and enables us to perform robust inference for various models.
In this study, we resolve the lack of smoothness by implementing a technique for $L_1$ regularization (often used in optimization studies), namely majorizer minimization \cite{Beck2009book,Beck2009}.
The technique reduces a "many-body" interaction problem to a "one-body" problem by introducing the majorizer of the original optimization problem with $L_1$ regularization.
This is a type of mean-field analysis used in statistical mechanics.
We must emphasize that this method does not change the optimal solution, and thus yields the exact optimal point under several optimized cost function conditions.
The remaining sections of the paper are organized as follows.
In the second section, we briefly review Boltzmann machine learning and the recent developments in this area.
In the third section, we introduce majorizer minimization, and obtain the algorithm to resolve the Boltzmann machine learning optimization problem, using $L_1$ regularization.
In the fourth section, we test our method with numerical experiments.
In the last section, we summarize our study.
\section{Boltzmann machine learning}
We assume that the generative model of the data ${\bf x} \in \{-1,1\}^N$ takes the form of the Ising model as
\begin{equation}
P({\bf x}|J,{\bf h}) = \frac{1}{Z(J,{\bf h})} \exp\left( \sum_{i=1}^N \sum_{j \in \partial i} J_{ij} x_i x_j + \sum_{i=1}^N h_ix_i \right),\label{GB}
\end{equation}
where $J_{ij}$ is a pair-wise interaction, $h_i$ is a bias, and $Z(J,{\bf h})$ is the partition function.
The sets of $J_{ij}$ and $h_i$ are denoted as $J$ and ${\bf h}$.
The number of components is represented by $N$.
The summation $j \in \partial i$ is calculated by summing the adjacent components to one denoted by $i$.
Boltzmann machine learning is used to estimate $J_{ij}$ and $h_i$ from snapshots of spin configurations, namely the given data, ${\bf x}^{(k)}$ for $k=1,2,\cdots,D$ by use of the Gibbs-Boltzmann distribution of the Ising model as in Eq. (\ref{GB}).
The standard method to estimate the parameters $J$ and ${\bf h}$ is the maximum-likelihood estimation \cite{Bishop2006} as
\begin{equation}
\left\{ J, {\bf h} \right\} = \arg \max_{J,{\bf h}}\left\{ \sum_{k=1}^D \log P({\bf x}^{(k)}|J,{\bf h})\right\}.
\end{equation}
In other words, we minimize the KL divergence between the generative model's distribution and the empirical distribution of the given data defined as
\begin{equation}
P_{\mathcal{D}}({\bf x}) = \frac{1}{D} \sum_{k=1}^D \delta({\bf x}- {\bf x}^{(k)}).
\end{equation}
The minimization of KL divergence
\begin{equation}
\min_{J,{\bf h}} {\rm KL}(P_{\mathcal{D}}({\bf x})|P({\bf x}|J,{\bf h})) = \min_{J,{\bf h}}\left\{ \sum_{\bf x} P_{\mathcal{D}}({\bf x}) \log \left(\frac{P_{\mathcal{D}}({\bf x})}{P({\bf x}|J,{\bf h})}\right)\right\}
\end{equation}
yields the maximum-likelihood estimation.
However, the computational time is excessive, because the method demands evaluation of the partition function depending on $J$ and ${\bf h}$.
Therefore, we require an effective technique to either approximate the partition function or avoid the computation of the partition function.
In the present study, we selected the latter technique.
One of the simplest methods to mitigate the computation of the log-likelihood function in Boltzmann machine learning is the pseudo-likelihood estimation \cite{Besag1975,Ekeberg2013}.
We change the cost function in the maximum-likelihood estimation, which has no terms in common with the partition function, as
\begin{equation}
\sum_{k=1}^D \log P({\bf x}^{(k)}|J,{\bf h} ) \approx \sum_{k=1}^D \log \prod_{i=1}^N P(x_i|J,{\bf h},{\bf x}^{(k)}_{/i}),
\end{equation}
where
\begin{equation}
P(x_i|J,{\bf h},{\bf x}_{/i}) = \frac{1}{Z_i(J,{\bf h}|{\bf x}_{/i})}\exp\left\{ \sum_{j \in \partial i} J_{ij} x_i x_j + h_i x_i \right\}
\end{equation}
and
\begin{equation}
Z_i(J,{\bf h}|{\bf x}_{/i}) = \sum_{x_i}\exp\left( \sum_{j \in \partial i} J_{ij} x_i x_j + h_i x_i \right) = 2\cosh\left( \sum_{j \in \partial i} J_{ij} x_j + h_i\right).
\end{equation}
In the following, we deal with the minimization problem and take the negative of the approximated quantity as the cost function, that is
\begin{equation}
\mathcal{L}_{\rm PL}(J,{\bf h}) = - \sum_{k=1}^D \log \prod_{i=1}^N P(x_i|J,{\bf h},{\bf x}^{(k)}_{/i}).
\end{equation}
This appears to be a type of mean-field analysis, but the pseudo-likelihood estimation asymptotically (large amount of training data) coincides with the maximum-likelihood estimation.
This method is very simple and easy to implement, but requires an excessive amount of data.
Another technique for changing the cost function is the minimum probability flow \cite{Sohl-Dickstein2011}.
This method was inspired by relaxation dynamics, starting from the empirical distribution determined by the given data toward the distribution, using tentative parameters.
Relaxation dynamics are implemented by a master equation as
\begin{equation}
\frac{dP_t({\bf x})}{dt} = \sum_{\bf y}W({\bf x}|{\bf y})P_t({\bf y}),
\end{equation}
where $W({\bf x}|{\bf y})$ is the transition rate matrix.
We impose a one-spin flip at each update and detailed balance condition as
\begin{eqnarray}
W({\bf x}^{(l)}|{\bf x}^{(k)}) = \exp\left\{- \frac{1}{2}\left(E({\bf x}^{(l)}|J,{\bf h}) - E({\bf x}^{(k)}|J,{\bf h})\right)\right\} && {\rm for}~ \sum_{i=1}^Nx^{(k)}_ix^{(l)}_i = N-2,\label{TM}
\end{eqnarray}
where
\begin{equation}
E({\bf x}|J,{\bf h}) = - \sum_{i=1}^N \sum_{j \in \partial i} J_{ij} x_i x_j - \sum_{i=1}^N h_ix_i.
\end{equation}
The choice of the transition matrix is very important in the following manipulation of the minimum probability flow.
The maximum likelihood estimation is computationally intractable due to the computation of the partition function.
We remove the dependence on the partition function by choosing the local update rule in the transition matrix as in Eq. (\ref{TM}).
For instance, the Metropolis method and heat-bath method can be applied to the minimum probability flow.
In the present study, we follow the original formulation of the minimum probability flow in the literature\cite{Sohl-Dickstein2011} for its symmetric form in computation as shown below.
If we tune the parameters adequately for the empirical distribution of the given data, the change from the initial distribution, namely the empirical distribution $P_{\mathcal{D}}({\bf x})$, is expected to be small; otherwise, it becomes large.
To capture this expectation, we then compute the following infinitesimal change of the KL divergence as
\begin{equation}
{\rm KL}(P_0({\bf x})|P_t({\bf x})) \approx {\rm KL}(P_0({\bf x})|P_0({\bf x})) + dt \frac{d}{dt}\left.{\rm KL}(P_0({\bf x})|P_t({\bf x}))\right\rvert_{t=0}.
\end{equation}
The combination of elementary algebra and the master equation leads up to the first order of $dt$ as
\begin{equation}
{\rm KL}(P_0({\bf x})|P_t({\bf x})) \approx \frac{dt}{D}\sum_{k=1}^D \sum_{l \notin \mathcal{D}|l \in \partial k}\exp\left\{ \frac{1}{2}\left(E({\bf x}^{(k)}|J,{\bf h})-E({\bf x}^{(l)}|J,{\bf h})\right)\right\}.\label{MPF}
\end{equation}
The true parameters are then estimated by minimization of this quantity.
This is the minimum probability flow method.
Notice that we do not require to manipulate the Markov chain Monte Carlo (MCMC) simulation, although the method is inspired by stochastic dynamics.
This is different from contrastive divergence, which requires computation by MCMC \cite{Welling2002}.
Once we impose the stochastic dynamics rule and the detailed balanced condition, we immediately compute the above quantity.
Thus, we utilize Eq. (\ref{MPF}) as the cost function to be minimized for estimating the parameters, instead of the log-likelihood function as in the maximum likelihood estimation; that is
\begin{eqnarray}
\mathcal{L}_{\rm MPF}(J,{\bf h}) = \frac{1}{D}\sum_{k=1}^D \sum_{l \notin \mathcal{D}}\exp\left\{ \frac{1}{2}\left(E({\bf x}^{(k)}|J,{\bf h})-E({\bf x}^{(l)}|J,{\bf h})\right)\right\},
\end{eqnarray}
where the summation over $l$ results in the case satisfying ${\sum_{i=1}^N x_i^{(k)}x_i^{(l)} = N-2}$.
The performance, estimation precision, and computational efficiency often exceed those of the pseudo-likelihood estimation for the same amount of data.
In the present study, we employ these methods to estimate the parameters; the following discussion can be straightforwardly applied to them.
Above all, we assume that parameters $J$ and ${\bf h}$ are assigned to all pairs and all components.
However, in order to elucidate the most relevant pair-wise interactions and biases from the given data, we employ an additional technique to prune less significant parameters.
A candidate is required to utilize the regularization of the $L_1$ norm \cite{Bishop2006}.
Let us then minimize the cost function $\mathcal{L}$ (=$\mathcal{L}_{\rm MPF}$ or $\mathcal{L}_{\rm PL}$) with $L_1$ norm as
\begin{equation}
\min_{J,{\bf h}}\left\{\lambda_J \sum_{(ij)}|J_{ij}| + \lambda_h \sum_{i=1}^N |h_i| + \mathcal{L}(J,{\bf h}) \right\}.
\end{equation}
The regularization technique was originally designed to obtain a unique estimation from underdetermined equations by imposing additional conditions.
Therefore, estimations that utilize regularization lead to stable solutions, even from small amounts of data.
As compensation, the entire cost function is not smooth, owing to the existence of the absolute value function.
The non-smoothness impedes the simple application of the gradient method, which identifies the minimal point of the cost function.
For the absolute value function, we may prepare several types of imitating functions.
However, this type of approximation does occasionally generate incorrect estimations, and reduces the convergence rate.
Instead of the original optimization problem with a non-smooth term, let us utilize a different function sharing the same optimal point below, that is the majorizer minimization.
\section{Majorizer minimization}
We briefly review majorizer minimization for convenience.
In general, we consider the optimization problem by minimizing a convex function $f$ with $N$-dimensional variables, which is assumed to be differentiable; its derivative $\nabla f({\bf x})$ is Lipschitz.
When the derivative is Lipschitz, there is a constant $L \ge 0$ for any ${\bf a}$ and ${\bf b}$
\begin{equation}
\sum_{k=1}^N\left( \left.\frac{\partial f}{\partial x_k}\right\rvert_{{\bf x}={\bf a}} - \left.\frac{\partial f}{\partial x_k}\right\rvert_{{\bf x}={\bf b}} \right)^2 \le L\sum_{k=1}^N\left(a_k - b_k \right)^2,
\end{equation}
where $L$ is termed as the Lipschitz constant and $a_k$ and $b_k$ are the $k$th component of $N$-dimensional vectors ${\bf a}$ and ${\bf b}$.
The majorizer of the function $f$ is then given by the following quadratic function
\begin{equation}
g({\bf x},{\bf v}) = f({\bf v}) + \sum_{k=1}^N \left.\frac{\partial f}{\partial x_k}\right\rvert_{{\bf x}={\bf v}}(x_k -v_k) + \frac{L}{2}\sum_{k=1}^N \left( x_k - v_k \right)^2.
\end{equation}
The majorizer always satisfies
\begin{equation}
f({\bf x}) \le g({\bf x},{\bf v}) \le f({\bf v}).
\end{equation}
Let us then consider the following optimization problem.
\begin{equation}
{\bf x}^{t+1} = \arg \min_{\bf x}\left\{ g({\bf x},{\bf x}^t) \right\}.
\end{equation}
The sequence of the optimal solutions $[{\bf x}^0,{\bf x}^1,\cdots,{\bf x}^T]$ satisfies
\begin{equation}
f({\bf x}^{t+1}) \le g({\bf x}^{t+1},{\bf x}^t) \le f({\bf x}^t)
\end{equation}
for $t=0,1,\cdots,T-1$.
This property of the majorizer gradually approaches the optimal solution of the original minimization problem.
This technique is referred to as the majorizer minimization approach, which is one of the gradient methods.
The convergence rate is known as $f({\bf x}^t) - f({\bf x}^*) = O(1/t)$, where the asterisk stands for the optimal solution.
When we utilize the regularization obtained with the $L_1$ norm, we solve the following optimization problem
\begin{equation}
{\bf x}^{t+1} = \arg \min_{\bf x}\left\{ g({\bf x},{\bf x}^t) + \lambda \left\| {\bf x}\right\|_1 \right\},
\end{equation}
where $\left\|{\bf x} \right\|_1 = \sum_{k=1}^N|x_k|$.
Because the majorizer is quadratic and the $L_1$ norm is separable, the optimal solution can be analytically obtained as
\begin{equation}
x_k^{t+1} = \eta_{\lambda/L}\left( x_k^{t} + \frac{1}{L} \left.\frac{\partial f}{\partial x_k}\right\rvert_{{\bf x}={\bf x}^t} \right),
\end{equation}
where
\begin{equation}
\eta_a(x)= {\rm sign}(x) (|x| - a).
\end{equation}
Therefore, solving alternative optimization problems is reduced to a simple substitution using the tentative solution ${\bf x}^t$.
The majorizer minimization method is broadly used in compressed sensing methods, which reconstruct original inputs from undersampled outputs.
In this problem, the original inputs should be sparse.
$L_1$-regularization enforces a sparse solution for the inference problem of the original signals.
Similarly, let us utilize the majorizer minimization method for estimation of the Boltzmann machine learning parameters.
Let us remark the role of the majorizer in short.
The majorizer modifies the original optimization problem into quadratic form.
The quadratic form separates the dependence on each component.
In other words, the many-body interaction system with the original function $f$ is changed into a one-body independent system consisting of the majorizer.
This is a type of mean-field analysis, which approximates the many-body interactions into an effective one-body description.
In statistical mechanics, the law of large numbers is imposed on the number of components $N$ to perform mean-field analysis and validation.
However, in this method, we do not require a large number of components; we only require the property of function $f$.
In this sense, it is a very generic yet powerful technique.
Let us apply the majorizer minimization approach to Boltzmann machine learning with $L_1$ regularization.
Because the pseudo-likelihood function and cost function in the minimum probability flow are differentiable and convex \cite{Sohl-Dickstein2011}, the majorizer minimization method can be applied.
The majorizer for Boltzmann machine learning is given as
\begin{eqnarray}\nonumber
G(J',{\bf h}';J,{\bf h}) &=& \mathcal{L}(J,{\bf h}) + \sum_{(ij)}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial J_{ij}}\right\rvert_{J,{\bf h}}\left(J'_{ij} - J_{ij} \right) + \frac{L_J}{2}\sum_{(ij)}\left(J'_{ij} - J_{ij}\right)^2
\\
&& \quad + \sum_{i}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial h_i}\right\rvert_{J,{\bf h}}\left(h'_{i} - h_{i} \right) + \frac{L_h}{2}\sum_{i}\left(h'_i - h_i \right)^2,
\end{eqnarray}
where $L_J$ and $L_h$ satisfy
\begin{eqnarray}
\sum_{(ij)}\left(\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial J_{ij}}\right\rvert_{A,{\bf h}} - \left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial J_{ij}}\right\rvert_{B,{\bf h}} \right)^2 \le L_J \sum_{(ij)} \left(A_{ij} - B_{ij} \right)^2\\
\sum_{i}\left(\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial h_{i}}\right\rvert_{J,{\bf a}} - \left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial h_{i}}\right\rvert_{J,{\bf b}} \right)^2 \le L_h \sum_{i} \left(a_{i} - b_{i} \right)^2.
\end{eqnarray}
Following the prescription of the majorizer minimization approach, let us iteratively solve the optimization problem
\begin{equation}
\left\{J^{t+1},{\bf h}^{t+1}\right\} = \arg \min_{J,{\bf h}}\left\{ \lambda_{J}\sum_{(ij)}|J_{ij}| + \lambda_h \sum_{i}|h_i| + G(J,{\bf h};J^{t},{\bf h}^t)\right\}.
\end{equation}
Because the dependence of $J$ and ${\bf h}$ on the majorizer is separate, we independently solve the optimization problem for each parameter as
\begin{eqnarray}
J_{ij}^{t+1} &=& \eta_{\lambda_J/L_J}\left( J_{ij}^{t} + \frac{1}{L_J}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial J_{ij}}\right\rvert_{J^t,{\bf h}^t}\right)\\
h_{i}^{t+1} &=& \eta_{\lambda_h/L_h}\left( h_{i}^{t} + \frac{1}{L_h}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial h_{i}}\right\rvert_{J^t,{\bf h}^t}\right).
\end{eqnarray}
The majorizer minimization method is a generic technique for reaching a minimum point by recursive manipulation, under the assumption that the cost function is convex and its derivative is Lipschitz.
These conditions are satisfied in the cost functions of the pseudo-likelihood function and minimum probability flow.
The derivatives of the pseudo-likelihood function yield
\begin{eqnarray}
- \frac{\partial \mathcal{L}_{\rm PL}(J,{\bf h})}{\partial J_{ij}} &=& \frac{1}{D}\sum_{k=1}^D x^{(k)}_ix^{(k)}_j - \frac{1}{D}\sum_{k=1}^D\sum_{i=1}^Nx_j^{(k)}\tanh\left( \sum_{j \in \partial i} J_{ij}x_j^{(k)} + h_i \right) \\
- \frac{\partial \mathcal{L}_{\rm PL}(J,{\bf h})}{\partial h_{i}} &=& \frac{1}{D}\sum_{k=1}^D x^{(k)}_i - \frac{1}{D}\sum_{k=1}^D\sum_{i=1}^N\tanh\left( \sum_{j \in \partial i} J_{ij}x_j^{(k)} + h_i \right).
\end{eqnarray}
In these cases, it is difficult to compute the Lipschitz constant.
We may use the backtracking technique, in which we gradually tune $L_J$ and $L_h$ by some rule such that
\begin{equation}
\mathcal{L}_{\rm PL}(J^{t+1},{\bf h}^{t+1}) \le G(J^{t+1},{\bf h}^{t+1}|J^t,{\bf h}^t).
\end{equation}
In addition, the case of the minimum probability flow is evaluated as
\begin{eqnarray}
\frac{\partial \mathcal{L}_{\rm MPF}(J,{\bf h})}{\partial J_{ij}} &=& \frac{1}{D}\sum_{k=1}^D \sum_{l \notin \mathcal{D} }\left( x^{(k)}_ix^{(k)}_j - x^{(l)}_ix^{(l)}_j \right)\exp\left\{\frac{1}{2}\left(E({\bf x}^{(k)}|J,{\bf h}) - E({\bf x}^{(l)}|J,{\bf h})\right)\right\} \label{MPF1}\\
\frac{\partial \mathcal{L}_{\rm MPF}(J,{\bf h})}{\partial h_{i}} &=& \frac{1}{D}\sum_{k=1}^D \sum_{l \notin \mathcal{D} }\left( x^{(k)}_i - x^{(l)}_i \right)\exp\left\{\frac{1}{2}\left(E({\bf x}^{(k)}|J,{\bf h}) - E({\bf x}^{(l)}|J,{\bf h})\right)\right\}.\label{MPF2}
\end{eqnarray}
These gradients are reduced for one-spin flips, using $\sum_{i=1}^N x_i^{(k)}x_i^{(l)} = N -2$
\begin{eqnarray}
\frac{\partial \mathcal{L}_{\rm MPF}(J,{\bf h})}{\partial J_{ij}} &=& \frac{2}{D}\sum_{k=1}^D\sum_{i =1}^N \sum_{j \in \partial i} x^{(k)}_ix^{(k)}_j \exp\left\{\sum_{n \in \partial i}J_{ij}x^{(k)}_ix^{(k)}_n + h_i x^{(k)}_i\right) \\
\frac{\partial \mathcal{L}_{\rm MPF}(J,{\bf h})}{\partial h_{i}} &=& \frac{2}{D}\sum_{k=1}^D \sum_{i =1}^N x^{(k)}_i \exp\left\{\sum_{j \in \partial i}J_{ij}x^{(k)}_ix^{(k)}_j + h_i x^{(k)}_i\right).
\end{eqnarray}
where we assume that $i$th spin is flipped from the $k$th spin configuration (this is the $l$th configuration in the summation in Eqs. (\ref{MPF1}) and (\ref{MPF2})).
Similarly, we may use the backtracking technique such that $L_J$ and $L_h$ hold
\begin{equation}
\mathcal{L}_{\rm MPF}(J^{t+1},{\bf h}^{t+1}) \le G(J^{t+1},{\bf h}^{t+1}|J^t,{\bf h}^t).
\end{equation}
An acceleration technique is available for the majorizer minimization method \cite{Beck2009}.
We modify the update rule into
\begin{eqnarray}\nonumber
J_{ij}^{t+1} &=& \eta_{\lambda_J/L_J}\left( J_{ij}^{t} + \frac{1}{L_J}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial J_{ij}}\right\rvert_{J^t,{\bf h}^t}\right) + \left( \frac{\beta_{t} - 1}{\beta_{t+1}}\right)\left(\eta_{\lambda_J/L_J}\left( J_{ij}^{t} + \frac{1}{L_J}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial J_{ij}}\right\rvert_{J^t,{\bf h}^t}\right) - J_{ij}^t\right) \\
\\ \nonumber
h_i^{t+1} &=& \eta_{\lambda_h/L_h}\left( h_{i}^{t} + \frac{1}{L_h}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial h_{i}}\right\rvert_{J^t,{\bf h}^t}\right) + \left( \frac{\beta_{t} - 1}{\beta_{t+1}}\right)\left(\eta_{\lambda_h/L_h}\left( h_{i}^{t} + \frac{1}{L_h}\left.\frac{\partial \mathcal{L}(J,{\bf h})}{\partial h_{i}}\right\rvert_{J^t,{\bf h}^t}\right) - h_{i}^t\right),\\
\end{eqnarray}
where
\begin{equation}
\beta_{t+1} = \frac{1+\sqrt{1+4\beta_t^2}}{2}.
\end{equation}
The initial condition is $\beta_0 = 1$.
In this update rule, the convergence speed is improved as $\sum_{(ij)}\left( J^t_{ij} - J^*_{ij}\right)^2$ and $\sum_{i}\left( h^t_{i} - h^*_{i}\right)^2 \sim O(1/t^2)$, where the asterisk denotes the optimal solution.
\section{Numerical test}
We conducted several numerical experiments to test the estimation of sparse interactions.
The spin configurations were generated from the Markov chain Monte Carlo simulations.
The linear size $N_{L}=5$; that is, the entire spin $N = N_L^2 = 25$.
The number of interactions was $N^2=625$; the number of biases was $N=25$.
The true parameters for the biases were given by a Gaussian distribution with zero mean and unit variance.
In contrast, the true parameters for the interactions were restricted to (i) the random sparse pairs (the non-zero interactions is restricted to $10\%$ of all pairs, namely $62$) and (ii) the nearest neighboring pairs on the square lattice (the number of non-zero interactions $100$).
We assumed that the interactions should be symmetric, namely $J_{ij} = J_{ji}$.
The values for the interactions used random variables that follow a Gaussian distribution with zero mean and unit variance.
The estimation had no prior knowledge of the structure of $J$ and ${\bf h}$.
In other words, the estimator did not know the lattice, and did not know that the non-zero interaction was restricted to specific pairs.
For each method, we estimated the parameters while changing $D$ as $D=100, 500$, $1000$, $2000$, $3000$, and $5000$.
The optimal selection of the coefficient $\lambda$ could not be known a priori.
We then tested several values of $\lambda$ for the estimations of the parameters.
In Fig. \ref{fig1}, we show the averaged performance over $100$ samples after $200$ iterations for the pseudo-likelihood estimation, and $50$ iterations for the minimum probability flow, for a case in which the pair-wise interactions were distributed randomly.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=140mm]{fig1.eps}
\end{center}
\caption{{\protect\small (Color online) Average performance of $L_1$-regularized inference in the case of random sparse interactions.
The horizontal axis denotes the amount of data. The vertical axis stands for the summation of the errors on estimations of $J$ and ${\bf h}$, ${\rm Err}_J+{\rm Err}_h$, which are defined as ${\rm Err}_J = \sqrt{\sum_{(ij)}\left(J_{ij} - J^{(\rm true)}_{ij}\right)^2/\sum_{(ij)}J_{ij}^2}$ and ${\rm Err}_h = \sqrt{\sum_{i}\left(h_i - h^{(\rm true)}_i\right)^2/\sum_{i}h_i^2}$.
The data amounts were $D=100$ (magenta), $D=500$ (yellow), $D=1000$ (cyan), $D=2000$ (red), $D=3000$ (green), and $D=5000$ (blue) from top to bottom.
}}
\label{fig1}
\end{figure}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=160mm]{fig2.eps}
\end{center}
\caption{{\protect\small (Color online) Profile (absolute value) of the pair-wise interactions in the random sparse case (one example).
The left panel shows the original configuration of the pair-wise interactions.
The center panel shows the results of the pseudo-likelihood estimation ($\lambda=0.2$); the right panel shows the results of the minimum probability flow ($\lambda=0.02$).
}}
\label{fig2}
\end{figure}
We note that the convergence speed of the minimum probability flow was significantly faster than the pseudo-likelihood estimation, although the precision of the convergent solutions was comparable.
The numbers of iterations used in both methods were sufficient to obtain the convergent estimations.
Both of the methods could estimate the correct values of the biases and interactions.
In Fig. \ref{fig2}, we show the profile of the estimated interactions for a single sample.
We confirmed that the estimation of the non-zero interactions had been achieved, although their absolute values tended to be smaller than the original values.
This is a characteristic property of the $L_1$ regularization.
We compared the pair-wise interactions and biases to the true parameters, as shown in Fig. \ref{fig3}.
We observe a fairly good performance for the nonzero components of the pair-wise interactions and biases.
The zeros of the pair-wise interactions are obtained as extremely small valued estimations.
We may set some thresholds to prune the irrelevant interactions in the estimation.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=120mm]{fig3.eps}
\end{center}
\caption{{\protect\small (Color online) Comparison of the pair-wise interactions and biases to the true parameters in random sparse interactions (one example).
The vertical axis denotes the true parameters and the horizontal axis stands for the estimated values.
The upper left panel shows the results for $J_{ij}$ in the pseudo-likelihood estimation ($\lambda=0.2$) and the upper right one shows that of the minimum probability flow ($\lambda=0.02$).
The lower panels describe the results for $h_i$ by the pseudo-likelihood estimation (left) and the minimum probability flow (right).
The red lines have a unit slope as a guide to the eye.}}
\label{fig3}
\end{figure}
Figure \ref{fig4} shows the performance averaged over $100$ samples after $200$ iterations of the pseudo-likelihood estimation, and $50$ iterations of the minimum probability flow for a case in which pair-wise interactions were set on the square lattice.
An increase in $D$ improved the precision of the estimation in both methods.
Both methods could lead to precise estimations of the pair-wise interactions and biases.
The profile of the estimated interactions is shown in Fig. \ref{fig5}.
A comparison of the estimated interactions and biases with the true parameters is shown in Fig. \ref{fig6}.
We emphasize that the estimator did not have any prior knowledge of the structure of the interactions.
In this sense, we have succeeded in deriving the relevant structure of the pair-wise interactions from a type of microscopic degrees of freedom snapshot.
This indicates that the microscopic behavior observation characterized the generative model through the estimation, by use of $L_1$ regularization.
In addition, we truncated insignificant parameters with the aid of $L_1$ regularization.
In both cases of the random sparse interactions and the square lattice, we succeeded in reproducing the structure of the pair-wise interactions and estimating the magnitude of the interactions.
We emphasize that the gradient method with majorizer minimization method was replaced by the simple iterative substitution.
The technique we showed is expected to be applied to wide range of applications to seek the relevant interactions and biases generating the data.
In these numerical experiments, we demonstrate the case when we intend to apply our technique to the actual data.
Thus we prepare the specific pair-wise interactions a priori and generate the numerous data.
To further investigate the precision of our method, the hyperparameters $\lambda_J$ and $\lambda_h$ may be assumed to be distributed following the hyperprior distribution.
As shown above, we would find the least square error in the optimal hyperparameters, which correspond to the distributed ones.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=140mm]{fig4.eps}
\end{center}
\caption{{\protect\small (Color online) Average performance of the $L_1$-regularized inference for a case in which a square lattice was used (one example).
The axes are the same as those in Fig. \ref{fig2}.
In this case, we further investigated the dependence on the amount of given data $D$.
The data amounts were $D=100$ (magenta), $D=500$ (yellow), $D=1000$ (cyan), $D=2000$ (red), $D=3000$ (green), and $D=5000$ (blue) from top to bottom.
}}
\label{fig4}
\end{figure}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=160mm]{fig5.eps}
\end{center}
\caption{{\protect\small Profile (absolute value) of the pair-wise interactions for a case in which a square lattice was used (one example).
The left panel shows the original configuration of the pair-wise interactions.
The center panel describes the estimation derived by the pseudo-likelihood estimation ($\lambda=0.1$) and the right panel shows the estimation derived by the minimum probability flow ($\lambda=0.018$).}}
\label{fig5}
\end{figure}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=120mm]{fig6.eps}
\end{center}
\caption{{\protect\small Comparison of the pair-wise interactions and biases to the true parameters for a case in which a square lattice was used.
The symbols and axes are the same as those in Fig. \ref{fig3}}}
\label{fig6}
\end{figure}
\section{Summary}
In this study, we analyzed Boltzmann machine learning in terms of pseudo-likelihood estimation and minimum probability flow.
In order to elucidate the most relevant parameters generating the data, we sought a sparse solution in the present study.
This task was very important for determining the structure of the data while pruning irrelevant parameters.
$L_1$ regularization was beneficial in obtaining a sparse solution by solving a given cost function.
However, in general, the non-smoothness of the $L_1$ norm hampered the direct manipulation of the gradient method, which is intended to minimize the cost function.
This study featured the implementation of the majorizer minimization method into the Boltzmann machine learning technique.
The majorizer minimization method is a type of mean-field analysis, which enabled us to express a many-body interacting system in terms of an effective one-body independent system.
We tested our method to elucidate the randomly distributed interactions, and those between the adjacent spins on the square lattice, without any prior knowledge.
The performance of our method is fairly satisfactory, as expected.
Increasing the amount of given data improved the precision of the estimations and enhanced the efficacy of the $L_1$ regularization.
In present study, the cost functions are given by the pseudo likelihood function as well as the minimum probability flow.
The former one is generalized to the composite pseudo likelihood function inspired by the cluster variational method \cite{Yasuda2012proc}.
In this kind of generalization, the majorizer minimization is applicable.
In this sense, our scheme is very flexible.
Notice that our numerical experiments were assumed to be an extremely generic case, that is with in homogenous pair-wise interactions and biases.
One might intend to infer the homogeneous property from the given data.
The necessary number for precise estimations should then be extremely reduced.
The recent study improves precision of the Boltzmann machine learning with the comparable number of the data by aid of the Belief propagation to estimate the average and variance from the empirical data \cite{Yasuda2015}.
We anticipate that future studies will apply our present method to actual observed data, to elucidate the essential property from nature.
\section*{Acknowledgement}
The present work is performed by the financial support from the JST-CREST, MEXT KAKENHI Grants No. 251200008 and 24740263 and the Kayamori Foundation of Informational Science Advancement.
|
\section{Introduction}
\label{sec:introduction}
Tensor decompositions and their applications in mining multi-aspect datasets are ubiquitous and ever increasing in popularity. Data Mining application of these techniques has been largely pioneered by the work of Kolda et al. \cite{kolda2005higher} where the authors introduce a topical aspect to a graph between webpages, and extend the popular HITS algorithm in that scenario. Henceforth, the field of multi-aspect/tensor data mining has witnessed rich growth with prime examples of applications being citation networks \cite{kolda2008scalable}, computer networks\cite{kolda2008scalable,maruhashi2011multiaspectforensics,papalexakis2012parcube}, Knowledge Base data \cite{kang2012gigatensor,papalexakis2012parcube,chang2013multi,chang2014typed}, and social networks \cite{bader2007temporal,kolda2008scalable,lin2009metafac,papalexakis2012parcube,jiang2014fema}, to name a few.
Tensor decompositions are undoubtedly a very powerful analytical tool with a rich variety of applications. However there exist research challenges in the field of data mining that need to be addressed, in order for tensor decompositions to claim their position as a de-facto tool for practicioners.
One challenge, which has received considerable attention, is the one of making tensor decompositions scalable to today's web scale. For instance, Facebook has around 2 billion users at the time of writing of this paper and is ever growing, and making tensor decompositions able to work on even small portions of the entire Facebook network is imperative for the adoption of these techniques by such big players. Very frequently, data that fall under the aforementioned category turn out to be highly sparse; the reason is that, e.g. each person on Facebook interacts with only a few hundreds of the users. Computing tensor decompositions for highly sparse scenarios is a game changer, and exploiting sparsity is key in scalability. The work of Kolda et al. \cite{kolda2005higher,SIAM-67648} introduced the first such approach of exploiting sparsity for scalability. Later on, distributed approaches based on the latter formulation \cite{kang2012gigatensor}, or other scalable approaches \cite{papalexakis2012parcube,erdos2013walk,beutelflexifact,de2014distributed} have emerged. By no means do we claim that scalability is a solved problem, however, we point out that there has been significant attention to it.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width = 0.45\textwidth]{crown_jewel.eps}
\end{center}
\caption{Starting from an unsupervised, exploratory application, \textsc{Scoup-Smt}\xspace automatically determines a solution with high quality, outperforming existing baselines, and enables discoveries in real data.}
\end{figure}
The main focus of this work, however, is on another, relatively less explored territory; that of assessing the {\em quality} of a tensor decomposition. In a great portion of tensor data mining, the task is exploratory and unsupervised: we are given a dataset, usually without any sort of ground truth, and we seek to extract {\em interesting} patterns or concepts from the data. It is crucial, therefore, to know whether a pattern that we extract actually models the data at hand, or whether it is merely modelling noise in the data. Especially in the age of Big Data, where feature spaces can be vast, it is imperative to have a measure of quality and avoid interpreting noisy, random variation that always exists in the data. Determining the ``right'' number of components in a tensor is a very hard problem \cite{hillar2013most}. This is why, many seminal exploratory tensor mining papers, understandably, set the number of components manually \cite{kolda2006tophits,sun2006beyond,bader2007temporal,kolda2008scalable}. When there is a specific task at hand, e.g. link prediction \cite{dunlavy2011temporal}, recommendation \cite{rendle2010pairwise}, and supervised learning \cite{tao2005supervised,he2014dusk}, that entails some measure of success, then there is some procedure (e.g. cross-validation) for selecting a good number of latent components which unfortunately cannot generalize to the case where labels or ground truth are absent.
However, not all hope is lost. There have been very recent approaches following the Minimum Description Length (MDL) principle \cite{araujo2014com2,metzler2015clustering}, where the MDL cost function usually depends heavily on the application at hand (e.g. community detection or boolean tensor clustering respectively). Additionally, there have been Bayesian approaches \cite{zhaobayesian} that, as in the MDL case, do not require the number of components as input. These approaches are extremely interesting, and we reserve their deeper investigation in future work, however in this work, we choose to operate on top of a different, very intuitive approach which takes into account properties of the PARAFAC decomposition \cite{PARAFAC} and is {\em application independent}, requiring no prior knowledge about the data; there exists highly influential work in the Chemometrics literature \cite{bro1998multi} that introduces heuristics for determining a good rank for tensor decompositions. Inspired by and drawing from \cite{bro1998multi}, we provide a comprehensive method for mining multi-aspect datasets using tensor decompositions.
Our contributions are:
\begin{itemize}\setlength{\itemsep}{0.0mm}
\item{\bf Algorithms} We propose \textsc{Scoup-Smt}\xspace, a comprehensive methodology on mining multi-aspect datasets using tensors, which minimizes manual trial-and-error intervention and provides quality characterization of the solution (Section \ref{sec:method_final}). Furthermore, we extend the quality assessment heuristic of \cite{bro1998multi} assuming KL-divergence, which has been shown to be more effective in highly sparse, count data \cite{chi2012tensors} (Section \ref{sec:method_corcondia}).
\item{\bf Evaluation \& Discovery} We conduct a large scale study on 10 real datasets, exploring the structure of hidden patterns within these datasets. To the best of our knowledge, this is the first such broad study. (Section \ref{sec:study1}). As a data mining case study, we apply \textsc{Scoup-Smt}\xspace to two real datasets discovering meaningful patterns (Section \ref{sec:study2}). Finally, we extensively evaluate our proposed method in synthetic data (Section \ref{sec:experiments}).
\end{itemize}
In order to encourage reproducibility, most of the datasets used are public, and we make our code publicly available at {\small \\ \url{http://www.cs.cmu.edu/~epapalex/src/AutoTen.zip}}.
\section{Background}
\label{sec:formulation}
Table \ref{tab:symbols} provides an overview of the notation used in this and subsequent sections.
\begin{table}[htbp]
\small
\begin{center}
\begin{tabular}{l | l}
{\bf Symbol} & {\bf Definition} \\ \hline
$\tensor{X}, \mathbf{X}, \mathbf{x}, x$ & Tensor, matrix, column vector, scalar \\ \hline
$\circ$ & outer product \\ \hline
$vec(~)$ & vectorization operator \\\hline
$\otimes$ & Kronecker product \\ \hline
$\mathbf{X}^\dag$ & Moore-Penrose pseudoinverse \\\hline
$\ast ~~\oslash$ & element-wise multiplication and division \\\hline
$\mathbf{A}^\dag$ & Moore-Penrose Pseudoinverse of $\mathbf{A}$ \\\hline
$D_{KL}(a||b)$ & KL-Divergence \\\hline
$\|\mathbf{A}\|_F$ & Frobenius norm \\\hline
\multirow{2}{*}{\textsc{KronMatVec}\xspace} & efficient computation of \\ & $\mathbf{y} = \left(\mathbf{A}_1 \otimes \mathbf{A}_2 \otimes \cdots \otimes \mathbf{A}_n \right)\mathbf{x} $
\cite{buis1996efficient} \\\hline
$\mathbf{x}(i)$ & $i$-th entry of $\mathbf{x}$ (same for matrices and tensors)\\ \hline
$\mathbf{X}(:,i)$ & spans the entire $i$-th column of $\mathbf{X}$ (same for tensors) \\ \hline
$\mathbf{x}^{(k)}$ & value at the $k$-th iteration \\ \hline
\texttt{CP\_NMU}\xspace & non-negative, Frobenius norm PARAFAC \cite{tensortoolbox}\\ \hline
\texttt{CP\_APR}\xspace & KL-Divergence PARAFAC \cite{chi2012tensors} \\ \hline
\end{tabular}
\caption{Table of symbols}
\label{tab:symbols}
\end{center}
\end{table}
\subsection{Brief Introduction to Tensor Decompositions}
Given a tensor $\tensor{X}$, we can decompose it according to the CP/PARAFAC decomposition \cite{PARAFAC} (henceforth referred to as PARAFAC) as a sum of rank-one tensors:
\[
\tensor{X} \approx \displaystyle{ \sum_{f=1}^{F} \mathbf{a}_f \circ \mathbf{b}_f \circ \mathbf{c}_f }
\]
where the $(i,j,k)$ entry of $\mathbf{a}_r\circ \mathbf{b}_r \circ \mathbf{c}_r$ is $\mathbf{a}_r(i)\mathbf{b}_r(j)\mathbf{c}_r(k)$. Usually, PARAFAC is represented in its matrix form $[\mathbf{A,B,C} ]$, where the columns of matrix $\mathbf{A}$ are the $\mathbf{a}_r$ vectors (and accordingly for $\mathbf{B,C}$).
The PARAFAC decomposition is especially useful when we are interested in extracting the true latent factors that generate the tensor. In this work, we choose the PARAFAC decomposition as our tool, since it admits a very intuitive interpretation of its latent factors; each component can be seen as soft co-clustering of the tensor, using the high values of vectors $\mathbf{a}_r, \mathbf{b}_r, \mathbf{c}_r$ as the membership values to co-clusters.
Another very popular Tensor decomposition is Tucker3 \cite{kroonenberg1980principal}, where a tensor is decomposed into rank-one factors times a core tensor:
\[
\tensor{X} \approx \displaystyle{ \sum_{p=1}^{P} \sum_{q=1}^{Q} \sum_{r=1}^{R} \tensor{G}(p,q,r)\mathbf{u}_p \circ \mathbf{v}_q \circ\mathbf{w}_r }
\]
where $\mathbf{U,V,W}$ are orthogonal.
The Tucker3 model is especially used for compression. Furthermore, PARAFAC can be seen as a restricted Tucker3 model, where the core tensor $\tensor{G}$ is super-diagonal, i.e. non-zero values are only in the entries where $i=j=k$. This observation will be useful in order to motivate the CORCONDIA diagnostic.
Finally, there also exist more expressive, but harder to interpret models, such as the Block Term Decomposition (BTD) \cite{de2008decompositions} , however, we reserve future work for their investigation.
\subsection{Brief Introduction to CORCONDIA}
As outlined in the Introduction, in the chemometrics literature, there exists a very intuitive heuristic by the name of CORCONDIA \cite{bro2003new}, which can serve as a guide in judging how well a given PARAFAC decomposition is modelling a tensor.
In a nutshell, the idea behind CORCONDIA is the following: Given a tensor $\tensor{X}$ and its PARAFAC decomposition $\mathbf{A,B,C}$, one could imagine fitting a Tucker3 model where matrices $\mathbf{A,B,C}$ are the factors of the Tucker3 decomposition and $\tensor{G}$ is the core tensor (which we need to solve for). Since, as we already mentioned, PARAFAC can be seen as a restricted Tucker3 decomposition with super-diagonal core tensor, if our PARAFAC modelling of $\tensor{X}$ using $\mathbf{A,B,C}$ is modelling the data well, the core tensor $\tensor{G}$ should be as close to super-diagonal as possible. If there are deviations from the super-diagonal, then this is a good indication that our PARAFAC model is somehow flawed (either the decomposition rank is not appropriate, or the data do not have the appropriate structure).
We can pose the problem as the following least squares problem:
\[
\min_{\tensor{G}} \| vec\left( \tensor{X} \right) - \left( \mathbf{A}\otimes \mathbf{B} \otimes \mathbf{C} \right) vec \left( \tensor{G} \right)\|_F^2
\]
with the least squares solution:
$
vec\left(\tensor{G} \right) = \left( \mathbf{A}\otimes \mathbf{B} \otimes \mathbf{C} \right)^\dag vec \left( \tensor{X} \right)
$
After computing $\tensor{G}$, the CORCONDIA diagnostic can be computed as
{\small
$
\displaystyle{ c = 100\left(1 - \frac{\sum_{i=1}^{F} \sum_{j=1}^{F} \sum_{k=1}^{F} \left(\tensor{G}(i,j,k) - \tensor{I}(i,j,k)\right)^2 }{F} \right)},
$
}
where $\tensor{I}$ is a super-diagonal tensor with ones on the $(i,i,i)$ entries. For a perfectly super-diagonal $\tensor{G}$ (i.e. perfect modelling), $c$ will be 100.
One can see that for rank-one models, the metric will always be 100, because the rank one component can trivially produce a single element ``super-diagonal'' core; thus, CORCONDIA is applicable for rank two or higher.
According to \cite{bro2003new}, values below 50 show some imperfection in the modelling or the rank selection; the value can also be negative, showing severe problems with the modelling. In \cite{bro2003new}, some of the chemical data analyzed have perfect, low rank PARAFAC structure, and thus expecting $c>50$ is reasonable. In many data mining applications, however, due to high data sparsity, data cannot have such perfect structure, but an {\em approximation} thereof using a low rank model is still very valuable. Thus, in our case, we expand the spectrum of acceptable solutions with reasonable quality to include smaller, positive values of $c$ (e.g. 20 or higher).
\subsection{Scaling Up CORCONDIA}
As we mention in the Introduction CORCONDIA as it's introduced in \cite{bro1998multi} is suitable for small and dense data. However, this contradicts the area of interest of the vast majority of data mining applications. To that end, very recently \cite{papalexakis2015fastcorcondia} we extended CORCONDIA to the case where our data are large but sparse, deriving a fast and efficient algorithm. Key behind \cite{papalexakis2015fastcorcondia} is avoiding to pseudoinvert $\left( \mathbf{A \otimes B \otimes C} \right)
In order to achieve the above, we need to reformulate the computation of CORCONDIA.
The pseudoinverse
$
\left( \mathbf{A} \otimes \mathbf{B} \otimes \mathbf{C} \right)^\dag
$
can be rewritten as
{\small
\[
\left( \mathbf{V_a} \otimes \mathbf{V_b} \otimes \mathbf{V_c} \right) \left( \mathbf{\Sigma_a}^{-1} \otimes \mathbf{\Sigma_b}^{-1} \otimes \mathbf{\Sigma_c}^{-1} \right) \left( \mathbf{U_a}^T \otimes \mathbf{U_b}^T \otimes \mathbf{U_c}^T \right)
\]
}
where $\mathbf{A} = \mathbf{U_a \Sigma_a {V_a}^T}$, $\mathbf{B} = \mathbf{U_b \Sigma_b {V_b}^T}$, and $\mathbf{C} = \mathbf{U_c \Sigma_c {V_c}^T}$ (i.e. the respective Singular Value Decompositions).
After we rewrite the least squares problem in the above form, we can efficiently carry out a series of Kronecker products times a vector very efficiently, {\em without} materializing the (potentially big) Kronecker product. In \cite{papalexakis2015fastcorcondia} we use the algorithm proposed in \cite{buis1996efficient} to do this, which we will henceforth refer to as \textsc{KronMatVec}\xspace operation:
\[
\mathbf{y} = \left(\mathbf{A}_1 \otimes \mathbf{A}_2 \otimes \cdots \otimes \mathbf{A}_n \right)\mathbf{x}
\]
What we have achieved thus far is extending CORCONDIA to large and sparse data, assuming Frobenius norm. This assumption postulates that the underlying data distribution is Gaussian. However, recently \cite{chi2012tensors} showed that for sparse data that capture counts (e.g. number of messages exchanged), it is more beneficial to postulate a Poisson distribution, therefore using the KL-Divergence as a loss function. This has been more recently adopted in \cite{ho2014marble} showing very promising results in biomedical applications. Therefore, one natural direction, which we follow in the first part of the next section, is to extend CORCONDIA for this scenario.
\section{Proposed Methods}
\label{sec:method}
In exploratory data mining applications, the case is very frequently the following: we are given a piece of (usually very large) data that is of interest to a domain expert, and we are asked to identify regular and irregular patterns that are potentially useful to the expert who is providing the data. During this process, very often, the analysis is carried out in a completely unsupervised way, since ground truth and labels are either very expensive or impossible to obtain. In our context of tensor data mining, here is the problem at hand:
\begin{infproblem}
Given a tensor $\tensor{X}$ without ground truth or labelled data, how can we analyze it using the PARAFAC decomposition so that we can also:
\begin{enumerate}
\item Determine automatically a good number of components for the decomposition
\item Provide quality guarantees for the resulting decomposition
\item Minimize human intervention and trial-and-error testing
\end{enumerate}
\end{infproblem}
In order to attack the above problem, first, in Section \ref{sec:method_corcondia} we describe how we can derive a fast and efficient metric of the quality of a decomposition, assuming the KL-Divergence. Finally, in \ref{sec:method_final}, we introduce \textsc{Scoup-Smt}\xspace, our unified algorithm for automatic tensor mining with minimal user intervention and quality characterization of the solution.
\label{sec:method}
\subsection{Quality Assessment with KL-Divergence}
\label{sec:method_corcondia}
As we saw in the description of CORCONDIA with Frobenius norm loss, its computation requires solving the least squares problem:
\[
\min_{\tensor{G}} \| vec\left( \tensor{X} \right) - \left( \mathbf{A}\otimes \mathbf{B} \otimes \mathbf{C} \right) vec \left( \tensor{G} \right)\|_F^2
\]
In the case of the \texttt{CP\_APR}\xspace modelling, where the loss function is the KL-Divergence, the minimization problem that we need to solve is:
\begin{equation}
\min_{\mathbf{x}} D_{KL}(\mathbf{y} || \mathbf{W}\mathbf{x})
\label{eq:kl_regression}
\end{equation}
where in our case, $\mathbf{W} = \mathbf{A\otimes B \otimes C}$.
Unlike the Frobenius norm case, where the solution to the problem is the Least Squares estimate, in the KL-Divergence case, the problem does not have a closed form solution. Instead, iterative solutions apply. The most prominent approach to this problem is via an optimization technique called {\em Majorization-Minimization} (MM) or {\em Iterative Majorization} \cite{heiser1995convergent}. In a nutshell, in MM, given a function that is hard to minimize directly, we derive a ``majorizing'' function, which is always greater than the function to be minimized, except for a support point where it is equal; we minimize the majorizing function, and iteratively updated the support point using the minimizer of that function. This procedure converges to a local minimum. For the problem of Eq. \ref{eq:kl_regression}, \cite{fevotte2011algorithms} and subsequently \cite{chi2012tensors}, employ the following update rule for the problem, which is used iteratively until convergence to a stationary point.
\begin{equation}
\mathbf{x}(j)^{(k)} = \mathbf{x}(j)^{(k-1)} ( \frac{ \sum_i \mathbf{W}(i,k) ( \frac{\mathbf{y}(j)}{\tilde{\mathbf{y}}}(j)^{(k-1)} )}{ \sum_i \mathbf{W}(i,j) } )
\label{eq:update_rule_naive}
\end{equation}
where $\tilde{\mathbf{y}}^{(k-1)} = \mathbf{W}\mathbf{x}^{(k-1)}$, and $k$ denotes the $k$-th iteration index.
The above solution is generic for any structure of $\mathbf{W}$. Remember, however, that $\mathbf{W}$ has very specific Kronecker structure which we should exploit. Additionally, suppose that we have a $10^4 \times 10^4 \times 10^4$ tensor; then, the large dimension of $\mathbf{W}$ will be $10^{12}$. If we attempt to materialize, store, and use $\mathbf{W}$ throughout the algorithm, that can prove catastrophic to the algorithm's performance. We can exploit the Kronecker structure of $\mathbf{W}$ so that we break down Eq. \ref{eq:update_rule_naive} into pieces, each one which can be computed efficiently, given the structure of $\mathbf{W}$.
The first step is to decompose the expression of the numerator of Eq. \ref{eq:update_rule_naive}. In particular, we equivalently write
$
\mathbf{x}^{(k)} = \mathbf{x}^{(k-1)} \ast \mathbf{z}_2
$
where
$
\mathbf{z}_2 = \mathbf{W}^T \mathbf{z}_1
$
and
$\mathbf{z}_1 = \mathbf{y}\oslash \tilde{\mathbf{y}}$.
Due to the Kronecker structure of $\mathbf{W}$:
\[
\mathbf{z}_2 = \textsc{KronMatVec}\xspace( \{ \mathbf{A}^T, \mathbf{B}^T, \mathbf{C}^T \},\mathbf{z}_1)
\]
Therefore, the update to $\mathbf{x}^{(k)} $ is efficiently calculated in the three above steps.
The normalization factor of the equation is equal to:
$
\mathbf{s}(j) = \sum_i \mathbf{W}(i,j).
$
Given the Kronecker structure of $\mathbf{W}$ however, the following holds:
\begin{claim}
The row sum of a Kronecker product matrix $\mathbf{A}\otimes\mathbf{B}$ can be rewritten as $\left( \sum_{i=1}^I \mathbf{A}(i,:) \right) \otimes \left( \sum_{j=1}^J \mathbf{B}(j,:)\right)$
\end{claim}
\begin{proof}
We can rewrite the row sums $\sum_{i=1}^I \mathbf{A}(i,:) =\mathbf{i}_I^T\mathbf{A}$ and $\sum_{j=1}^J \mathbf{B}(j,:) =\mathbf{i}_J^T\mathbf{B}$ where $\mathbf{i}_I$ and $\mathbf{i}_J$ are all-ones column vectors of size $I$ and $J$ respectively. For the Kronecker product of the row sums and by using properties of the Kronecker product, and calling $\mathbf{A\otimes B} = \mathbf{W}$ we have
\[
\left( \mathbf{i}_I^T\mathbf{A} \right) \otimes \left( \mathbf{i}_J^T\mathbf{B} \right) = \left(\mathbf{i}_I \otimes \mathbf{i}_J \right)^T \left( \mathbf{A\otimes B} \right) = \mathbf{i}_{IJ}^T\mathbf{W} = \sum_{i=1}^{IJ}\mathbf{W}(i,:)
\]
which concludes the proof.
\end{proof}
Thus, $
\mathbf{s} = \left( \sum_i \mathbf{A}(i,:) \right) \otimes \left( \sum_j \mathbf{B}(j,:) \right) \otimes \left(\sum_n \mathbf{C}(n,:) \right).
$
Putting everything together, we end up with Algorithm \ref{alg:efficient_kl_regression} which is an efficient solution to the minimization problem of Equation \ref{eq:kl_regression}. As in the naive case, we also use Iterative Majorization in the efficient algorithm; we iterate updating $\mathbf{x}^{(k)}$ until we converge to a local optimum. Finally, Algorithm \ref{alg:efficient_kl} shows the steps to compute CORCONDIA under KL-Divergence efficiently.
\begin{center}
\begin{algorithm} [!htp]
\small
\caption{Efficient Quality Assesment with KL-Divergence loss} \label{alg:efficient_kl}
\begin{algorithmic}[1]
\REQUIRE Tensor $\tensor{X}$ and \texttt{CP\_APR}\xspace factor matrices $\mathbf{A,B,C}$.
\ENSURE CORCONDIA diagnostic $c$.
\STATE some more stuff
\STATE some stuff here
\STATE $\displaystyle{ c = 100\left(1 - \frac{\sum_{i=1}^{F} \sum_{j=1}^{F} \sum_{k=1}^{F} \left(\tensor{G}(i,j,k) - \tensor{I}(i,j,k)\right)^2 }{F} \right)}$
\end{algorithmic}
\end{algorithm}
\end{center}
\begin{center}
\begin{algorithm} [!htp]
\small
\caption{Efficient Majorization Minimization for KL-Divergence Regression} \label{alg:efficient_kl_regression}
\begin{algorithmic}[1]
\REQUIRE Vector $\mathbf{y}$ and matrices $\mathbf{A,B,C}$.
\ENSURE Vector $\mathbf{x}$
\STATE Initialize $\mathbf{x}^{(0)}$ randomly
\STATE $\tilde{\mathbf{y}} = \textsc{KronMatVec}\xspace(\{ \mathbf{A, B, C} \},\mathbf{x}^{(0)})$
\STATE $\mathbf{s} = \left( \sum_i \mathbf{A}(i,:) \right) \otimes \left( \sum_j \mathbf{B}(j,:) \right) \otimes \left(\sum_n \mathbf{C}(n,:) \right)$
\STATE Start loop:
\STATE $\mathbf{z}_1 = \mathbf{y}\oslash \tilde{\mathbf{y}}$
\STATE $\mathbf{z}_2 = \textsc{KronMatVec}\xspace( \{ \mathbf{A}^T, \mathbf{B}^T, \mathbf{C}^T \},\mathbf{z}_1)$
\STATE $\mathbf{x}^{(k)} = \mathbf{x}^{(k-1)} \ast \mathbf{z}_2$
\STATE $\tilde{\mathbf{y}} = \textsc{KronMatVec}\xspace(\{ \mathbf{A, B, C} \},\mathbf{x}^{(k)})$
\STATE End loop
\STATE Normalize $\mathbf{x}^{(k)} $ using $\mathbf{s}$
\end{algorithmic}
\end{algorithm}
\end{center}
\subsection{Scoup-SMT\xspace: Automated Unsupervised Tensor Mining}
\label{sec:method_final}
At this stage, we have the tools we need in order to design an automated tensor mining algorithm that minimizes human intervention and provides quality characterization of the solution. We call our proposed method \textsc{Scoup-Smt}\xspace, and we view this as a step towards making tensor mining a fully automated tool, used as a black box by academic and industrial practicioners.
\textsc{Scoup-Smt}\xspace is a two step algorithm, where we first search through the solution space and at the second step, we automatically select a good solution based on its quality and the number of components it offers.
A sketch of \textsc{Scoup-Smt}\xspace follows, and is also outlined in Algorithm \ref{alg:auto_tensor}.
\paragraph{Solution Search}
The user provides a data tensor, as well as a maximum rank that reflects the budget that she is willing to devote to \textsc{Scoup-Smt}\xspace's search. We neither have nor require any prior knowledge whether the tensor is highly sparse, or dense, contains real values or counts, hinting whether we should use, say, \texttt{CP\_NMU}\xspace postulating Frobenius norm loss, or \texttt{CP\_APR}\xspace postulating KL-Divergence loss.
Fortunately, our work in this paper, as well as our previous work \cite{papalexakis2015fastcorcondia} has equipped us with tools for handling all of the above cases. Thus, we follow a data-driven approach, where we let the data show us whether using \texttt{CP\_NMU}\xspace or \texttt{CP\_APR}\xspace is capturing better structure.
For a grid of values for the decomposition rank (bounded by the user provided maximum rank), we run both \texttt{CP\_NMU}\xspace and \texttt{CP\_APR}\xspace, and we record the quality of the result as measured by the CORCONDIA diagnostic into vectors $\mathbf{c}_{Fro}$ and $\mathbf{c}_{KL}$ (using the algorithm in \cite{papalexakis2015fastcorcondia} and Algorithm \ref{alg:efficient_kl} respectively), truncating negative values to zero.
\paragraph{Result Selection}
At this point, for both \texttt{CP\_NMU}\xspace and \texttt{CP\_APR}\xspace we have points in two dimensional space $(F_i,c_i)$, reflecting the quality and the corresponding number of components.
Informally, our problem here is the following:
\begin{infproblem}
Given points $(F_i,c_i)$ we need to find one that maximizes the quality of the decomposition, as well as finding as many hidden components in the data as possible.
\end{infproblem}
Intuitively, we are seeking a decomposition that discovers as many latent components as possible, without sacrificing the quality of those components.
Essentially, we have a multi-objective optimization problem, where we need to maximize both $c_i$ and $F_i$. However, if we, say, get the Pareto front of those points (i.e. the subset of all non-dominated points), we end up with a family of solutions without a clear guideline on how to select one. We propose to use the following, effective, two-step maximization algorithm that gives an intuitive {\em data-driven} solution:
\begin{itemize}
\item {\bf Max c step}: Given vector $\mathbf{c}$, run 2-means clustering on its values. This will essentially divide the vector into a set of good/high values and a set of low/bad ones. If we call $m_1 , m_2$ the means of the two clusters, then we select the cluster index that corresponds to the maximum between $m_1$ and $m_2$.
\item {\bf Max F step}: Given the cluster of points with maximum mean, we select the point that maximizes the value of $F$. We call this point $(F^*,c^*)$.
\end{itemize}
Another alternative is to formally define a function of $c,F$ that we wish to maximize, and select the maximum via enumeration. Coming up with the particular function to maximize, considering the intuitive objective of maximizing the number of components that we can extract with reasonably high quality ($c$), is a hard problem, and we risk biasing the selection with a specific choice of a function. Nevertheless, an example such function can be $g(c,F) = logc logF$ for $c>0$, and $g(0,F) = 0$; this function essentially measures the area of the rectangle formed by the lines connecting $(F,c)$ with the axes (in the log-log space) and intuitively seeks to find a good compromise between maximizing $F$ and $c$. This function performs closely to the proposed data-driven approach and we defer a detailed discussion and investigation to future work.
\hide{
Figure \ref{fig:pareto_heuristic} shows pictorially the output of this two-step algorithm for a set of points taken from a real dataset.
\begin{figure}[!ht]
\begin{center}
\includegraphics[width = 0.4\textwidth]{FIG/pareto_heuristic_example.eps}
\end{center}
\caption{Example of choosing a good point}
\label{fig:pareto_heuristic}
\end{figure}
}
After choosing the ``best'' points $(F_{Fro}^*,c_{Fro}^*)$ and $(F_{KL}^*,c_{KL}^*)$, at the final step of \textsc{Scoup-Smt}\xspace, we have to select between the results of \texttt{CP\_NMU}\xspace and \texttt{CP\_APR}\xspace. In order do so, we can use the following strategies:
\begin{enumerate}
\item Calculate $s_{Fro} = \displaystyle{\sum_f \mathbf{c}_{Fro}(f)}$ and $s_{KL} = \displaystyle{\sum_f \mathbf{c}_{KL}(f)}$, and select the method that gives the largest sum. The intuition behind this data-driven strategy is choosing the loss function that is able to discover results with higher quality on aggregate, for more potential ranks.
\item Select the results that produce the maximum value between $c_{Fro}^*$ and $c_{KL}^*$. This strategy is conservative and aims for the highest quality of results, possibly to the expense of components of lesser quality that could still be acceptable for exploratory analysis.
\item Select the results that produce the maximum value between $F_{Fro}^*$ and $F_{KL}^*$. Contrary to the previous strategy, this one is more aggressive, aiming for the highest number of components that can be extracted with acceptable quality.
\end{enumerate}
Empirically, the last strategy seems to give better results, however they all perform very closely in synthetic data. Particular choice of strategy depends on the application needs, e.g. if quality of the components is imperative to be high, then strategy 2 should be preferred over strategy 3.
\begin{center}
\begin{algorithm} [!htp]
\small
\caption{\textsc{Scoup-Smt}\xspace: Automatic Unsupervised Tensor Mining} \label{alg:auto_tensor}
\begin{algorithmic}[1]
\REQUIRE Tensor $\tensor{X}$ and maximum budget for component search $F$
\ENSURE PARAFAC decomposition $\mathbf{A,B,C}$ of $\tensor{X}$ and corresponding quality metric $c^*$.
\FOR{$f = 2\cdots F$}
\STATE Run \texttt{CP\_NMU}\xspace for $f$ components. Update $\mathbf{c}_{Fro}(f)$ with the result of Algorithm in \cite{papalexakis2015fastcorcondia}.
\STATE Run \texttt{CP\_APR}\xspace for $f$ components. Update $\mathbf{c}_{KL}(f)$ with the result of Algorithm \ref{alg:efficient_kl}.
\ENDFOR
\STATE Find $(F_{Fro}^*,c_{Fro}^*)$ and $(F_{KL}^*,c_{KL}^*)$ using the two-step maximization as described in the text.
\STATE Choose between \texttt{CP\_NMU}\xspace and \texttt{CP\_APR}\xspace using one of the three strategies described in the text.
\STATE Output the chosen $c^*$ and the corresponding decomposition.
\end{algorithmic}
\end{algorithm}
\end{center}
We point out that lines 1-5 of Algorithm \ref{alg:auto_tensor} are embarrassingly parallel.
Finally, it is important to note that \textsc{Scoup-Smt}\xspace not only seeks to find a good number of components for the decomposition, combining the best of both worlds of \texttt{CP\_NMU}\xspace and \texttt{CP\_APR}\xspace, but furthermore is able to provide quality assessment for the decomposition: if for a given $F_{max}$ none of the solutions that \textsc{Scoup-Smt}\xspace sifts through yields a satisfactory result, the user will be able to tell because of the very low (or zero in extreme cases) $c^*$ value.
\section{Experimental Evaluation}
\label{sec:experiments}
We implemented \textsc{Scoup-Smt}\xspace in Matlab, using the Tensor Toolbox \cite{tensortoolbox}, which provides efficient manipulation and storage of sparse tensors. We make our code publicly available\footnote{Download our code at \\ \url{http://www.cs.cmu.edu/~epapalex/src/AutoTen.zip}}. The online version of our code contains a test case that uses the same code that we used for the following evaluation.
All experiments were run on a workstation with 4 Intel(R) Xeon(R) E7- 8837 and 1TB of RAM.
\subsection{Evaluation on Synthetic Data}
\label{sec:synthetic}
In this section, we empirically measure \textsc{Scoup-Smt}\xspace's ability to uncover the true number of components hidden in a tensor. The experimental setting is as follows: We create synthetic tensors of size $50\times50\times50$, using the function \texttt{create\_problem} of the Tensor Toolbox for Matlab as a standardized means of generating synthetic tensors, we create two different test cases: 1) sparse factors, with total number of non-zeros equal to 500, and 2) dense factors.
In both cases, we generate random factors with integer values.
We generate these three test cases for true rank $F_o$ ranging from 2-5. For both test cases, we distinguish a noisy case (where Gaussian noise with variance $0.1$ is by default added by \texttt{create\_problem}) and a noiseless case.
We compare \textsc{Scoup-Smt}\xspace against three baselines:
\begin{itemize}\setlength{\itemsep}{0.0mm}
\item {\bf Baseline 1}: A Bayesian tensor decomposition approach, as introduced very recently in \cite{zhaobayesian} which automatically determines the rank.
\item {\bf Baseline 2}: This is a very simple heuristic approach where, for a grid of values for the rank, we run \texttt{CP\_NMU}\xspace and record the Frobenius norm loss for each solution. If for two consecutive iterations the loss does not improve more than a small positive number $\epsilon$ (set to $10^{-6}$ here), we declare as output the result of the previous iteration.
\item {\bf Baseline 3}: Same as Baseline 2 with sole difference being that we use \texttt{CP\_APR}\xspace and accordingly instead of the Frobenius norm reconstruction error, we measure the log-likelihood, and we stop when it stops improving more than $\epsilon$. We expect Baseline 3 to be more effective than Baseline 2 in sparse data, due to the more delicate and effective treatment of sparse, count data by \texttt{CP\_APR}\xspace.
\end{itemize}
\textsc{Scoup-Smt}\xspace as well as Baselines 2 \& 3 require a maximum bound $F_{max}$ on the rank; for fairness, we set $F_{max}=2F_o$ for all three methods.
In Figures \ref{fig:synthetic_errors} and \ref{fig:synthetic_errors_noiseless} we show the results for both test cases, for noisy and noiseless data respectively. The error is measured as $|F_{est} - F_o|$ where $F_{est}$ is the estimated number of components by each method. Due to the randomized nature of the synthetic data generation, we ran 1000 iterations and we show the average results.
In the noisy case (Fig. \ref{fig:synthetic_errors}) we observe that in both scenarios and for all chosen ranks, \textsc{Scoup-Smt}\xspace outperforms the baselines, having lower error.
In the noiseless case of Fig. \ref{fig:synthetic_errors_noiseless}, we observe consistent behavior, with all methods experiencing a small boost to their performance, due to the absence of noise.
We calculated statistical significance of our results ($p<0.01$) using a two-sided sign test.
Overall, we conclude that \textsc{Scoup-Smt}\xspace largely outperforms the baselines. The problem at hand is an extremely hard one, and we are not expecting any {\em tractable} method to solve it perfectly. Thus, the results we obtain here are very encouraging.
\begin{figure}[!htf]
\begin{center}
\subfigure[Sparse]{ \includegraphics[width = 0.23\textwidth]{FIG/experiments/sparse_errors.eps}}
\subfigure[Dense]{ \includegraphics[width = 0.23\textwidth]{FIG/experiments/dense_errors.eps}}
\caption{Error for \textsc{Scoup-Smt}\xspace and the baselines, for noisy synthetic data.}
\label{fig:synthetic_errors}
\end{center}
\end{figure}
\begin{figure}[!htf]
\begin{center}
\subfigure[Sparse]{ \includegraphics[width = 0.23\textwidth]{FIG/experiments/noiseless/sparse_errors.eps}}
\subfigure[Dense]{ \includegraphics[width = 0.23\textwidth]{FIG/experiments/noiseless/dense_errors.eps}}
\caption{Error for \textsc{Scoup-Smt}\xspace and the baselines, for noiseless synthetic data.}
\label{fig:synthetic_errors_noiseless}
\end{center}
\end{figure}
\section{Data Mining Case Study}
\label{fig:case_study}
After establishing that \textsc{Scoup-Smt}\xspace is able to perform well in a control, synthetically generated setting, the next step is to see its results ``in the wild''. To that end, we are conducting two case studies. Section \ref{sec:study1} takes 10 diverse real datasets shown in Table \ref{tab:datasets} and investigates their rank structure. In Section \ref{sec:study2} we apply \textsc{Scoup-Smt}\xspace to two of the datasets of Table \ref{tab:datasets} and we analyze the results, as part of an exploratory data mining study.
\begin{table*}[!ht]
\small
\begin{center}
\caption{Datasets analyzed \label{tab:datasets}}
{
\rowcolors{1}{lightgray}{white}
\begin{tabular}{llllll}
& \textbf{Name} & \textbf{Description} & \textbf{Dimensions} & {\bf Number of nonzeros} \\
& \textsc{Enron}\xspace & (sender, recipient, month) & $186\times186\times44$ & 9838 \\ \hiderowcolors
& \texttt{Reality Mining}\xspace\cite{eagle2009inferring} & (person, person, means of communication) &$88\times88 \times 4$ & 5022 \\
& \textsc{Facebook}\xspace \cite{viswanath-2009-activity} & (wall owner, poster, day) & $63891\times 63890 \times 1847$ & 737778 \\
& \texttt{Taxi}\xspace\cite{yuan2011driving,Wang2014travel} & (latitude, longitude,minute) & $100\times100\times 9617$ & 17762489 \\
& \texttt{DBLP}\xspace \cite{papalexakis2013more} & (paper, paper, view)&$7317\times7317\times3$ & 274106\\
& \texttt{Netflix}\xspace & (movie, user, date) & $17770\times252474\times88$ & 50244707 \\
& \texttt{Amazon co-purchase}\xspace\cite{snapnets} & (product, product, product group)& $256\times256\times5$ & 5726 \\
& \texttt{Amazon metadata}\xspace\cite{snapnets} & (product, customer, product group) & $10000\times263011\times5$ & 441301\\
& \texttt{Yelp}\xspace & (user, business, term) &$43872\times11536\times10000$ & 10009860\\
& \texttt{Airport}\xspace & (airport, airport, airline) & $9135\times9135\times19305$ & 58443 \\
\end{tabular}
}
\end{center}
\end{table*}
\subsection{Rank Structure of Real Datasets}
\label{sec:study1}
Since exploration of the rank structure of a dataset, using the CORCONDIA diagnostic, is an integral part of \textsc{Scoup-Smt}\xspace, we deem necessary to dive deeper into that process.
In this case study we are analyzing the rank structure of 10 real datasets, as captured by CORCONDIA with Frobenius norm loss (using our algorithm from \cite{papalexakis2015fastcorcondia}, as well as CORCONDIA with KL-Divergence loss (introduced here).
Most of the datasets we use are publicly available and can be obtained by either following the link within the original work that introduced them, or (whenever applicable) a direct link.
\textsc{Enron}\xspace\footnote{\url{http://www.cs.cmu.edu/~enron/}} is a social network dataset, recording the number of emails exchanged between employees of the company for a period of time, during the company crisis. \texttt{Reality Mining}\xspace \cite{eagle2009inferring} is a multi-view social network dataset, recording relations between MIT students (who calls whom, who messages whom, who is close to whom and so on). \textsc{Facebook}\xspace \cite{viswanath-2009-activity} is a time evolving snapshot of Facebook, recording people posting on other peoples' Walls. \texttt{Taxi}\xspace\footnote{\small\url{http://research.microsoft.com/apps/pubs/?id=152883}} is a dataset of taxi trajectories in Beijing; we discretize latitude and longitude to a $100\times 100$ grid. \texttt{DBLP}\xspace is a dataset recording which researched published what paper under three different views (first view shows co-authorship, second view shows citation, and third view shows whether two authors share at least three keywords in their title or abstract of their papers). \texttt{Netflix}\xspace comes from the Netflix prize dataset and records movie ratings by users over time. \texttt{Amazon co-purchase}\xspace data records items bought together, and the category of the first of the two products. \texttt{Amazon metadata}\xspace records customers who reviewed a product, and the corresponding product category. \texttt{Yelp}\xspace contains reviews of Yelp users for various businesses (from the data challenge\footnote{\url{https://www.yelp.com/dataset_challenge/dataset}}). Finally, \texttt{Airport}\xspace\footnote{\url{http://openflights.org/data.html}} contains records of flights between different airports, and the operating airline.
\begin{figure*}[!ht]
\begin{center}
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/enron_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/reality_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/facebook_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/taxis_100_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/dblp_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/netflix_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/amazon_co_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/amazon_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/yelp_fro.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/airport_fro.eps} }
\end{center}
\caption{CORCONDIA for \texttt{CP\_NMU}\xspace}
\label{fig:real_fro}
\end{figure*}
\begin{figure*}[!ht]
\begin{center}
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/enron_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/reality_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/facebook_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/taxis_100_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/dblp_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/netflix_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/amazon_co_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/amazon_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/yelp_kl.eps} }
\subfigure[]{ \includegraphics[width = 0.18\textwidth]{FIG/real-data-figs/airport_kl.eps} }
\end{center}
\caption{CORCONDIA for \texttt{CP\_APR}\xspace}
\label{fig:real_kl}
\end{figure*}
We ran our algorithms for $F=2\cdots50$, and truncated negative values to zero. For KL-Divergence and datasets \textsc{Facebook}\xspace, \texttt{Netflix}\xspace, \texttt{Yelp}\xspace, and \texttt{Airport}\xspace we used smaller versions (first 500 rows for \texttt{Netflix}\xspace and \texttt{Yelp}\xspace, and first 1000 rows for \textsc{Facebook}\xspace and \texttt{Airport}\xspace), due to high memory requirements of Matlab; this means that the corresponding figures describe the rank structure of a smaller dataset, which might be different from the full one. Figure \ref{fig:real_fro} shows CORCONDIA when using Frobenius norm as a loss, and Fig. \ref{fig:real_kl} when using KL-Divergence. The way to interpret these figures is the following: assuming a \texttt{CP\_NMU}\xspace (Fig. \ref{fig:real_fro}) or a \texttt{CP\_APR}\xspace (Fig. \ref{fig:real_kl}) model, each figure shows the modelling quality of the data for a given rank. This sheds light to the rank structure of a particular dataset (although that is not to say that it provides a definitive answer about its true rank). For the given datasets, we observe a few interesting differences in structure: for instance, \textsc{Enron}\xspace and \texttt{Taxi}\xspace in Fig. \ref{fig:real_fro} seem to have good quality for a few components. On the other hand, \texttt{Reality Mining}\xspace, \texttt{DBLP}\xspace, and \texttt{Amazon metadata}\xspace have reasonably acceptable quality for a larger range of components, with the quality decreasing as the number gets higher. Another interesting observation, confirming recent results in \cite{zhang2014understanding}, is that \texttt{Yelp}\xspace seems to be modelled better using a high number of components. Figures that are all-zero merely show that no good structure was detected for up to 50 components, however, this might indicate that such datasets (e.g. \texttt{Netflix}\xspace) have an even higher number of components.
Finally, contrasting Fig. \ref{fig:real_kl} to Fig. \ref{fig:real_fro}, we observe that in many cases using the KL-Divergence is able to discover better structure than the Frobenius norm (e.g. \textsc{Enron}\xspace and \texttt{Amazon co-purchase}\xspace).
\subsection{Scoup-SMT\xspace in practice}
\label{sec:study2}
We used \textsc{Scoup-Smt}\xspace to analyze two of the datasets shown in Table \ref{tab:datasets}. In the following lines we show our results.
\subsubsection{Analyzing \texttt{Taxi}\xspace}
The data we have span an entire week worth of measurements, with temporal granularity of minutes. First, we tried quantizing the latitude and longitude into a $1000\times1000$ grid; however, \textsc{Scoup-Smt}\xspace warned us that the decomposition was not able to detect good and coherent structure in the data, perhaps due to the extremely sparse variable space of our grid. Subsequently, we modelled the data using a $100\times100$ grid and \textsc{Scoup-Smt}\xspace was able to detect good structure. In particular, \textsc{Scoup-Smt}\xspace output 8 rank-one components, choosing Frobenius norm as a loss function.
In Figure \ref{fig:beijing_taxi} we show 4 representative components of the decomposition. In each sub-figure, we overlay the map of Beijing with the coordinates that appear to have high activity in the particular component; every sub-figure also shows the temporal profile of the component. The first two components (Fig. \ref{fig:beijing_taxi}(a), (b)) spatially refer to a similar area, roughly corresponding to the tourist and business center in the central rings of the city. The difference is that Fig. \ref{fig:beijing_taxi}(a) shows high activity during the weekdays and declining activity over the weekend (indicated by five peaks of equal height, followed by two smaller peaks), whereas Fig. \ref{fig:beijing_taxi}(b) shows a slightly inverted temporal profile, where the activity peaks over the weekend; we conclude that Fig. \ref{fig:beijing_taxi}(a) most likely captures business traffic that peaks during the week, whereas Fig. \ref{fig:beijing_taxi}(b) captures tourist and leisure traffic that peaks over the weekend. The third component (Fig. \ref{fig:beijing_taxi}(c)) is highly active around the Olympic Center and Convention Center area, with peaking activity in the middle of the week.
Finally, the last component (Fig. \ref{fig:beijing_taxi}(d) ) shows high activity only outside of Beijing's international airport, where taxis gather to pick-up customers; on the temporal side, we see daily peaks of activity, with the collective activity dropping during the weekend, when there is significantly less traffic of people coming to the city for business.
By being able to analyze such trajectory data into highly interpretable results, we can help policymakers to better understand the traffic patterns of taxis in big cities, estimate high and low demand areas and times and optimize city planning in that respect. There has been very recent work \cite{Wang2014travel} towards the same direction, and we view our results as complementary.
\begin{figure*}[!ht]
\begin{center}
\subfigure[{\bf Tourist \& Business Center}: High activity during weekdays, low over the weekend]{\includegraphics[width = 0.495\textwidth]{FIG/beijing-taxi/business-tourist-center.eps}}
\subfigure[{\bf Downtown}: Consistent activity over the week]{\includegraphics[width = 0.495\textwidth]{FIG/beijing-taxi/mid-rings.eps}}
\subfigure[{\bf Olympic Center}: Activity peak during the week]{\includegraphics[width = 0.495\textwidth]{FIG/beijing-taxi/olympic-center.eps}}
\subfigure[{\bf Airport}: High activity during weekdays, low over the weekend]{\includegraphics[width = 0.495\textwidth]{FIG/beijing-taxi/airport.eps}}
\end{center}
\caption{Latent components of the \texttt{Taxi}\xspace dataset, as extracted using \textsc{Scoup-Smt}\xspace.}
\label{fig:beijing_taxi}
\end{figure*}
\subsubsection{Analyzing \texttt{Amazon co-purchase}\xspace}
\begin{table*}[!htf]
\begin{center}
\scriptsize
\begin{tabular}{|c|c|c|}
\hline
{\bf Cluster type} & {\bf Products} & {\bf Product Types} \\\hline
\multirow{4}{*}{\#1 Self Improvement} & Resolving Conflicts At Work : A Complete Guide for Everyone on the Job & Book \\
& How to Kill a Monster (Goosebumps) & Book \\
& Mensa Visual Brainteasers & Book \\
& Learning in Overdrive: Designing Curriculum, Instruction, and Assessment from Standards : A Manual for Teachers & Book \\\hline
\multirow{2}{*}{\#2 Psychology, Self Improvement} & Physicians of the Soul: The Psychologies of the World's Greatest Spiritual Leaders & Book \\
& The Rise of the Creative Class: And How It's Transforming Work, Leisure, Community and Everyday Life & Book \\ \hline
\multirow{3}{*}{\#3 Technical Books} & Beginning ASP.NET Databases using C\# & Book \\
& BizPricer Business Valuation Manual w\/Software & Book \\
& Desde Que Samba E Samba & Music \\ \hline
\multirow{3}{*}{\#4 History} & War at Sea: A Naval History of World War II & Book \\
& Jailed for Freedom: American Women Win the Vote & Book \\
& The Perfect Plan (7th Heaven) & Book \\ \hline
\end{tabular}
\end{center}
\caption{Latent components of the \texttt{Amazon co-purchase}\xspace dataset, as extracted using \textsc{Scoup-Smt}\xspace}
\label{tab:amazon}
\end{table*}
This dataset records pairs of products that were purchased together by the same customer on Amazon, as well as the category of the first product in the pair. This dataset, as shown in Figures \ref{fig:real_fro}(g) and \ref{fig:real_kl} does not have perfect trilinear structure, however a low rank trilinear approximation still offers reasonably good insights for product recommendation and market basket analysis.
By analyzing this dataset, we seek to find coherent groups of products that people tend to purchase together, aiming for better product recommendations and suggestions. For the purposes of this study, we extracted a small piece of the co-purchase network of 256 products. \textsc{Scoup-Smt}\xspace was able to extract 24 components by choosing KL-Divergence as a loss. On Table \ref{tab:amazon} we show a representative subset of our resulting components (which were remarkably sparse, due to the KL-Divergence fitting by \texttt{CP\_APR}\xspace). We observe that products of similar genre and themes tend to naturally cluster together. For instance, cluster \#1 contains mostly self improvement books. We also observe a few topical outliers, such as the book \texttt{How to Kill a Monster (Goosebumps)} in cluster \#1, and CD \texttt{Desde Que Samba E Samba} in cluster \#3 that contains Technical / Software Development books.
\section{Conclusions}
\label{sec:conclusions}
In this paper, we work towards an automatic, unsupervised tensor mining algorithm that minimizes user intervention. Our main contributions are:
\begin{itemize}\setlength{\itemsep}{0.0mm}
\item {\bf Algorithms} We propose \textsc{Scoup-Smt}\xspace, a novel automatic and unsupervised tensor mining algorithm, which can provide quality characterization of the solution. We extend the highly intuitive heuristic of \cite{bro1998multi} for KL-Divergence loss, providing an efficient algorithm.
\item{\bf Evaluation \& Discovery} We evaluate our methods in synthetic data, showing their superiority compared to the baselines, as well as a wide variety of real datasets. Finally, we apply \textsc{Scoup-Smt}\xspace to two real datasets discovering meaningful patterns (Section \ref{sec:study2}).
\end{itemize}
\section{Related Work}
\label{sec:related}
\paragraph{Tensors and their data mining applications}
One of the first applications was on web mining, extending the popular HITS algorithm \cite{kolda2005higher}. There has been work on analyzing citation networks (such as DBLP) \cite{kolda2008scalable}, detecting anomalies in computer networks\cite{kolda2008scalable,maruhashi2011multiaspectforensics,papalexakis2012parcube}, extracting patterns from and completing Knowledge Bases \cite{kang2012gigatensor,papalexakis2012parcube,chang2014typed} and analyzing time-evolving or multi-view social networks. \cite{bader2007temporal,kolda2008scalable,lin2009metafac,papalexakis2012parcube,araujo2014com2,jiang2014fema},
The long list of application continues, with extensions of Latent Semantic Analysis \cite{cai2006tensor,chang2013multi}, extensions of Subspace Clustering to higher orders \cite{huang2008simultaneous}, Crime Forecasting \cite{mu2011empirical}, Image Processing \cite{liu2013tensor}, mining Brain data \cite{davidson2013network,he2014dusk,papalexakisturbo}, trajectory and mobility data \cite{zheng2010collaborative,Wang2014travel}, and bioinformatics \cite{ho2014marble}.
\paragraph{Choosing the right number of components}
As we've mentioned throughout the text, CORCONDIA \cite{bro1998multi} is using properties of the PARAFAC decomposition in order to hint towards the right number of components. In \cite{papalexakis2015fastcorcondia}, we introduce a scalable algorithm for CORCONDIA (under the Frobenius norm). Moving away from the PARAFAC decompostion, Kiers and Kinderen \cite{kiers2003fast} introduce a method for choosing the number of components for Tucker3.
There has been recent work using Minimum Description Length (MDL): In \cite{araujo2014com2} the authors use MDL in the context of community detection in time-evolving social network tensors, whereas in \cite{metzler2015clustering}, Metzler and Miettinen use MDL to score the quality of components for a binary tensor factorization. Finally, there have also been recent advances using Bayesian methods \cite{zhaobayesian} in order to automatically decide the number of components.
\section{Acknowledgements}
{\scriptsize
Research was supported by the National Science Foundation Grant No. IIS-1247489. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the funding parties.
The author would like to thank Professors Christos Faloutsos, Nicholas Sidiropoulos, and Rasmus Bro for valuable conversations, and Miguel Araujo for sharing the \texttt{Airport}\xspace dataset.
}
\balance
\bibliographystyle{IEEEbib}
|
\section{Introduction}
A \emph{perfect matching} in a hypergraph $H$ is a collection of vertex-disjoint edges of $H$ which cover the vertex set $V(H)$ of $H$.
Given a $k$-uniform hypergraph $H$ with an $\ell$-element vertex set $S$ (where $0 \leq \ell \leq k-1$) we define
$d_H (S)$ to be the number of edges containing $S$. The \emph{minimum $\ell$-degree $\delta _{\ell}
(H)$} of $H$ is the minimum of $d_H (S)$ over all $\ell$-element sets of vertices in $H$.
In recent years the problem of determining the minimum $\ell$-degree threshold that ensures a perfect matching in a $k$-uniform hypergraph has received much attention
(see e.g.~\cite{ko1, rrs1, pik, rrs, hps, mark, KOTmatch, khan1, khan2, afh, zhao, czy, zhao2, mycroft, tim, jie}).
See~\cite{rrsurvey} for a survey on matchings (and Hamilton cycles) in hypergraphs.
Suppose that $\ell, k, n \in \mathbb N$ such that $k \geq 3$, $\ell \le k-1$ and $k$ divides $n$.
Let $m_{\ell}(k, n)$ denote the smallest integer $m$ such that every $k$-uniform hypergraph $H$ on $n$ vertices with $\delta _{\ell}(H)\ge m$ contains a perfect matching.
The conjectured value of $m_{\ell}(k, n)$ comes from two types of constructions.
The first type of constructions are referred to as \emph{divisibility barriers}.
Given a set $V$ of $n$ vertices with a partition $A, B$,
let $E_{\text{odd}}(A, B)$ ($E_{\text{even}}(A, B)$) denote the family of all $k$-element subsets of $V$ that intersect $A$ in an odd (even) number of vertices.
Define $\mathcal B_{n,k}(A,B)$ to be the $k$-uniform hypergraph with vertex set $V$ and edge set $E_{\text{odd}} (A,B)$. Note that the complement $\overline{\mathcal B}_{n,k} (A,B)$ of $\mathcal B_{n,k} (A,B)$ has edge set $E_{\text{even}} (A,B)$.
Define $\mathcal H_{\text{ext}} (n,k)$ to be the collection of the following hypergraphs. First, $\mathcal H_{\text{ext}} (n,k)$ contains all hypergraphs $\overline{\mathcal B} _{n,k} (A,B)$ where $|A|$ is odd. Second, if $n/k$ is odd
then $\mathcal H_{\text{ext}} (n,k)$ also contains all hypergraphs $\mathcal B _{n,k} (A,B)$ where $|A|$ is even; if $n/k$ is even then $\mathcal H_{\text{ext}} (n,k)$ also contains all hypergraphs $\mathcal B _{n,k} (A,B)$ where $|A|$ is odd.
It is easy to see that no hypergraph in $\mathcal H_{\text{ext}} (n,k)$ contains a perfect matching.
Define $\delta (n,k, \ell)$ to be the maximum of the minimum $\ell$-degrees among all the hypergraphs in $\mathcal H_{\text{ext}} (n,k)$. Note that
$\delta (n,k, \ell)= (1/2+o(1))\binom{n-\ell}{k-\ell}$ but the general formula of
$\delta (n,k, \ell)$ is unknown
(see more discussion in \cite{zhao}).
The other type of extremal constructions are referred to as \emph{space barriers}.
Let $H^*(n,k)$ be the $k$-uniform hypergraph on $n$ vertices whose vertex set is
partitioned into two vertex classes
$A$ and $B$ of sizes $n/k-1$ and $(1-1/k)n +1$ respectively and whose edge set consists precisely
of all those edges with at least one endpoint in $A$. Then $H^*(n,k)$ does not have a perfect matching and
$\delta _{\ell} (H^*(n,k)) = \binom{n-\ell}{k-\ell}-\binom{(1-1/k)n -\ell +1}{k-\ell}\approx \left(1-\left(\frac{k-1}{k}\right) ^{k-\ell} \right ) \binom{n-\ell}{k-\ell}$.
An asymptotic version of the following conjecture appeared in~\cite{hps, survey} and the minimum vertex degree version was stated in~\cite{KOTmatch}.
\begin{conj}\label{generalconj}
Let $k, \ell \in \mathbb N$ such that $\ell \leq k-1$. Then for sufficiently large $n\in k\mathbb{N}$,
\[
m_{\ell}(k, n)= \max \left \{ \delta (n,k,\ell), \ \binom{n-\ell}{k-\ell}-\binom{(1-1/k)n -\ell +1}{k-\ell}\right \} + 1.
\]
\end{conj}
Note that for all $1\leq \ell \leq k-1$,
\[
\left( \frac{k-1}{k} \right) ^{k-\ell} < \left ( \frac{1}{e} \right)^{1-\frac{\ell}{k}} \ \ \text{ and } \ \
1- \left (\frac{k-1}{k} \right) ^{k \ln 2} \rightarrow \frac{1}{2} \ \ \text{
as } \ \ k \rightarrow \infty,
\]
where $\ln$ denotes the natural logarithm function.
Thus, for $1\ll k\ll n$, if $\ell$ is significantly bigger than $(1-\ln 2)k \approx 0.307k$
then $\delta(n,k, \ell) > \binom{n-\ell}{k-\ell}-\binom{(1-1/k)n -\ell +1}{k-\ell}$.
On the other hand, if $\ell$ is smaller than $(1-\ln 2)k$ then $\delta(n,k, \ell) < \binom{n-\ell}{k-\ell}-\binom{(1-1/k)n -\ell +1}{k-\ell}$ for sufficiently large $n$.
Conjecture~\ref{generalconj} has been proven in a number of special cases. Indeed, R\"odl, Ruci\'nski and Szemer\'edi~\cite{rrs} proved the conjecture for $\ell= k-1$.
The authors~\cite{zhao, zhao2} generalized this result by showing $m_{\ell}(k, n)= \delta (n,k,\ell)+1$ for all $k/2 \leq \ell \leq k-1$
(independently Czygrinow and Kamat~\cite{czy} proved this for $(k, \ell)=(4,2)$).
In the case when $(k, \ell)=(3,1)$, Conjecture~\ref{generalconj} was confirmed by K\"uhn, Osthus and Treglown~\cite{KOTmatch} and independently
Khan~\cite{khan1}. Khan~\cite{khan2} also resolved the case when $(k, \ell)=(4,1)$. Alon, Frankl, Huang, R\"odl, Ruci\'nski and Sudakov~\cite{afh} determined $m_{\ell}(k, n)$ asymptotically in the case when
$(k, \ell)= (5, 1)$, $(5, 2)$, $(6, 2)$, and $(7, 3)$. Other than these results, no other
asymptotic or exact results are known (the best known general bounds are due to K\"uhn, Osthus and Townsend~\cite{tim}).
A connection between $m_{\ell}(k, n)$ and the minimum $\ell$-degree that forces a \emph{perfect fractional matching} was discovered in \cite{afh}.
Let $H$ be a $k$-uniform hypergraph on $n$ vertices. A \emph{fractional matching} in $H$ is a function $w: E(H) \rightarrow [0,1]$ such that for each $v \in V(H)$ we have that $\sum _{e \ni v} w(e)\leq1$. Then $\sum _{e \in E(H)} w(e)$ is the \emph{size} of $w$. If the size of the largest fractional matching $w$ in $H$ is $n/k$ then we say that $w$ is a \emph{perfect fractional matching}.
Given $k,\ell \in \mathbb N$ such that $\ell \leq k-1$, define $c^*_{k,\ell}$ to be the smallest number $c$ such that every $k$-uniform hypergraph $H$ on $n$ vertices with $\delta _{\ell} (H) \geq (c +o(1)) \binom{n-\ell}{k-\ell}$ contains a perfect fractional matching. It is easy to see that
the hypergraph $H^* (n,k)$ defined earlier contains no perfect fractional matching. Thus $c^*_{k,\ell}\ge 1-\left(\frac{k-1}{k}\right) ^{k-\ell}$.
Alon et al. \cite[Theorem 1.1]{afh} showed that for fixed $k, \ell$, as $n\in k\mathbb{N}$ tends to infinity,
\begin{equation}
\label{eq:afh}
m_{\ell}(k, n)= \left( \max \left \{ \frac12, \ c^*_{k, \ell} \right \} + o(1) \right) \binom{n-\ell}{k-\ell}.
\end{equation}
Furthermore, in~\cite{afh} the authors conjectured that $c^*_{k,\ell}= 1-\left(\frac{k-1}{k}\right) ^{k-\ell}$ and confirmed this for $\ell \ge k-4$.
Together with \eqref{eq:afh}, this gives the aforementioned (asymptotic) results on $m_{\ell}(k, n)$ for $(k, \ell)= (5, 1)$, $(5, 2)$, $(6, 2)$ and $(7, 3)$.
In this note we prove the following refinement of \eqref{eq:afh}.
\begin{thm} \label{mainthm}
Fix $k, \ell \in \mathbb N$ with $\ell \leq k-1$ and let $n\in k\mathbb{N}$. Then
\[
m_{\ell}(k, n)= \max \left \{ \delta (n,k,\ell)+1, \ (c^*_{k, \ell} + o(1)) \binom{n-\ell}{k-\ell} \right \}.
\]
\end{thm}
Although it looks like a small improvement, Theorem~\ref{mainthm} enables us to determine $m_{\ell}(k, n)$ \emph{exactly} whenever $c^*_{k, \ell}< 1/2$.
A recent result of K\"uhn, Osthus and Townsend~\cite[Theorem 1.9]{tim} showed that
\[
c^*_{k, \ell} \le \frac{k- \ell}{k} - \frac{k - \ell - 1}{k^{k-\ell}}
\]
for all $\ell \le k -2$. Together with $c^*_{k, k-1}= 1/k$ (see \cite{rrs1}), this implies that $c^* _{k, \ell} < 1/2$ for all $k/2 \le \ell \le k -1$. Consequently Theorem~\ref{mainthm} implies the aforementioned results of \cite{zhao, zhao2}: $m_{\ell}(k, n)= \delta (n,k, \ell)+1$ for all $k/2 \leq \ell \leq k-1$.
Furthermore, in~\cite{afh} it was shown that $c^*_{5,2}=61/125<1/2$ and $c^*_{7,3}=1105/2401<1/2$. Therefore an immediate consequence of Theorem~\ref{mainthm} is the following corollary.
\begin{cor}\label{cor1}
Suppose that $(k,\ell)=(5,2)$ or $(7,3)$. Then $m_{\ell}(k, n)= \delta (n,k, \ell)+1$ for sufficiently large $n$.
\end{cor}
After this paper was submitted, Han~\cite{hanerdos} showed that $c^* _{k,\ell}<1/2$ in the case when $0.42k\leq \ell <k/2$ or $(k,\ell)=(12,5),(17,7)$. Thus, together with Theorem~\ref{mainthm} this resolves Conjecture~\ref{generalconj} in these cases.
Let us highlight the ideas behind the proof of Theorem~\ref{mainthm}. It is informative to first recall the proof of \eqref{eq:afh}, which is an application of the absorbing method. The authors of \cite{afh} first applied a lemma of H\`an, Person and Schacht~\cite[Lemma 2.4]{hps}, which states that every $k$-uniform hypergraph $H$ with $\delta_{\ell}(H)\ge (1/2 + o(1)) \binom{n}{k - \ell}$ contains a small matching $M$ that can absorb any vertex set $W\subset V(H)\setminus V(M)$ of size much smaller than $M$ (that is, there is matching in $H$ on $V(M)\cup W$ exactly). Next they found an almost perfect matching in $H[V(H)\setminus V(M)]$ by first finding a fractional matching and then converting it to an integer matching. This immediately provides the desired perfect matching of $H$ because $M$ can absorb all uncovered vertices in $V(H)\setminus V(M)$. In order to prove Theorem~\ref{mainthm}, we prove a result stronger than \cite[Lemma 2.4]{hps}, Theorem~\ref{thm:abs}, which implies that every $k$-uniform hypergraph $H$ with $\delta_{\ell}(H)\ge (1/2 - o(1)) \binom{n}{k - \ell}$ either contains the aforementioned $M$ or looks like a hypergraph in $\mathcal H_{\text{ext}} (n,k)$. If $H$ contains $M$, then we proceed as in \cite{afh} (except that we apply a lemma from \cite{tim} when converting a fractional matching to an integer matching); if $H$ looks like a hypergraph in $\mathcal H_{\text{ext}} (n,k)$ and $\delta_{\ell}(H)\ge \delta(n, k, \ell) + 1$, then we obtain a perfect matching by applying \cite[Theorem 4.1]{zhao}.
Theorem~\ref{thm:abs} is the main contribution of this note -- it is stronger than two absorbing theorems in our previous papers \cite[Theorem 3.1]{zhao} and \cite[Theorem 3.1]{zhao2}, in which we assume that $\ell \ge k/2$. The proof of Theorem~\ref{thm:abs} is actually shorter than those of the two previous absorbing theorems because 1) we use a different absorbing structure which allows us to apply a lemma from \cite{lo}; 2) when proving Lemma~\ref{lem:main}, we avoid using auxiliary hypergraphs and obtain the structure of $H$ by considering the neighborhoods of the vertices of $H$ directly.
\vspace{2mm}
\noindent\textbf{Notation:}
Given a set $X$ and $r\in \mathbb N$, we write $\binom{X}{r}$ for the set of all $r$-element subsets of $X$.
Let $H$ be a $k$-uniform hypergraph.
We write $V(H)$ for the vertex set and $E(H)$ for the edge set of $H$.
Define $e(H):=|E(H)|$. Given ${v}\in V(H)$, we write $N_H(v)$ to denote the \emph{neighborhood of $v$}, that is, the family of those $(k-1)$-subsets of $V(H)$ which, together with $v$, form an edge in $H$. Given $X \subseteq V(H)$, we write $H[X]$ for the \emph{subhypergraph of $H$ induced by $X$}, namely, $H[X] := (X, E(H)\cap \binom{X}{k})$. We denote the \emph{complement of $H$} by $\overline{H}$. That is, $\overline{H} := (V(H), \binom{V(H)}{k}\setminus E(H))$.
Suppose that $n,k \in \mathbb N$. When $|A|=\lfloor n/2 \rfloor$ and $|B|=\lceil n/2 \rceil$, we define $\mathcal B_{n,k}:=\mathcal B_{n,k}(A,B)$ and $\overline{\mathcal B}_{n,k} := \overline{\mathcal B}_{n,k}(A, B)$.
We will often write $0<a_1 \ll a_2 \ll a_3$ to mean that we can choose the constants
$a_1,a_2,a_3$ from right to left. More
precisely, there are increasing functions $f$ and $g$ such that, given
$a_3$, whenever we choose some $a_2 \leq f(a_3)$ and $a_1 \leq g(a_2)$, all
calculations needed in our proof are valid.
Hierarchies with more constants are defined in the obvious way.
\section{Proof of Theorem~\ref{mainthm}}
The lower bound for $m_{\ell}(k, n)$ in Theorem~\ref{mainthm} follows from the definitions of $\delta(n, k, \ell)$ and $c^*_{k, \ell}$ immediately. The following (more general) result provides the desired upper bound for $m_{\ell}(k, n)$.
\begin{thm}\label{main2}
Given any $\theta >0$, $k, \ell, \ell ' \in \mathbb N$ where $1\leq \ell, \ell ' \leq k-1$ there is an $n_0 \in \mathbb N$ such that the following holds. Let $n \geq n_0$ where $k$ divides $n$.
If $H$ is a $k$-uniform hypergraph on $n$ vertices with
\begin{equation}
\label{eq:2deg}
\delta _{\ell} (H) > \delta (n,k,\ell) \ \ \text{ and } \ \ \delta_{\ell '} (H) > (c^*_{k,\ell'} +\theta) \binom{n-\ell '}{k-\ell '} ,
\end{equation}
then $H$ contains a perfect matching.
\end{thm}
The proof of Theorem~\ref{main2} splits into \emph{extremal} and \emph{non-extremal} cases, the former case being when $H$ looks like an element of $\mathcal H_{\text ext} (n,k)$. To make this precise we introduce more notation.
Let $\varepsilon>0$. Suppose that $H$ and $H'$ are $k$-uniform hypergraphs on $n$ vertices. We say that $H$ is \emph{$\varepsilon$-close to $H'$} if $H$ becomes a copy of $H'$ after adding and deleting at most $\varepsilon n^k$ edges. More precisely, $H$ is $\varepsilon$-close to $H'$ if there is an isomorphic copy $\tilde{H}$ of $H$ such that $V(\tilde{H}) = V(H')$ and $|E(\tilde{H})\triangle E(H')| \le \varepsilon n^k$.
Our proof of the non-extremal case uses the absorbing method. Given a $k$-uniform hypergraph $H$, a set $S\subseteq V(H)$ is called an \emph{absorbing set for $Q\subseteq V(H)$}, if both $H[S]$ and $H[S\cup Q]$ contain perfect matchings. In this case, if the matching covering $S$ is $M$, we also say \emph{$M$ absorbs $Q$}.
Our main result, Theorem~\ref{thm:abs}, extends \cite[Theorem 3.1]{zhao2}. It ensures that if $H$ is as in Theorem~\ref{main2} then $H$ contains a small absorbing matching or $H$ is close to one of $\mathcal B_{n,k}$ and $\overline{\mathcal B}_{n,k}$. We postpone its proof to the next subsection.
\begin{thm}\label{thm:abs}
Given any $\varepsilon >0$ and integer $k \geq 2$, there exist $0< \alpha, \xi < \varepsilon$
and $n_0 \in \mathbb N$ such that the following holds. Suppose that $H$ is a $k$-uniform hypergraph on $n \geq n_0$ vertices. If
$$\delta _{1} (H) \geq \left( \frac{1}{2}-\alpha \right) \binom{n - 1}{k- 1}$$
then $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$, or $H$ contains
a matching $M$ of size $|M| \le \xi n/k$ that absorbs any set $W\subseteq V(H) \setminus V(M)$ such that $|W| \in k\mathbb{N}$ with $|W| \le \xi^2 n$.
\end{thm}
The next result from~\cite{zhao} ensures a perfect matching when our hypergraph $H$ is close to one of the
extremal hypergraphs $\mathcal B_{n,k}$ and $\overline{\mathcal B}_{n,k}$.
\begin{thm}\cite[Theorem 4.1]{zhao}\label{extthm}
Given $1\le \ell \le k-1$, there exist $\varepsilon > 0$ and $n_0 \in \mathbb{N}$ such that the following holds. Suppose that $H$ is a $k$-uniform hypergraph on $n\ge n_0$ vertices such that $n$ is divisible by $k$. If $\delta_{\ell} (H) > \delta(n, k, \ell)$ and $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$, then $H$ contains a perfect matching.
\end{thm}
The final tool required for the proof of Theorem~\ref{main2} is a weaker version of Lemma 5.6 in~\cite{tim}.
\begin{lemma}\cite{tim} \label{lem:frac}
Let $k \geq 2$ and $1 \leq \ell \leq k-1$ be integers, and let $\varepsilon >0$. Suppose that for some $b,c \in (0,1)$ and some $n_0 \in \mathbb N$, every $k$-uniform hypergraph $H$ on $n \geq n_0$ vertices with $\delta _{\ell} (H) \geq cn^{k-\ell}$ has a fractional matching of size $(b+\varepsilon)n$. Then there exists an $n_0 ' \in \mathbb N$ such that any $k$-uniform hypergraph $H$ on $n \geq n'_0$ vertices with $\delta _{\ell} (H) \geq (c+ \varepsilon) n^{k-\ell}$ contains a matching of size at least $bn$.
\end{lemma}
\medskip
\noindent
\textbf{Proof of Theorem~\ref{main2}.}
Choose $\varepsilon >0$ from Theorem~\ref{extthm}. We may additionally assume that $\varepsilon \ll \theta ,1/k$.
Let $0<\alpha, \xi<\varepsilon$ be as in Theorem~\ref{thm:abs}.
Let $n$ be sufficiently large and divisible by $k$. Assume that $H$ is a $k$-uniform hypergraph on $n$ vertices satisfying \eqref{eq:2deg}.
Since $\delta _{\ell} (H) > \delta (n,k,\ell) = (1/2 - o(1)) \binom{n- \ell}{k- \ell}$, it follows that $\delta_1(H)\ge (1/2 - \alpha) \binom{n-1}{k-1}$. By Theorem~\ref{thm:abs}, $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$, or $H$ contains a matching $M$ of size $|M| \le \xi n/k$ that absorbs any set $W\subseteq V(H) \setminus V(M)$ satisfying $|W| \in k\mathbb{N}$ with $|W| \le \xi^2 n$.
In the former case, since $\delta _{\ell} (H) > \delta (n,k,\ell)$, Theorem~\ref{extthm} provides a perfect matching in $H$. In the latter case, set $c^* := c^*_{k, \ell'}$, $H':=H[V(H)\setminus V(M)]$ and $n_1:=|V(H')|$.
Since $|V(M)| \le \xi n$,
\[
\delta _{\ell'} (H') \geq \delta _{\ell'} (H) - |V(M)|\binom{n - \ell' - 1}{k-\ell'-1} \geq (c^* + \theta - \xi k) \binom{n- \ell'}{k- \ell'} > (c^* /(k-\ell ')! +2\xi ^2 )n_1 ^{k- \ell'},
\]
where the last inequality follows since $\xi \ll \theta , 1/k$.
Let $c:= c^*/(k- \ell')! + \xi ^2$. By the definition of $c^*$, for sufficiently large $\tilde{n}$, every $k$-uniform hypergraph $F$ on $\tilde{n}$ vertices with $\delta_{\ell'}(F)\ge c \tilde{n} ^{k- \ell'}$ contains a perfect fractional matching.
Applying Lemma~\ref{lem:frac} with $ \xi ^2/k$ and $(1 - \xi ^2)/k$ playing the roles of $\varepsilon$ and $b$ respectively, we conclude that $H'$ contains a matching $M'$ of size at least $(1- \xi^2) n_1/k$. Let $W$ be the uncovered vertices of $H'$. Then $|W|\le \xi^2 n$. We finally absorb $W$ using the absorbing property of $M$.
\noproof\bigskip
\subsection{Proof of Theorem~\ref{thm:abs} }
The proof of Theorem~\ref{thm:abs} follows from the following three lemmas.
Lemma~\ref{lo} is a special case of \cite[Lemma~1.1]{lo} and gives a sufficient condition for a hypergraph $H$ to contain a small matching that absorbs \emph{any} much smaller set of vertices from $H$.
\begin{lemma}\cite{lo} \label{lo}
Let $k\in \mathbb N$ and $\gamma ' >0$.
Then there exists an $n_0 \in \mathbb N$ such that the following holds. Suppose that $H$ is a $k$-uniform hypergraph
on $n \geq n_0$ vertices so that, for any $x,y \in V(H)$, there are at least $\gamma ' n^{2k-1}$ $(2k-1)$-sets $X \subseteq V(H)$ such that both $H[X \cup \{x\}]$ and $H[X \cup \{y\}]$ contain perfect matchings.
Then $H$ contains a matching $M$ so that
\begin{itemize}
\item $|M|\leq (\gamma '/2)^k n/(8k^2(k-1))$;
\item $M$ absorbs any $W \subseteq V(G) \setminus M$ such that $|W| \in k \mathbb N$ and $|W|\leq (\gamma '/2)^{2k} n/(128k(k-1)^2) $.
\end{itemize}
\end{lemma}
\begin{lemma}\label{absorbing}
Let $k\in \mathbb N$ and $0< \gamma ' \ll \gamma \ll 1/k$. There exists an $n_0 \in \mathbb N$ such that the following holds.
Let $H=(V,E)$ be a $k$-uniform hypergraph on $n \geq n_0$ vertices. Suppose that for every $x,y \in V$
at least one of the following conditions holds.
\begin{itemize}
\item[(i)] $|N_H (x) \cap N_H (y) | \geq \gamma n^{k-1}$;
\item[(ii)] There exists at least $\gamma n$ vertices $z \in V$ such that
$|N_H(x) \cap N_H(z)| \geq \gamma n^{k-1}$ and $|N_H(y) \cap N_H(z)| \geq \gamma n^{k-1}$.
\end{itemize}
There there are at least $\gamma ' n^{2k-1}$ $(2k-1)$-sets $X \subseteq V$ such that both $H[X \cup \{x\}]$ and $H[X \cup \{y\}]$ contain perfect matchings.
\end{lemma}
\begin{lemma} \label{lem:main}
Let $k\in \mathbb N$ and $0< \alpha \ll \gamma \ll \varepsilon , 1/k$.
Then there exists an $n_0 \in \mathbb N$ such that the following holds. Let $H=(V, E)$ be a $k$-uniform hypergraph on $n \geq n_0$ vertices such that $\delta_1(H)\ge (1/2 - \alpha) \binom{n-1}{k-1}$.
Suppose that there exists $x_0, y_0\in V$ such that
\begin{itemize}
\item[(i)] $|N_H (x_0) \cap N_H (y_0) | < \gamma n^{k-1}$;
\item[(ii)] at most $\gamma n$ vertices $z \in V$ satisfy $|N_H(z) \cap N_H(x_0)| \geq \gamma n^{k-1}$ and $|N_H(z) \cap N_H(y_0)| \geq \gamma n^{k-1}$.
\end{itemize}
Then $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$.
\end{lemma}
We postpone the proof of Lemmas~\ref{absorbing} and \ref{lem:main} and prove Theorem~\ref{thm:abs} first.
\smallskip
{\noindent \bf Proof of Theorem~\ref{thm:abs}.} Given $\varepsilon >0$ and $k \geq 2$, choose constants $\alpha , \gamma ', \gamma$ so that
$0<\alpha \ll \gamma ' \ll \gamma \ll \varepsilon, 1/k$. Set $\xi := (\gamma '/2)^k / \sqrt{128k(k-1)^2}$. Let $n$ be sufficiently large and $H$ be a $k$-uniform hypergraph as in the statement of the theorem.
By Lemmas~\ref{absorbing} and~\ref{lem:main}, $H$ is $\varepsilon$-close to $\mathcal B_{n,k}$ or $\overline{\mathcal B}_{n,k}$ or for every $x, y\in V(H)$ there are at least $\gamma ' n^{2k-1}$ $(2k-1)$-sets $X \subseteq V(H)$ such that both $H[X \cup \{x\}]$ and $H[X \cup \{y\}]$ contain perfect matchings. In the former case we are done. In the latter case, Lemma~\ref{lo} implies that $H$ contains a matching $M$ so that
\begin{itemize}
\item $|M|\leq (\gamma '/2)^k n/(8k^2(k-1)) \leq\xi n/k$;
\item $M$ absorbs any $W \subseteq V(H) \setminus M$ such that $|W| \in k \mathbb N$ and $|W|\leq (\gamma '/2)^{2k} n/(128k(k-1)^2) =\xi ^2 n $,
\end{itemize}
as required.
\noproof\bigskip
\noindent
{\bf Proof of Lemma~\ref{absorbing}.}
Note that (i) and (ii) imply that $\delta _1 (H) \geq \gamma n^{k-1}$ and
so $e(H) \geq \gamma n^k/k$.
Consider any $x,y \in V$. First assume that (i) holds. Fix any $X' \subseteq V$ where $|X'|=k-1$ and $X' \cup \{x\}, X' \cup \{y\} \in E$. By (i) there are at least $\gamma n^{k-1}$ choices for $X'$.
Next choose some $X'' \subseteq V\setminus (X' \cup \{x,y\})$ such that $|X''|=k$ and $X''\in E$. There are at least $\gamma n^k/k -(k+1) \binom{n}{k-1} \geq \gamma n^k /(2k)$ choices for $X''$.
Set $X:=X' \cup X''$. Note that both $H[X \cup \{x\}]$ and $H[X \cup \{y\}]$ contain perfect matchings. Further, since there are at least
$\gamma n^{k-1}$ choices for $X'$, at least $ \gamma n^k /(2k)$ choices for $X''$ and each $(2k-1)$-set may be counted at most $\binom{2k-1}{k-1}$ times, there are at least
\[
\gamma n^{k-1} \times \frac{\gamma n^k }{2k} \times \frac{1}{ \binom{2k-1}{k-1}} > \gamma ' n^{2k-1}
\]
choices for $X$ (as $\gamma ' \ll \gamma \ll 1/k$), as desired.
Now suppose that (ii) holds. Fix any $z \in V$ such that $|N_H(x) \cap N_H(z)| \geq \gamma n^{k-1}$ and $|N_H(y) \cap N_H(z)| \geq \gamma n^{k-1}$. There are at least $\gamma n $ choices for $z$. Next fix some $X' \in N_H(x) \cap N_H(z)$ that is disjoint from $y$. There are at least
$\gamma n^{k-1}-\binom{n}{k-2} \geq \gamma n^{k-1}/2$ choices for $X'$. Finally, fix some $X'' \in N_H(y) \cap N_H(z)$ so that $X''$ is disjoint from $X' \cup \{x\}$.
There are at least $\gamma n^{k-1} -k \binom{n}{k-2} \geq \gamma n^{k-1} /2$ choices for $X''$. Set $X:=X' \cup X''\cup \{z\}$.
So $|X|=2k-1$ and both $H[X \cup \{x\}]$ and $H[X \cup \{y\}]$ contain perfect matchings.
Further, there are at least
$$\gamma n \times \frac{\gamma n^{k-1}}{2} \times \frac{\gamma n^{k-1}}{2} \times \frac{1}{(2k-1)\binom{2k-2}{k-1}} > \gamma ' n^{2k-1}$$ choices for $X$ (as $\gamma ' \ll \gamma \ll 1/k$), as desired.
\noproof\bigskip
The rest of this section is devoted to the proof of Lemma~\ref{lem:main}. We draw on ideas used in the proof of Lemma 5.4 in~\cite{zhao2}. We need two results from \cite{zhao2}. The first one implies that if any two vertices in a hypergraph have roughly the same neighborhood, then the hypergraph is near complete or empty.
\begin{lemma} \cite[Lemma 2.2]{zhao2}
\label{lem:KK}
Given any $k \in \mathbb N$ and $\rho >0$ there exists an $n_0 \in \mathbb N$ such that the following holds.
Let $F=(V, E)$ be a $k$-uniform hypergraph on $n \geq n_0$ vertices with edge density $|E|/\binom{n}{k}\in [\rho, 1 - \rho]$.
Then there exist two vertices $v, v'\in V$ such that $| N_F(v)\triangle N_F(v') | \ge \rho (1- \rho) n^{k-1}/(k+1)!$.
\end{lemma}
\begin{prop}\cite[Proposition 2.3]{zhao2}
\label{evensum}
For $r \in \mathbb N$, $0\le c\le 1$ and $n\to \infty$,
\begin{align*}
\sum_{0\le i\le r, \, i \, \rm{even}} \binom{cn}{r- i} \binom{ (1-c)n }{i} & = \frac{n^r}{2r!} ( 1+ (2c-1)^r ) - O(n^{r-1}), \\
\sum_{0\le i\le r, \, i \,\rm{odd}} \binom{cn}{r- i} \binom{ (1-c)n }{i} & = \frac{n^r}{2r!} ( 1- (2c-1)^r ) - O(n^{r-1}).
\end{align*}
\end{prop}
\smallskip
\noindent
{\bf Proof of Lemma~\ref{lem:main}.}
Define
\[
X := \{ v\in V: |N_H(y_0) \cap N_H(v)| < \gamma n^{k-1} \} \quad \text{and} \quad
Y := \{v\in V: |N_H(x_0) \cap N_H(v)| < \gamma n^{k-1} \}.
\]
Then by Lemma~\ref{lem:main} (i), $x_0\in X$ and $y_0\in Y$. Let $V_0:= V\setminus (X\cup Y)$. We have $|V_0| \le \gamma n$ by Lemma~\ref{lem:main} (ii).
Roughly speaking, our goal is to show that $|X| \approx |Y| \approx n/2$ and $H \approx \mathcal B _{n,k}(X, Y)$ or $H \approx \overline{\mathcal B} _{n,k}(X, Y)$.
We first provide several properties of $X$ and $Y$, for example, $X\cap Y = \emptyset$, and
$N_H(v)\approx N_H(v')$ whenever $v, v'\in X$ or $v, v'\in Y$.
\begin{claim}\label{clm:vv'} The following conditions hold.
\begin{enumerate}[{\rm (i)}]
\item For all $v\in X\cup Y$, we have $d_H(v) \le (1/2 + \alpha) \binom{n-1}{k-1} + \gamma n^{k-1}$.
\item $X\cap Y = \emptyset$.
\item For any two vertices $x, x' \in X$, we have $|N_H(x) \triangle N_H(x')| < 5\gamma n^{k-1}$. The same holds for all $y, y' \in Y$.
\item For any $x\in X$ and $y\in Y$, we have $|N_H(x) \cap N_H(y)| \le 4\gamma n^{k-1}$ and $|\overline{N}_H(x) \cap \overline{N}_H(y)| \le 4\gamma n^{k-1}$, where $\overline{N}_H(x) := \binom{V\setminus \{x\} }{ k-1}\setminus N_H(x)$ consists of \emph{non-neighbors} of $x$.
\end{enumerate}
\end{claim}
\begin{proof}
To see (i), suppose $x\in X$ and $y\in Y$. Since $d_H(x_0), d_H(y_0)\ge (1/2 - \alpha) \binom{n-1}{k-1}$, by the definition of $X$ and $Y$,
\begin{align} \label{eq:Nx0}
|N_H(y_0)\setminus N_H(x)|, |N_H(x_0)\setminus N_H(y)| \ge \left(\frac12 - \alpha \right) \binom{n-1}{k-1} - \gamma n^{k-1}.
\end{align}
Consequently, we have $d_H(x), d_H(y) \le (1/2 + \alpha) \binom{n-1}{k-1} + \gamma n^{k-1}$.
\smallskip
To see (ii), suppose that there exists $v\in X\cap Y$. Then by \eqref{eq:Nx0},
\begin{align*}
|( N_H( x_0)\cup N_H(y_0) ) \setminus N_H(v) | &\ge | N_H(x_0)\setminus N_H(v) | + | N_H(y_0) \setminus N_H(v)| - | N_H(x_0) \cap N_H(y_0) |\\
&\ge 2 \left(\frac{1}2 - \alpha \right) \binom{n-1}{k-1} - 3\gamma n^{k-1} \ge \binom{n-1}{k-1} - 4\gamma n^{k-1},
\end{align*}
which implies that $|N_H(v)|\le 4\gamma n^{k-1}$, contradicting the minimum degree condition of $H$.
\smallskip
To see (iii), consider $x\in X$. By the definition of $X$ and the minimum degree condition of $H$, we have
$| N_H(x) \cup N_H(y_0)| \ge 2 \left(\frac{1}2 - \alpha \right) \binom{n-1}{k-1} - \gamma n^{k-1}$.
Let $x'\in X\setminus \{x\}$. Then
\begin{align}
| N_H(x') \setminus N_H(x) | & = | (N_H(x') \setminus N_H(x) )\setminus N_H(y_0) ) | + | (N_H(x') \setminus N_H(x)) \cap N_H(y_0)| \nonumber\\
&\le \left | \binom{V\setminus \{x'\} }{k-1}\setminus ( N_H(x) \cup N_H(y_0) ) \right | + | N_H(x')\cap N_H(y_0) | \nonumber\\
&\le \binom{n-1}{k-1} - 2 \left(\frac{1}2 - \alpha \right) \binom{n-1}{k-1} + \gamma n^{k-1} + \gamma n^{k-1} \nonumber\\
&= 2\alpha \binom{n-1}{k-1} + 2\gamma n^{k-1}. \label{eq:difN}
\end{align}
The same bound holds for $|N_H(x) \setminus N_H(x') |$. Hence
\begin{align*}
|N_H(x) \triangle N_H(x')| &= | N_H(x') \setminus N_H(x) | + |N_H(x) \setminus N_H(x') | \le 2\left( 2\alpha \binom{n-1}{k-1} + 2\gamma n^{k-1} \right) < 5\gamma n^{k-1}.
\end{align*}
Analogously we can derive that $|N_H(y) \triangle N_H(y')| < 5\gamma n^{k-1}$ for any $y, y' \in Y$.
\smallskip
To see (iv), consider $x\in X$ and $y\in Y$. By \eqref{eq:difN}, we have $|N_H(x)\setminus N_H(x_0)| \le 2\alpha \binom{n-1}{k-1} + 2\gamma n^{k-1}$. By the definition of $Y$, we have $|N_H(y) \cap N_H(x_0)| < \gamma n^{k-1}$. Thus
\begin{align}\label{eq:comN}
| N_H(x) \cap N_H(y) | \le | N_H(y) \cap N_H(x_0) | + | N_H(x) \setminus N_H(x_0) | \le 2\alpha \binom{n-1}{k-1} + 3\gamma n^{k-1} \le 4\gamma n^{k-1},
\end{align}
which proves the first assertion of (iv). By the minimum degree condition and \eqref{eq:comN}, we have
\[
|N_H(x) \cup N_H(y)| \ge 2 \left(\frac{1}2 - \alpha \right) \binom{n-1}{k-1} - 3\gamma n^{k-1} - 2\alpha \binom{n-1}{k-1}.
\]
It follows that
\begin{align*}
|\overline{N}_H(x) \cap \overline{N}_H(y)| &\le \binom{|V\setminus \{x, y\}|}{k-1} - | N_H(x) \cup N_H(y) |\\
&\le \binom{n-2}{k-1} - (1 - 2\alpha)\binom{n-1}{k-1} + 3\gamma n^{k-1} + 2\alpha \binom{n-1}{k-1} \le 4\gamma n^{k-1},
\end{align*}
which proves the second assertion of (iv).
\end{proof}
Since $|V_0|\leq \gamma n$ and $X \cap Y =\emptyset$ we have $|X|\geq (1-\gamma)n/2$ or $|Y|\geq (1-\gamma )n/2$. Without loss of generality we may assume that
$|X|\geq (1-\gamma)n/2\ge n/3.$
Let $0< \gamma_0< 1/2$ such that $\frac{\gamma_0(1-\gamma_0)}{ (k+1)! } = 5\gamma \cdot 3^{k-1}$. We apply Lemma~\ref{lem:KK} to $F= H[X]$ with $\rho = \gamma_0$. Since
\[
|N_F(v) \triangle N_F(v')|\leq |N_H(v) \triangle N_H(v')| < 5\gamma n^{k-1} \le \frac{\gamma_0(1-\gamma_0)}{ (k+1)! }|X|^{k-1}
\]
for any $v, v'\in X$ (Claim~\ref{clm:vv'} (iii)), there are two possible cases:
\begin{description}
\item[Case 1] $e(H[X]) \le \gamma_0\binom{|X|}{k} $,
\item[Case 2] $e(H[X]) \ge (1- \gamma_0)\binom{|X|}{k}$.
\end{description}
In the rest of the proof we assume that one of the two cases holds. Once we have obtained more information we will prove that $|X|$ and $|Y|$ are close to $n/2$. At present we require the following weaker lower bounds on $|X|$ and $|Y|$.
\begin{claim}\label{clm:XY}
$|X|,|Y| \ge (1 - (\frac12 + 2\gamma_0)^{\frac{1}{k-1}} - \gamma) n$.
\end{claim}
\begin{proof} The bound on $|X|$ follows since $|X|\geq (1-\gamma)n/2$.
Since $|X|+ |Y| + |V_0| = n$ and $|V_0| \le \gamma n$, to prove the bound on $|Y|$, it suffices to show that $|X|\leq (\frac12 + 2\gamma_0)^{\frac{1}{k-1}} n$. In Case 1, there exists a vertex $x\in X$ such that $d_{H[X]}(x) \le \gamma_0 \binom{|X| -1}{k-1}$ and consequently, $|\overline{N}_{H[X]}(x)| \ge (1- \gamma_0) \binom{|X| -1}{k-1}$. Together with the minimum degree condition, this gives
\[
\left (\frac12 - \alpha \right ) \binom{n-1}{k-1} \le d_{H}(x) \le \binom{n-1}{k-1} - (1- \gamma_0) \binom{|X| -1}{k-1}.
\]
In Case 2, there exists a vertex $x\in X$ such that $d_{H[X]}(x) \ge (1- \gamma_0) \binom{|X| -1}{k-1}$. By Claim~\ref{clm:vv'} (i),
\[
(1- \gamma_0) \binom{|X| -1}{k-1} \le d_{H[X]}(x) \le \left(\frac12 + \alpha \right) \binom{n-1}{k-1} + \gamma n^{k-1}.
\]
In either case we have $(1- \gamma_0) \binom{|X| -1}{k-1} \le (\frac12 + \alpha) \binom{n-1}{k-1} + \gamma n^{k-1}$, which implies that $\binom{|X| -1}{k-1} \le (1+ 2\gamma_0)((1/2 + \alpha) \binom{n-1}{k-1} + \gamma n^{k-1})$. Letting $|X| = c n$, it follows that
\[
c^{k-1} \binom{n-1}{k-1} - O(n^{k-2}) \le \left(\frac12 + \gamma_0 + 2\alpha \right) \binom{n-1}{k-1}+ 2\gamma n^{k-1}
\]
Since $\gamma n^{k-1} \le \frac{\gamma_0}{5} \binom{n-1}{k-1}$, we conclude that $c\le (\frac12 + 2\gamma_0)^{\frac{1}{k-1}}$.
\end{proof}
Given two disjoint subsets $A, B\subset V$ and two integers $i, j\ge 0$, we call an $(i+j)$-set $S\subseteq V$ an $A^i B^j$-set if $|S\cap A|=i$ and $|S\cap B|=j$, and let $A^i B^j$ denote the family of all $A^i B^j$-sets.
Let $c_0 := 1 - (\frac12 + 2\gamma _0)^{\frac{1}{k-1}} - \gamma $, and $\gamma_i := \gamma_{i-1} + 5\gamma k!/ c_0^{k-1}$ for $i=1, \dots, k$.
\begin{claim} \label{clm:dXY}
\begin{enumerate}[{\rm (i)}]
\item In Case 1, for $1\le i\le k$, for any $y\in Y$, at least $(1- \gamma_i) \binom{|X|}{k-i} \binom{|Y|}{i-1}$ $X^{k-i} Y^{i-1}$-sets are neighbors (respectively, non-neighbors) of $y$ if $i$ is odd (respectively, even).
\item In Case 2, for $1\le i\le k$, for any $y\in Y$, at least $(1- \gamma_i) \binom{|X|}{k-i} \binom{|Y|}{i-1}$ $X^{k-i} Y^{i-1}$-sets are neighbors (respectively, non-neighbors) of $y$ if $i$ is even (respectively, odd).
\end{enumerate}
\end{claim}
\begin{proof} We prove both cases by induction on $i$. In Case 1 there exists a vertex $x_1\in X$ such that $d_{H[X]}(x_1) \le \gamma_0 \binom{|X| -1}{k-1}$ and consequently $|\overline{N}_H(x_1)\cap \binom{X}{k-1}| \ge (1 - \gamma_0) \binom{|X| -1}{k-1}$.
Fix a vertex $y\in Y$. By Claim~\ref{clm:vv'} (iv), $|\overline{N}_H(x_1) \cap \overline{N}_H(y)| \le 4\gamma n^{k-1}$. Thus at least $(1 - \gamma_0) \binom{|X| -1}{k-1} - 4\gamma n^{k-1}$ $X^{k-1}$-sets are neighbors of $y$. By Claim~\ref{clm:XY}, $|X|, |Y|\ge c_0 n$. Then for any $0\le i\le k-1$,
\begin{equation}\label{eq:XYi}
\binom{|X|}{k-i-1} \binom{|Y|}{i} \ge \frac{(c_0 n)^{k-1}}{(k-1)!} - O(n^{k-2}) \ge \frac{c_0^{k-1} n^{k-1}}{k!}.
\end{equation}
Together with the definition of $\gamma_1$, we conclude that at least
\[
(1 - \gamma_0) \binom{|X| -1}{k-1} - 4\gamma n^{k-1} \ge (1 - \gamma_0) \binom{|X|}{k-1} - O(n^{k-2}) - \frac{4\gamma k!}{c_0^{k-1}} \binom{|X|}{k-1} \ge (1 - \gamma_1) \binom{|X|}{k-1}
\]
$X^{k-1}$-sets are neighbors of $y$. This confirms (i) for $i=1$. In Case 2, by averaging, there exists a vertex $x_1\in X$ such that $d_{H[X]}(x_1) \ge (1- \gamma_0) \binom{|X| -1}{k-1}$. Fix a vertex $y\in Y$. By Claim~\ref{clm:vv'} (iv), $| N_H(x_1) \cap N_H(y)| \le 4\gamma n^{k-1}$. Thus at least
$(1 - \gamma_0) \binom{|X| -1}{k-1} - 4\gamma n^{k-1} \ge (1 - \gamma_1) \binom{|X|}{k-1}$
$X^{k-1}$-sets are non-neighbors of $y$. This confirms (ii) for $i=1$.
For the induction step, we first assume that for some $1\le i\le k$, every $y\in Y$ has at least $(1- \gamma_i) \binom{|X|}{k-i} \binom{|Y|}{i-1}$ $X^{k-i} Y^{i-1}$-sets in its neighborhood. Consequently at least $(1- \gamma_i) \binom{|X|}{k-i} \binom{|Y|}{i}$ $X^{k-i} Y^{i}$-sets are edges of $H$. By averaging, there exists $x_i\in X$ whose neighborhood contains at least $(1- \gamma_i) \binom{|X| - 1}{k- i-1} \binom{|Y|}{i}$ $X^{k-i-1} Y^{i}$-sets. Fix $y\in Y$. By Claim~\ref{clm:vv'} (iv),
\[
|(N_H(x_i)\cap X^{k-i-1} Y^i) \setminus N_H(y)|\ge (1 - \gamma_i) \binom{|X| - 1}{k- i- 1} \binom{|Y|}{i} - 4\gamma n^{k-1} .
\]
Since $|\overline{N}_H(y)\cap X^{k-i-1} Y^i| \ge |(N_H(x_i)\cap X^{k-i-1} Y^i) \setminus N_H(y)| - O(n^{k-2})$, we conclude that at least
\[
(1 - \gamma_i) \binom{|X| - 1}{k- i- 1} \binom{|Y|}{i} - 4\gamma n^{k-1} - O(n^{k-2})\ge (1 - \gamma_{i+1}) \binom{|X|}{k-i -1} \binom{|Y|}{i}
\]
$X^{k-i-1} Y^{i}$-sets are non-neighbors of $y$, where we use \eqref{eq:XYi} and the definition of $\gamma_{i+1}$.
Analogously we can show that if for some $1\le i\le k$, every $y\in Y$ has at least $(1- \gamma_i) \binom{|X|}{k-i} \binom{|Y|}{i-1}$ $X^{k-i} Y^{i-1}$-sets as non-neighbors, then at least $(1 - \gamma_{i+1}) \binom{|X|}{k-i -1} \binom{|Y|}{i}$ $X^{k-i-1} Y^{i}$-sets are neighbors of $y$.
This completes our induction proof.
\end{proof}
\begin{claim}\label{clm:n2}
$|X|, |Y| \ge (1- \eta) n/2$, where $\eta := (2\gamma_k)^{1/(k-1)} + \gamma$.
\end{claim}
\begin{proof}
Suppose that Claim~\ref{clm:dXY} (i) holds (the proof when Claim~\ref{clm:dXY} (ii) holds is analogous).
Let $\tilde{n} := |X\cup Y|$. Note that $\tilde{n} \ge (1- \gamma)n$ because $|V_0| \leq \gamma n$.
Let $c:= |X| / \tilde{n}$. It suffices to show that $(1 - (2\gamma_k)^{1/(k-1)})/2 \le c\le (1 + (2\gamma_k)^{1/(k-1)})/2 $ because this implies that
\[
|X| = c\tilde{n}\ge \frac12 \big(1 - (2\gamma_k)^{1/(k-1)} \big) (1- \gamma)n > (1 - \eta) \frac{n}{2}
\]
and $|Y| = (1-c)\tilde{n} \ge \frac12 (1 - (2\gamma_k)^{1/(k-1)}) (1- \gamma)n > (1 - \eta) {n}/{2}$.
For any $y\in Y$, by Claim~\ref{clm:dXY} (i),
\[
d_H(y) \ge \sum_{1\le i\le k, \, i \,\rm{odd} } (1- \gamma_i) \binom{|X| }{k-i} \binom{|Y|}{i-1} \ge (1-\gamma_k) \sum_{0\le j\le k-1, \, j \,\rm{even} } \binom{|X|}{k-1-j} \binom{|Y|}{j} .
\]
Hence, by Proposition~\ref{evensum}, ${d}_H(y) \ge (1- \gamma_k) \frac{\tilde{n}^{k-1}}{2(k-1)!} ( 1+ (2c-1)^{k-1} ) -O(n^{k-2})$. If $(2c-1)^{k-1} \ge 2\gamma_k$, then
\[
d_H(y) \ge (1- \gamma_k)\frac{(1- \gamma)^{k-1} n^{k-1}}{2(k-1)!} (1+ 2\gamma_k) - O(n^{k-2}) \ge \left (1 + \frac{\gamma_k}2\right ) \frac{n^{k-1}}{2(k-1)!},
\]
as $5\gamma(k-1)! < \gamma_k \ll 1$.
This contradicts Claim~\ref{clm:vv'} (i).
Thus $(2c-1)^{k-1} < 2\gamma_k$. If $c\ge 1/2$, then $c< (1+ (2\gamma_k)^{1/(k-1)})/2$; if $c< 1/2$ and $k-1$ is even, then $(1-2c)^{k-1} = (2c- 1)^{k-1} < 2\gamma_k$ and thus $c> (1 - (2\gamma _k)^{1/{(k-1)}})/2$. In either case we are done. Otherwise assume that $c< 1/2$ and $k-1$ is odd. By Claim~\ref{clm:dXY} (i),
\[
\overline{d}_H(y) \ge \sum_{1\le i\le k, \, i \,\rm{even} } (1- \gamma_i) \binom{|X| }{k-i} \binom{|Y|}{i-1} \ge (1-\gamma_k) \sum_{0\le j\le k-1, \, j \,\rm{odd} } \binom{|X|}{k-1-j} \binom{|Y|}{j},
\]
where $\overline{d}_H(y):= |\overline{N} _H (y)|$.
By Proposition~\ref{evensum}, we have
$\overline{d}_H(y) \ge (1- \gamma_k) \frac{\tilde{n}^{k-1}}{2(k-1)!} ( 1- (2c-1)^{k-1} ) -O(n^{k-2})$.
If $1- (2c-1)^{k-1} = 1 + (1-2c)^{k-1} \ge 1 + 2\gamma_k$, then we obtain a contradiction as before because $ \overline{d}_H(y) \le(\frac12 + \alpha)\binom{n-1}{k-1}$. Hence, $(1- 2c)^{k-1} < 2\gamma_k$ and consequently $c> (1 - (2\gamma_k)^{1/(k-1)})/2$, as desired.
\end{proof}
By Claim~\ref{clm:n2}, there exists a partition $ X', Y'$ of $V$ such that $|X'|= \lceil n/2 \rceil$, $|Y' | = \lfloor n/2 \rfloor$ and $| X\cap X'|, |Y\cap Y'| \ge (1-\eta) n/2$. We claim that $H$ is $\varepsilon$-close to $\mathcal B _{n,k}(Y', X')$ in Case~1 and $\varepsilon$-close to $\overline{\mathcal B} _{n,k}(Y', X')$ in Case~2.
Indeed, set $\mathcal{B} := \mathcal{B}_{\tilde{n}, k}(Y, X)$, where $\tilde{n} := |X\cup Y|$, and $H' := H[X\cup Y]$.
By definition, $E(\mathcal{B})$ consists of all $X^{k-i} Y^{i}$-sets for all odd $0\le i\le k$. If Claim~\ref{clm:dXY} (i) holds, then $|E(\mathcal{B})\cap E(H')| \ge (1- \gamma_k) |E(\mathcal{B})| $ and $|E( \overline{\mathcal{B}} )\cap E(\overline{H'})| \ge (1- \gamma_k) |E( \overline{\mathcal{B}} )| $. Thus
\[
| E(\mathcal{B}) \triangle E(H') | = | E(\mathcal{B})\setminus E(H')| + | E( \overline{\mathcal{B}} )\setminus E(\overline{H'})| \le \gamma_k |E(\mathcal{B})| + \gamma_k |E( \overline{\mathcal{B}} )|\le \gamma_k \binom{n}{k}.
\]
Let $V' := (X\cap X')\cup (Y\cap Y')$. Then $|V'|\ge (1- \eta)n$ and
\[
\left| \big(E(\mathcal B_{n,k}(Y', X')) \triangle E(H) \big) \setminus \binom{V'}{k} \right| \le \eta n \binom{n-1}{k-1}.
\]
Since $\gamma \ll \varepsilon ,1/k$, we have that $\gamma _k , \eta \ll \varepsilon$.
Therefore,
\[
| E(\mathcal B_{n,k}(Y', X')) \triangle E(H)| \le {\eta} n \binom{n-1}{k-1} + | E(\mathcal{B}) \triangle E(H') | \le {\eta} n \binom{n-1}{k-1} + \gamma_k \binom{n}{k}\le \varepsilon \binom{n}{k}.
\]
which implies that $H$ is $\varepsilon$-close to $\mathcal B_{n,k}(Y', X')$.
Analogously we can show that $H$ is $\varepsilon$-close to $\overline{\mathcal B}_{n,k}(Y', X')$ in Case~2.
This completes the proof of Lemma~\ref{lem:main}.
\noproof\bigskip
\section{An application of Lemma~\ref{absorbing}}
The following simple application of Lemma~\ref{absorbing} implies that the minimum $\ell$-degree condition that forces a perfect fractional matching also forces a perfect matching in a $k$-uniform hypergraph $H$, if we additionally assume that $H$ has a small number of vertices of large degree.
\begin{thm}\label{conc}
Given any $0<\varepsilon \le \delta'$ and $k, \ell \in \mathbb N$ where $\ell<k$, there is an $n_0 \in \mathbb N$ such that the following holds. Let $H$ be a $k$-uniform hypergraph on $n\geq n_0$ vertices where $k$ divides $n$. Suppose that $\delta _{1} (H)\geq \delta' \binom{n-1}{k-1}$ and $\delta_{\ell}(H) \ge (c^*_{k, \ell} + \varepsilon) \binom{n-\ell}{k-\ell}$. If there are at least $\varepsilon n$ vertices $x \in V(H)$ so that $d_H (x) \geq (1-\delta '+\varepsilon)\binom{n-1}{k-1}$ then $H$ contains a perfect matching.
\end{thm}
\noindent
\textbf{Sketch proof.}
It is easy to see that $H$ satisfies Lemma~\ref{absorbing} (ii) (where we choose $0<\gamma \ll \varepsilon$) and so by Lemma~\ref{lo}, $H$ contains a small absorbing matching $M$. Let $H':=H\setminus V(M)$. Then
$\delta _{\ell}(H') \geq (c^*_{k, \ell} + \varepsilon/2) \binom{n-\ell}{k-\ell}$ and so by Lemma~\ref{lem:frac}, $H'$ contains a matching covering all but a very small set of vertices. After absorbing the uncovered vertices
by $M$, we obtain a perfect matching in $H$.
\noproof\bigskip
\section*{Acknowledgements}
This research was partially carried out whilst the authors were visiting the Institute for Mathematics and its Applications at the University of Minnesota.
The authors would like to thank the institute for the nice working environment. The authors are also grateful
to the referees for their careful reviews.
|
\section{Introduction}
\subsection{New Breakthrough: Full Duplex Radios}
How much does it cost to purchase a wireless spectrum? In a wireless spectrum auction in January of 2015, the Federal Communications Commission (FCC) raised, for a 65 MHz bandwidth, a record-breaking $\$$44.9 billion. This illustrates how valuable the wireless spectrum has become; it offers more high-speed connectivity and satisfies more user demand for data within a limited wireless spectrum. The FCC is considering releasing more spectrum for wireless broadband usage. For the endless surge in wireless data traffic, however, this cannot be the ultimate solution.
Mobile devices with advanced wireless network capabilities, such as smartphones and tablets, are becoming ubiquitous, and keeping pace with their growth is the ever-increasing demand for bandwidth. Global mobile data traffic will increase nearly tenfold between 2014 and 2019. In that time, mobile data traffic is expected to grow at a compounded annual growth rate of 57~\%, reaching 24.3 exabytes per month by 2019 \cite{cisco2015}. These trends could create a {\it spectrum crunch} as the frequencies used to carry this traffic become exhausted.
Although the laws of physics prohibit the production of more spectrum, there is a lot of potential for {\it aggressive expansion} in scarce resource, that is, boosting spectral efficiency using novel technologies. A candidate, for creating a new breakthrough to alleviate the {\it spectrum crunch}, is full duplex. It theoretically doubles spectral efficiency, making it worth billions of dollars. Full duplex thus holds the tremendous potential to carry out the solutions needed in the future evolution of wireless systems.
\subsection{Key Challenge: Self-interference} \label{Key Challenge}
Since Guglielmo Marconi developed the wireless telegraph in 1895, the bane of wireless networks has been self-interference. It is the presence of self-interference that represents the key challenge to implementing full duplex wireless systems. Self-interference is the phenomenon where, through the coupling of transceivers in a wireless network, a signal is transmitted from a transmitter to its own receiver while that receiver is attempting to receive a signal sent by the other device. It compels the fundamental assumption that a wireless network has to be operated in half-duplex mode on the same channel. For example, Long Term Evolution (LTE) frequency-division duplex (FDD) today is operated so that the downlink and uplink transmission take place in two different frequency bands. In other words, the existence of self-interference cuts in half the amount of resources available, such as time and frequency, for wireless communications. For this reason, it is essential to manage self-interference to achieve the highest throughput performance with limited radio resources.
\subsection{The Beginning of Aggressive Expansion: SDR Platform-based Prototyping} \label{Aggressive Expansion}
\begin{figure*}
\centering
\includegraphics[width=6in]{Fig1.jpg}
\caption{(a) Real-time full duplex radio experiment setup in our laboratory/the exhibition hall at IEEE Globecom in Austin, TX, USA, December 2014 \;(b) the categorized comparison of implementation characteristics by each research group. In the SDR platform, VSG and VSA denote vector signal generator and analyzer, respectively.} \label{fig:Exp_Setup}
\end{figure*}
Up to this point, researchers have mostly depended on software simulations to test their theories that exploit simplified channel models (e.g., additive white Gaussian noise (AWGN), Rayleigh fading, etc.). In real-world wireless systems, however, impairments occur that are often overlooked in simulations, such as amplifier nonlinearity, gain/phase offset, I/Q imbalance, quantization effects, and timing jitter. Such impairments make prototyping imperative if the feasibility and commercial viability of any new wireless standard or technology are to be validated.
For next generation wireless research, a viable prototyping option has emerged known as software-defined radio (SDR)~\cite{sdr_Mitola}. SDR enables researchers to rapidly prototype a system. Researchers at Stanford~\cite{Bharadia2013}~\cite{hong2014}, Rice~\cite{Duarte2012}\cite{Duarte2014}, and Princeton~\cite{aryafar2012} have implemented various testbeds to build in-band full-duplex radios using combined radio frequency (RF) antennas and SDR platform~\cite{Hong_FD_2015}. As shown in Fig.~\ref{fig:Exp_Setup}(a), a real-time full duplex LTE system was also demonstrated at IEEE Globecom in Austin, TX, USA in December 2014.\footnote{The demo video is available at http://www.cbchae.org/.}
The categorized comparison of implementation characteristics by each research group is summarized in Fig. \ref{fig:Exp_Setup}(b). The two sections that follow elaborate on how to solve the key challenge and implement real-time full duplex radios.
\section{Prototype Settings: System Specifications $\&$ Hardware Architecture} \label{Settings}
The demonstrated full duplex prototype~\cite{globecom2014} is based on the LTE downlink standard \cite{sesia2009lte} with the following system specifications: a transmission bandwidth of 20 MHz, 30.72 MHz sampling rate, 15 kHz subcarrier spacing, 2048 fast Fourier transform (FFT) size, and variable 4/16/64 quadrature amplitude modulation (QAM). The prototype is implemented, as shown in Fig. \ref{fig:Exp_Setup}(a), using LabVIEW system design software and state-of-the-art PXIe SDR platform, where two full duplex nodes consist of the following four main components.
$\bullet$ {\bf Dual-Polarization Full Duplex RF Antenna} Dual-polarization slot antenna with high cross-polarization discrimination (XPD) in all directions \cite{oh2014}.
$\bullet$ {\bf PXIe-8133} Real-time (RT) controller equipped with a 1.73 GHz quad-core Intel Core i7-820 processor and 8 GB of dual-channel 1333 MHz DDR3 random access memory (RAM)~\cite{ni8133}.
$\bullet$ {\bf NI 5791R} 100 MHz bandwidth baseband transceiver module equipped with dual 130 MS/s analog-to-digital converter (ADC) with 14-bit accuracy, and dual 130 MS/s digital-to-analog converter (DAC) with 16-bit accuracy\cite{ni5791}.
$\bullet$ {\bf PXIe-7965R} Field-programmable gate array (FPGA) module equipped with a Virtex-5 SX95T FPGA optimized for digital signal processing, 512 MB of onboard RAM, and 16 direct memory access (DMA) channels for high-speed data streaming at more than 800 MB/s\cite{ni7965}.
\begin{figure*}
\centering
\includegraphics[width = 5.5in]{Fig2.jpg}
\caption{Block diagram of the proposed full duplex radio architecture. }
\label{fig:BD_FD}
\end{figure*}
In addition, all these modules, except for the analog cancellation part including a dual-polarization full duplex RF antenna, sit in the NI PXIe-1075 chassis. The chassis plays a role in data aggregation with both FPGA processors and a RT controller for real-time signal processing. As explained above, for transmitting and receiving simultaneously, the NI 5791R transceiver includes both transmit (Tx) and receive (Rx) ports connected with DAC and ADC, respectively.
As can be seen in Fig. \ref{fig:Exp_Setup}(a), we constructed a link for full duplex radios in an exhibition hall (a severe channel environment), where a great crowd of people was present, as well as in an indoor open space environment. The distance between full duplex communicating nodes was about 1.2~m. Note that in fact much longer ranges are possible. In this demo/experiment, one transceiver is connected with a dual-polarization full duplex RF antenna (in the white box in Fig.~\ref{fig:Exp_Setup}(a)), and the other is connected with an omni-antenna for simplicity. In other words, the transceiver connected with the omni-antenna only transmits an uplink signal. We then observe results at the transceiver equipped with a full duplex RF antenna, where both the Tx and Rx ports are connected.
\section{Proposed Full Duplex System}
In this section, we elaborate, in processing order, on our design blocks for the real-time full duplex LTE system, from transmission to reception and self-interference cancellation.
The block diagram of our full duplex radio architecture is illustrated in Fig. \ref{fig:BD_FD}.
\begin{figure*}
\centering
\includegraphics[width = 7in]{Fig3.jpg}
\caption{(a) Frame structure of our prototype (b) arrangement of primary synchronization symbol in frequency domain (c) reference symbol patterns for a full duplex link (d) measurement of synchronization and channel estimation in front panel of our prototype.}
\label{fig:Transmission}
\end{figure*}
\subsection{Transmission}
As illustrated in Fig. \ref{fig:Transmission}(a), we follow the frame structure of the LTE downlink with a frame duration of 10~ms for transmission. Each frame is divided into 20 slots, each being 0.5 ms in duration. Each slot contains 6 orthogonal frequency division multiplexing (OFDM) symbols with 512 cyclic prefix (CP) length (extended mode).
The data bit is generated on the PXIe-8133 RT controller. After the modulation block, the data symbols are interleaved with reference symbols (RSs) stored in a look-up table. An array of interleaved symbols is padded with zeros to form an array of 2048 samples. The 2048 samples are passed through a 2048-point inverse FFT (IFFT) block transforming the frequency domain samples into the time domain. The 2048 IFFT with 512 CP insertion block is executed on the PXIe-7965R FPGA module. To operate the discrete Fourier transform (DFT), it uses Xilinx {\it fast Fourier transform} intellectual property (IP) core.
\subsection{Analog Self-interference Cancellation}
Conventional approaches to deal with self-interference as passive analog cancellation are 1) isolation between Tx and Rx signals~\cite{Bharadia2013}~\cite{hong2014}, 2) antenna separation between the Tx and Rx antennas~\cite{Duarte2012}\cite{Duarte2014}, or 3) signal inversion with a $\pi$-phase shifter~\cite{aryafar2012}. Although these strategies have been extensively studied and adapted to full duplex radios as a good solution, We focus on a simpler, more compact strategy that provides outstanding self-interference cancellation performance. Furthermore, we discuss a solution that provides more robustness to environmental effects, such as the Doppler effect and multi-path, to achieve stable analog cancellation performance.
For analog self-interference cancellation, we introduce a novel RF antenna. Our approach is based on a dual-polarization antenna with a high XPD characteristic. XPD is defined as the ratio of the co-polarized average received power to the cross-polarized average received power. It represents, in other words, the ability to maintain radiated or received polarization purity between horizontally and vertically polarized signals. As shown in Fig. \ref{fig:Exp_Setup}(a), the proposed RF unit is a compact antenna with two poles. One pole is used as a radiated Tx output; the other is used as a received Rx input in a full duplex radio. XPD is an important characteristic, particularly in full duplex systems, where cross-talk between Tx and Rx ports can curb the system's throughput performance. Since XPD has a relationship to inter-port isolation, the dual-polarization antenna with high XPD is, in full duplex systems, an excellent solution. We find that the dual-polarization antenna itself achieves 42 dB of isolation. Active analog cancellation provides an additional cancellation gain up to 18 dB by tuning the attenuation, phase shift, and delay parameters, i.e., totally 60 dB by analog cancellation.
\subsection{Digital Self-interference Cancellation}
The goal of digital self-interference cancellation is to suppress, after canceling self-interference in the analog domain, any residual self-interference. Digital self-interference cancellation consists of rebuilding self-interference and subtracting it from the received signal. A key parameter to consider in the real-time digital self-interference canceler is the guaranteed throughput of digital data, in a given time, between transmitting and receiving streams in a node. Unlike unidirectional communications that uses a radio only for transmission or reception, a digital self-interference canceler of full duplex node requires, without a bottleneck, high speed computation/data throughput between transmit and receive elements. To perform rebuilding self-interference and subtracting it from the received signal in the real-time, we use handshake protocols, shift registers, shared registers with scheduled access, and dedicated first-in first-out (FIFO) buffers.
In \cite{choi2013}, the authors implemented a full duplex solution, which needs no additional synchronization or channel estimation for self-interference in the digital domain. Our prototype, however, focuses on an independent system that operates in real-time to maximize cancellation performance in the analog/digital domain, respectively. Further, additional synchronization and estimation is exploited to reduce the complexity in our digital self-interference canceler.
At the moment of decoding the desired symbol, it is critical to know the perfect timing between self-interference and the received symbol in full duplex mode. Thus, key issues include designing synchronization and channel estimation strategies for residual self-interference as well as for a desired link. We produce a process for implementing a digital self-interference canceler from synchronization and channel estimation. In order to operate a real-time digital self-interference canceler with high performance, we focus on FPGA implementation using LabVIEW system design software and PXIe SDR platform.
\begin{figure*}
\centering
\includegraphics[width = 6.3in]{Fig4.jpg}
\caption{A part of source code for a digital self-interference canceller that operates in real-time on the software-defined radio platform. }
\label{fig:LV}
\end{figure*}
$\bullet$ {\bf Synchronization:} Synchronization is one of the key blocks in real-time full duplex radios. Under synchronization for full duplex, there are two operations: synchronization for decoding the desired symbol and rebuilding self-interference. In the synchronization block for decoding the desired symbol, we estimate time offset by random propagation delays and sampling clock offsets between two full duplex radios. In the synchronization block for rebuilding self-interference, we estimate the time offset between Tx port and Rx port of a full duplex radio.
To facilitate timing synchronization, the LTE downlink standard specifies a primary synchronization signal (PSS). Accordingly, the receiver can successfully perform timing synchronization in half-duplex mode. Note, however, that we need to keep performing synchronization for the self-interference signal as well as for the desired signal. Thus, we use a property, where the ZC sequence with a different root index is orthogonal to each other. The PSS is modulated by a ZC sequence given as, $P\left[ k \right] = {e^{ - j\frac{\pi }{N}uk\left( {k + 1} \right)}}, - 31 \le k \le - 1$, and $P\left[ k \right] = {e^{ - j\frac{\pi }{N}u\left( {k + 1} \right)\left( {k + 2} \right)}}, 1 \le k \le 31$, where $k$ is the subcarrier index, $u$ is the root index, and $N$ it the sequence length ($N$ = 63). We use a different root index relatively prime to $N$ for the PSS of each full duplex radio, i.e, $u_1$ = 25, $u_2$ = 29. These symbols are located on the 62 subcarriers, symmetrically arranged around the DC-carrier in the last OFDM symbol of the first and eleventh slots of each frame as shown in Fig. \ref{fig:Transmission}(a) and (b). As the duration of a frame is 10 ms, the PSS is therefore transmitted after every 5 ms time intervals or once per half-frame.
To calculate the correlation between the ideal sequence and the estimated PSS signal, it is necessary to extract the PSS subcarrier from the received signal. For this reason, we design a low-pass filter (LPF) using Xilinx's finite impulse response (FIR) IP core, as shown in Fig. \ref{fig:Transmission}(b). The designed LPF has a cut-off frequency of 1.4 MHz, a stop-band attenuation of 50~dB, and a pass-band ripple of 0.1 dB. After the received signal samples are passed through the LPF, each synchronization block for decoding the desired symbol and rebuilding the self-interference is executed to calculate, independently, the correlation with its own PSS. As a result, a maximum peak is detected at the sample index of the first sample of the OFDM symbol following the PSS symbol, as illustrated in Fig. \ref{fig:Transmission}(d). A starting index of the desired signal is delivered into FFT block, and a starting index of self-interference signal is delivered into the digital cancellation block.
\begin{figure*}
\centering
\includegraphics[width = 6.3in]{Fig5.jpg}
\caption{Measurement results of (a) analog self-interference cancellation and (b) digital self-interference cancellation measurement.}
\label{fig:AD_SIC}
\end{figure*}
$\bullet$ {\bf Channel Estimation:} Channel estimation also has, for full duplex, two operations: estimations for 1) the channels between two full duplex nodes and 2) the channel between the Tx and Rx ports in its own full duplex node. The former operates to decode the desired symbol after digital self-interference cancellation, while the latter operates to rebuild the self-interference using the known Tx data. In order to handle the two operations simultaneously, we design RS patterns that are orthogonal between two full duplex nodes, exploiting the pattern of cell-specific reference signals for multiple antenna port. The RS patterns are shown in Fig.~\ref{fig:Transmission}(c).
Both channel estimation blocks have two steps in common: 1) RS extraction and 2) interpolation in order. After the received samples are passed through the FFT block, the RS subcarriers of each channel are extracted from an OFDM symbol in a data and reference demultiplexing block. A channel coefficient of each RS subcarrier is calculated using original RSs stored in block memory. To estimate the channel coefficients of RS subcarriers, a least-square method is exploited. After passing through the data and reference demultiplexing block, the channel estimates of the RS subcarriers are split into two groups. One is for the channel estimation between nodes; the other is for the channel estimation between antenna ports. In each channel estimation block, we implemented a linear interpolator using Xilinx's FIR IP core. The linear interpolator in each block estimates the channel coefficients of data subcarriers as well as RS subcarriers.
In Fig. \ref{fig:Transmission}(d), the bottom left and right figures are screen shots of the instantaneous channel estimation result between nodes and ports, respectively, in frequency domain.
$\bullet$ {\bf Digital Cancellation:} In most research on full duplex implementation, digital self-interference cancellation is performed in the time domain. To carry out digital cancellation in the time domain, an additional IFFT block is needed for rebuilding self-interference. As mentioned above, in the real-time digital self-interference canceler, it is critical to operate the process for rebuilding self-interference and canceling it out without a bottleneck. For lower computational complexity and faster rebuilding self-interference, we execute it in the frequency domain. One might argue that if the SI is not overlapped coherently with the received signal, it would be difficult to cancel the SI in the frequency domain. If the difference, however, in the received time of the SI and the desired signal is less than the CP length, which is common in practice, there is no problem in canceling the SI after FFT. We will also investigate this issue in our future work.
Digital cancellation utilizes the baseband samples of the transmitted signal to rebuild self-interference in the digital domain and subtracts them from the received samples. Note that we know the baseband samples of the transmitted signal from its own node. Self-interference can be rebuilt in the digital domain using the baseband samples of the transmitted signal and the channel estimates between the ports of its own node. As mentioned above, we should know which self-interference (a sample index) is mixed in the received sample at the moment of decoding the desired symbol. Accordingly, we include a counter in the digital cancellation block. As soon as the starting index of the self-interference signal arrives in the digital cancellation block from the synchronization block to rebuild self-interference, the counter operates to choose a rebuilt digital sample for subtraction processing. After digital cancellation, a zero-forcing channel equalizer operates to decode the desired symbols. Illustrated in Fig. \ref{fig:LV} is a part of the source code for the digital self-interference canceler operated in the frequency domain.
\section{Prototype Test Results}
Using the real-time full duplex LTE prototype as described in the previous section, we measure the level of analog and digital self-interference and calculate the bit error rate (BER) and throughput performance. In this prototype, the carrier frequency is the 2.52 GHz in LTE bands. As shown in Fig.~\ref{fig:AD_SIC}(a), we find that the dual-polarization antenna provides about 42 dB of isolation from our experiments, i.e., the self-interference that is leaking to the Rx port is reduced by about 42 dB. Moreover, by tuning the attenuation, phase shift, and delay parameters, we achieve 60 dB of analog self-interference cancellation with analog active cancellation. In the digital domain, we calculate error vector magnitude (EVM) for self-interference to measure the average level of digital self-interference cancellation. As a result, we achieve 43 dB of self-interference cancellation in the digital domain as shown in Fig. \ref{fig:AD_SIC}(b).
In order to compare the throughput improvement, we also implemented the LTE-FDD prototype. Our design is based upon the LTE downlink standard with system specifications that include a transmission bandwidth of 10 MHz, 15.36 MHz sampling rate, 15 kHz subcarrier spacing, 1024 FFT size, 256 CP length, and variable 4/16/64 QAM.
Fig.~\ref{fig:Cons_TP}(a) shows the constellation, taken during an over-the-air test of the full duplex communications link. One full duplex radio transmits a 4~QAM modulated signal as the downlink, and receives a 64 QAM modulated signal as the uplink. As a result, the goal of this full duplex radio is to decode 64 QAM, the desired symbol after perfectly canceling out the 4~QAM symbol as self-interference. In Fig. \ref{fig:Cons_TP}(a), the left constellation shows that, with only analog cancellation, self-interference is not perfectly cancelled out, while the right constellation shows that with both analog and digital cancellation self-interference is perfectly cancelled out.
\begin{figure}
\centering
\includegraphics[width = 3.5in]{Fig6.jpg}
\caption{(a) Constellation result with only analog passive self-interference cancellation (left) constellation result with analog and digital self-interference cancellation (right) (b) throughput performances for each constellation.}
\label{fig:Cons_TP}
\end{figure}
As shown in Fig. \ref{fig:Cons_TP}(b), our prototype provides high throughput performance in real-time. It delivers a throughput increase of 1.9x on the 4, 16~QAM and 1.89x on the 64~QAM compared to the conventional half duplex mode.\footnote{Half duplex means out-of-band full duplex, i.e., FDD throughout this paper.}
\section{Research Challenges}
Notwithstanding our focus on designing more practical full duplex radios, several research challenges remain before the most viable solution for next generation communication systems is achieved.
\subsection{Hardware Impairments}
The performance of a full duplex system depends heavily on hardware impairments: amplifier non-linearity, gain/phase offset, I/Q imbalance, quantization effects, and timing jitter. For example, a nonlinearly amplified OFDM signal occurs intermodulation distortion (IMD), which the amplitude modulation of signals containing two or more different frequencies in a system. The IMD raises the noise floor and causes inter-carrier interference, which induce performance degradation of the full duplex system. Because most analog components in the system have the non-linearity property, the cancellation of all non-linear components as well as linear components is a significant burden on a real-time system. Thus, some pre-processing strategies such as pre-distortion for reducing hardware impairments represent an interesting research topic.
\subsection{Joint PHY/MAC Prototyping}
Most implementations of full duplex radios have mainly focused on the physical layer design, which enables bi-directional communications between a single pair link. There exists apparent limitations in translating the performance gains obtained from the demonstration of a single pair link into network performance.
Transmissions on full duplex mode create potential interference outside the full duplex link. This calls for the prototyping of media access control (MAC) layer protocols, including discovering and exploiting full duplex opportunities in a distributed manner. Another interesting area for future work is the joint PHY/MAC approach for prototyping.
\subsection{Full Duplex System with OFDM and SC-FDMA}
Since single carrier frequency division multiple access (SC-FDMA) has a peak-to-average power
ratio (PAPR) lower than that of OFDMA, it is used for the uplink multiple access scheme in the LTE of cellular systems. Most implementation studies of full duplex, however, deal with only OFDM frame structures. There are many potential challenges in asymmetric uplink/downlink frame structures in LTE.
\subsection{Comparison with LTE-TDD}
LTE-TDD (time division duplexing), generally, has been known to have many benefits, such as low latency, spectrum flexibility, uplink/downlink flexibility, and lower cost per bit. To discuss the various performance characteristics such as latency, throughput, power consumption, and flexibility between TDD and full duplex, we believe that it is worth comparing full duplex prototype with a comparable LTE-TDD prototype.
\subsection{Novel Solution for RF/Analog Cancellation}
RF/Analog cancellation plays a critical part in attenuating high-powered self-interference sufficiently such that Rx saturation and dynamic range are not an issue when operating digital cancelation. We showed the analog cancellation based on the dual polarization would be a good option. It does have, however, two main weaknesses. First, it struggles to perform active analog cancellation in real-time. While high XPD makes the passive analog cancellation level high, to estimate the coefficients for active cancellation is difficult. Second, channel reciprocity that can simplify the link overhead may not be assumed for the polarization antennas. The novel solutions for these issues will be an interesting research topic.
\section{Conclusion}
Full duplex radio technologies could be a major contributor to increasing spectrum efficiency in areas of explosive traffic demand where there are limited radio resources. To validate the feasibility and commercial viability of any new wireless standard or technology like full duplex radio, SDR-based prototyping is imperative. We prototyped a design that combines dual-polarization full duplex RF and the digital self-interference canceler that operates in real-time on the SDR platform. We focused on a more practical prototype that exhibited outstanding self-interference cancellation performance. The main portion of this article is dedicated to presenting the design, implementation, and evaluation of a real-time full duplex LTE system, a candidate for next generation wireless communication systems. We expect our prototype design to provide worthwhile insights into developing the most viable solution for future wireless communication systems.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\renewcommand{\baselinestretch}{1.0}
\bibliographystyle{IEEEtran}
|
\section{Introduction and main result}
\subsection{General setup}
Let $L_0$ be a self-adjoint operator on a Hilbert space $\mathcal H$ and consider
\begin{equation}
\label{2}
L=L_0+\d I,
\end{equation}
where $\d\in\mathbb R$ is a perturbation parameter and $I$ is a self-adjoint operator. It is assumed that $L$ is self-adjoint. We suppose that the spectrum of $L_0$ is absolutely continuous (possibly, but not necessarily semi-bounded) and that $L_0$ has finitely many eigenvalues $e$ with finite multiplicities $m_e$. All the eigenvalues of $L_0$ are embedded in the continuous spectrum. The general problem we consider is how the stability or partial stability or instability of these embedded eigenvalues under the perturbation affect the dynamics generated by $L$. One may readily incorporate into our results and proofs the case where $L_0$ has also isolated eigenvalues by using ordinary analytic perturbation theory on the corresponding subspaces.
In the setting of usual analytic perturbation theory \cite{Kato} an isolated eigenvalue $e$ of $L_0$ with multiplicity $m_e$ splits, under perturbation, into a group of eigenvalues $E_{e,1},\ldots,E_{e,\ell_e}$ of $L$ ($1\le \ell_e\le m_e$), in the sense that $E_{e,j}=E_{e,j}(\d)\rightarrow e$ as $\d\rightarrow 0$, for $j=1,\ldots,\ell_e$. For fixed $e$, the sum of the multiplicities of the eigenvalues $E_{e,j}$ equals $m_e$. On the other hand, it is well known that {\em embedded} eigenvalues can be unstable, or partially stable, under perturbation. Instability means that $L$ does not have any eigenvalues in a neighbourhood of $e$ for small $\d$. Partial stability means that the embedded eigenvalue $e$ of $L_0$ splits into a group of eigenvalues of $L$ whose sum of multiplicities are strictly smaller than that of $e$.
In this paper, we consider the situation where $L_0$ has unstable and partially stable eigenvalues, and where the partially stable ones undergo a reduction to dimension one under perturbation. Namely, close to any eigenvalue $e$ of $L_0$, the operator $L$ either does not have any eigenvalue ($e$ unstable) or $L$ has exactly one simple eigenvalue $E_e$ close to $e$, meaning that $\lim_{\d\rightarrow 0}E_e=e$. It is supposed that all eigenvalues of $L$ are of this form. One may develop the arguments of the present paper in the more general setting where close to every $e$, $L$ has several eigenvalues $E_1,\ldots,E_\ell$ and each of them may be degenerate. We do not do this here to keep the exposition simpler.
The dependence of $E_e$ on $\d$ is not governed by usual analytic perturbation theory, since the unperturbed $e$ is an embedded eigenvalue of $L_0$. However, suitably modified expressions from analytic perturbation theory of isolated eigenvalues will still play a role in the present setting. Let $P_e$ be the spectral projection of $L_0$ associated to the eigenvalue $e$. If $e$ was an isolated eigenvalue of $L_0$, then the first and the second order corrections (in $\d$) of eigenvalues would be given, according to analytic perturbation theory \cite{Kato}, by the eigenvalues of $P_e IP_e$ and of $P_e IP_e^\perp(L_0-e)^{-1}IP_e$, respectively.
We assume that
\begin{itemize}
\item[{\bf (A1)}] For all eigenvalues $e$ of $L_0$,
\begin{equation}
\label{n1}
P_e I P_e =0.
\end{equation}
\end{itemize}
{}For embedded $e$, the resolvent $P^\perp_e(L_0-e)^{-1}$ does not exist as a bounded operator, so $P_e IP_e^\perp(L_0-e)^{-1}IP_e$ is not defined, typically. Nevertheless, we can replace $e$ by $e-\i\epsilon$ and consider $\epsilon$ small. This suggests that the second order eigenvalue corrections to $e$ are linked to the {\em level shift operator}
\begin{equation}
\label{n2}
\Lambda_e = -P_e IP_e^\perp (L_0-e+\i0_+)^{-1} IP_e,
\end{equation}
where $\i0_+$ indicates the limit
of the resolvent $(L_0-e+\i \epsilon)^{-1}$,
as $\epsilon\rightarrow 0_+$. The existence of the limit is guaranteed by assumption (A2)
below (take $\d=0$ in the resolvent in \eqref{nlap}). The operator $\Lambda_e$ is represented by an $m_e\times m_e$ matrix.
Let $Q$ be an orthogonal projection and denote
\begin{equation}
\label{n4}
R_z=(L-z)^{-1}\quad \mbox{and}\quad R_z^{Q}= (Q^\perp L Q^\perp-z)^{-1}\upharpoonright_{{\rm Ran} Q^\perp} .
\end{equation}
In the following, we denote by $C(\phi,\psi)$ a constant which is independent of $z$ and $\Delta$ but which may depend on $\phi, \psi\in \mathcal H$.
\begin{itemize}
\item[{\bf (A2)}] (Limiting Absorption Principle.)
There is a dense set ${\mathcal D}\subset{\mathcal H}$ with ${\rm Ran}\, IP_e\subset\mathcal D$ ($\forall e$) and there is an $\alpha>0$ such that the following hold.\\
{\bf (1)} Let $S_e=\{ z\in{\mathbb C}_-\ :\ |{\rm Re}z-e|\le\alpha\}$.
Here, ${\mathbb C}_-$ denotes the (open) lower complex half plane. For all $e$ and all $\phi,\psi\in{\mathcal D}$, we have
\begin{equation}
\label{nlap}
\sup_{z\in S_e}| \tfrac{d^k}{dz^k} \scalprod{\phi}{R_z^{P_e}\psi}| \le C(\phi,\psi)<\infty,\ \quad k=0,\ldots 3.
\end{equation}
{\bf (2)} Let $S_\infty=\{z\in{\mathbb C}_-\ :\ |{\rm Re}z-e|>\alpha \mbox{ for all $e$}\}$. For all $\phi,\psi\in{\mathcal D}$, we have
\begin{equation}
\label{nlap1}
\sup_{z\in S_\infty}| \tfrac{d^k}{dz^k} \scalprod{\phi}{ R_z\psi}| \le C(\phi,\psi)<\infty,\ \quad k=0,1.
\end{equation}
\end{itemize}
Assumption (A2)(1)
implies that $P_e I R_z^{P_e} IP_e$ extends as a twice continuously differentiable function
to real $z$ with $|z-e|\le\alpha$. Moreover, the estimates \eqref{nlap} and \eqref{nlap1} with $k=0$ imply that $P^\perp_e LP^\perp_e$ and $L$ have purely absolutely continuous spectrum in the interval $(e-\alpha,e+\alpha)$ and in the region $\{x\in{\mathbb R}\ :\ |x-e|>\alpha \mbox{ for all $e$}\}$, respectively (see e.g. \cite{KoMeSo} or \cite{CFKS}, Proposition 4.1).
\begin{itemize}
\item[({\bf A3})] The operators $\tfrac{d}{d \Delta} P_e I R_z^{P_e} IP_e$ is bounded,
uniformly for $z\in{\mathbb C}_-$ with $|{\rm Re}z-e|\le \alpha$ and for all $|\Delta|\le \Delta_0$.
\end{itemize}
The Feshbach map associated to an orthogonal projection $Q$, applied to $L-z$, is defined by
\begin{equation}
\label{n3}
{\mathfrak F}(L-z;Q)= Q( L-z- L R_z^Q L)Q.
\end{equation}
It follows from assumptions (A1) and (A2)
that
$$
\lim_{\epsilon\rightarrow 0_+} {\mathfrak F}(L-E_e+\i\epsilon;P_e) \equiv {\mathfrak F}(L-E_e;P_e) = P_e(e-E_e-\d^2 IR_{E_e-\i0_+}^{P_e}I)P_e.
$$
Let $\psi_{E_e}=\psi_{E_e}(\d)$ be s.t. $L\psi_{E_e}=E_e\psi_{E_e}$. Note that $P_e\psi_{E_e}\neq 0$ for otherwise $P^\perp_eLP^\perp_e\psi_{E_e}=E_e P_e^\perp\psi_{E_e}$, which cannot hold for small $\d$, since $P^\perp_eLP^\perp_e$ has purely
absolutely continuous spectrum in a neighbourhood of $e$ due to Condition (A2)(1). We normalize $\psi_{E_e}$ as $\|P_e\psi_{E_e}\|=1$.
By the isospectrality property of the Feshbach map (see Appendix \ref{FeshbachAppendix}, Proposition \ref{Feshprop}), we have
\begin{equation}
\label{n5}
0={\mathfrak F}(L-E_e;P_e) P_e\psi_{E_e}=\big(\tfrac{e-E_e}{\d^2}-P_eIR_{E_e-\i 0_+}^{P_e}IP_e)P_e\psi_{E_e}.
\end{equation}
This, together with Conditions (A2)(1) and (A3), implies that $\xi_\d:= \tfrac{e-E_e}{\d^2}$ is bounded in $\d$ for small $\d$ and that $P_eIR_{E_e-\i 0_+}^{P_e}IP_e=\Lambda_e+O(|\d|+ |e-E_e|)=\Lambda_e+O(|\d|)$. On a suitable sequence $\d_n\rightarrow 0$, we have $\xi_{\d_n}\rightarrow \xi_0$ and $P_e\psi_{E_e}(\d_n)\rightarrow P_e\psi_0$ for some $\xi_0\in\mathbb R$ and some unit vector $P_e\psi_0$ (Bolzano-Weierstrass). Consequently, taking $\d\rightarrow 0$ in \eqref{n5} along this sequence gives
\begin{equation}
\label{nnn1}
\Lambda_eP_e\psi_0 = \xi_0P_e\psi_0,
\end{equation}
showing that $\xi_0$ is a real eigenvalue of $\Lambda_e$. For ease of presentation, we assume the following.
\medskip
\begin{itemize}
\item[{\bf (A4)}] (Fermi Golden Rule Condition.) The eigenvalues of all the level shift operators $\Lambda_e$ are simple. Moreover,
{\bf (1)} If $e$ is an unstable eigenvalue, then all the eigenvalues $\lambda_{e,0},\ldots,\lambda_{e,m_e-1}$ of $\Lambda_e$ have strictly positive imaginary part.
{\bf (2)} If $e$ is a partially stable eigenvalue, then $\Lambda_e$ has a single real eigenvalue $\lambda_{e,0}$. All other eigenvalues $\lambda_{e,1},\ldots,\lambda_{e,m_e-1}$ have strictly positive imaginary part.
\end{itemize}
\medskip
Under condition (A4)(2), the set $\{\xi_\d=\tfrac{e-E_e}{\d^2}\}$ for $\d$ small has a unique limit point $\xi_0$ and we have $\lambda_{e,0}=\xi_0$.
Having only simple eigenvalues, $\Lambda_e$ is diagonalizable and has the spectral representation
\begin{equation}
\label{n7}
\Lambda_e =\sum_{j=0}^{m_e-1} \lambda_{e,j}P_{e,j},
\end{equation}
where $P_{e,j}$ are the (rank one) spectral projections. We introduce the notation
\begin{equation}
\label{defple}
a\prec b,
\end{equation}
where $a$ is a complex number, a vector or a bounded operator and $b>0$, to mean that $|a|\le {\rm const.} b$, where $|\cdot|$ is the appropriate norm and $\rm const.$ is a constant which does not depend on the coupling parameter $\d$, nor on time $t$.
\begin{thm}[Resonance expansion of propagator]
\label{nthm1}
There is a constant $c>0$ s.t. for $0<|\Delta|<c$ the following holds. Denote the projection onto the eigenvalue $E_e$ of $L$ by $\Pi_{E_e}$. Let $t>0$, $\phi,\psi\in\mathcal D$ s.t. $L\phi$, $L\psi\in{\mathcal D}$. Then
\begin{eqnarray}
\label{nmainreseqn}
\scalprod{\phi}{{\rm e}^{\i t L}\psi} &=& \sum_{e\ {\rm partially\, stable}}\Big\{ {\rm e}^{\i t E_e} \scalprod{\phi}{\Pi_{E_e}\psi} + \sum_{j=1}^{m_e-1} {\rm e}^{\i t (e+ \d^2 a_{e,j})}\scalprod{\phi}{\Pi'_{e,j}\psi} \Big\}\\
&& + \sum_{e\ {\rm unstable}} \sum_{j=0}^{m_e-1} {\rm e}^{\i t (e+ \d^2 a_{e,j})}\scalprod{\phi}{\Pi'_{e,j}\psi} +R(t),
\nonumber
\end{eqnarray}
where
\begin{equation}
\label{nmthm1}
R(t) \prec \frac{1}{t}.
\end{equation}
The exponents $a_{e,j}$ and the operators $\Pi'_{e,j}$ are close to the spectral data of the level shift operator $\Lambda_e$, \eqref{n7}. Namely,
\begin{equation}
\label{nmthm2}
a_{e,j} = \lambda_{e,j}+O(\d), \qquad \Pi'_{e,j} =P_{e,j}+O(\d).
\end{equation}
\end{thm}
\medskip
\noindent
{\bf Remarks and discussion.\ }
\smallskip
1. For an expansion of $\scalprod{\phi}{{\rm e}^{-\i tL}\psi}$ for $t>0$, simply take the adjoint of \eqref{nmainreseqn}.
2. The exponents $a_{e,j}$ are the eigenvalues of an explicit matrix. They can be calculated to all orders in $\d$ (see Lemma \ref{lemma1}). The operators $\Pi'_{e,j}$ have also expressions calculable to all orders in $\d$ (see \eqref{qtilde}). Those ``complex energies'' $a_{e,j}$ are called resonances. They coincide with the eigenvalues of the spectrally deformed generator of dynamics in situations where the latter exists.
3. The remainder term is small relative to the contributions of the exponentially decaying terms in \eqref{nmainreseqn} for times $t$ satisfying ${\rm e}^{-\gamma \d^2 t}>\!\!>1/t$, where
\begin{equation}
\label{decayrate}
\gamma=\min_{e,j}\big\{{\rm Im} a_{e,j} \big\}.
\end{equation}
The inequality ${\rm e}^{-\gamma \d^2 t}\ge C/t$, for some (large) $C$ is equivalent to $\frac{\ln(t)-\ln(C)}{t}\ge\gamma\d^2$. For small $\d$, it is valid for {\em intermediate times}, $t_0 < t < t_1$, with $t_0=C+O(\gamma\d^2)$ and $t_1\sim 1/(\gamma\d^2)$. During this time-interval, the decay of \eqref{nmainreseqn} behaves as exponential, to leading order.
4. The Fermi Golden Rule Condition (A4) guarantees that instability (and partial stability) of eigenvalues is visible at lowest order, $O(\Delta^2)$, in the perturbation. It may happen that resonances acquire non-vanishing imaginary parts only at higher orders in $\Delta$. Our method can be adapted to describe this situation. The dynamical consequence is a slower decay of the corresponding directions in Hilbert space, see also \cite{MeBeSo}.
5. Increasing the regularity assumptions on the vectors $\phi$, $\psi$ allows to show a faster polynomial decay of the remainder than \eqref{nmthm1} (this amounts to taking higher $z$ derivatives of the resolvent, c.f. \eqref{14}).
\subsection{History, relation to other work.}
The analysis of resonance phenomena has a long history and plays an important role in quantum physics \cite{BW,KP,S,WW}. Its modern description, involving dilation analytic Hamiltonians (\cite{AC,BC}), was given in \cite{Simon} and further developed in \cite{Simon2,Sig1,Hun2}. We refer to \cite{HiSig} for a textbook presentation and many more references. A time-dependent theory of quantum resonances was established in \cite{MS}, inspired by \cite{SW} and further developed in \cite{KRW}. In these works, as well as in \cite{CGH}, a variant of the Mourre theory in combination with the Feshbach projection method is used to link dynamical properties of quantum systems to spectral objects. The approach of the present work is, in spirit, similar to \cite{CGH}; see the end of this paragraph for a comparison. Nevertheless, all the above-mentioned works require regularity conditions within Mourre theory that do not allow the treatment of open quantum systems at positive temperature. In the context of open quantum systems at positive temperature, the link between quantum resonances and approach of an equilibrium state has been pioneered, using complex deformation theory, in \cite{JP,BFSrte}. The work \cite{BFSrte} is based on a sophisticated renormalization group method initiated in \cite{BFS1,BFS2,BFS3}. Recently, a method based on graph expansions of the propagator rather than purely spectral considerations was given in \cite{DeKu}. The spectral approach has been further developed to yield a detailed description of open systems dynamics in terms of resonances in \cite{MSB1,MSB2,MSB3}, with applications to quantum information theory \cite{M,MBBG} and quantum chemistry \cite{MBS}. The spectral analysis and its consequences for ``return to equilibrium'' based on Mourre theory and positive commutators was carried out in \cite{DJ,Mthesis,FM,DJ2}. However, these papers are limited to the study of the spectrum of the Liouville operator with the goal (typically) of showing that it has a single, simple eigenvalue at zero (and absolutely continuous spectrum otherwise). This information alone does not provide any detail about the dynamics other than ergodicity. However, one is interested in information such as directions of decay and decay rates which describe, for example the speed of thermalization, decoherence and the dynamics of entanglement. In the method of complex deformation, complex resonance energies are linked ``automatically'' to the decay rates of reduced density matrix elements \cite{MSB2}. The same expressions describing those decay rates appear as well in Mourre theory as a consequence of the Fermi Golden Rule (see also \cite{FaMoSk}), however, linking them to the dynamics, and in particular to time decay, is more delicate and has not been done previously for open systems. We show in the present paper how to extract the detailed dynamical information from the Mourre theory in a technical setup that includes positive temperature open quantum systems. It is important to be able to handle these questions using a softer approach than the spectral deformation one. Indeed, the applicability of the latter demands much more regularity from the models and, in some physically relevant situations, the spectral deformation technique is not applicable at all. This happens for the spin-boson model at arbitrary coupling, whose ergodicity has been shown recently in \cite{KoMeSo} using Mourre theory. As an application of our method, we give a detailed expansion of the propagator of this model in the present paper.
The philosophies of \cite{CGH} and the present paper are similar, in that the common main idea is to write the propagator as a contour integral over the resolvent and subsequently use the Feshbach map to analyze the latter. However, right from the start, the technical assumptions are very different. A core assumption of \cite{CGH} is that multicommutators of the Hamiltonian $H$ with the (Mourre theory) conjugate operator $A$ are relatively $H$-bounded. While this is typically true for, say, for Schr\"odinger operators, it is not so for open quantum systems. The problem comes from the fact that the number operator is not bounded relatively to the free field Hamiltonian (so already relation (2) of \cite{CGH} is not valid). The situation even is worse for the spin-boson model at arbitrary coupling, where each successive commutation of the Liouvillian with the conjugate operator produces a more singular operator, as explained in \cite{KoMeSo} (this is the reason why the spectral deformation theory fails).
The dissipative character of the system is guaranteed in \cite{CGH} by assuming a Mourre estimate, localized spectrally on a subspace around the embedded eigenvalue in question. Accordingly, the main result of \cite{CGH} describes the dynamics of an initial state (wave function) which is spectrally localized close to the embedded eigenvalue. This makes good physical sense in the context of, say, Schr\"odinger operators, where long lived initial states are expected to lie close to unperturbed bound states. However, in open systems problems, one considers initial states which are spatial perturbations of equilibrium states and which are not at all spectrally localized relative to the Liouville operator. Our dissipation assumption (A2) is thus a Limiting Absorption Principle which is {\em not} spectrally localized and which produces results for initial vectors which are not spectrally localized. Condition (A2) can be heuristically understood in the context of open systems as saying that the reservoir stays essentially in its equilibirum state during the dynamical process (the Born approximation).
We mention that the main result of \cite{CGH} is stated and proven for a simple unperturbed eigenvalue. In open systems however, the origin is {\em always} a degenerate eigenvalue of the unperturbed Liouville operator and so we have put in place a formalism that works for the degenerate case as well.
\subsection{Outline of the proof of Theorem \ref{nthm1}}
For $\psi\in\mathrm{dom}(L)$, we have (\cite{EngelNagel}, Corollary II 3.6)
\begin{equation}
\label{4}
{\rm e}^{\i tL} \psi= -\frac{1}{2\pi\i}\int_{{\mathbb R}-\i w}{\rm e}^{\i tz}R_z\psi \,dz,
\end{equation}
where $w>0$ is arbitrary and $R_z=(L-z)^{-1}$, see \eqref{n4}. We write ${\rm e}^{\i tz}=\frac{1}{\i t}\frac{d}{dz}{\rm e}^{\i tz}$ and integrate by parts in \eqref{4} to obtain
\begin{equation}
\label{14}
e^{\i tL} \psi= \frac{1}{\i t}\frac{1}{2\pi\i}\int_{{\mathbb R}-\i w}{\rm e}^{\i tz}\textstyle\frac{d}{dz}R_z\psi \,dz.
\end{equation}
The boundary terms in the integration by parts vanish since $R_z\psi\rightarrow 0$ as $|z|\rightarrow\infty$.
We analyze separately the contributions to the integral in \eqref{14} coming from $z$ in different regions on the line of integration. Define the gap of all the clusters of resonances by
\begin{equation}
\label{delta}
\delta = \min_{e,i,j}\big\{ |\lambda_{e,i}-\lambda_{e,j}|\ :\ i\neq j \big\}>0
\end{equation}
and denote the eigenvalue gap of $L_0$ by
\begin{equation}
\label{gapdef}
g = \min_{e,e'}\{ |e-e'|\ :\ e\neq e'\} > 0.
\end{equation}
Let
\begin{equation}
\label{13}
\alpha=\tfrac12\min\big\{c \delta,\, g, \, {\rm Im}\lambda_{e,j}\ :\ (e,j) \ {\rm s.t.\ } {\rm Im}\lambda_{e,j}>0 \big\} > 0.
\end{equation}
Here, $c$ is a constant not depending on $\d,w$ (its origin is explained in Lemma \ref{lemma1}).
For any eigenvalue $e$ of $L_0$, set
\begin{equation}
{\mathcal G}_e=\{x-\i w\ :\ |x-e|\le \alpha\}
\label{6}
\end{equation}
and set
\begin{equation}
{\mathcal G}_\infty= \big\{x-\i w\ :\ x\in{\mathbb R}\big\} \backslash \cup_{e} {\mathcal G}_e.
\label{8}
\end{equation}
It follows from \eqref{14} that
\begin{equation}
\label{9}
\scalprod{\phi}{{\rm e}^{\i tL}\psi}= \sum_{e} J_e(t) +J_\infty(t),
\end{equation}
where
\begin{equation}
\label{10}
J_\#(t) =\frac{1}{\i t}\frac{1}{2\pi\i}\int_{{\mathcal G}_\#}{\rm e}^{\i t z} \scalprod{\phi}{\textstyle\frac{d}{dz}R_z\psi} dz.
\end{equation}
We now apply a suitable Feshbach map to the resolvent $R_z$ in \eqref{10}, with a projection depending on the region of integration. Let $P$ be an orthogonal projection and recall the notation \eqref{n4}. The resolvent has the representation
\begin{equation}
\label{0}
R_z={\mathfrak F}(z)^{-1} +{\mathcal B}(z)+ R_z^Q,
\end{equation}
where ${\mathfrak F}(z)\equiv {\mathfrak F}(L-z,Q)$, see \eqref{n3},
and
\begin{equation}
{\mathcal B}(z) = - {\mathfrak F}(z)^{-1} Q L R^Q_z - R_z^Q L Q{\mathfrak F}(z)^{-1} +R_z^Q L Q {\mathfrak F}(z)^{-1} Q L R_z^Q.
\label{0.1}
\end{equation}
We explain these relations and some properties of the Feshbach map in Appendix \ref{FeshbachAppendix}. For $z\in{\mathcal G}_e$, we choose the projection $Q$ in the Feshbach map to be $P_e$. For $z\in{\mathcal G}_\infty$, the argument is simpler, see Section \ref{inftysect}.
Let us assume that the unperturbed, partially stable eigenvalue of $L_0$ is at the origin, $e=0$. (Otherwise see section \ref{proofthmnthm1sect}.) Then $L$ has a simple eigenvalue $E\equiv E_0$
with $E\rightarrow 0$ as $\d\rightarrow 0$. To analyze $J_0(t)$, we write, according to \eqref{0},
\begin{equation}
\label{p8}
J_0(t) = \frac{1}{\i t}\frac{1}{2\pi\i} \int_{{\mathcal G}_0} {\rm e}^{\i tz} \left\{ \scalprod{\phi}{\tfrac{d}{dz}{\mathfrak F}(z)^{-1}\psi}
+\scalprod{\phi}{\tfrac{d}{dz}{\mathcal B}(z)\psi}
+\scalprod{\phi}{\tfrac{d}{dz}R_z^{P_0}\psi} \right\} dz.
\end{equation}
The Feshbach term is ${\mathfrak F}(z) = -z+\d^2A_z$, where $A_z=-P_0IR_z^{P_0}IP_0$. For $z=0$ and $\d=0$, $A_z$ is just the level shift operator $\Lambda_0$, \eqref{n2}. We show in Lemma \ref{lemma1} that $A_z$ is diagonalizable,
$A_z = \sum_{j=0}^{m_e-1} a_{0,j}(z)Q_j(z)$, and that the eigenvalues $a_{0,j}(z)$ of $A_z$ satisfy $a_{0,0}(E)=E/\d^2$
for all $\d\neq 0$ (this follows from the isospectrality property of the Feshbach map and the fact that $E$ is an eigenvalue of $L$) and $a_{0,j}(z) = \lambda_{0,j}+O(\d^2+|z|)$, $j=1,\ldots,m_e-1$ (since $A_z$ is close to $\Lambda_0$). Then we can write
\begin{equation}
\label{p3}
\tfrac{d}{dz}{\mathfrak F}(z)^{-1} = \tfrac{d}{dz} \sum_{j=0}^{m_e-1} \frac{Q_j(z)}{-z+\d^2 a_{0,j}(z)}.
\end{equation}
We are interested in the singularities of this function as $z$ is close to the real axis. They come from the denominator. To understand the nature of the singularities, and since $z\mapsto Q_j(z)$ is regular, consider $Q_j(z)\approx Q_j(0)$
for a moment. Then
\begin{equation}
\label{p4}
\tfrac{d}{dz}{\mathfrak F}(z)^{-1}\approx \sum_{j=0}^{m_e-1} \frac{1-\d^2a'_{0,j}(z)}{(z-\d^2a_{0,j}(z))^2} Q_j(0).
\end{equation}
For $j=0$ we have $-E+\d^2a_{0,0}(E)=0$ (see above) and the corresponding summand is
\begin{eqnarray*}
\lefteqn{
\frac{1}{(z-E)^2} \frac{1-\d^2a'_{0,0}(z)}{\{1-\d^2[a_{0,0}(z)-a_{0,0}(E)]/(z-E)\}^2} Q_0(0)}\\
&& \qquad\qquad\qquad\approx \frac{1}{(z-E)^2} \frac{1-\d^2a'_{0,0}(E)}{(1-\d^2a'_{0,0}(E))^2} Q_0(0) = \frac{1}{(z-E)^2} \frac{Q_0(0)}{1-\d^2a'_{0,0}(E)}.
\end{eqnarray*}
By using that the projection associated to the eigenvalue $E$ of $L$ is given by
\begin{equation}
\label{p12}
\Pi_E=\lim_{\epsilon\rightarrow 0_+}(\i\epsilon)(L-E+\i\epsilon)^{-1}
\end{equation}
and decomposing the resolvent in this limit according to \eqref{0} with projection $P_0$, we identify (see \eqref{034})
$$
\frac{Q_0(E)}{1-\d^2a'_{0,0}(E)} = P_0\Pi_EP_0.
$$
For $j>0$ we have $a_{0,j}(0)=\lambda_{0,j}+O(\d^2)$ which is in the open upper complex half plane and the corresponding summand in \eqref{p4} is
$$
\frac{1-\d^2a'_{0,j}(z)}{(z-\d^2a_{0,j}(z))^2} Q_j(0) \approx \frac{Q_j(0)}{(1-\d^2a_{0,j}(0))^2}.
$$
In Section \ref{proofthmnthm1sect} we make these arguments rigorous. Namely, we show that
\begin{equation}
\label{p2}
\tfrac{d}{dz}{\mathfrak F}(z)^{-1} =\frac{1}{(z-E)^2}P_0\Pi_EP_0 +\sum_{j=1}^{m_e-1} \frac{Q_j(0)}{(z-\d^2 a_{0,j}(0))^2} + \widetilde T(z),
\end{equation}
where $\int_{{\mathcal G}_0}{\rm e}^{\i tz} \widetilde T(z)dz\prec 1$. Now we have to multiply \eqref{p2} by ${\rm e}^{\i t z}$ and integrate over $z\in{\mathcal G}_0 = [-\alpha,\alpha]-\i w$. Having in mind a standard argument from complex analysis, we complete the path ${\mathcal G}_0$ into a closed contour (a rectangle with a `roof' parallel to ${\mathcal G}_0$ but shifted far into the upper complex half plane). We then use the Cauchy formula for contour integrals to get
\begin{equation}
\label{p5}
\int_{{\mathcal G}_0}\frac{{\rm e}^{\i tz}}{(z-E)^2}dz = 2\pi\i\cdot \i t {\rm e}^{\i t E}+O(1/t).
\end{equation}
The $O(1/t)$ term is the contribution from the parallel vertical sides of the rectangular closed integration path (see \eqref{017}). In a similar way, we treat the sum in \eqref{p2}. Here the poles are at $z=\d^2 a_{0,j}(0)$ and so
\begin{equation}
\label{p6}
\int_{{\mathcal G}_0}\frac{{\rm e}^{\i tz}}{(z-\d^2a_{0,j}(0))^2} dz = 2\pi\i \cdot\i t\, {\rm e}^{\i t\d^2a_{0,j}(0)} +O(1/t).
\end{equation}
Combining \eqref{p2} with \eqref{p5} and \eqref{p6} yields
\begin{equation}
\label{p7}
\frac{1}{\i t}\frac{1}{2\pi\i} \int_{{\mathcal G}_0}{\rm e}^{\i tz} \scalprod{\phi}{\tfrac{d}{dz}{\mathfrak F}(z)^{-1}\psi} dz = {\rm e}^{\i t E}\scalprod{\phi}{P_0\Pi_EP_0\psi} +\sum_{j=1}^{m_e-1} {\rm e}^{\i t\d^2 a_{0,j}(0)} \scalprod{\phi}{Q_j(0)\psi} +O(1/t).
\end{equation}
Next we deal with the second integrand in \eqref{p8}. Using again the spectral representation of ${\mathfrak F}(z)^{-1}$, we have from \eqref{0.1}
\begin{equation}
\label{p10}
\tfrac{d}{dz}{\mathcal B}(z) = \tfrac{d}{dz} \sum_{j=0}^{m_e-1} \frac{q_j(z)}{-z+\d^2a_{0,j}(z)},
\end{equation}
where
\begin{equation}
\label{p9}
q_j(z) =-\d \left[ Q_j(z)P_0IR_z^{P_0} +R_z^{P_0}IP_0Q_j(z) -\d R_z^{P_0} IP_0Q_j(z)P_0IR_z^{P_0}\right].
\end{equation}
The expression \eqref{p10} has the same structure as \eqref{p3}. We readily obtain, in analogy with \eqref{p7},
\begin{equation*}
\label{p11}
\frac{1}{\i t}\frac{1}{2\pi\i} \int_{{\mathcal G}_0}{\rm e}^{\i tz} \scalprod{\phi}{\tfrac{d}{dz}{\mathcal B}(z)\psi} dz ={\rm e}^{\i tE} \scalprod{\phi}{\tfrac{q_0(E)}{1-\d^2a'_{0,0}(E)}\psi} +\sum_{j=1}^{m_e-1} {\rm e}^{\i t\d^2 a_{0,j}(0)} \scalprod{\phi}{q_j(0)\psi} +O(|\d|/t).
\end{equation*}
Proceeding as above, after \eqref{p12}, we identify $P_0\Pi_EP_0 +\tfrac{q_0(E)}{1-\d^2a'_{0,0}(E)}=\Pi_E$ (see also \eqref{039}). Finally, since by Assumption (A2)(1),
$$
\int_{{\mathcal G}_0}{\rm e}^{\i tz}\scalprod{\phi}{\tfrac{d}{dz}R_z^{P_0}\psi} dz\prec 1,
$$
we obtain
\begin{equation}
\label{p13}
J_0(t) = \scalprod{\phi}{\Pi_E\psi} + \sum_{j=1}^{m_e-1}{\rm e}^{\i t\d^2 a_{0,j}(0)}\scalprod{\phi}{\Pi'_{0,j}\psi} +O(1/t),
\end{equation}
where $\Pi'_{0,j} = Q_0(0)+O(|\d|)=P_{0,j}+O(|\d|)$. This explains the contribution of a term on the right side of \eqref{nmainreseqn} coming from a partially stable eigenvalue $e$ ($=0$). The analysis for unstable $e$ follows using the same arguments. Finally, to deal with $J_\infty(t)$, we write
\begin{equation}
\label{i0}
R_z\psi = (z+\i)^{-2} R_z(L+\i)^2\psi -(z+\i)^{-1}\psi - (z+\i)^{-2}(L+\i)\psi,
\end{equation}
which is valid for $\psi\in\mathrm{dom}(L^2)$. The negative powers of $z$ help the convergence of the $z$-integral over ${\mathcal G}_\infty$. The bound $J_\infty(t)\ple1/t$ is then easily reached using (A2)(2).
\section{Application to open quantum systems}
\subsection{Setup} The Hilbert space is the product of a system and a reservoir part,
\begin{equation}
\label{1}
{\mathcal H} ={\mathcal H}_\S\otimes{\mathcal H}_\mathcal{R}.
\end{equation}
The self-adjoint generator of dynamics, called {\em Liouvillean}, is of the form \eqref{2}, where $L_0$, the free (non interacting) Liouvillean, is a sum of a system and a reservoir contribution,
\begin{equation}
\label{3}
L_0=L_\S+L_\mathcal{R},
\end{equation}
and $I$ is the system-reservoir interaction operator. We consider the system to be finite-dimensional and the reservoir to be an infinitely extended free Bose gas at positive temperature, as we explain now.
Let $\S$ be a quantum system with pure state space ${\mathfrak H}_\S$ of dimension $d_0<\infty$. For instance, for a spin $1/2$, $d_0=2$. Then the Hilbert space ${\mathcal H}_\S$ in \eqref{1} is the GNS space (Liouville space)
\begin{equation}
\label{particleGNS}
{\mathcal H}_\S = {\mathfrak H}_\S\otimes{\mathfrak H}_\S,
\end{equation}
so that $d=\dim {\mathcal H}_\S=d_0^2$. The doubling of the pures state system Hilbert space in \eqref{particleGNS} allows to represent any (pure or mixed) state of $\S$ by a vector. Namely, let $\rho$ be a density matrix on ${\mathfrak H}_\S$. It has the diagonalized form $\rho=\sum_i p_i |\psi_i\rangle\langle\psi_i|$, to which we associate the vector $\Psi_\rho = \sum_i \sqrt{p_i}\psi_i\otimes\overline\psi_i\in {\mathfrak H}_\S\otimes {\mathfrak H}_\S$
(complex conjugation in any fixed basis -- we will choose the eigenbasis of the system Hamiltonian). Then ${\rm Tr}(\rho A)=\scalprod{\Psi_\rho}{(A\otimes\mathbf{1}_\S)\Psi_\rho}$ for all $A\in{\mathcal B}({\mathfrak H}_\S)$ and where $\mathbf{1}_\S$ is the identity in ${\mathfrak H}_\S$. This is the GNS representation of the state given by $\rho$ \cite{BR,MSB2}. Let $H_\S=\sum_j E_j |\varphi_j\rangle\langle\varphi_j|$ be the Hamiltonian of $\S$, acting on ${\mathfrak H}_\S$. The equilibrium density matrix is $\rho_\S = {\rm e}^{-\beta H_\S}/{\rm Tr\,}{\rm e} ^{-\beta H_\S}$, which is represented on ${\mathcal H}_\S$ by the vector
\begin{equation}
\label{systemkms}
\Omega_{\S,\beta} = ({\rm Tr\,}{\rm e}^{-\beta H_\S})^{-1} \sum_j {\rm e}^{-\beta E_j/2} \varphi_j\otimes\varphi_j.
\end{equation}
The (GNS) Hilbert space of the spatially infinitely extended free bose gas, for states normal w.r.t. the equilibrium (KMS) state, is the Fock space
\begin{equation}
{\mathcal H}_\mathcal{R} = {\mathcal F}_\beta= \bigoplus_{n\geq 0} L^2_{\rm sym}(({\mathbb R}\times S^2)^{n},(d u\times d\Sigma)^{n}),
\label{2.1}
\end{equation}
taken over the single-particle space $L^2({\mathbb R}\times S^2,d u\times d\Sigma)$, where $d\Sigma$ is the uniform measure on $S^2$ \cite{AW,JP}. ${\mathcal F}_\beta$ carries a representation of the CCR algebra in which the Weyl operators are given by $W(f_\beta) = e^{i\phi(f_\beta)}$, where $\phi(f_\beta)=\frac{1}{\sqrt{2}}(a^*(f_\beta)+a(f_\beta))$. Here, $a^*(f_\beta)$ and $a(f_\beta)$ denote creation and annihilation operators on ${\mathcal F}_\beta$, smoothed out with the function
\begin{equation}
f_\beta(u,\Sigma) = \sqrt{\frac{u}{1-e^{-\beta u}}}\ |u|^{1/2} \left\{
\begin{array}{ll}
f(u,\Sigma), & u\geq 0\\
-\overline{f}(-u,\Sigma), & u<0
\end{array}
\right.
\label{2.3}
\end{equation}
belonging to $L^2({\mathbb R}\times S^2,d u\times d\Sigma)$. It is easy to see that the CCR are satisfied, namely,
\begin{equation}
W(f_\beta)W(g_\beta) = e^{-\frac{i}{2}{\rm Im}\scalprod{f}{g}} W(f_\beta+g_\beta).
\label{ccr}
\end{equation}
The vacuum vector $\Omega\in{\mathcal F}_\beta$ represents the infinite-volume equilibrium state of the free Bose field, determined by the formula
\begin{equation}
\label{thav}
\scalprod{\Omega}{W(f_\beta)\Omega} = \exp\left\{ \textstyle-\frac14 \scalprod{f}{\coth(\beta|k|/2)f}\right\}.
\end{equation}
The Weyl algebra is represented on ${\mathcal F}_\beta$ as $W(f)\mapsto W(f_\beta)$, for functions $f\in L^2({\mathbb R}^3)$ such that $ \scalprod{f}{\coth(\beta|k|/2)f}<\infty$. We denote the von Neumann algebra of the represented Weyl operators by ${\mathcal W}_\beta$.
The combined system-reservoir Hilbert space is then $\mathcal H$, \eqref{1}, and the von Neumann algebra of observables is
\begin{equation}
\label{vna}
{\frak M}={\mathcal B}({\mathfrak H}_\S)\otimes\mathbf{1}_\S\otimes{\mathcal W}_\beta \subset {\mathcal B}({\mathcal H}).
\end{equation}
The coupled dynamics is given by
\begin{equation}
\label{dyn}
\alpha^t(A) ={\rm e}^{\i tL}A{\rm e}^{- \i tL}, \qquad A\in\frak M.
\end{equation}
It is generated by the self-adjoint Liouville operator acting on $\mathcal H$,
\begin{eqnarray}
L &=& L_0+\d I\label{2.4}\\
L_0 &=& L_\S+L_\mathcal{R}, \label{2.4'}\\
I &=& V - JVJ.\label{2.4''}
\end{eqnarray}
Here, $L_\S=H_\S\otimes \one_\S-\one_\S\otimes H_\S$ and $H_\S$ is the system Hamiltonian acting on ${\mathfrak H}_\S$. $L_\r=d\Gamma(u)$ is the second quantization of multiplication by the radial variable $u$. The interaction $I$ in \eqref{2.4''} is ``in standard form", involving a self-adjoint interaction operator $V$ acting on $\mathcal H$ and the modular conjugation $J$, which acts as
\begin{equation}
J( A\otimes\one_\S\otimes W(f_\beta(u,\Sigma)) )J = \one_\S\otimes\overline A\otimes W(\overline f_\beta(-u,\Sigma)),
\label{2.6}
\end{equation}
where $\overline A$ is the matrix obtained from $A$ by taking entrywise complex conjugation (matrices are represented in the eigenbasis of $H_\S$). Note that by \eqref{2.3}, we have $\overline f_\beta(-u,\Sigma) = -e^{-\beta u/2}f_\beta (u,\Sigma)$. By the Tomita-Takesaki theorem \cite{BR}, conjugation by $J$ maps the von Neumann algebra of observables \eqref{vna} into its commutant. In particular, $V$ and $JVJ$ commute (strongly on a suitable domain). For more detail about this well-known setup we refer to \cite{JP,BFSrte,MSB2} and references therein. We have in mind two commonly used forms for $V$,
\begin{equation}
\label{V}
V_1 = G\otimes\mathbf{1}_\S\otimes \phi(h_\beta)+\textrm{h.c.}\qquad \mbox{or} \qquad V_2=G\otimes\mathbf{1}_\S\otimes W(h_\beta)+\textrm{h.c.},
\end{equation}
for some matrix $G$ on ${\mathfrak H}_\S$ and where $h_\beta$ is a (represented) {\em form factor}, obtained from an $h\in L^2({\mathbb R}^3)$ by \eqref{2.3}. The interaction $V_1$ is standard. $V_2$ comes about when considering the spin-boson system at arbitrary coupling strength \cite{KoMeSo}, see Section \ref{sbsect}.
The vector representing the uncoupled $(\alpha^t,\beta)$-KMS state ($\d=0$) on $\mathfrak M$ is
\begin{equation}
\Omega_{0,\rm KMS} = \Omega_{\S,\beta}\otimes \Omega,
\label{2.13}
\end{equation}
where $\Omega_{\S,\beta}$ is given in \eqref{systemkms}.
In this setup of open systems, one can derive Condition (A2) from a global limiting absorption principle as follows.
\begin{thm}
\label{lapos}
Let $P_\mathcal{R}=\mathbf{1}_\S\otimes|\Omega\rangle\langle\Omega|$.
Suppose that there is a dense set ${\mathcal D}\subset{\mathcal H}$ with ${\rm Ran} IP_\mathcal{R} \subset {\mathcal D}$, s.t. for all $\varphi,\psi\in\mathcal D$,
\begin{equation}
\label{nlapos1}
\sup_{z\in{\mathbb C}_-} |\tfrac{d^k}{dz^k}\scalprod{\phi}{R_z^{P_\mathcal{R}}\psi}|\le C(\phi,\psi),\quad k=0,\ldots,3.
\end{equation}
Then Condition (A2) holds with $\alpha=g/2=\min_{e\neq e'}\{|e-e'|\}/2$.
\end{thm}
We prove Theorem \ref{lapos} in Appendix \ref{appendix1}. (The constant $C(\phi,\psi)$ in \eqref{nlapos1} may differ from that in (A2).)
\subsection{The spin-boson model for arbitrary coupling strength}
\label{sbsect}
The spin-boson Hamiltonian is \cite{Leggett}
\begin{equation}
\label{1.3}
H=-\tfrac12\d\sigma_x+\tfrac12\varepsilon\sigma_z+H_\mathcal{R}+\tfrac12q_0\sigma_z\otimes\phi(h),
\end{equation}
where $\sigma_x$ and $\sigma_z$ are the Pauli matrices
$$
\sigma_x=
\begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix}, \qquad
\sigma_z=
\begin{pmatrix}
1 & 0\\
0 & -1
\end{pmatrix}
$$
and $\d, \varepsilon\in\mathbb R$ are the `tunneling matrix element' and the `detuning parameter', respectively. (We use units so that $\hbar$ takes the value one.) The reservoir Hamiltonian is $H_\mathcal{R}=\int_{{\mathbb R}^3} |k| a^*(k)a(k) d^3k$. The coupling constant is $q_0\in\mathbb R$. Associated to the Hamiltonian $H$ is the Liouvillean
\begin{equation}
\label{sbliouv}
L = L_\S+L_\mathcal{R}+q_0I
\end{equation}
with
\begin{eqnarray}
L_S&=& H_\S\otimes\mathbf{1}_{{\mathbb C}^2} -\mathbf{1}_{{\mathbb C}^2}\otimes H_\S,\label{-=1}\\
L_\mathcal{R}&=&d\Gamma(u)\label{-=2}
\end{eqnarray}
where $H_\S=-\tfrac12\d\sigma_x+\tfrac12\varepsilon\sigma_z$, and $I$ given in \eqref{2.4''} with $V=\tfrac12 \sigma_z\otimes\phi(h)$. The total Hilbert space is given by \eqref{1}, \eqref{particleGNS} with ${\mathfrak H}_\S={\mathbb C}^2$ and \eqref{2.1}.
In order to be able to analyze the spectrum of $L$ for arbitrarily large couplings $q_0\in\mathbb R$, one applies the unitary (`polaron'-) transformation $U$ (see \cite{KoMeSo})
\begin{equation}
\label{unitary}
U = u\, JuJ,\qquad \mbox{where}\qquad u={\rm e}^{\i\sigma_z\otimes\mathbf{1}_\S\otimes\phi(f_\beta)},
\end{equation}
resulting in a new Liouvillean \cite{Leggett,KoMeSo}
\begin{equation}
\label{nn1}
{\mathcal L} = ULU^* = {\mathcal L}_0+\d I,
\end{equation}
where
\begin{equation}
\label{nn2}
{\mathcal L}_0={\mathcal L}_\S+{\mathcal L}_\mathcal{R} = \tfrac{\varepsilon}{2}\big( \sigma_z\otimes\mathbf{1}_\S - \mathbf{1}_\S\otimes\sigma_z\big) +L_\mathcal{R}
\end{equation}
and
\begin{equation}
\label{nn3}
I = -\tfrac12 \big({\mathcal V}-J{\mathcal V}J\big),\qquad {\mathcal V}=\sigma_+\otimes\mathbf{1}_\S\otimes W(2f_\beta) + \sigma_-\otimes\mathbf{1}_\S\otimes W(-2f_\beta).
\end{equation}
Here, $\sigma_+$ and $\sigma_-$ are the raising and lowering operators and
\begin{equation}
\label{2.9'''}
f_\beta=(-\tfrac{i}{2}q_0h/u)_\beta.
\end{equation}
The non-interacting KMS state associated to $\L_0$ is
\begin{equation}
\label{nonintkms}
\Psi_{0,{\rm KMS}} = \Psi_{\S,\beta}\otimes\Omega,\qquad \Psi_{\S,\beta}=\frac{{\rm e}^{-\beta\varepsilon/4}\varphi_+\otimes\varphi_+ +{\rm e}^{\beta\varepsilon/4}\varphi_-\otimes\varphi_-}{\sqrt{{\rm e}^{-\beta\varepsilon/2}+{\rm e}^{\beta\varepsilon/2}}}
\end{equation}
and the interacting KMS state associated to $\L$ is (Araki's perturbation theory of KMS states)
\begin{equation}
\label{intkms}
\Psi_{\rm KMS} = \frac{{\rm e}^{-\beta(\L_0+\d{\mathcal V})/2}\Psi_{0,{\rm KMS}}}{\|{\rm e}^{-\beta(\L_0+\d{\mathcal V})/2}\Psi_{0,{\rm KMS}} \|}.
\end{equation}
Note that the spectrum of ${\mathcal L}_0$ consists of a purely absolutely continuous part covering all of $\mathbb R$, in which are embedded the eigenvalues $e=\pm\varepsilon$ (each simple) and the doubly degenerate eigenvalue $e=0$. The following is the main result of \cite{KoMeSo}:
\begin{thm}[\cite{KoMeSo}]
\label{KoMeSothm}
Recall that $(\cdot)_\beta$ is defined in \eqref{2.3}. Assume that $(1+|\i\partial_u|^\eta)(\i h/u)_\beta \in L^2({\mathbb R}\times S^2, du\times d\Sigma)$ for some $\eta>2$. Then, for any $q_0\in\mathbb R$, $q_0\neq 0$, there is a constant $\d_0>0$ s.t. if $0<|\d|\le\d_0$, then ${\mathcal L}$ has purely absolutely continuous spectrum covering $\mathbb R$ and a single, simple eigenvalue at zero. The associated eigenvector is $\Psi_{\rm KMS}$, \eqref{intkms}.
\end{thm}
{\em Remarks\ } 1. The eigenvalues $e=\pm\varepsilon$ of ${\mathcal L}_0$ are unstable, while $e=0$ is partially stable. The associated (simple) eigenvalue of $\mathcal L$ is $E=0$.
2. The spectral properties of $L$ and $\mathcal L$ are the same, as the operators are unitarily equivalent to each other.
3. The KMS state associated to $L$ is given by
\begin{equation}
\label{equilstate}
\Omega_{\rm KMS} = U^*\Psi_{\rm KMS} = \frac{{\rm e}^{-\beta(L_0+q_0\sigma_z\otimes\mathbf{1}_{{\mathbb C}^2}\otimes \phi(h_\beta))/2}\Omega_{0,{\rm KMS}}}{\| {\rm e}^{-\beta(L_0+q_0\sigma_z\otimes\mathbf{1}_{{\mathbb C}^2}\otimes \phi(h_\beta))/2}\Omega_{0,{\rm KMS}}\|},
\end{equation}
where $L_0=L_\S(\d)+L_\mathcal{R}$, see \eqref{-=1}, \eqref{-=2}. One shows that $\Omega_{0,\rm KMS}$ is in the domain of ${\rm e}^{-\beta(L_0+q_0\sigma_z\otimes\mathbf{1}_{{\mathbb C}^2}\otimes \phi(h_\beta))/2}$ for any $q_0$, $\d\in\mathbb R$ (see e.g. \cite{DJP,BFSrte,BR}).
\medskip
We now verify assumptions (A1)-(A4) for the spin-boson system, i.e., for the operator $\mathcal L$, \eqref{nn1}.
The eigenprojections of ${\mathcal L}_0$ are given, for $e\in{\rm spec}({\mathcal L}_S)$
by
$$
P_e=\mathbf{1}[{\mathcal L}_\S=e]\otimes |\Omega\rangle\langle\Omega|.
$$
Since $\mathbf{1}[{\mathcal L}_\S=e]\big(\sigma_\pm\otimes\mathbf{1}_\S\big) \mathbf{1}[{\mathcal L}_\S=e]=0=\mathbf{1}[{\mathcal L}_\S=e]\big(\mathbf{1}_\S\otimes\sigma_\pm\big) \mathbf{1}[{\mathcal L}_\S=e]$, condition (A1) holds.
To verify the limiting absorption principle (A2), let $N=d\Gamma(\mathbf{1})$ be the number operator on Fock space \eqref{2.1} and put $\bar N=P^\perp_\mathcal{R} N$. Let $A=d\Gamma(\i\partial_u)$ and put $\bar A=P^\perp_\mathcal{R} A$. For $\alpha,\nu\ge 0$, define the norms
\begin{equation}
\label{alphanunorms}
\|\xi\|_{\alpha,\nu}= \|\bar{N}^{\nu/2}(1+\bar A^2)^{\alpha/2}\xi\|.
\end{equation}
We have the following regularity properties of the resolvent $R^{P_\mathcal{R}}_z$.
\begin{thm}
\label{MK-Lem1}
Let $\mu\ge 1$ and suppose that $\partial^j_uf_\beta\in L^2({\mathbb R}\times S^2,du\times d\Sigma)$, for $j=0,\ldots,2\mu+1$. We have
\begin{eqnarray}
\sup_{z\in{\mathbb C}_-} |\tfrac{d^{\mu-1}}{dz^{\mu-1}}\scalprod{\phi}{ R^{ P_\mathcal{R}}_z\psi}|&
\prec & \|\phi\|_{\mu,2\mu}\,\|\psi\|_{\mu,2\mu}\label{eq-w17} \\
\sup_{z\in{\mathbb C}_-}|\tfrac{d}{d\d}\scalprod{\phi}{R^{ P_\mathcal{R}}_z\psi}|& \prec & \|\phi\|_{3,1}\|\psi\|_{3,1}.
\label{MK-eq19}
\end{eqnarray}
\end{thm}
We give a proof of Theorem \ref{MK-Lem1} in Appendix \ref{appendix1}. The bound \eqref{eq-w17} with $\mu=4$ implies \eqref{nlapos1}, with the dense set \begin{equation}
\label{deh}
{\mathcal D}=\{\psi\ :\ \|\psi\|_{4,8}<\infty\}.
\end{equation}
To see that ${\rm Ran}IP_\mathcal{R}\subset{\mathcal D}$, it suffices to check that
\begin{equation}
\label{acheck}
\| N^4 (1+A^4)W(2f_\beta)\Omega\|<\infty.
\end{equation}
It is not hard to use the relation
\begin{equation}
\label{commut} {}[d\Gamma(D),W(f)]=W(f)(\phi(iDf)+\tfrac12\scalprod{f}{Df})
\end{equation}
(where $\phi$ is the field operator, see the proof of Lemma 3.1 in \cite{KoMeSo} for technical details) for $D=-\imath \partial_u$ and $D=\mathbf{1}$ to see that \eqref{acheck} holds provided $\partial_u^jf_\beta\in L^2({\mathbb R}\times S^2)$, $j=0,\ldots,4$.
Therefore, Theorem \ref{MK-Lem1} combined with this last observation shows that the assumptions of Theorem \ref{lapos} are satisfied. Thus, by the latter theorem, assumption (A2) holds.
Next, assumption (A3) is shown to hold in Theorem \ref{lape}, \eqref{ddlocal}. Namely, the regularity in $\d$ of $\tfrac{d^k}{d\d^k} P_eIR^{P_e}_z IP_e$ is derived from that of $\tfrac{d^k}{d\d^k} P_eIR^{P_\mathcal{R}}_z IP_e$, given in Theorem \ref{MK-Lem1}, \eqref{MK-eq19}.
The Fermi Golden Rule Assumption (A4) is verified by examining the level shift operators $\Lambda_0$ and $\Lambda_{\pm \varepsilon}$. $\Lambda_0$ is two-dimensional, given by (see \cite{KoMeSo}, Proposition 3.5)
\begin{equation}
\label{lambdanot}
\Lambda_0 = \i\tau^{-1} P^\perp_{\S,\beta}
\end{equation}
where $P^\perp_{\S,\beta}$ is the complement of $P_{\S,\beta}=|\Omega_{\S,\beta}\rangle\langle\Omega_{\S,\beta}|$ in ${\rm Ran} \,\mathbf{1}[\L_\S=0]$ and where $\Omega_{\S,\beta}\propto {\rm e}^{-\beta \varepsilon/4}\varphi_+\otimes\varphi_++{\rm e}^{\beta\varepsilon/4}\varphi_-\otimes\varphi_-$ (see \eqref{systemkms} and \eqref{nn2}). Also,
\begin{equation}
\tau^{-1} = \int_0^\infty d t\cos(\varepsilon t) \cos\left[\frac{q_0^2}{\pi}\, Q_1(t)\right]e^{-\frac{q_0^2}{\pi}\, Q_2(t)}
\label{t2}
\end{equation}
with
$$
Q_1(t) = \int_0^\infty d\omega \frac{J(\omega)}{\omega^2}\sin(\omega t)\quad \mbox{and}\quad
Q_2(t) = \int_0^\infty d\omega \frac{J(\omega)(1-\cos(\omega t))}{\omega^2} \coth(\beta\omega/2).
$$
Here, $J(\omega)$ is the {\em spectral density} of the reservoir, defined by
\begin{equation}
J(\omega) = \textstyle\frac{\pi}{2}\omega^2\int_{S^2} |h(\omega,\Sigma)|^2 d\Sigma,\qquad \omega\geq 0,
\label{1.10}
\end{equation}
the integral being taken over the angular part in ${\mathbb R}^3$. The function $h$ is the form factor in the interaction \eqref{1.3}.\footnote{The spectral density is related to the Fourier transform of the reservoir correlation function $C(t)=\omega_{R,\beta}(e^{i tH_R}\varphi(h)e^{-i tH_R}\varphi(h))$ by $J(\omega)=\sqrt{\pi/2}\tanh(\beta\omega/2)[\widehat{C}(\omega)+\widehat{C}(-\omega)]$. Of course, it is assumed here, as it is in \cite{Leggett}, that the integral in \eqref{t2} does not vanish, so that $\tau<\infty$ is a finite relaxation time.} Relation \eqref{lambdanot} gives
\begin{equation}
\label{lambdanotnot}
\lambda_{0,0}=0\qquad \mbox{and}\qquad \lambda_{0,1}=\i\tau^{-1}.
\end{equation}
Hence assumption (A4)(2) holds. The resonances $\lambda_{\pm\epsilon,0}$ are the eigenvalues of the one-dimensional level shift operators $\Lambda_{\pm\varepsilon}$, which are easily calculated to be
\begin{equation}
\label{lambdanotepsilon}
\lambda_{\pm\varepsilon,0} = \pm x+\tfrac12 \i\tau^{-1},
\end{equation}
where $\pm x$ is the real part. Assumption (A4)(1) thus holds.
Set $\varphi_{+-}= \varphi_+\otimes\varphi_-$ etc. and
\begin{eqnarray*}
X_0&=&W(f_\beta) JW(f_\beta)\Omega,\qquad \ \, X_0^*=W(f_\beta)^*J W(f_\beta)^*\Omega,\\
X_+&=&W(f_\beta)^* JW(f_\beta)\Omega,\qquad X_-=JX_+=W(f_\beta)JW(f_\beta)^*\Omega.
\end{eqnarray*}
The dynamics of the spin-boson system at arbitrary coupling is then explicitly given as follows.
\begin{cor}[Dynamics of the spin-boson system at arbitrary coupling strength]
\label{cor1}
Suppose that $uf_\beta, \partial_u^j f_\beta\in L^2({\mathbb R}\times S^2)$ for $j=0,\ldots,4$. For any $q_0\in\mathbb R$ there is a $\d_0>0$ s.t. if $0<|\d|<\d_0$ then the following holds.
Denote by $\Pi_0=|\Omega_{\rm KMS}\rangle\langle\Omega_{\rm KMS}|$ the projection onto the coupled KMS state, \eqref{equilstate}. Let $\phi,\psi\in\mathrm{dom}(L_\mathcal{R})\cap \mathrm{dom}(N^{17/2}+1)(A^4+1)$. We have for all $t>0$
\begin{eqnarray}
\scalprod{\phi}{{\rm e}^{\i tL}\psi} &=& \scalprod{\phi}{\Pi_{0}\psi} + {\rm e}^{\i t \d^2 a_0}\scalprod{\phi}{\Pi'_0\psi}\label{dynthm1}\\
&& + {\rm e}^{\i t (\varepsilon+ \d^2 a_\varepsilon)}\scalprod{\phi}{\Pi'_\varepsilon\psi} + {\rm e}^{\i t (-\varepsilon+ \d^2 a_{-\varepsilon})}\scalprod{\phi}{\Pi'_{-\varepsilon}\psi} +R(t),
\nonumber
\end{eqnarray}
where $R(t) \prec \frac{1}{t}$ and
\begin{equation*}
a_0 = \i\tau^{-1} +O(\d), \qquad a_{\pm\varepsilon} =\pm x+\tfrac{\i}{2}\tau^{-1} +O(\d),
\end{equation*}
\begin{eqnarray}
\Pi'_0 &=& ({\rm e}^{\beta\varepsilon/2}+{\rm e}^{-\beta\varepsilon/2})^{-1}\big( {\rm e}^{\beta\varepsilon/2}|\varphi_{++}\rangle\langle\varphi_{++}|\otimes|X_0\rangle\langle X_0| -|\varphi_{++}\rangle\langle\varphi_{--}|\otimes|X_0\rangle\langle X^*_0|\nonumber\\ &&-|\varphi_{--}\rangle\langle\varphi_{++}|\otimes|X_0^*\rangle\langle X_0| +{\rm e}^{-\beta\varepsilon/2} |\varphi_{--}\rangle\langle\varphi_{--}|\otimes|X_0^*\rangle\langle X_0^*| \big) +O(\d),\label{--1}\\
\Pi'_{\varepsilon} &=& |\varphi_{+-}\rangle\langle\varphi_{+-}|\otimes |X_+\rangle\langle X_+|+O(\d),\label{--2}\\
{} \qquad \Pi'_{-\varepsilon} &=&|\varphi_{-+}\rangle\langle\varphi_{-+}|\otimes |X_-\rangle\langle X_-|+O(\d).\label{--3}
\end{eqnarray}
\end{cor}
Note that the dynamics in Corollary \ref{cor1} is expressed with respect to the original Liouville operator $L$ (not the unitarily equivalent $\L$).
{\em Proof of Corollary \ref{cor1}. } We have $\scalprod{\phi}{{\rm e}^{\i tL}\psi}=\scalprod{U\phi}{{\rm e}^{\i t\L}U\psi}$. Theorem \ref{nthm1} gives the resonance expansion for $\scalprod{U\phi}{{\rm e}^{\i t\L}U\psi}$ provided $U\phi, U\psi\in{\mathcal D}$ s.t. $\L U\phi,\L U\psi\in{\mathcal D}$, where $\mathcal D$ is the set of vectors with finite $\|\cdot\|_{4,8}$ norm (see after Theorem \ref{MK-Lem1}).One can easily show the bound
\begin{eqnarray}
\nonumber
\|U\psi\|_{2\alpha,\nu}&\prec& \|\psi\|_{2\alpha,\nu}.
\label{UD}
\end{eqnarray}
We conclude from \eqref{UD} that if $\psi\in \mathrm{dom}(N^{\nu+1/2}+1)(A^{2\alpha}+1)$ for some $\nu\ge 1/2$ and $\alpha\ge0$, and if $(\i\partial_u)^jf_\beta\in L^2({\mathbb R}\times S^2)$ for $j=0,\ldots,2\alpha$, then $\|U\psi\|_{2\alpha,\nu}<\infty$.
We infer that if $\psi\in\mathrm{dom}(L_\mathcal{R})\cap \mathrm{dom}(N^{17/2}+1)(A^4+1)$, then $U\psi\in{\mathcal D}$ and $\L U\psi\in{\mathcal D}$. (To see the latter inclusion, we use \eqref{commut} with $D=u$ and proceed as in \eqref{UD}.)
The operators $\Pi'_0$ and $\Pi'_{\pm\varepsilon}$ are given by
\begin{equation}
\label{projU}
\Pi'_0= U^*P^\perp_{\S,\beta}P_\mathcal{R} U +O(\d),\qquad \Pi'_{\pm\varepsilon}=U^* P_{\pm\varepsilon}U +O(\d),
\end{equation}
where $P_{\pm\varepsilon}$ are the eigenprojections of $\L_0$ associated to the eigenvalues $\pm\varepsilon$, $P^\perp_{\S,\beta}$ is defined after \eqref{lambdanot} and $P_\mathcal{R}=\mathbf{1}_\S\otimes|\Omega\rangle\langle\Omega|$. It is easy to calculate $U\varphi_{+-}\otimes\Omega = \varphi_{+-}\otimes W(f_\beta)^* JW(f_\beta)J\Omega$, which shows \eqref{--2}. Similarly one obtains \eqref{--3}. For \eqref{--1}, we first calculate
$$
P^\perp_{\S,\beta} = ({\rm e}^{\beta\varepsilon/2}+{\rm e}^{-\beta\varepsilon/2})^{-1}\big( {\rm e}^{\beta\varepsilon/2}|\varphi_{++}\rangle\langle\varphi_{++}| -|\varphi_{++}\rangle\langle\varphi_{--}|-|\varphi_{--}\rangle\langle\varphi_{++}| +{\rm e}^{-\beta\varepsilon/2}|\varphi_{--}\rangle\langle\varphi_{--}|\big).
$$
Then we use \eqref{projU} to arrive at \eqref{--1}. This completes the proof of Corollary \ref{cor1}. \hfill \qed
\bigskip
Being a KMS state, $\Omega_{\rm KMS}$ given in \eqref{equilstate} is {\em separating} for $\mathfrak M$, which means that ${\mathfrak M}'\Omega_{\rm KMS}$ is dense in $\mathcal H$, where ${\mathfrak M}'$ is the commutant of $\mathfrak M$, see \cite{BR}. Any (normal) state $\omega$ on $\mathfrak M$ is given by a normalized vector $\Psi\in\mathcal H$ via $\omega(A)=\scalprod{\Psi}{A\Psi}$. We introduce the dense set
\begin{equation}
\label{dehnot}
{\mathcal D}_0 =\mathrm{dom}(L_\mathcal{R})\cap \mathrm{dom}(N^{17/2}+1)(A^4+1).
\end{equation}
The set of states $\omega$ arising from vectors in
\begin{equation}
\label{regularstates}
\big\{\Psi\in{\mathcal H}\ :\ \|\Psi\|=1, \, \Psi = B\Omega_{\rm KMS} \mbox{\ for some $B\in{\mathfrak M}'$, s.t.\ } B^*\Psi\in{\mathcal D}_0\big\}
\end{equation}
is dense (in the norm of states on $\mathfrak M$). We call it the set of regular states, ${\mathcal S}_{\rm reg}$. We also introduce the regular observables,
\begin{equation}
\label{regobs}
{\mathfrak M}_{\rm reg} = \big\{ A\in{\mathfrak M}\ :\ A\Omega_{\rm KMS}\in{\mathcal D}_0\big\}.
\end{equation}
Let us denote the coupled equilibrium state by
\begin{equation}
\label{jointKMSstate}
\omega_{\rm KMS}(A) = \scalprod{\Omega_{\rm KMS}}{A\Omega_{\rm KMS}}.
\end{equation}
\begin{cor}[Return to equilibrium]
\label{cor2}
For any $\omega_0\in{\mathcal S}_{\rm reg}$, $A\in{\mathfrak M}_{\rm reg}$, $t\ge 0$, we have
\begin{equation}
\label{mm63}
\left| \omega_0(\alpha^t(A)) - \omega_{\rm KMS}(A)\right| \le C_{A,\omega_0} \left[ {\rm e}^{-\d^2t/2\tau} + 1/t \right].
\end{equation}
The constant $C_{A,\omega_0}$ depends on the initial state $\omega_0$ and the observable $A$, but not on $t,\d$.
\end{cor}
One readily verifies that all states of the form $\omega_\S\otimes\omega_\mathcal{R}$, where $\omega_\S$ is arbitrary and $\omega_\mathcal{R}$ is the reservoir equilibrium, belong to ${\mathcal S}_{\rm reg}$. Moreover, all observables on the system alone belong to ${\mathfrak M}_{\rm reg}$, since $\Omega_{\rm KMS}\in{\mathcal D}_0$.
\bigskip
{\em Proof of Corollary \ref{cor2}.\ } Let $\Psi=B\Omega_{\rm KMS}$, $B\in{\mathfrak M}'$, be the vector representing $\omega_0$. Since $B$ commutes with $\alpha^t(A)$ and $\Omega_{\rm KMS}$ is in the kernel of $L$, we have $\alpha^t(A)\Psi=B\alpha^t(A)\Omega_{\rm KMS} = B{\rm e}^{\i tL}A\Omega_{\rm KMS}$. Thus,
\begin{equation}
\label{p1}
\omega_0(\alpha^t(A)) = \scalprod{\Psi}{{\rm e}^{\i tL} A{\rm e}^{-\i tL}\Psi} = \scalprod{B^*\Psi}{{\rm e}^{\i tL}A\Omega_{\rm KMS}}.
\end{equation}
Now we apply \eqref{dynthm1} and, using that $\Pi_0=|\Omega_{\rm KMS}\rangle\langle\Omega_{\rm KMS}|$, obtain directly \eqref{mm63}.\hfill \qed
\section{Proof of Theorem \ref{nthm1}}
\label{proofthmnthm1sect}
{}For $z\in{\mathbb C}_-$, $|{\rm Re}z-e|\le \alpha$ we define the operator
\begin{equation}
\label{Az}
A_z=-P_eI R_z^{P_e} IP_e.
\end{equation}
As $z\rightarrow e$ and $\d\rightarrow 0$, $A_e$ approaches the level shift operator $\Lambda_e$. More precisely, we have the following result, in which $\delta$ is, recall, given by \eqref{delta}.
\begin{lem}
\label{lemma1}
There is a constant $c$ such that if $|\d|$, $|{\rm Re}z-e| <c \delta$, and $z\in{\mathbb C}_-$, then
1. All eigenvalues of $A_z$ are distinct. Call them $a_{e,j}=a_{e,j}(z)$, $j=0,\ldots,m_e-1$. Each $a_{e,j}$ satisfies $|\lambda_{e,j}-a_{e,j}|<\delta/2$ for exactly one eigenvalue $\lambda_{e,j}$ of $\Lambda_e$.
2. The eigenvalues $a_{e,j}(z)$ of $A_z$, and the associated Riesz projections $Q_j(z)$ are analytic in $z\in{\mathbb C_-}$, $|{\rm Re}z-e|<c\delta$ and continuous as ${\rm Im}z\rightarrow 0_-$. They satisfy the bounds $\tfrac{d^k}{dz^k}a_{e,j}(z)$, $\tfrac{d^k}{dz^k}Q_j(z)\prec 1$ for $k=0,\ldots,3$, uniformly for $|{\rm Re}z-e|<c\delta$ and ${\rm Im} z \le 0$.
\end{lem}
The simplicity of the spectrum implies the spectral representation
\begin{equation}
\label{naz}
A_z = \sum_{j=0}^{m_e-1} a_{e,j}(z) Q_j(z).
\end{equation}
{\em Proof of Lemma \ref{lemma1}.\ } When necessary, we display the $\d$-dependence of $A_z$ by $A_z(\d)$. (Here, $A_x$ for $x\in\mathbb R$ is understood as the limit of $A_z$, as $z\rightarrow x$, $z\in{\mathbb C}_-$.) We have $A_e(0)=\Lambda_e$ and, by assumption (A3),
\begin{equation}
\label{04}
\|\Lambda_e-A_z(\d)\|\prec |\d|+|z-e|.
\end{equation}
Assumption (A4) implies that
\begin{equation}
\label{03}
\|(\Lambda_e-\zeta)^{-1}\| \prec \frac{1}{{\rm dist}(\zeta,{\rm spec}(\Lambda_e))}.
\end{equation}
Using the standard Neumann series for resolvents, together with the estimates \eqref{04} and \eqref{03}, yields
\begin{equation}
\label{05}
(A_z(\d)-\zeta)^{-1} = (\Lambda_e-\zeta)^{-1} +O\left(\frac{|\d|+|z-e|}{{\rm dist}(\zeta,{\rm spec}(\Lambda_e))^2}\right),
\end{equation}
provided $|\d|$, $|z-e| < c_0\, {\rm dist}(\zeta,{\rm spec}(\Lambda_e))$, for some constant $c_0$ independent of $\d$, $z$. Let $\mathcal{C}_j$ be the circle centered at $\lambda_{e,j}$ with radius $\delta/2$ and define
\begin{equation}
\label{06}
Q_j(\d,z) = \frac{-1}{2\pi \i}\oint_{\mathcal{C}_j}(A_z(\d)-\zeta)^{-1}d \zeta.
\end{equation}
Note that $Q_j(0,e)$ is the Riesz eigenprojection of $\Lambda_e$ associated to the eigenvalue $\lambda_{e,j}$. Using \eqref{05} and \eqref{06} we obtain that $\|Q_j(\d,z)-Q_j(0,e)\|<1$, provided $|\d|$, $|z-e|<c_0\delta/2$ and $|\d|+|z-e| < c_1\delta^2/4$, for some $c_1$ independent of $\d$, $z$. This proves point 1. of Lemma \ref{lemma1}.
Next,
\begin{equation}
\label{07}
Q_j'=\tfrac{d}{dz}Q_j=\frac{1}{2\pi\i}\oint_{\mathcal{C}_j}(A_z(\d)-\zeta)^{-1}A'_z(\d) (A_z(\d)-\zeta)^{-1}d \zeta \ \prec \,1.
\end{equation}
Since $a_{e,j}(z)={\rm Tr} A_zQ_j(z)$ we get $a'_{e,j}(z)={\rm Tr}(A'_zQ_j(z)+A_zQ'_j(z))\prec 1$. The statements about the higher derivatives follow in the same manner. This shows point 2 and completes the proof of Lemma \ref{lemma1}.\hfill \qed
\subsection{Estimates for $z$ in a vicinity of a partially stable eigenvalue $e$.}
\label{nestzero}
We introduce the operators
\begin{equation}
\label{qtilde}
\widetilde Q_j = -Q_j(e)P_eIR^{P_e}_{e-\i 0_+}-R^{P_e}_{e-\i 0_+}IP_eQ_j(e) +\d R^{P_e}_{e-\i 0_+} IP_eQ_j(e)P_eIR^{P_e}_{e-\i 0_+},\qquad j\ge 1,
\end{equation}
where $Q_j(e)$ is determined by \eqref{Az} and $R^{P_e}_{e-\i 0_+}$ is the limit of $R^{P_e}_{z}=(\bar L-z)^{-1}$ as $z$ approaches $e$ through the lower half plane. Note that $Q_j(e)=P_{e,j}+O(|\d|)$, see also \eqref{n7}.
\begin{prop}
\label{propzero}
Let $\Pi_{E_e}$ be the orthogonal projection associated to the eigenvalue $E_e$ of $L$. We have, for $\phi,\psi\in\mathcal D$,
\begin{equation}
\label{13.01}
J_e(t) = {\rm e}^{\i t E_e} \scalprod{\phi}{\Pi_{E_e}\psi} + \sum_{j=1}^{m_e-1} {\rm e}^{\i t(e+\d^2 a_{e,j}(e))}\scalprod{\phi}{(Q_j(e)+\d\, \widetilde Q_j)\psi} +R_e(t)
\end{equation}
where (recall that $w>0$ is the arbitrary parameter in \eqref{14})
$$
|R_e(t)|\le C \,\frac{1+{\rm e}^{wt}/t}{t},
$$
for a constant $C$ independent of $\d,t,w$, and where
\begin{equation}
\label{13.1}
a_{e,j}(e) = \lambda_{e,j} +O(|\d|),\qquad Q_j(e)=P_{e,j}+O(|\d|).
\end{equation}
\end{prop}
{\em Proof of Proposition \ref{propzero}.\ } We apply the Feshbach map with projection $P_e$ (having rank $m_e$),
\begin{equation}
\label{01}
{\mathfrak F}(z) \equiv{\mathfrak F}(L-z;P_e)= e-z+\d^2A_z,
\end{equation}
where $A_z$ is given in \eqref{Az}. Due to Condition (A2), $z\mapsto A_z$ is analytic for $z\in{\mathbb C}_-$, $|z-e|<\alpha$, and its $z$-derivatives up to degree 3 stay bounded as ${\rm Im}z\rightarrow 0_-$. According to the decomposition \eqref{0}, we have
\begin{equation}
\label{08}
\int_{{\mathcal G}_e}{\rm e}^{\i tz}\scalprod{\phi}{\textstyle\frac{d}{dz}R_z\psi} dz = \int_{{\mathcal G}_e}{\rm e}^{\i tz}\left[ \scalprod{\phi}{\textstyle\frac{d}{dz}{\mathfrak F}(z)^{-1}\psi} + \scalprod{\phi}{\textstyle\frac{d}{dz}{\mathcal B}(z)\psi}+ \scalprod{\phi}{\textstyle\frac{d}{dz}R^{P_e}_z\psi} \right] dz,
\end{equation}
where
\begin{equation}
{\mathcal B}(z) = -\d {\mathfrak F}(z)^{-1} P_e I R_z^{P_e} - \d R_z^{P_e} I P_e{\mathfrak F}(z)^{-1} +\d^2 R_z^{P_e} I P_e {\mathfrak F}(z)^{-1} P_e I R_z^{P_e} .
\label{09'}
\end{equation}
To examine ${\mathfrak F}(z)^{-1}$, we use the spectral representation of the operator $A_z$.
\subsubsection{The contribution to \eqref{08} from $\tfrac{d}{dz}{\mathfrak F}(z)^{-1}$} In this subsection, we will simply write
$$
a_j \equiv a_{e,j},\qquad 0\le j\le m_e-1,
$$
to ease the notation. Due to \eqref{01} and Lemma \ref{lemma1},
\begin{equation}
\label{033}
{\mathfrak F}(z)^{-1} = \sum_{j=0}^{m_e-1}\frac{1}{e-z+\d^2a_j}Q_j.
\end{equation}
We analyze the first term on the right side of \eqref{08}, using that
\begin{equation}
\tfrac{d}{dz}{\mathfrak F}(z)^{-1} =\sum_{j=0}^{m_e-1} T_j,\qquad \mbox{where}\qquad T_j=\frac{1-\d^2a_j'}{(z-e-\d^2a_j)^2}Q_j + \frac{1}{e-z+\d^2a_j}Q'_j .
\label{09}
\end{equation}
We examine the singularities of $T_j$ in $z$. By the isospectrality of the Feshbach map, we know that $e-E_e+\d^2A_{E_e}$ has an eigenvalue zero (see also \eqref{n5}). Therefore, $e-E_e+\d^2a_0(E_e)=0$. Also, $a_j(z)=\lambda_{e,j}+O(|\d|+|z-e|)$ for $j=1,\ldots,m_e-1$. Consider first $T_0$. We have
$$
z-e-\d^2a_0(z)=z-e-\d^2a_0(E_e)+\d^2(a_0(E_e)-a_0(z))=z-E_e+\d^2(a_0(E_e)-a_0(z))
$$
and so
\begin{eqnarray}
\label{012}
\frac{1-\d^2a_0'}{(z-e-\d^2a_0)^2} &=&\frac{1}{(z-E_e)^2}\frac{1-\d^2a'_0(z)}{[1-\d^2(a_0(z)-a_0(E_e))/(z-E_e)]^2}\\
& =& \frac{1}{(z-E_e)^2}\frac{1}{1-\d^2 a'_0(E_e)} +\frac{h(z)}{(z-E_e)^2},
\nonumber
\end{eqnarray}
where
\begin{equation}
\label{013}
h(z) = \frac{1-\d^2a'_0(z)}{[1-\d^2(a_0(z)-a_0(E_e))/(z-E_e)]^2} -\frac{1-\d^2a'_0(E_e)}{[1-\d^2a_0'(E_e)]^2} =O(\d^2|z-E_e|^2).
\end{equation}
To arrive at \eqref{013}, we expand $h(z)$ around a point $z_0\in {\mathbb C}_-$ which is very close to $E_e$,
$$
h(z)=h(z_0) +(z-z_0)h'(z_0)+\int_{[z_0,z]}d s\int_{[s,z_0]}ds'h''(s').
$$
The integrals are over paths (straight lines) in the lower complex plane. Then, sending $z_0\rightarrow E_e$, using that $h(E_e)=h'(E_e)=0$ and controlling the double integral with the third derivative of $h$, we arrive at \eqref{013}. In this argument, we assume the derivatives up to order three to have a continuous extension as ${\rm Im} z\rightarrow 0_-$.
Thus
\begin{eqnarray}
\label{014}
T_0 &=& \frac{1}{(z-E_e)^2}\frac{Q_0(E_e)}{1-\d^2 a'_0(E_e)} + \frac{1}{z-E_e}\frac{[Q_0(z)-Q_0(E_e)]/(z-E_e)}{1-\d^2 a'_0(E_e)}\\
&&-\frac{1}{z-E_e}\frac{Q'_0(z)}{1-\d^2[a_0(z)-a_0(E_e)]/(z-E_e)} +O(\d^2).
\nonumber
\end{eqnarray}
An expansion of the {\em sum} of the second and third term on the right side of \eqref{014} shows that this term is $O(1)$ uniformly in $z\in{\mathcal G}_e$, giving the bound
\begin{equation}
\label{015}
T_0= \frac{1}{(z-E_e)^2}\frac{Q_0(E_e)}{1-\d^2 a'_0(E_e)} +O(1+\d^2).
\end{equation}
Therefore,
\begin{equation}
\label{016}
\int_{{\mathcal G}_e}{\rm e}^{\i tz}\scalprod{\phi}{T_0\psi}=\frac{\scalprod{\phi}{Q_0(E_e)\psi}}{1-\d^2 a'_0(E_e)} \int_{{\mathcal G}_e} \frac{{\rm e}^{\i tz}}{(z-E_e)^2}dz + O(1).
\end{equation}
The remaining integral on the right side is estimated using the standard Cauchy formula from complex analysis. Namely, we complete ${\mathcal G}_e$ into a rectangular closed path, adding the vertical pieces ${\mathcal C}_\pm=\{e\pm\alpha+\i y\ :\ y\in [-w,R]\}$ and the horizontal roof $\{x+\i R\ :\ e-\alpha\le x\le e+\alpha\}$. Then we obtain from the Cauchy integral formula of basic complex analysis, upon taking $R\rightarrow\infty$, that
\begin{equation}
\label{017}
\int_{{\mathcal G}_e}\frac{{\rm e}^{\i tz}}{(z-E_e)^2}dz = 2\pi\i ({\rm e}^{\i tz})'|_{z=E_e} + O({\rm e}^{wt}/t) = \i t\, 2\pi\i\, {\rm e}^{\i tE_e} + O({\rm e}^{wt}/t).
\end{equation}
The remainder term comes from the integrals along the two vertical pieces of the path, which are bounded above by $\int_{-w}^\infty {\rm e}^{-yt}dy$. Note that $w>0$ is arbitrary (see \eqref{4}) and we will take $w\rightarrow 0$ which will make the remainder in \eqref{017} to be $O(1/t)$. Combining \eqref{016} and \eqref{017} yields
\begin{equation}
\label{018}
\frac{1}{\i t}\frac{1}{2\pi\i} \int_{{\mathcal G}_e}{\rm e}^{\i tz}\scalprod{\phi}{T_0\psi}dz = {\rm e}^{\i tE_e} \frac{ \scalprod{\phi}{Q_0(E_e)\psi}}{1-\d^2 a'_0(E_e)}+O(1/t+{\rm e}^{wt}/t^2).
\end{equation}
Next, we analyze $T_j$ in \eqref{09}, for $j\ge1$. Recall that $a_j(e)|_{\d=0}=\lambda_{e,j}$ are the eigenvalues with strictly positive imaginary part of the level shift operator $\Lambda_e$. The integrands behave in a different way now, since ${\rm Im}a_j(e)>0$ (while before, $a_0(E_e)=0$). The following bound is useful,
\begin{eqnarray}
|z-e-\d^2a_j(z)| &=& \left| z-e-\d^2\lambda_{e,j} -\d^2(a_j(e) - \lambda_{e,j})-\d^2(a_j(z)-a_j(e)) \right|\nonumber\\
&\ge& |z-e-\d^2\lambda_{e,j}| -c_1\d^2(|\d|+|z-e|)\nonumber\\
&\ge& \tfrac12 |z-e-\d^2\lambda_{e,j}|,
\label{020}
\end{eqnarray}
provided that $|\d|$, $|z-e|\le c_2 {\rm Im}\lambda_{e,j}$, where $c_1$ and $c_2=1/(2c_1)$ are independent of $\d$ and $z$. To arrive at the last inequality, \eqref{020}, we proceed as follows: the inequality is equivalent to $2c_1\d^2(|\d|+|z-e|)\le |z-e-\d^2\lambda_{e,j}|$. Now $|z-e-\d^2\lambda_{e,j}| \ge |{\rm Im}z-\d^2{\rm Im}\lambda_{e,j}|\ge \d^2{\rm Im}\lambda_{e,j}$, since ${\rm Im} z<0$ and ${\rm Im}\lambda_{e,j}>0$.
We have
\begin{eqnarray}
\label{019}
T_j &=& \frac{Q_j(e)}{(z-e-\d^2a_j(e))^2} - \frac{\d^2a'_j(z)Q_j(z)}{(z-e-\d^2a_j(z))^2}+ R_j+S_j,
\end{eqnarray}
where
\begin{eqnarray}
R_j &=& \frac{1}{z-e-\d^2a_j(z)}\left[\frac{z-e}{z-e-\d^2a_j(z)} \frac{Q_j(z)-Q_j(e)}{z-e}-Q_j'(z)\right] \nonumber\\
&=& \frac{1}{z-e-\d^2a_j(z)}\left[\frac{Q_j(z)-Q_j(e)}{z-e}-Q_j'(z) +\frac{\d^2a_j(z)}{z-e-\d^2a_j(z)}\frac{Q_j(z)-Q_j(e)}{z-e} \right] \nonumber\\
&=& O\left(\frac{|z-e|}{|z-e-\d^2\lambda_{e,j}|} + \frac{\d^2}{|z-e-\d^2\lambda_{e,j}|^2}\right)
\label{021}
\end{eqnarray}
and
\begin{equation}
\label{022}
S_j = \frac{-2\d^2 (z-e)(a_j(e)-a_j(z)) +\d^4[a_j(e)^2-a_j(z)^2]}{ [z-e-\d^2a_j(z)]^2 [z-e-\d^2a_j(e)]^2} Q_j(0)= O\left( \frac{\d^2|z-e|^2+\d^4|z-e|}{|z-e-\d^2\lambda_{e,j}|^4} \right).
\end{equation}
To arrive at the estimates \eqref{021} and \eqref{022} we have used \eqref{020}. Similarly, the second term on the right side of \eqref{019} is $O(\d^2/|z-e-\d^2\lambda_{e,j}|^2)$ and so we obtain
\begin{equation}
\label{023}
T_j = \frac{Q_j(e)}{(z-e-\d^2a_j(e))^2} +\widetilde T_j,
\end{equation}
with
\begin{equation}
\label{023.1}
\|\widetilde T_j\|\prec \frac{|z-e|}{|z-e-\d^2\lambda_{e,j}|} + \frac{\d^2}{|z-e-\d^2\lambda_{e,j}|^2}+ \frac{\d^2|z-e|^2+\d^4|z-e|}{|z-e-\d^2\lambda_{e,j}|^4}.
\end{equation}
Note that, with $\lambda_{e,j}=\xi_j+\i\eta_j$ and $w, \eta_j>0$, we have
$$
\frac{|z-e|}{|z-e-\d^2\lambda_{e,j}|}\le 1+\frac{\d^2|\lambda_{e,j}|}{|z-e-\d^2\lambda_{e,j}|}\le1+\frac{|\lambda_{e,j}|}{\eta_j}\prec 1,
$$
so
\begin{equation}
\label{024}
\int_{{\mathcal G}_e}\frac{|z-e|}{|z-e-\d^2\lambda_{e,j}|}dx \prec 1.
\end{equation}
Also,
\begin{equation}
\label{025}
\int_{{\mathcal G}_e} \frac{\d^2}{|z-e-\d^2\lambda_{e,j}|^2} dx \le2\int_0^{\alpha+\d^2|\xi_j|}\frac{\d^2}{x^2 +\d^4\eta_j^2} dx\le 2\int_0^\infty\frac{dy}{y^2+ \eta_j^2}\prec 1.
\end{equation}
Very similarly, one sees that
\begin{equation}
\label{0.26}
\int_{{\mathcal G}_e}\frac{\d^2|z-e|^2+\d^4|z-e|}{|z-e-\d^2\lambda_{e,j}|^4}dx\prec 1
\end{equation}
and it follows that
\begin{equation}
\label{027}
\int_{{\mathcal G}_e}{\rm e}^{\i tz} \widetilde T_j \,dz \prec 1.
\end{equation}
Next, using Cauchy's integral formula as above (\eqref{015}-\eqref{017}), we obtain that
\begin{equation}
\label{028}
\int_{{\mathcal G}_e} {\rm e}^{\i tz}\frac{Q_j(e)}{(z-e-\d^2a_j(e))^2}dz = \i t\, 2\pi\i \, {\rm e}^{\i t (e+\d^2 a_j(e))} Q_j(e) +O({\rm e}^{wt}/t).
\end{equation}
Combining \eqref{028} with \eqref{023} and \eqref{027}, we see that for all $j\geq 1$,
\begin{equation}
\label{029}
\frac{1}{\i t}\frac{1}{2\pi \i} \int_{{\mathcal G}_e} {\rm e}^{\i tz} \scalprod{\phi}{T_j\psi} dz = {\rm e}^{\i t(e+\d^2a_j(e))}\scalprod{\phi}{Q_j(e)\psi} +O(1/t+{\rm e}^{w t}/t^2).
\end{equation}
At this point, it is instructive to explain the coefficient $(1-\d^2a'_0(E_e))^{-1}$ in front of the non-decaying term in \eqref{018}. Let $\Pi_{E_e}$ be the projection onto the embedded eigenvalue $E_e$ of $L$. We have (in the strong sense) $\Pi_{E_e}=\lim_{\epsilon \rightarrow 0_+}(\i \epsilon) (L-E_e+\i \epsilon)^{-1}$, so by \eqref{0},
\begin{equation}
\label{031}
\Pi_{E_e}=\lim_{\epsilon \rightarrow 0_+}(\i \epsilon)\left\{ {\mathfrak F}(L-E_e+\i \epsilon;P_e)^{-1} +{\mathcal B}(E_e-\i\epsilon) +R_{E_e-\i\epsilon}^{P_e}\right\}.
\end{equation}
The $P_e$-block of the decomposition is, by \eqref{033},
\begin{equation}
P_e\Pi_{E_e}P_e = \lim_{\epsilon\rightarrow 0_+}(\i\epsilon){\mathfrak F}(L-E_e+\i\epsilon;P_e)^{-1}
=\lim_{\epsilon\rightarrow 0_+}\sum_{j=0}^{m_e-1}\frac{\i\epsilon}{e-E_e+\i\epsilon+\d^2a_j(E_e-\i\epsilon)} Q_j(E_e-\i\epsilon).
\end{equation}
{}For $j\ge 1$, we have ${\rm Im}a_j(E_e)>0$ and the corresponding term in the sum vanishes in the limit $\epsilon\rightarrow 0$. Hence
\begin{equation}
\label{032}
P_e\Pi_{E_e}P_e=\lim_{\epsilon\rightarrow 0_+}(\i\epsilon) \frac{Q_0(E_e-\i\epsilon)}{e-E_e+\i\epsilon+\d^2a_0(E_e-\i\epsilon)} =\frac{Q_0(E_e)}{1-\d^2a_0'(E_e)}.
\end{equation}
We have used the relation $e-E_e+\d^2a_0(E_e)=0$ (see after \eqref{09}). Therefore, the non-decaying, oscillating term on the right side in \eqref{018} is
\begin{equation}
\label{034}
\frac{{\rm e}^{\i tE_e}}{1-\d^2a_0'(E_e)} \scalprod{\phi}{Q_0(E_e)\psi} ={\rm e}^{\i tE_e} \scalprod{\phi}{P_e\Pi_{E_e}P_e\psi}.
\end{equation}
Finally, we combine \eqref{018}, \eqref{029} and \eqref{034} to arrive at
\begin{eqnarray}
\label{035}
\frac{1}{\i t}\frac{1}{2\pi\i}\int_{{\mathcal G}_e}{\rm e}^{\i tz}\scalprod{\phi}{\tfrac{d}{dz}{\mathfrak F}(z)^{-1}\psi} d z &=&{\rm e}^{\i tE_e} \scalprod{\phi}{P_e\Pi_{E_e}P_e \psi}+\sum_{j=1}^{m_e-1} {\rm e}^{\i t(e+\d^2a_j(e))}\scalprod{\phi}{Q_j(e)\psi}\nonumber\\
&& + O(1/t+{\rm e}^{w t}/t^2).
\end{eqnarray}
On the right side appears the $P_e$-block $P_e\Pi_{E_e}P_e$ of the projection $\Pi_{E_e}$. The contributions of the terms in \eqref{08} with ${\mathcal B}(z)$ and $R_{z}^{P_e}$ will add the remaining blocks to finally give the full expression $\scalprod{\phi}{\Pi_{E_e}\psi}$.
\subsubsection{The contribution to \eqref{08} from $\tfrac{d}{dz}{\mathcal B}(z)$} From \eqref{09'} and \eqref{033}, we have
\begin{equation}
\label{036}
\tfrac{d}{dz}\scalprod{\phi}{{\mathcal B}(z)\psi} = \d \ \sum_{j=0}^{m_e-1} \frac{1-\d^2a'_j}{(z-e-\d^2a_j)^2}\, q_j+ \frac{1}{e-z+\d^2a_j} \, q'_j,
\end{equation}
where
\begin{equation}
\label{037}
q_j(z) = - \scalprod{\phi}{[Q_jP_e I R_z^{P_e} + R_z^{P_e} I P_eQ_j -\d R_z^{P_e} I P_e Q_j P_e I R_z^{P_e} ]\psi}.
\end{equation}
The summand in \eqref{036} is of the same form as $T_j$ in \eqref{09}, with $Q_j$ replaced by $q_j$. We may thus repeat the analysis leading to \eqref{018} and \eqref{029}, giving
\begin{eqnarray}
\label{038}
\frac{1}{\i t}\frac{1}{2\pi\i} \int_{{\mathcal G}_e}{\rm e}^{\i tz}\tfrac{d}{dz}\scalprod{\phi}{{\mathcal B}(z)\psi}dz &=& {\rm e}^{\i tE_e}\frac{\d\, q_0(E_e)}{1-\d^2 a'_0(E_e)} +\d\,\sum_{j=1}^{m_e-1} {\rm e}^{\i t(e+\d^2a_j(e))}q_j(e) \\
&&\ \ + C(\phi,\psi) \cdot \,O\left(|\d|/t+|\d|{\rm e}^{wt}/t^2\right).\nonumber
\end{eqnarray}
Recalling \eqref{031}, \eqref{032} and using that $R_{E_e-\i\epsilon}^{P_e}$ stays bounded as $\epsilon\rightarrow 0_+$, we get
\begin{eqnarray}
\scalprod{\phi}{\Pi_{E_e}\psi} &=& \lim_{\epsilon\rightarrow 0_+} (\i\epsilon) \scalprod{\phi}{\left\{{\mathfrak F}(L-E_e+\i\epsilon;P_e)^{-1} +{\mathcal B}(E_e-\i\epsilon)\right\}\psi}\nonumber\\
&=& \scalprod{\phi}{P_e\Pi_{E_e}P_e\psi} + \frac{\d\,q_0(E_e)}{1-\d^2 a'_0(E_e)}.
\label{039}
\end{eqnarray}
\subsubsection{The contribution to \eqref{08} from $\tfrac{d}{dz}R_z^{P_e}$} By assumption (A2), \eqref{nlap}, we have $\scalprod{\phi}{\tfrac{d}{dz}R_z^{P_e}\psi}\prec 1$ and so we get
\begin{equation}
\label{040}
\frac{1}{\i t}\frac{1}{2\pi \i} \int_{{\mathcal G}_e} {\rm e}^{\i tz} \scalprod{\phi}{\tfrac{d}{dz}R_z^{P_e}\psi}d z \prec \frac{C(\phi,\psi)}{t}.
\end{equation}
Combining \eqref{08}, \eqref{035}, \eqref{038}, \eqref{039} and \eqref{040} gives \eqref{13.01}. This concludes the proof of Proposition \ref{propzero}.\hfill \qed
\subsection{Estimates for $z$ in a vicinity of an unstable eigenvalue $e$.}
\label{nestunstable}
The analysis is the same, actually somewhat easier, than the one presented in Section \ref{nestzero}. Indeed, for an unstable eigenvalue $e$, all the $\lambda_{e,j}$ have strictly positive imaginary part (see Assumption (A4), (1)). Therefore, we can proceed as in Section \ref{nestzero}, and in \eqref{09}, all the terms $T_j$ (even for $j=0$) are now treated as above, after \eqref{018}. We immediately obtain the following result.
\begin{prop}
\label{prope}
We have, for $\phi,\psi\in\mathcal D$,
\begin{equation}
\label{n13}
J_e(t) = \sum_{j=0}^{m_e-1} {\rm e}^{\i t(e+\d^2 a_{e,j}(e))}\scalprod{\phi}{(Q_j(e)+\d\, \widetilde Q_j)\psi} +R_e(t)
\end{equation}
where (recall that $w>0$ is the arbitrary parameter in \eqref{14})
$$
|R_e(t)|\le C \,\frac{1+{\rm e}^{wt}/t}{t},
$$
for a constant $C$ independent of $\d,t,w$ and
\begin{equation}
\label{n13.1}
a_{e,j}(e) = \lambda_{e,j} +O(|\d|),\qquad Q_j(e)=P_{e,j}+O(|\d|).
\end{equation}
Also $\widetilde{Q}_j$ is defined in \eqref{qtilde}.
\end{prop}
\subsection{Estimates for $z$ away from the eigenvalues $e$}
\label{inftysect}
On the unbounded set ${\mathcal G}_\infty$ we use the relation \eqref{i0}. This will help to ensure that the integrand is decaying sufficiently quickly at infinity. Setting $\widetilde\phi=(L-\i)\phi$ and $\widetilde\psi = (L+\i)\psi$,
\begin{eqnarray}
\int_{{\mathcal G}_\infty} {\rm e}^{\i tz} \scalprod{\phi}{\tfrac{d}{dz}R_z\psi} dz &=& \int_{{\mathcal G}_\infty} {\rm e}^{\i tz}\left\{ \frac{-2\scalprod{\widetilde\phi}{R_z\widetilde\psi}}{(z+\i)^3} +\frac{\scalprod{\widetilde\phi}{\tfrac{d}{dz}R_z\widetilde\psi}}{(z+\i)^2} \right\} dz \label{i4}\\
&&+\int_{{\mathcal G}_\infty} {\rm e}^{\i tz}\left\{ \frac{\scalprod{\phi}{\psi}}{(z+\i)^2} + \frac{2\scalprod{\phi}{\widetilde\psi}}{(z+\i)^3} \right\} dz. \label{i5}
\end{eqnarray}
The integral \eqref{i5} is $\prec \|\phi\|\,\|L\psi\|$. The terms on the right side of \eqref{i4} involving $R_z$ and $\tfrac{d}{dz}R_z$ are estimated using Assumption (A2)(2). Thus
\begin{equation}
\label{i6}
\frac{1}{\i t}\frac{1}{2\pi\i}\int_{{\mathcal G}_\infty} {\rm e}^{\i t z}\scalprod{\phi}{\tfrac{d}{dz} R_z\psi} dz \prec \frac{C(\widetilde\phi,\widetilde\psi)}{t}.
\end{equation}
We have proven the following result.
\begin{prop}
\label{propaway}
We have
\begin{equation}
\label{13'''}
J_\infty(t) \prec \frac{C((L-\i)\phi,(L+\i)\psi)}{t}.
\end{equation}
\end{prop}
\subsection{Proof of Theorem \ref{nthm1}}
We combine the estimates in Propositions \ref{propzero}, \ref{prope} and \ref{propaway} to obtain the expansion \eqref{nmainreseqn} with the bound
$$
|R(t)|\prec \frac{1+{\rm e}^{w t}/t}{t}
$$
for the remainder. Since $w>0$ is arbitrary, we have $|R(t)|\prec 1/t$. \hfill \qed
|
\section{Introduction}
\label{sec:intro}
Phase diagrams of materials describe the regions of stability and equilibria of structurally distinct phases and are fundamental in both materials science and industry. In order to augment experiments, computer simulations and theoretical calculations are often used to provide reference data and describe phase transitions.
A plethora of methods exist to determine individual phase boundaries, including Gibbs ensemble Monte Carlo~\cite{GEMC}, Gibbs--Duhem integration~\cite{G-D_int}, thermodynamic integration and even direct molecular dynamics simulations of coexistence.
Each of these algorithms requires the user to specify at least the identity and approximate location of the phase transition under investigation.
Moreover, in the case of the solid phases, where much of the interest lies, advance knowledge of the crystal structure of each phase is required.
Calculating an entire phase diagram by combining the results of such methods therefore demands a high degree of prior knowledge of the result.
This in turn poses a barrier to the discovery of unexpected phases and phase transitions.
Furthermore, such algorithms require specific expertise and separate setup for each type of phase transition.
In this paper we introduce a single algorithm, based on nested sampling (NS)~\cite{bib:skilling,bib:skilling2}, that enables the efficient calculation of {\em complete} pressure-temperature phase diagrams, including the solid region.
This algorithm requires no prior knowledge of the phase diagram, and takes only the potential energy function together with the desired pressure and temperature ranges as inputs.
Moreover, the direct output of the simulation is the partition function as an explicit function of its natural variables, so calculating thermodynamic observables, such as the heat capacity, is straightforward.
Nested sampling systematically explores the entire potential energy landscape, and in this way is related to parallel tempering (also known as replica exchange)~\cite{bib:partemp_swendsen,bib:partemp} and Wang-Landau sampling~\cite{bib:wang_landau}.
However, those algorithms encounter a particular convergence problem at first order phase transitions
because
the probability distributions (parametrised in terms of temperature in case of
parallel tempering or energy in case of Wang-Landau) on the two sides of the phase transition have very little overlap~\cite{AllenTildesley}.
This results
in poor
equilibration between the distributions on either side of the phase transition
and large errors (both random and systematic) in the predicted locations of phase transitions.
The NS algorithm was designed to solve this problem.
It constructs a sequence of decreasing potential energy levels, $\{E_i\}$, each of which bounds from above a volume of configuration space $\chi_i$, with the property that $\chi_i$ is approximately a constant factor smaller than the volume, $\chi_{i-1}$, corresponding to the level above. Each volume is sampled uniformly, and therefore each distribution will have an approximately constant fractional overlap with the one immediately before and after, ensuring fast convergence of the sampling and allowing an accurate evaluation of phase space integrals. In particular, the energy levels near the phase transition, where phase volumes change rapidly, will be very narrowly spaced. The sequence of energy levels comprise a discretisation of the cumulative density of states $\chi(E)$, which allows the evaluation of the partition function at arbitrary temperatures,
\begin{eqnarray}
Z(N,V,\beta) &=& \frac1{N!}\left(\frac{2\pi m}{\beta h^2}\right)^{3N/2}\int dE \,\chi'(E) e^{-\beta E} \label{eq:Z_NVT}\\
&\approx&Z_m(N,\beta)\sum_i (\chi_{i-1}-\chi_i) e^{-\beta E_i}\label{eq:Z_NS}
\end{eqnarray}
where $N$ is the number of particles of mass $m$, $V$ is the volume, $\beta$ is the inverse temperature, $h$ is Planck's constant, the density of
states $\chi'$ is the derivative of $\chi$, and we labelled the factor resulting from the momentum integral as $Z_m$. The total phase space volume is $\chi_0=V^N$ corresponding to the ideal gas limit. Note that the sequence of energies $\{E_i\}$ and configuration space volumes $\{\chi_i\}$ are independent of temperature, so the partition function can
be evaluated {\em a posteriori} at any temperature by changing $\beta$ in \eqref{eq:Z_NS}.
The basic NS algorithm is as follows. We initialise by generating a pool of $K$ uniformly random configurations and iterate the following loop starting at $i=1$.
\begin{enumerate}
\item Record the energy of the sample with the highest energy as $ E_{i} $, and use it as the new energy limit, $E_{\mathrm{limit}} \leftarrow E_i$. The corresponding phase space volume is $\chi_{i}\approx \chi_0 [K/(K+1)]^i$.
\item Remove the sample with energy $E_i$ from the pool and generate a new configuration uniformly random in the configuration space, subject to the constraint that its energy is less than $E_{\mathrm{limit}}$. One way to do this is to clone a randomly chosen existing configuration and make it undergo a random walk of $L$ steps, subject only to the energy limit constraint.
\item Let $i \gets i +1$, and return to step 1.
\end{enumerate}
At each iteration, the pool of $K$ samples are uniformly distributed in configuration space with energy $E<E_{\mathrm{limit}}$. The finite sample size leads to a statistical error in $\log \chi_i$, and also in the computed observables, that is asymptotically proportional to $1/\sqrt{K}$, so any desired accuracy can be achieved by increasing $K$. Note that for any given $K$, the sequence of energies and phase volumes converge exponentially fast (the number of
iterations required to obtain results shown below never exceeded $2000\cdot
K$), and increasing $K$ necessitates a new simulation from scratch.
Since its inception NS has been used successfully for Bayesian model selection in astrophysics~\cite{multinest1}, and also to investigate the potential energy landscapes of atomistic systems ranging from clusters to proteins~\cite{bib:our_NS_paper,bib:our_NSHS_paper,Burkoff1,bib:wheatley1,bib:wheatley2,bib:wheatley3, neilsen_npt, bib:Frenkel_NS,bib:diffns}.
The structure of this paper is as follows.
In section~\ref{sec:npt_ns} we modify the NS algorithm to enable its application at constant isotropic pressure with fully flexible periodic boundary conditions~\cite{martyna1994constant} where the periodic simulation cell is allowed to change shape.
In sections~\ref{sec:calc_phase_dia} and~\ref{sec:results} we show that this development enables the determination of pressure-temperature phase diagrams of materials directly from the potential energy function without recourse to any other{\em\ a priori} knowledge.
In particular, in section~\ref{sec:results} we calculate phase diagrams for aluminium and NiTi.
Finally in section~\ref{sec:conc_outlook} we conclude this paper, discussing some consequences of the capability to calculate entire phase diagrams with a single method and in a highly automated fashion.
\section{Nested sampling with fully flexible periodic boundary conditions at constant pressure}
\label{sec:npt_ns}
Nested sampling produces new samples by cloning an existing sample and then evolving the clone using a Markov chain Monte Carlo (MCMC) random walk~\cite{brooks2011handbook}. Although one could work in the NVT ensemble and use equations~\ref{eq:Z_NVT} and \ref{eq:Z_NS}, that would be very inefficient.
MCMC simulations performed at fixed pressure require just a fraction of the computational expense as equivalent calculations performed at fixed volume.
There are two reasons for this.
First, allowing the system to change volume by dilating or contracting expedites the cooperative freeing of jammed atoms.
In contrast, at fixed volume, atoms that have become jammed are only freed by the coincidental movement of all atoms to separate them.
Consequently MCMC simulations at fixed pressure explore configuration space far more rapidly than simulations at fixed volume.
The second reason arises from the thermodynamic behaviour of systems at a first order phase transition.
At a phase transition under constant volume conditions the two phases coexist and an interface forms between them.
Such interfaces are large on the atomic scale~\cite{frenkel2001understanding}
and
the behaviour of atoms at an interface is not representative of the behaviour of atoms in the equilibrium phases.
As a result the interface introduces a systematic error that is only overcome by simulating very large numbers of atoms.
Such interfaces also occur under constant pressure conditions in the infinite system size limit.
The contribution to the Gibbs Free Energy from an interface is proportional to $\gamma N^{\frac{2}{3}}$, where $\gamma$ is the interfacial tension.
In contrast, the Gibbs Free Energies of each of the pure phases are extensive (proportional to $N$).
Therefore the Gibbs Free Energy cost of the interface is negligible for thermodynamic systems.
Conversely, for the relatively small system sizes amenable to density of states calculation methods such as nested sampling, the Gibbs Free Energy cost of the interface is appreciable, provided $\gamma$ is not close to zero.
Consequently, at a constant pressure phase transition between phases with identical atomic compositions, configurations containing an interface have negligible statistical weight in such simulations, and a discontinuous transition is observed from one equilibrium phase to the other.
This enables the accurate simulation of phase transitions using much smaller numbers of atoms.
Using small numbers of atoms to simulate a phase transition naturally introduces new finite size errors.
In particular, for a fixed number of atoms, it is not possible to represent all crystal structures in a simulation cell of fixed shape.
This representational bias is removed by the use of fully flexible periodic boundary conditions~\cite{martyna1994constant}, which allow the simulation cell to deform smoothly and thus take any shape.
However, using fully flexible periodic boundary conditions allows the formation of very thin simulation cells containing unphysical quasi one and two dimensional configurations, characterised by interacting periodic images.
In subsection~\ref{subsec:mindepth_theory} we describe a rigorous solution to this new finite size problem.
Later, in subsection~\ref{subsec:part_fn_and_thermodynamics} we describe the calculation of the constant pressure partition function and heat capacity, both as explicit functions of temperature, using nested sampling.
\subsection{Constraint on the simulation cell to exclude unphysical quasi one and two dimensional configurations}\label{subsec:mindepth_theory}
The partition function at fixed isotropic pressure $p$ with fully flexible periodic boundary conditions~\cite{martyna1994constant} is
\begin{equation}
\begin{aligned}
\Delta(N,p,\beta) & = Z_m \beta p \int d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \times \\
& \int_{0}^{\infty} \!\!d V V^N \int_{\left(0,1\right)^{3N}}\! d\mathbf{s} \, e^{-\beta H\left( \mathbf{s},\mathbf{h}_{0},V,p \right)}.
\end{aligned}
\label{eq:partition_fn_NPT}
\end{equation}
Here $H\left( \mathbf{s},\mathbf{h}_{0},V,p \right) = E\left( \mathbf{s},\mathbf{h}_{0},V \right) + pV $, $\mathbf{h}$ is the $3\times3$ matrix of lattice vectors relating the Cartesian positions of the atoms $\mathbf{r}$ to the fractional coordinates $\mathbf{s}$ via $\mathbf{r}=\mathbf{hs}$, $ V=\det\mathbf{h}$ is the volume, and $\mathbf{h_{0}}=\mathbf{h}V^{-1/3}$ is the image of the unit cell normalised to unit volume.
The partition function~\eqref{eq:partition_fn_NPT} corresponds to integration over all nine elements of the matrix $\mathbf{h}_{0}$, and the $\delta$-function restricts the integration to matrices satisfying $ \det\mathbf{h}_{0} = 1$.
This partition function is formally correct in the thermodynamic limit~\cite{martyna1994constant,tuckerman2008statistical}.
However, finite systems in this description can adopt configurations for which the simulation cell becomes very thin.
In this case, periodic boundary conditions give rise to a quasi one or two dimensional system.
The prevalence of such configurations leads to a poor approximation of the three dimensional atomic system due to excessively large finite size effects.
We exclude such thin configurations by changing the limits for integration over elements of $ \mathbf{h}_{0}$, so that the perpendicular distances between opposite faces of the simulation cell $ \mathbf{h}_{0}$ are greater than some ``minimum cell depth'' value $d_0$.
The perpendicular distance between faces of the unit cell $\mathbf{h}$ made by lattice vectors $ \mathbf{h}^{\left( i \right)} $ and $ \mathbf{h}^{\left( j \right)} $ is given by
\begin{equation} \label{eq:d_perp}
d^{\perp}_{\mathbf{h}^{\left( k \right)}} = \frac{ \det \mathbf{h} }{| \mathbf{h}^{\left( i \right)}\times\mathbf{h}^{\left( j \right)} | }.
\end{equation}
The cell depth $D\!\left(\mathbf{h}_0\right)$, which measures how ``thin'' the cell has become, is defined as the minimum value of $d^{\perp}_{\mathbf{h}^{\left( k \right)}}$, for the cell at normalised (unit) volume $\mathbf{h}_0$.
\begin{equation} \label{eq:cell_depth}
D\!\left(\mathbf{h}_0\right) = \min_{i=1,2,3}\left(d^{\perp}_{\mathbf{h}_0^{\left( i \right)}} \right)
\end{equation}
Thus we integrate over elements of $ \mathbf{h}_{0}$ such that
\begin{equation} \label{eq:min_cell_depth_crit}
D\left(\mathbf{h}_0\right) > d_0.
\end{equation}
The minimum cell depth $d_0$ is a real number on the interval $[0,1]$ where $d_0=1$ restricts the simulation cell to a cube.
Smaller values of $d_0$ are accordingly less restrictive on the shape of the simulation cell, and $d_0=0$ corresponds to no restrictions on the simulation cell.
Incorporating this change of integration limits into the partition function~\eqref{eq:partition_fn_NPT} yields a new partition function
\begin{equation}
\begin{aligned}
\widetilde{\Delta}(N,p,\beta,d_0) & = Z_m \beta p \int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \times \\
& \int_{0}^{\infty} \!\!d V V^N \int_{\left(0,1\right)^{3N}}\! d\mathbf{s} \, e^{-\beta H\left( \mathbf{s},\mathbf{h}_{0},V,p \right) }.
\end{aligned}
\label{eq:partition_fn_NPT_w_mch}
\end{equation}
In the thermodynamic limit~\eqref{eq:partition_fn_NPT_w_mch} is equal to~\eqref{eq:partition_fn_NPT} up to a factor which depends only on $d_0$.
The two partition functions are equal if and only if $d_0=0$.
In tests with 64 atoms we verified that the heat capacity curves were independent of $d_0$ at values of 0.65, 0.7 and 0.8, in Lennard-Jonesium and aluminium.
The window of independence from $d_0$ grows wider as the number of particles is increased.
For larger numbers of atoms, there are more ways to arrange those atoms into a given crystal structure, including in simulation cells that are closer to a cube.
Similarly, unphysical correlations are introduced when the \emph{absolute} number of atoms between faces of the cell becomes too small, and therefore larger simulations can tolerate ``thinner'' simulation cells $\mathbf{h}_{0}$.
The nickel-titanium calculations were performed with $d_0=0.7$.
\subsection{Partition function and thermodynamic variables}\label{subsec:part_fn_and_thermodynamics}
The partition function we seek to calculate is given in equation~\eqref{eq:partition_fn_NPT_w_mch}.
Above some sufficiently large volume $V_0$, we approximate the system as an ideal gas, neglecting interatomic interactions, which corresponds to the condition $E\left( \mathbf{s},\mathbf{h}_{0},V \right) \ll pV$.
In this approximation the volume integral in~\eqref{eq:partition_fn_NPT_w_mch} is the sum of two parts
\begin{equation}
\begin{aligned}
& \widetilde{\Delta} (N,p,\beta,d_0) \approx Z_m \beta p \Bigg[ \Delta_{\mathrm{NS}} (N,p,\beta,V_0,d_0) \\
& + \int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \int_{V_0}^{\infty} \!\!d V V^N \int_{\left(0,1\right)^{3N}} \! d\mathbf{s} \, e^{-\beta pV } \Bigg]
\end{aligned}
\label{eq:int_split}
\end{equation}
where
\begin{equation}
\begin{aligned}
\Delta_{\mathrm{NS}} (N,p,\beta,V_0,d_0) & = \int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \times \\
& \int_{0}^{V_0} \!\!d V V^N \int_{\left(0,1\right)^{3N}}\! d\mathbf{s} \, e^{-\beta\left[ E\left( \mathbf{s},\mathbf{h}_{0},V \right) + pV \right]}
\end{aligned}
\label{eq:pf_NS}
\end{equation}
We calculate $\Delta_{\mathrm{NS}}$ using nested sampling.
Calculations are performed at fixed pressure to generate a sequence of {\em enthalpies}, $H_i$, where $H=E\left( \mathbf{s},V,\mathbf{h}_{0}\right) + pV$.
The NS approximation for $\Delta_{\mathrm{NS}}$, is
\begin{equation}\label{eq:DD_NS}
\begin{split}
\Delta_{\mathrm{NS}} (N,p,\beta,V_0,d_0) \approx& \sum_{i=1}^{i_{\mathrm{max}}}{\left(\chi_{i-1}-\chi_{i}\right) e^{-\beta H_i} } \\
\approx& \sum_{i=1}^{i_{\mathrm{max}}}{\Delta\chi_i e^{-\beta H_i} }
\end{split}
\end{equation}
where $\chi_{i} \approx \chi_0\left(\frac{K}{K+1}\right)^i$, $\chi_0 = \frac{V_0^{N+1}}{N+1}$, and $\Delta\chi_i \approx \chi_{i-1}-\chi_{i}$.
We use single atom Monte Carlo (MC) moves in fractional coordinates with the amplitude updated every $\frac{K}{2}$ iterations to maintain a good acceptance rate.
Uniform sampling of lattice shape matrices $\mathbf{h}_{0}$ subject to equation~\eqref{eq:min_cell_depth_crit} was achieved by independent shearing and stretching moves which do not change the volume.
The ratios of atom, volume, shear and stretch moves were $N\!\!:\!\!10\!\!:\!\!1\!\!:\!\!1$.
Further details of the MC moves and parallelisation scheme are given in the Supplemental Material (SM)~\cite{SM}.
We show in appendix~\ref{app:ig_pf} that volumes greater than $V_0$ make a negligible contribution to the partition function~\eqref{eq:int_split}, provided $ k_{\mathrm{B}}T \ll pV_0 $.
In this case we have
\begin{equation} \label{eq:pf_notail}
\widetilde{\Delta}(N,p,\beta,d_0) \approx \frac { \beta p } {N!} \Biggl( \frac {2\pi m } {\beta h^2}\Biggr)^{3N/2} \Delta_{\mathrm{NS}}(N,p,\beta,V_0,d_0)
\end{equation}
where we have expanded $Z_m$.
One can always assert the condition $ k_{\mathrm{B}}T \ll pV_0 $, and in practice it is easy to find values of $V_0$ suitable for physically relevant conditions.
We found $V_0 = 10^7 N \: \mathrm{\AA^3}$ to be suitable for all conditions considered in this paper.
From~\eqref{eq:pf_notail} we obtain the expected enthalpy
\begin{equation} \label{eq:expected_H}
\begin{split}
\langle H \rangle =& -\frac {\partial \log \widetilde{\Delta}( N , p ,\beta,d_0)} {\partial \beta} \\
=& \left( \frac{3 N }{2} - 1 \right)\frac{1}{\beta} + \langle H_{\mathrm{configurations}} \rangle
\end{split}
\end{equation}
and the heat capacity at constant pressure
\begin{align}
C_p =& - k_{\mathrm{B}}\beta^2 \frac {\partial \langle H \rangle } {\partial \beta} \label{eq:cp_line1} \\
=& \left(\frac{3 N }{2}-1\right)k_{\mathrm{B}} \label{eq:cp_line2} \\
& + { k_{\mathrm{B}}\beta^2} \left({\langle H^2_{\mathrm{configurations}} \rangle - { \langle H_{\mathrm{configurations}} \rangle}^2 }\right) \nonumber
\end{align}
where
\begin{equation} \label{eq::moments_tail_cont_zero}
\begin{split}
\langle H_{\mathrm{configurations}}\rangle &\approx \frac{ \sum_{i=1}^{i_{\mathrm{max}}}{ \Delta\chi_i \: H_i \: e^{-\beta H_i}} }{ \sum_{i=1}^{i_{\mathrm{max}}}{ \Delta\chi_i \: e^{-\beta H_i}} }, \\
\langle H^2_{\mathrm{configurations}}\rangle &\approx \frac{ \sum_{i=1}^{i_{\mathrm{max}}}{ \Delta\chi_i \: H^2_i \: e^{-\beta H_i}} }{ \sum_{i=1}^{i_{\mathrm{max}}}{ \Delta\chi_i \: e^{-\beta H_i}} }.
\end{split}
\end{equation}
This form~\eqref{eq::moments_tail_cont_zero} naturally does not depend on the contribution made by the low density configurations omitted from the NS calculation, or explicitly on the value of $d_0$.
We used equations~\eqref{eq::moments_tail_cont_zero} when calculating the heat capacities presented in this paper.
\section{Calculating phase diagrams}
\label{sec:calc_phase_dia}
In this section we describe a method for calculating the phase diagram of a material from the output of nested sampling.
We then benchmark the performance of nested sampling on the periodic Lennard-Jones system, and find nested sampling to be orders of magnitude more efficient than Parallel Tempering (PT) for resolving the melting and evaporation transitions.
Given the partition function~\eqref{eq:pf_notail}, phase transitions can be easily located by finding the peaks of response functions such as the heat capacity~\eqref{eq:cp_line2}.
By performing separate NS simulations at a number of pressures and combining the pressure and temperature values corresponding to the heat capacity peaks one can straightforwardly construct the \emph{entire} phase diagram including all thermodynamically stable phases.
This process is illustrated in Figure~\ref{fig:lj_phase_dia_from_cps}.
\begin{figure
\begin{center}
\includegraphics[width=8.5cm]{LJ_phased_3D.pdf}
\end{center}
\vspace{-15pt}
\caption {Demonstration of how NS can be used to calculate phase diagrams, using the case of the periodic Lennard-Jones model.
NS calculations are performed at a series of pressures and phase transitions are located by peaks of the heat capacity curves (blue).
The red lines show values from the literature for the melting (solid)~\cite{McNeilWatson2006}, boiling (dashed)~\cite{Kofke1993} and sublimation (dotted)~\cite{AgrawalKofke1995} curves.}
\label{fig:lj_phase_dia_from_cps}
\end{figure}
In Figure~\ref{fig:lj_nspt} we compare the performance of NS to that of PT for calculating the melting and evaporation transitions.
NS provides a reasonable estimate of the melting and boiling points using only $\sim 10^{8}$ energy evaluations, while parallel tempering needs many orders of magnitude more computational effort than NS to find the evaporation transition and almost two orders of magnitude more computational effort to find the melting transition. (A similar increase in computational efficiency compared with parallel tempering was found for LJ clusters~\cite{bib:our_NS_paper} and hard spheres~\cite{bib:our_NSHS_paper,bib:HS_PT}.)
\begin{figure
\begin{center}
\includegraphics[width=8.5cm]{NS_vs_PT_effort_rot90.pdf}
\end{center}
\vspace{-15pt}
\caption {Performance comparison of NS and PT.
Sixty-four Lennard-Jones particles were simulated at a pressure of $0.027$ (Lennard-Jones units).
Both NS and PT simulations were initialised from the vapour phase.
PT was performed using 128 equispaced temperature values in the range $[0.4,1.4]$.
The left panel shows the estimated transition temperatures as a function of computational cost while the right panel shows the mean enthalpy as a function of temperature corresponding to three selected values of the cost.
}
\label{fig:lj_nspt}
\end{figure}
Finally, in Figure~\ref{fig:lj_pd} we show the phase diagram for 64 particles of Lennard Jonesium as calculated using NS with $K=640$, $L=1.6\times10^5$.
Comparison with the literature phase diagrams for $\sim 500$ particles confirms excellent agreement with the literature values for the evaporation transition~\cite{Kofke1993} and also the solid-liquid and high pressure solid-vapour transitions~\cite{McNeilWatson2006}.
Below the triple point, we observe slower convergence with respect to $L$ towards literature values of the sublimation transition~\cite{AgrawalKofke1995}.
We also find the beginning of the Widom-line: the shallow line of heat capacity maxima that extends into the supercritical region.
The Widom-line and our method for estimating the critical point are described in the SM~\cite{SM}.
\begin{figure
\begin{center}
\includegraphics[width=8.5cm]{lj_pd.pdf}
\end{center}
\vspace{-15pt}
\caption {Phase diagram for $N=64$ Lennard-Jones particles as calculated using NS, with comparison to the literature ($N\approx500$) phase diagram, as described in the text.
}
\label{fig:lj_pd}
\end{figure}
\section{Results}\label{sec:results}
\subsection{Aluminium}\label{subsec:al}
\begin{figure*}[ht]
\includegraphics[width=17.5cm]{Al_phase_diagram_all4_2x2.pdf}
\vspace{-40pt}
\caption {Phase diagrams corresponding to four
EAM models of aluminium. Red symbols show the NS results, the error bars are
calculated as the width at half maximum of the peaks on the heat capacity
curves. On the boiling line points are connected by a solid line up to the critical point.
(The method we used to estimate the critical point is described in the SM~\cite{SM}.)
Black symbols show experimental melting points measured with Bridgman cells~\cite{Errandonea:2010ct}, with Diamond anvil cells
(DAC (a)~\cite{Boehler:1997vv} and (b)~\cite{Hanstrom:2000tr}) and shock waves (SW)~\cite{bib:Shaner_SWAl}. Different square symbols show estimates
of the critical point from experiments, (a)~\cite{Al_Tc}, (b)~\cite{CT_Fortov}, (c)~\cite{CT_Alder} and (d)~\cite{CT_Renaudin}. For NPB-EAM and MD-EAM large black squares
show the critical point and smaller black squares show the evaporation temperatures, all calculated using Gibbs ensemble Monte Carlo~\cite{Bhatt:2006fw}. At pressures below the critical point, NS parameters $K=800$ and $L=3000$ were used (the total number of energy evaluations was $3\times10^{9}$ for each pressure), while runs at pressures where solid-solid transitions are present required $K=3200$ and $L=15000$ (total number of energy evaluations were $4 \times 10^{10}$).
}
\label{fig:Al_phase_diagram}
\end{figure*}
In this section we apply the new algorithm to several empirical models of aluminium in order to demonstrate the capability of nested sampling to find solid-solid phase transitions without any prior knowledge of the crystal structures or even the existence of multiple stable phases. Furthermore, although the particular off-the-shelf models we use here do not reproduce the experimentally determined phase diagram of the material everywhere, the fact that nested sampling allows a direct calculation of the entire phase diagram means that in the future one could automate the optimisation of potentials to match the experimental phase diagram.
As one of the most commonly used metals, the thermodynamic properties of
aluminium have been extensively studied. The melting line of aluminium has
been measured up to 125 GPa~\cite{Errandonea:2010ct,Boehler:1997vv,Hanstrom:2000tr,bib:Shaner_SWAl},
with good agreement between the different experimental techniques. Theoretical
calculations have also been performed using embedded-atom type potentials~\cite{Foiles_EAM,Voter_EAM,OhJoh_EAM,Mei:1992uc,Morris:1994uz,Ercol_EAM,Liu:2004js,Mishin_EAM}
and \emph{ab initio} methods~\cite{deWijs:1998vm,Vocadlo:2002ih,Alfe:2004iq},
the latter providing melting temperatures up to 350 GPa~\cite{Bouchet:2009fo}.
At ambient conditions aluminium crystallises in the face-centred-cubic (fcc) structure, but a phase transition to the hexagonal-close-packed (hcp) structure at 217 GPa has been revealed by X-ray diffraction experiments~\cite{Akahama:2006iz} and the body-centred-cubic (bcc) phase has been also produced in laser-induced microexplosions~\cite{Vailionis:2011bs}.
The critical points of most metals are not amenable to conventional experimental study and thus estimation of their properties is usually based upon empirical relationships between the critical temperature and other measured thermodynamic properties.
In the case of aluminium these result in predictions in a wide temperature and pressure range~\cite{CT_Renaudin,Al_Tc,CT_Fortov,CT_Alder}.
We chose four widely used models all based on
the embedded-atom method (EAM): (1) the model developed by Liu {\em et al.}~\cite{Liu:2004js} (LEA-EAM), which is an improved version of the original potential of Ercolessi and Adams~\cite{Ercol_EAM}, (2) the model developed by Mishin {\em et al.}~\cite{Mishin_EAM} using both experimental and \emph{ab initio} data (Mishin-EAM), (3) the EAM of Mei and Davenport~\cite{Mei:1992uc} (MD-EAM) and (4) the recently modified version of the MD-EAM, reparametrised by Jasper {\em et al.} to accurately reproduce the DFT energies for Al clusters and nanoparticles of various sizes (NPB-EAM)~\cite{NPB_EAM}.
The phase diagrams for all four models based on NS simulations with 64 particles are shown in Figure~\ref{fig:Al_phase_diagram}.
The resulting critical parameters vary over a wide range for the different models.
Above the critical point we observe the Widom-line, indicated by those points not linked by a solid line. Heat capacity maxima corresponding to the Widom-line become broader away from the critical point, as indicated by the larger error bars. The Widom-line and our method for estimating the critical point are described in the SM~\cite{SM}.
The melting lines are in a good agreement with the available experimental data up to the pressure value $p \approx 25$ GPa. Above that pressure, the melting curves diverge from the experimental results, except for the MD-EAM potential, which reproduces the melting curve remarkably well.
At higher pressures small peaks appear on the heat capacity curves below the melting temperature for all models indicating solid-solid phase transitions (see appendix~\ref{app:solid-solid_pt}).
We post-processed the samples from the NS simulations.
As expected, the fcc structure was found to be stable at low pressures in all four models.
However, the models differ markedly in their predictions at high pressures.
The only commonality between the predicted high pressure solid phase diagrams is that the maximum predicted stable pressure for the fcc structure is far too low, both in comparison with experiment and density functional theory~\cite{Akahama:2006iz,Boettger_hcp,Sinko_hcp}.
\subsection{NiTi}\label{subsec:niti}
Finally, in order to demonstrate that NS is applicable to more complex problems, we show results for a material of current scientific interest, the NiTi shape memory alloy~\cite{NiTi_Ji, NiTi_discover}.
The shape memory effect relies on the structural phase transition from the high temperature austenitic phase to the low temperature martensitic phase~\cite{NiTi_transition}.
Studying this transition is particularly challenging with traditional free energy methods because the austenitic phase does not correspond to a local minimum of the potential energy surface.
Figure~\ref{fig:NiTi} shows the pressure-temperature-composition phase diagram corresponding to a recent EAM model~\cite{NiTi_EAM, NiTi_strct_trans} as computed with NS.
The phase transition temperatures are within 50 K of the experimental values and reproduce the trend with compositional change.
We predict a decreasing transition temperature with increasing pressure.
It is notable that this potential successfully reproduces the
martensitic transition temperature, despite the fact that the minimum enthalpy structure for the potential is different to the structure observed both experimentally and in DFT: here the lowest enthalpy structure (which we label B19X) is orthorhombic (see the SM~\cite{SM} for a description of the low enthalpy structures we identified).
Thus it appears that the austenite-martensite transition temperature is not sensitive to the detailed geometry and ordering of the lowest enthalpy structures.
Such empirical potentials can therefore be useful tools for studying this transition in the future.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=8.5cm]{NiTi_PD_small.pdf}
\end{center}
\vspace{-10pt}
\caption {NiTi martensitic phase transition as a function of Ni content (at 0.66 GPa) and pressure (at 1:1 composition). The simulation cell contained 64 atoms in the cases of the 50\% and 51.6\% Ni compositions and 108 atoms in the case of 50.8\% Ni content. NS parameters were $K=1920$, $L=10^5$, each data point used $10^{10} $ energy evaluations, and Ni--Ti swap moves were also included in the MC. Experimental results are taken from~\cite{NiTi_exp}.}
\label{fig:NiTi}
\end{figure}
\section{Conclusion and outlook}\label{sec:conc_outlook}
In summary, we have extended the nested sampling algorithm to allow simulations
using fully flexible periodic boundary conditions at fixed pressure
and demonstrated how it can be used to determine pressure-temperature-composition phase diagrams. In contrast to existing methods for comparing specific phases, NS explores the entire configuration space without requiring any prior knowledge about the structures of different solid phases with the only necessary input
being the composition and the desired pressure and temperature ranges. This makes it the method of choice for exploring the pressure-temperature-composition space, which is the next unexplored realm naturally following much recent work in crystal structure exploration at zero temperature. Since the algorithm is run independently for different pressures and compositions, and also has excellent parallel scaling up to a number of processors equal to the number of simultaneous samples, it might even be possible to run it on ab initio models on exascale computers. Furthermore, we suggest NS is eminently suitable for validating materials models, and in the future could even play a role in the automatic optimisation of empirical models.
\begin{acknowledgements}
RJNB acknowledges support from the EPSRC. LBP acknowledges support from St. Catharine's College, Cambridge and to the Royal Society. APB acknowledges support from Magdalene College, Cambridge, the Leverhulme Trust and the Isaac Newton Trust. GC acknowledges EPSRC grant EP/J010847/1. Computer time was provided via the UKCP consortium funded by EPSRC grant ref. EP/K013564/1 and via the Darwin supercomputer in the University of Cambridge High Performance Computing Service funded under EPSRC grant EP/J017639/1.
Data from this publication can be found at \texttt{https://www.repository.cam.ac.uk/handle/\\1810/255091}.
\end{acknowledgements}
\begin{appendices}
\section{Ideal gas contribution to the partition function}\label{app:ig_pf}
In this appendix we show that the ideal gas contribution to the partition function~\eqref{eq:int_split} asymptotically approaches zero for any positive minimum cell depth $d_0$, in the limit $ k_{\mathrm{B}}T / p V_0 \rightarrow 0 $.
The ideal gas contribution to the partition function~\eqref{eq:int_split} is
\begin{equation}
\int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \int_{V_0}^{\infty} \!\!d V V^N \int_{\left(0,1\right)^{3N}} \! d\mathbf{s} \, e^{-\beta pV }
\end{equation}
We begin by noting that the exponential term does not depend on $E(\mathbf{s},\mathbf{h}_0,V)$, and therefore $\int_{(0,1)^{3N}} d \mathbf{s} = 1$.
Thus we have
\begin{equation}
\begin{aligned}
\int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \int_{V_0}^{\infty} d V V^N \int_{(0,1)^{3N}} \mathrm {d} \mathbf{s} e^{-\beta pV} \\
= \int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right) \int_{V_0}^{\infty} d V V^N e^{-\beta pV} .
\end{aligned}
\end{equation}
The integral over volume $V$ evaluates to
\begin{equation}
\int_{V_0}^{\infty} d V V^N e^{-\beta pV} = \frac {1} {(\beta p)^{N+1}} \Gamma(N+1,\beta p V_0)
\end{equation}
where $\Gamma(N+1,\beta p V_0)$ is the upper incomplete gamma function.
Finally
we
define the function $A\left(d_0\right)$ to be equal to the integral over $\mathbf{h}_{0}$
\begin{equation}
A\left(d_0\right) = \int_{D\left(\mathbf{h}_0\right) > d_0} \!\!\!\!\!\!\!\!\! d\mathbf{h}_{0} \delta\left(\det{\mathbf{h}_{0}}-1\right).
\end{equation}
The function $A\left(d_0\right)$ is finite for any positive value of $d_0$, $A\left( 1 \right)=0$ and $A\left(d_0\right)$ diverges in the limit $d_0 \rightarrow 0$.
In the orthorhombic case, where all angles of the simulation cell are equal to $\frac{\pi}{2}$,
$A\left(d_0\right) = \frac{9}{2}\left( \logd_0 \right)^2$, with $A=1$ at $d_0 \approx 0.62$.
However at any positive value of $d_0$ the contribution to the partition function~\eqref{eq:int_split} due to volumes greater than $V_0$ goes to zero in the limit $k_{\mathrm{B}}T / p V_0 \rightarrow 0$ because $\Gamma(N+1,\beta p V_0) \rightarrow 0$ in the same limit.
\section{Identifying solid-solid phase transitions}\label{app:solid-solid_pt}
The locations of phase transitions are determined solely by looking at the peaks in the heat capacity.
Next, we inspect the system at temperatures either side of the phase transition.
Specific phases can be identified in the following way.
If no appropriate order parameter is to hand, then
one picks a number of random configurations from the output of nested sampling, chosen according to their thermal weights $ \Delta\chi_i e^{-\beta H_i} $, and inspects them by eye.
If an appropriate order parameter {\em is} known, one can compute the free energy landscape for that order parameter.
Here one proceeds by binning the weights $ \Delta\chi_i e^{-\beta H_i} $ of all configurations, according to the order parameter, to create a partial sum $\Delta_j=\sum{\Delta\chi_i e^{-\beta H_i}}$ for each bin $j$.
The free energy for each bin can then be computed as $F_j = -\frac{1}{\beta}\left[\log(\Delta_j) + \log\left(\frac{\beta p}{N!}\right) + \frac{3N}{2}\log\left( \frac{2\pi m}{\beta h^2} \right) \right]$.
In fact, simply calculating the expected enthalpy at the phase transition, and then examining the order parameter values for output configurations around that enthalpy is often sufficient to identify the crystal structures.
An example of the latter approach is shown in Figure~\ref{fig:Al_NSQ6} for the Mishin-EAM potential, which compares the enthalpies and $Q_6$ bond order parameter values for nested sampling output configurations at three different pressures.
At $p=25.0$ GPa no phase transition occurs, and only fcc configurations are present.
At $p=34.9$ GPa a first order phase transition occurs at the average enthalpy marked by the vertical dashed line.
At that enthalpy there is a clear transition between two basins, from a first basin
that corresponds to the bcc structure, to a second
that corresponds to the hcp structure.
Finally, at $p=37.5$ GPa no phase transition occurs and so there is no peak in the heat capacity.
At this pressure the bcc structure is stable at all temperatures below the melting point.
Nevertheless, the hcp structure is clearly visible as a metastable structure.
\begin{figure}[h]
\begin{center}
\includegraphics[width=9cm]{NS_Q6_pressures.pdf}
\end{center}
\caption {Identification of solid phases from inspection of an order parameter: Mishin-EAM potential for aluminium.
These plots compare the enthalpies and $Q_6$ bond order parameter values for nested sampling output configurations at three different pressures.
Nested sampling proceeds towards lower enthalpies, thus from right to left in each plot.
Horizontal dotted lines show the bond order parameters for the perfect fcc, bcc and hcp phases, and the vertical dashed line shows the expected enthalpy at the solid-solid phase transition, which was located by inspecting the heat capacity and observing a peak.
These results are discussed in the text.
}
\label{fig:Al_NSQ6}
\end{figure}
\restoreapp
\end{appendices}
|
\section{Introduction }
Classical Morse theory \cite{Mat02,Mi62} analyzes the topology of
the Riemannian manifolds by studying critical points of smooth functions
defined on it. Morse theory has several applications and extensions.
It is an elegant extension of maximum and minimum principles for smooth
functions on compact manifolds. It can be considered to be a part
of differential topology wherein it gives a way to analyze the topology
of the Riemannian manifolds by studying critical points of smooth
functions defined on it. Marston Morse established this subject while
using critical point theory to study closed geodesics on smooth Riemannian
manifolds. Its place in field of geometry and topology got firmly
established through subsequent applications and extensions by Smale,
Bott, Thom, Novikov, Witten, Goresky and Macpherson, Arnold and Floer
among others. For instance, it was applied by Smale to resolve the
Poincar\'{e}\xspace conjecture in the higher dimensions and also to mathematically
formulate the Pareto optimality problem in economics \cite{Mat02,Mi62,Sm75}.
In the 90s, Robin Forman formulated a completely combinatorial analogue
of Morse theory, now known as discrete Morse theory. Since Lewiner's
doctoral work \cite{Le02}, discrete Morse theory has rapidly become
a popular tool in computational topology and visualization communities
\cite{Gh14,GNPH07,RWS11}. Forman provides an extremely readable and
a compelling introduction to discrete Morse theory in \cite{Fo02a} .
Please refer to \autoref{fig:HasseGraph} and \autoref{tab:dictionary}
in \autoref{sub:Graph-Theoretic-Reformulation}, for a quick overview
of the graph theory setting of discrete Morse theory so as to facilitate
a quick foray into the core computer science problem at hand.
\section{Background and Preliminaries}
\subsection{Approximation Algorithms \label{sub:Approximation-Algorithms}}
\cite{Va03} and\cite{WS11} are standard texts for approximation
algorithms.
\begin{definition}[Approximation Algorithm] Given a Problem $\varPi$, an $\alpha$-Approximation Algorithm $\mathscr{A}_\varPi$ is a polynomial time algorithm that produces a solution whose value $\mathscr{V}_\varPi$ is within a factor $\alpha$ of the value of the optimal solution $\mathscr{O}_\varPi$, for all instances $\varPi_\mathscr{I}$ of the problem.
\end{definition}
For a minimization problem, $\nicefrac{\mathscr{V}_{\varPi}}{\mathscr{O}_{\varPi}}\leq\alpha$
for some $\alpha>1$.
\begin{definition}[Approximation factor preserving reduction] Given two minimization problems $\varPi_1$ and $\varPi_2$ (with optimal solutions of values $\mathscr{O}_{\varPi_1}$ and $\mathscr{O}_{\varPi_2}$ respectively), an \textbf{approximation factor preserving reduction} from $\varPi_1$ to $\varPi_2$ consists of two polynomial time algorithms $f(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ and $g(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ s.t.,
\begin{enumerate}
\item For any instance $\mathscr{I}_1$ of $\varPi_1$, $\mathscr{I}_2=f(\mathscr{I}_1)$ is an instance of $\varPi_2$ s.t.
\begin{equation}\label{eq:apcond1}
\mathscr{O}_{\varPi_2}(\mathscr{I}_2) \leq \mathscr{O}_{\varPi_1}(\mathscr{I}_1)
\end{equation}
\item Also, for any solution $t$ of $\mathscr{I}_2$, $s=g(\mathscr{I}_1,t)$ is a solution $\mathscr{I}_1$ s.t.
\begin{equation} \label{eq:apcond2}
\mathscr{V}_{\varPi_1}(\mathscr{I}_1,s)\leq \mathscr{V}_{\varPi_2}(\mathscr{I}_2,t)
\end{equation}
\end{enumerate}
\end{definition}
It can be easily shown that such a reduction combined with an $\alpha$-approximation
algorithm for $\varPi_{2}$ gives an $\alpha$-approximation algorithm
for $\varPi_{1}.$ See \cite{Va03} pg.348.
\subsection{Discrete Morse theory \label{sub:Discrete-Morse-Theory}}
Discrete Morse theory has also found several applications in algorithmic
combinatorics, algebra and geometry, such as the study of evasiveness
of graph properties \cite{Fo00}, topological properties of graph
complexes \cite{Jo08}, minimal resolutions and Gröbner bases \cite{JW09},
poset topology \cite{Wa07}, homology of braid groups \cite{Fa06}
and tropical geometry analogues of Lefschetz section theorem \cite{AB14}.
We see our work to be thematically related to each of these works,
in the sense that, they explore discrete Morse theory's intersections
with algebra, geometry, combinatorics and algorithmics. Forman's discrete
Morse theory has been generalized in several directions including:
Morse theory for vector Fields \cite{Fo98c}, Witten-Morse theory
\cite{Fo98b}, Novikov-Morse theory \cite{Fo02b}, $L^{2}$-Morse
homology \cite{MY99} and equivariant Morse theory \cite{Fre07} and
tame flows \cite{Ni10}.
Forman's theory is known to hold for regular cell complexes \cite{Fo98a}
\footnote{Simplicial and cubical complexes are special families of regular cell complexes. }
\subsubsection{Elementary Discrete Morse Theory}
\begin{definition}[Simplex, Face, Coface, Simplicial Complex] \label{def:simp} A \textbf{simplex} $\sigma^{n}$\footnote{An n-dimensional simplex $\sigma^{n}$ may be denote either as $\sigma$ or $\sigma^{n}$ depending on whether or not we wish to emphasize its dimension. Moreover, \emph{vertices} of a complex are the same as its $0$-dimensional simplices.} is the convex hull of $n+1$ distinct points $x_0,x_1,\dots,x_n$ in $\mathbb{R}^{m}$ \, $n\leq m$. A \textbf{face} of a simplex is the convex hull of some subset of set of its vertices. If $\tau$ is a face of $\sigma$ then $\sigma$ is a \textbf{coface} of $\tau$. A finite collection of simplices in $\mathbb{R}^{m}$ is called a \textbf{simplicial complex} if any two of its simplices either have no common points or intersect along their common face.
\end{definition}
\begin{definition} [Boundary \& Coboundary of a simplex $\sigma$] Let $\mathcal{K}$ be a simplicial complex and let $\sigma,\tau$ be simplices of $\mathcal{K}$. The relation \textbf{'$\bm{\prec}$'} is defined as: $\tau \prec \sigma \Leftrightarrow \{\tau \subset \sigma \,\, \& \, \,\dim\tau = \dim\sigma - 1 \}$. The boundary $\eth\:\sigma$ and respectively coboundary $\delta\:\sigma$ of a simplex are simply defined as\\ $\delta \: \sigma \,= \: \{ \tau \, | \, \tau \prec \sigma \}$ $\eth \: \sigma = \{ \rho \, | \, \sigma \prec \rho \}$ \end{definition}
\begin{definition}[Hasse graph] The \textbf{Hasse graph} of a complex $\mathcal{K}$ is graph whose vertices are in 1-1 correspondence with the simplices of the complex and there is an edge between every node that represents simplex $\beta^{d}$ and a node that represent simplex $\alpha^{d-1}$ iff $\alpha\prec\beta$. Also, by the term \textbf{d-(d-1) level} of Hasse graph $\mathcal{H}$, we mean the subset of edges of the Hasse graph that join the d-dimensional cofaces to the (d-1) dimensional faces of Hasse graph. Please see \autoref{fig:HasseGraph}.
\end{definition}
\begin{definition}[Discrete Morse function] Let $\mathcal{K}$ denote a finite regular simplex complex and let $\mathcal{L}$ denote the set of simplices of $\mathcal{K}$. A function $\mathcal{F}:\mathcal{L}\rightarrow\mathbb{R}$ is called a \textbf{discrete Morse function} (DMF) if it usually assigns higher values to higher dimensional simplices, with at most one exception locally at each simplex. Equivalently, a function $\mathcal{F}:\mathcal{L}\rightarrow\mathbb{R}$ is a discrete Morse function if for every $\sigma \in \mathcal{L} $ we have:\\
(\textsf{A}.)\; $\mathcal{N}_1(\sigma)= \# \{\rho \in cbd \: \sigma | \mathcal{F}(\rho) \leq \mathcal{F}(\sigma) \} \leq 1 $\\
(\textsf{B}.)\; $\mathcal{N}_2(\sigma)= \# \{\tau \in bd \: \sigma \:\: | \mathcal{F}(\tau) \geq \mathcal{F}(\sigma) \} \leq 1 $.
\end{definition}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\includegraphics[scale=0.6]{figure1.png}%
\end{minipage}}
I\end{figure}
\begin{definition}[Critical/Regular simplices]
If $\mathcal{N}_1(\sigma) = \mathcal{N}_2(\sigma)= 0$ for a simplex $\sigma$ then it is \textbf{critical}, else it is \textbf{regular}.
\end{definition}
\begin{definition} [Combinatorial Vector Field] A \textbf{combinatorial vector field} (DVF) $\mathcal{V}$ on $\mathcal{L}$ is a collection of pairs of simplices $\{\langle\alpha,\beta\rangle\}$ such that $\{\alpha^{m}\prec\beta^{(m+1)}\}$ and each simplex occurs in at most one such pair of $\mathcal{V}$.
\end{definition}
\begin{definition}[Gradient Vector Field] A pair of simplices $\{\alpha^{m}\prec\beta^{(m+1)}\}$ s.t. $\mathcal{F}(\alpha) \geq \mathcal{F}(\beta)$ determines a \textbf{gradient pair}. Each simplex must occur in at most one gradient pair of $\mathcal{V}$. A \textbf{discrete gradient vector field} (DGVF) $\mathcal{V}$ corresponding to a \textsf{DMF} $\mathcal{F}$ is a collection of simplicial pairs $\{\alpha^{(p)}\prec\beta^{(p+1)} \}$ s.t. $\{\alpha^{(p)}\prec\beta^{(p+1)}\}\in \mathcal{V}$ iff $\mathcal{F}(\beta)\leq \mathcal{F}(\alpha)$.
\end{definition}
\begin{definition}[Gradient Path]
A $\bm{\mathcal{V}}$\textbf{-path} is a simplicial sequence $\{\sigma_{0}^{(m)},$ $\tau_{0}^{(m+1)},$ $\sigma_{1}^{(m)},$ $\tau_{1}^{(m+1)},$ $\dots\sigma_{q}^{(m)},$ $\tau_{q}^{(m+1)}, $ $\sigma_{q+1}^{(m)}\}$ s.t. for $i=0,\dots q,\{\sigma_{i}\prec\tau_{i}\}\in \mathcal{V}$, $\sigma_{i}\prec\tau_{i}\succ\sigma_{i+1}$ and $\sigma_{i}\neq\sigma_{i+1}$. A $\mathcal{V}$-path corresponding to a DMF $\mathcal{F}$ is a \textbf{gradient path} of $\mathcal{F}$.
\end{definition}
\begin{theorem} \textbf{[Forman~\cite{Fo98a}]:} Let $\mathcal{K}$ be a CW complex with a \textsf{DMF} $\mathcal{F}$ defined on it. Then $\mathcal{K}$ is homotopy equivalent to a CW complex $\Omega$, such that $\Omega$ has precisely one m-dimensional simplex for every m-dimensional critical simplex in $\mathcal{K}$ and no other simplices besides these. Moreover, let $c_{m}$ be the number of m-dimensional critical simplices, $b_{m}$ the $m^{th}$ Betti number w.r.t. some vector field $\mathcal{V}$ and $n$ the maximum dimension of $\mathcal{K}$. Then we have:
\begin{description}[style=nextline, leftmargin=1cm]
\item[The Weak Morse inequalities]
Let $\chi$ be the Euler characteristic.
\begin{enumerate}[labelindent=*, leftmargin=*, labelsep=1em, label=(\Roman*.), align=left]
\item $c_{0}- c_{1} \dots + (-1)^{n}c_{n} = b_{0} - b_{1} \dots + (-1)^{n}b_{n} = \chi(\mathcal{K}) $
\item For every $m \in \mleft\{0\dots n\mright\}$: we have $c_{m}\geq b_{m}$
\end{enumerate}
\item[The Strong Morse Inequalities]
For every $m \in \mleft\{0 \dots n\mright\}$: $c_{m}- c_{m-1}\dots + (-1)^{m}c_{0} \geq b_{m} - b_{m-1}\dots + (-1)^{m}b_{0} $
\end{description}
\end{theorem}
\subsubsection{Morse Homology \label{sec:Morse-Homology}}
\begin{theorem}\label{thm:MorsecellNone} If $a<b$, are real numbers, such that [a,b] contains no critical values of Morse function $\mathcal{F}$, then the sublevel set $\mathcal{M}(b)$ is homotopy equivalent to the sublevel set $\mathcal{M}(a)$.
\end{theorem}
\begin{theorem}[Forman~\cite{Fo98a}]\label{thm:Morsecell} Suppose $\sigma^{p}$ is a critical cell of index p with $\mathcal{F}(\sigma)\in[a,b]$ and $\mathcal{F}^{-1}(a,b)$ contains no other critical points. Then $M(b)$ is homotopy equivalent to
\[M(a)\bigcup_{e_b^{p}} e^{p} \]
where $e^{p}$ denotes a p-dimensional cell with boundary $e_b^{p}$.
\end{theorem}
\begin{notation}[$\mathfrak{B}(\mathcal{K}),\mathfrak{M}(\mathcal{K})$] Given a simplicial complex $\mathcal{K}$, $\mathfrak{B}(\mathcal{K})=\sum\limits_{i=1}^{D} \beta_{i}$ and $\mathfrak{M}(\mathcal{K})=\sum\limits_{i=1}^{D} c_{i}$. \end{notation}.
Let $\mathcal{F}$ be a discrete Morse function defined on simplicial complex
$W$. Let $C_{q}(W,\mathbb{\, Z})$ denote the space of $q$-simplicial
chains, and $\mathcal{M}_{q}$ which is a subset of $C_{q}(W,\mathbb{\, Z})$
denote the span of the critical $q$-simplices. Let $\mathcal{M_{\star}}$
denote the space of Morse chains. Let $c_{q}$ denote the number of
critical $q$-simplices. Then we have, $\mathcal{M}_{q}\cong\mathbb{Z}^{c_{q}}$.
\begin{theorem}[Forman~\cite{Fo98a}]\label{thm:morsehom}
There exist boundary maps $\hat{\partial_{q}}\,:\,\mathcal{M}_{q}\rightarrow\mathcal{M}_{q-1}$, for each $q$, which satisfy $ \hat{\partial}_{q}\circ\hat{\partial}_{q+1}=0 $ and s.t. the resulting differential complex \begin{eqnarray*} 0\stackrel{}{\longrightarrow}\mathcal{M}_{n}\stackrel{\hat{\partial}_{n}}{\longrightarrow}\mathcal{M}_{n-1}\stackrel{\hat{\partial}_{n-1}}{\longrightarrow}\dots\stackrel{\hat{\partial}_{q+2}}{\longrightarrow}\mathcal{M}_{q+1}\stackrel{\hat{\partial}_{q+1}}{\longrightarrow}\mathcal{M}_{q}\stackrel{\hat{\partial}_{q}}{\longrightarrow}\dots\stackrel{\hat{\partial}_{2}}{\longrightarrow}\mathcal{M}_{1}\stackrel{\hat{\partial}_{1}}{\longrightarrow}\mathcal{M}_{0}\longrightarrow0 \end{eqnarray*} calculates the homology of $W$. i.e. if we go with the natural definition, \begin{eqnarray*} H_{q}(\mathcal{M},\hat{\partial})=\frac{\mathrm{ker}\hat{\partial}_{q}}{\mathrm{im}\hat{\partial}_{q+1}} \end{eqnarray*} \\
Then for each $q$, we have $H_{q}(\mathcal{M},\hat{\partial})=H_{q}(W,\mathbb{Z})$.
\end{theorem}
\begin{theorem} [Boundary Operator Computation Forman~\cite{Fo98a}]\label{thm:boundop} Consider an oriented simplicial complex. Then for any critical (p+1)-simplex $\beta$ set: \\ \\
$\partial \beta = \sum\limits_{critical \: \alpha(p)}\: P_{\alpha\beta}\:\alpha $ \\ \\
$ P_{\alpha \beta} = \sum\limits_{\gamma \in \Gamma(\beta,\alpha)} N(\gamma) $ \\ \\
where $\Gamma(\beta,\alpha)$ is the set of discrete gradient paths which go from a face in $\eth\,\beta$ to $\alpha$. The multiplicity $N(\gamma)$ of any gradient path $\gamma$ is equal to $\pm1$ depending on whether given $\gamma$ the orientation on $\beta$ induces the chosen orientation on $\alpha$ or the opposite orientation. With the boundary operator above, the complex computes the homology of complex K.
\end{theorem}
In \autoref{thm:Morsecell}, Forman's establishes the existence of
a cell complex (let us call it the\emph{ `Morse Smale complex`})
that is homotopy equivalent to the original complex. For proof details
please refer to Forman \cite{Fo98a}. The boundary operator in \autoref{thm:boundop}
for the chain complex construction (that is referred to as simply
the \emph{`Morse complex`}) tells us how to use the new CW complex
that is built in construction described in proof of \autoref{thm:Morsecell}.
Note that the Morse complex itself is a chain complex and not a CW
complex. But the chain complex construction allows us to express the
simplicial homology of the input complex in terms of the Morse homology
of the Morse Smale complex.
\begin{thm}[Critical simplex cancellation, Forman~\cite{Fo98a,Fo02a}] \label{thm:red1} Let $\mathcal{F}$ be a discrete Morse function on a simplicial complex $\mathcal{K}$ such that $\sigma^{(p+1)}$ and $\tau^{p}$ are critical. Let there be a unique gradient path from $\partial\sigma$ to $\tau$. Then there is another Morse function $\mathcal{G}$ on $\mathcal{K}$ with the same set of critical simplices except $\tau$ and $\sigma$. Also, the gradient vector field associated to $\mathcal{G}$ is equal to the gradient vector field associated to $\mathcal{F}$ except along the unique gradient path from $\partial\sigma$ to $\tau$. \end{thm}
\subsubsection{Graph Theoretic Reformulation\label{sub:Graph-Theoretic-Reformulation}}
Given a simplicial complex $\mathcal{K}$, we construct its Hasse Graph representation $\mathcal{H_K}$ (an undirected, multipartite graph) as follows: To every simplex $\sigma_{\mathcal{K}}^{d}\in\mathcal{K}$ associate a vertex $\sigma_{\mathcal{H}}^{d}\in\mathcal{H_K}$. The dimension d of the simplex $\sigma_{\mathcal{K}}^{d}$ determines the \emph{vertex level} of the vertex $\sigma_{\mathcal{H}}^{d}$ in $\mathcal{H_K}$. Every face incidence $(\tau_{\mathcal{K}}^{d-1},\sigma_{\mathcal{K}}^{d})$ determines an undirected edge $\langle\tau_{\mathcal{H}}^{d-1},\sigma_{\mathcal{H}}^{d}\rangle$ in $\mathcal{H_K}$. Now orient the graph $\mathcal{H_K}$ to a form a new directed graph $\overline{\mathcal{H_K}}$. Initally all edges of $\overline{\mathcal{H_K}}$ have default orientation. The default orientation is a directed edge $\sigma_{\mathcal{H}}^{d}\rightarrow\tau_{\mathcal{H}}^{d-1}\in\overline{\mathcal{H_K}}$ that connects a k-dim node $\sigma_{\mathcal{H}}^{d}$ to a (k-1)-dim node $\tau_{\mathcal{H}}^{d-1}$. Finally, associate a matching $\mathcal{M}$ to graph $\mathcal{H_K}$. If an edge $\langle\tau_{\mathcal{H}}^{d-1},\sigma_{\mathcal{H}}^{d}\rangle\in\mathcal{M}$ then, \emph{reverse} the orientation of that edge to $\tau_{\mathcal{H}}^{d-1}\rightarrow\sigma_{\mathcal{H}}^{d}\in\overline{\mathcal{H_K}}$\textbf{.} The matching induced reorientation needs to be such that the graph $\overline{\mathcal{H}_{\mathcal{K}}}$ is a Directed Acyclic Graph. A graph matching on $\mathcal{H_K}$ that leaves the graph $\overline{\mathcal{H_K}}$ acyclic in the manner prescribed above is known as \emph{Morse Matching}. \autoref{tab:dictionary} provides a translating dictionary from simplicial complexes to their Hasse graphs. See \autoref{fig:HasseGraph} and Figure~\ref{fig:HasseGraph2}.
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\includegraphics[scale=0.55]{figure2.png}
\captionof{figure} {Matching induced orientation of Hasse Graph}
\label{fig:HasseGraph}%
\end{minipage}}
\captionof{table}{Graph Theoretic dictionary for Morse Matching}
\centering{}\begin{center}
\begin{center} \resizebox{.75\textwidth}{!}{%
\begin{tabular}{|c||c|c|} \hline & \textsf{Morse theory on cell complex $\mathcal{K}$} & \textsf{Graph theory on Hasse Graph $\overline{\mathcal{H}_{\mathcal{K}}}$}\tabularnewline \hline \hline
\textsf{1.} & \textsf{gradient Pair $\langle\alpha^{d-1},\beta^{d}\rangle\in\mathbb{V}$ } & \textsf{Matched pair of vertices $\mleft(\alpha,\beta\mright)\in\mathcal{H_K}$ }\tabularnewline \hline
\textsf{2.} & \textsf{Dimension d } & \textsf{Multipartite Graph Level d }\tabularnewline \hline
\textsf{3.} & \textsf{$\sigma^{d-1}\prec\tau^{d}$ s.t. $\langle\sigma^{d-1},\tau^{d}\rangle\notin\mathbb{V}$ } & \textsf{Default down-edge $\ensuremath{\tau}\rightarrow\ensuremath{\sigma}$}\tabularnewline \hline
\textsf{4.} & \textsf{$\sigma^{d-1}\prec\tau^{d}$ s.t. $\langle\sigma^{d-1},\tau^{d}\rangle\in\mathbb{V}$ } & \textsf{Matching up-edge $\sigma\rightarrow\tau$ }\tabularnewline \hline
\textsf{5.} & $\mathbb{V}$\textsf{-Path} & \textsf{Directed Path}\tabularnewline \hline
\textsf{6.} & \textsf{Non-trivial Closed }$\mathbb{V}$\textsf{-Path} & \textsf{Directed Cycle}\tabularnewline \hline
\textsf{7.} & \textsf{CVF} & \textsf{Matching on the Hasse Graph}\tabularnewline \hline
\textsf{8.} & \textsf{DGVF} & \textsf{Morse Matching (i.e. Acyclic Matching)}\tabularnewline \hline
\textsf{9.} & \textsf{Critical Cell $\zeta^{d}$ } & \textsf{Unmatched Vertex $\zeta$}\tabularnewline \hline
\textsf{10.} & \textsf{Regular Cell $\xi^{d}$ } & \textsf{Matched Vertex $\xi$}\tabularnewline \hline
\end{tabular}%
}
\par\end{center}%
\par\end{center}\begin{center}
\label{tab:dictionary}
\par\end{center}
\end{figure}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\includegraphics[scale=0.15]{figure3.png}%
\end{minipage}}
\protect\caption{Matching Induced reorientation. Example 2}
\label{fig:HasseGraph2}
\end{figure}
\subsection{Prior Work}
Since the two problems, \textsf{MMMP} and \textsf{MMUP} are NP-Hard
\cite{JP06}, a logical choice is to optimize the number of critical
cells in an efficient manner is by use of approximation algorithms.
Joswig et al. \cite{JP06} noted that \textsf{MMMP} and \textsf{MMUP}
are NP-Hard problems, and posed the approximability of \textsf{MMUP}
and\textsf{ MMMP} as open problems, by pointing out an error in Lewiner's
claim about inapproximability of \textsf{MMUP} in \cite{Le02}. Recently,
Burton et al. \cite{BLPS13} developed an FPT algorithm for optimizing
Morse functions. Some of the notable works that seek optimality of
Morse matchings either by restricting the problem to 2-manifolds or
by applying heuristics are \cite{AFV11,AFV12,BLW11,HMMNWJD10,He05,JP06,Le03a,Le03b,LB14}.
The approximablity of \textsf{MMMP} is established in \cite{Ra14}.
Also, we provide an FPT algorithm for counting Morse matchings using
graph polynomials in \cite{Ra14c}. It is also worth noting
that in in \cite{Ra14c}, we use the same \emph{rigid edges framework}
introduced in this work to formulate recurrence relations for Morse
polynomials (that count the number of gradient vector fields on a
simplicial complex.)
\subsection{Problem Definition \& Contributions \label{sub:Problem-Definition}}
In computer science terms, the max-Morse matching problem (\textsf{MMMP})
can be described as follows: Consider a bdd. degree multipartite graph
$\mathcal{H}$. Associate a matching induced reorientation to $\mathcal{H}$,
such that the oriented graph $\mathcal{H}_{m}$ is acyclic. The goal
is to maximize the cardinality of matched (regular) nodes. For the
\emph{min-Morse unmatched problem} (\textsf{MMUP}), we find a Matching
that keeps the graph $\mathcal{H}_{m}$ acyclic while minimizing the
no. of unmatched (critical) nodes. More formally:
\begin{definition}[Min-Morse Unmatched Problem] The vector field that minimizes the number of critical cells over the set of all DGVFs that can be defined on a regular cell complex say $\mathcal{K}$, is known as Min-Morse Matching on complex $\mathcal{K}$. The \textbf{Min-Morse Unmatched Problem} is to find such an optimal Morse Matching.
\end{definition}
\begin{notation}[$\lll$]
For any two functions $f, g$ if $f=\tilde{O}(1)$ i.e. $f = {\log^{O(1)} (g)}$, then we denote it as $\bm{f \lll g}$.
\end{notation}
\begin{definition}[Strong and Weak Morse Optimality Conditions(SMOC/WMOC)]\label{def:MOC} Aggregate Betti number is the sum of Betti numbers across dimensions. For any given discrete vector field, the aggregate Morse number is the sum of Morse numbers across dimensions. Given a simplicial complex $\mathcal{K}$ with $\mathfrak{M}(\mathcal{K})$ and $\mathfrak{B}(\mathcal{K})$ representing the aggregate Betti and the \emph{optimal} aggregate Morse numbers respectively, we say that:
\begin{enumerate}
\item We say that a family of simplicial complexes $\varOmega$ satisfies the strong Morse optimality condition when for each $\mathcal{K}\in\varOmega$, $\mathfrak{M}(\mathcal{K})=O(\mathfrak{B}(\mathcal{K}))=O(1)$.
\item We say that a family of simplicial complexes $\varOmega$ satisfies the weak Morse optimality condition when for each $\mathcal{K}\in\varOmega$, $\mathfrak{M}(\mathcal{K}) = \tilde{O}(1)$ i.e. when $\mathfrak{M}(\mathcal{K})\lll |\mathcal{K}|$.
\end{enumerate}
\end{definition}
\autoref{tab:algcontr} and \autoref{tab:appcontr} list the algorithms
and the applications related contributions respectively.
\begin{table}
\begin{center}
\protect\caption{MMUP-APX Algorithmic Contribution}
\begin{tabular}{|c|c|c|}
\hline
& ARV based rounding & MWUM based solution\tabularnewline
\hline
\hline
min-DBCRE-APX ratio & $O(\sqrt{\log n})$ & $O(\log n)$\tabularnewline
\hline
MMUP-APX ratio & $O(\log^{3/2}n)$ & $O(\log^{2}n)$\tabularnewline
\hline
Methodology & Leighton-Rao (Interior Point+ARV) & Leighton-Rao (MWUM)\tabularnewline
\hline
Time Complexity & Polynomial Time & Nearly linear time\tabularnewline
\hline
\end{tabular}
\end{center}
\label{tab:algcontr}
\end{table}
\begin{table}
\protect\caption{MMUP-APX Contribution to Applications}
\begin{center}
\begin{tabular}{|>{\centering}m{4cm}||c||>{\centering}m{4cm}|}
\hline
How & When & \multicolumn{1}{c|}{Why}\tabularnewline
\hline
\hline
MMUP-APX & $\widetilde{O}(n)$ & Ratio: $O(\log^{2}n)$. Resolves an open problem from \cite{JP06}\tabularnewline
\hline
\hline
$H(\mathcal{K},\mathbb{A})$ & $\widetilde{O}(n)$ & Assuming WMOC for arbitrary coeffs\tabularnewline
\hline
\hline
Persistent Homology & $\widetilde{O}(n^{2})$ & Assuming WMOC for arbitrary coeffs \tabularnewline
\hline
\hline
Scalar Field Topology & $\widetilde{O}(n)$ & Finds an APX-optimal \emph{compatible} \footnote{Motivated by need for more rigor, we introduce the notion of compatibility between DMFs and scalar fields. } WMDGVF. \tabularnewline
\hline
\end{tabular}
\end{center}
\label{tab:appcontr}
\end{table}
\section{Motivation \label{sec:Motivation-and-Applications}}
\subsection{Why is DMT central to Algorithmic Topology in the Big Data Era?}
The computer assisted proof of Lorenz equations \cite{MM95} by Mischaikow
and Mrozek (along with the Delfinado-Edelsbrunner paper) may very
well be seen as the founding stones of modern computational topology.
From the earliest days, Mischaikow, Mrozek (and collaborators) have
relied on reductions as opposed to working on reducing the complexity
of Smith normal form algorithms for improving runtime. The RedHom-ChoMP
application of discrete Morse theory \cite{HMMNWJD10}, report dramatic
reductions in complex size using discrete Morse theory (albeit on
a very limited dataset) while \cite{MN13} applies discrete Morse
theory to obtain reductions in persistent homology computations. The
approximability of Morse matching has been a well-known open problem
in computational topology over the last decade. Its importance can
be acutely gauged from the following remarks by the authors of \cite{MN13}:
\begin{quote}
\small{
`` \textsf{The efficiency of our approach depends crucially on m being much smaller than n} $\dots$''
}
\end{quote}
\begin{quote}
\small{
``$\dots$ \textsf{constructing an optimal acyclic matching} -- \textsf{that is, a matching which minimizes m} -- \textsf{is NP hard. Providing sharp bounds on optimal \textbf{m} values relative to \textbf{n} for arbitrary complexes would require major breakthroughs in algebraic Topology as well as graph theory.} $\ldots$''
}
\end{quote} \footnote{where \textbf{m} is the sum of Morse numbers across dimensions and \textbf{n} is the size of the complex.}
The first quotation justifies our weak Morse optimality condition
for deriving complexity bounds. The second of the two quotations is
a direct appraisal of the MMUP approximability problem. \footnote{The range of applications of the MMUP-APX algorithm should not come as a surprise to those in the know of computational topology (especially when one realizes that DMT is equivalent to looking for minimum size complex over a large set of complexes that are simple homotopy equivalent to it using entirely combinatorial methods.)}
These observations may elicit the following response: Why does discrete
Morse theory have so much computational flexibility when it comes
to reductions? The author believes that the answer to this lies in
the fact that the \emph{optimal} discrete Morse function finds the
minimal sized complex over a \emph{wide range} of complexes that are
simple homotopy equivalent to the input complex. We qualify our statement
with the words \emph{wide range of complexes} as opposed to saying
all complexes because the range \emph{does not} include the set of
all simple homotopy equivalent complexes. This can be concretely concluded
by observing the NP-hardness proof of Morse matchings provided in\cite{JP06}
which clearly allows us to see why the optimal collapsibility problem
and the optimal discrete Morse function are not the same problems.
\subsection{\textbf{\underline{}}Universal Beauty \& Applicability of DMT}
Forman's theory is unrivaled in its beauty and simplicity when compared
to earlier combinatorial adaptations of Morse theory. Its combinatorial
power comes from the fact that the function defined over \emph{a }combinatorial
space is also discrete. Its topological elegance comes from the fact
that each pair of matched simplices in Forman's framework corresponds
to an elementary collapse on the simplicial complex. Riding on a spate
of mathematical and computational applications, Forman's theory has
emerged as \emph{the definitive} combinatorial analogue of Morse theory.
Discrete Morse theory has evolved into an important tool in algebraic,
geometric and topological combinatorics \cite{Ko08,OW07,Wa07}.
\subsection{Why is DMT a prime target for methods from Algorithmic frontiers?}
The primary construct in discrete Morse theory, namely the discrete
gradient vector field can essentially be specified as a graph matching
induced reorientation of a directed Hasse graph such that the reoriented
graph has no cycles. graph matching is one of the most studied problems
in computer science and, in fact, the study of polynomial time algorithms
began with complexity analysis of a graph matching algorithm by Edmonds.
The acylic subgraph problem was one of the 20 original problems shown
by Karp to be NP-complete. Moreover, the Hasse graph of a simplicial
complex is a structured graph, namely a sparse multipartite graph.
The special structures of sparseness and bipartiteness often lead
to simpler, faster algorithms for variety of computer science problems.
In addition to that, discrete Morse theory is also intimately related
to evasiveness and monotone graph properties. Owing to the combinatorial
(in fact, graph theoretic) nature of its specification, discrete Morse
theory is a natural target for methods from classical non-numerical
algorithms.
\section{MMUP-APX Algorithm: A Pr\'{e}cis\xspace}
Now, in our case, we follow the following steps:
\begin{enumerate}
\item Reduce the min-Morse unmatched problem to a variant of min-feedback-arc-set
(min-FAS) problem (that we refer to as min-partially oriented problem
(min-POP)). Suppose we are given a Hasse graph $\mathcal{G}$ with
edge set $\mathcal{E_{G}}$. Then, analogous to min-FAS, the end goal
of min-POP is to minimize the number of edges removed from graph $\mathcal{\mathcal{G}}$,
namely $\mathcal{E_{X}}$ such that the graph $\mathcal{G}(\mathcal{V},\mathcal{E}_{\mathcal{G}}\setminus\mathcal{E_{\mathcal{X}}})$
is acyclic. There is one crucial difference (between min-FAS and min-POP)
however: the edge set $\mathcal{E_{G}}$ is composed of rigid edges
$\mathcal{E_{R}}$ and normal edges $\mathcal{E_{N}}$ i.e. $\mathcal{E_{G}}=\mathcal{E_{R}}+\mathcal{E_{N}}$.
Moreover, the set $\mathcal{E_{X}\subsetneqq\mathcal{E_{\mathcal{N}}}}$.
So rigid edges can not be deleted. In \autoref{sec:Reduction-of-MMUP}
(\autoref{thm:minred} and \autoref{thm:linred}), we show that such
a reduction can be achieved in nearly linear in space and time.
\item Having reduced the min-Morse unmatched problem to a variant of min-FAS
(i.e. min-POP), we formulate the min-POP problem as a vector program.
We then relax the 0/1-edge constraints for normal edges.
\item Instead of solving the min-POP directly, we interpret the SDP formulation
above as as a variant of the min-directed balanced cut (min-DBC) problem
namely the min-DBCRE problem. In case of the min-DBCRE problem, we
can not cut any of the rigid edges. Otherwise min-DBCRE and min-DBC
are essentially the same. It is worth noting that any solution to
the min-FAS (and hence the min-POP) can be interpreted as a series
of possibly sub-optimal directed balanced cuts. One can intuit this
by observing the cut-based mechanism in \autoref{fig:LRDC}. The
details are discussed in \autoref{sec:The-Leighton-Rao-Divide-=000026-Conque}.
\item Therefore, we use the Leighton Rao divide-\&-conquer scheme that relaxes
the min-POP constraints of the problem instance at that particular
recursion level by formulating it as a min-DBCRE SDP and then solving
this SDP.
\item Solve+Round
\begin{enumerate}
\item Method 1: ARV-rounding based. Please see \autoref{fig:LR}.
\begin{enumerate}
\item The SDP is solved using the Interior Point Method.
\item Decompose the SDP solution into individual vectors using Cholesky
factorization.
\item Following that, we round it using Agarwal et al. method (based on
ARV rounding). \autoref{sub:Handling-Forbidden-Edges:} describes
how forbidden edges are handled in context of ARV rounding procedure.
\autoref{sub:Extension-to-Agarwal} establishes an $O\sqrt{\log n}$-ratio
for min-directed balanced cut (with rigid edges) by extending Agarwal
et al.'s framework to accommodate rigid edges.
\begin{enumerate}
\item For the ARV rounding procedure to work, the SDP must geometrically
embed the vertex set into an $l_{2}^{2}$-metric space. (See \autoref{def:l22}
and \autoref{thm:arvst}).
\item The rounding procedure constitutes of projections on random hyperplanes
and the approximation ratio is an outcome of high-dimensional measure
concentration phonomena.
\end{enumerate}
\end{enumerate}
\item Method 2: Matrix Multiplicative Weights Update Method based. Please
see \autoref{fig:LR2}
\begin{enumerate}
\item Use binary search to reduce the optimization problem to a feasibility
problem (pg.6, Section 3 of \cite{AK07}). If $\alpha$ is the current
guess for optimum vlaue of SDP then we either try to construct a PSD
matrix that is primal feasible and has value $>\alpha$ or a dual
feasible solution whose value is at most $(1+\delta)\alpha$ for an
abitrarily small value of $\delta$. For every iteration, the violation-checking-Oracle
starts by applying the rounding algorithm on the current primal solution.
\begin{enumerate}
\item If the violation-checking-Oracle fails then the current iterate is
primal feasible with value $\leq\alpha$.
\item If the violation-checking-Oracle succeeds then the current iterate
is either primal infeasible or has optimal value $>\alpha$. (The
failure to round is so spectacular that the algorithm finds a definitive
way to move towards primal feasibility.)
\item For every guess $\alpha$, if the violation-checking-Oracle does not
fail for $\tilde{O}(\nicefrac{\rho^{2}\log n}{\alpha^{2}})$ iterations,
then the algorithm provides a feasible dual solution with value at
most $(1+\delta)\alpha$ .
\end{enumerate}
\item Matrix exponentiation: In each iteration, we compute the Cholesky
decomposition of the matrix exponential to implement the multiplicative
weights update rule which involves matrix exponentiation. It is sufficient
to compute a $(1+\epsilon)$-approximation of the Cholesky decomposition.
This is done using random projections onto an $O(\nicefrac{\log n}{\epsilon^{2}})$
space and subsequently using the Johnson-Lindenstrauss lemma. Computing
only an approximate value of matrix exponentiation leads to drastic
gains in complexity.
\item The principal computational bottleneck of the MWUM algorithm is a
max-flow subroutine. The flow subroutines are used in the violation-checking-oracle
to check if the triangle inequalities (along with the objective value
constraint) belonging to the SDP are satisfied. Interestingly, the
rationale of using flows to check for violations is developed as part
of 'expander flow' framework developed in \cite{ARV09,Lee05}. By
using a nearly-linear time max-flow algorithm, we can arrive at an
approximate solution of the SDP in nearly linear time.
\end{enumerate}
\end{enumerate}
\end{enumerate}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\vspace*{1.0\baselineskip}
\begin{small}
\begin{tikzpicture}[
grow=down,
level 1/.style={sibling distance=6cm,level distance=3cm},
level 2/.style={sibling distance=6cm, level distance=3cm},
edge from parent/.style={very thick,draw=myrobinegg!40!black!60, },
edge from parent path={(\tikzparentnode.south) -- (\tikzchildnode.north)}, kant/.style={text width=2cm, text centered, sloped}, every node/.style={text ragged, inner sep=2mm,font=\small},
punkt/.style={rectangle, rounded corners, shade, top color=white, bottom color=myrobinegg!50!black!20, draw=myrobinegg!40!black!60, very thick }
]
\node[punkt] [rectangle split, rectangle split, rectangle split parts=3, text ragged]
{
SDP Size=$N$
\nodepart{second}
Solve: Interior Point Method
\nodepart{third}
Round: ARV + Agarwal et al.
}
child {
node[punkt] [rectangle split, rectangle split, rectangle split parts=3, text ragged]
{
SDP Size=$c_1 \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: Interior Point Method
\nodepart{third}
Round: ARV+Agarwal et al.
}
edge from parent
node[kant, below, pos=.6] {}
}
child {
node [punkt,rectangle split, rectangle split,
rectangle split parts=3]
{
SDP Size=$(1-c_1)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: Interior point Method
\nodepart{third}
Round: ARV + Agarwal et al.
}
child {
node [punkt,rectangle split, rectangle split,
rectangle split parts=3]
{
SDP Size=$ (1-c_1)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} c_2 \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: Interior Point Method
\nodepart{third}
Round: ARV + Agarwal et al.
}
edge from parent node[below, kant, pos=.6] {} }
child {
node [punkt, rectangle split, rectangle split parts=3]{ SDP Size=$(1-c_1)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}(1-c_2)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: Interior Point Method
\nodepart{third}
Round: ARV + Agarwal et al.
}
edge from parent node[kant, above] {}}
edge from parent{ node[kant, above] {}}
};
\end{tikzpicture}
\end{small}%
\end{minipage}}
\protect\caption{Divide-\&-Conquer SDPAlgorithm 1: ARV Rounding based}
\label{fig:LR}
\end{figure}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\vspace*{1.0\baselineskip}
\begin{tikzpicture}[
grow=down,
level 1/.style={sibling distance=6cm,level distance=3cm},
level 2/.style={sibling distance=6cm, level distance=3cm},
edge from parent/.style={very thick,draw=myrobinegg!40!black!60, },
edge from parent path={(\tikzparentnode.south) -- (\tikzchildnode.north)}, kant/.style={text width=2cm, text centered, sloped}, every node/.style={text ragged, inner sep=2mm,font=\small},
punkt/.style={rectangle, rounded corners, shade, top color=white, bottom color=myrobinegg!50!black!20, draw=myrobinegg!40!black!60, very thick }
]
\node[punkt] [rectangle split, rectangle split, rectangle split parts=3, text ragged]
{
SDP Size=$N$
\nodepart{second}
Solve: MWUM + $\tilde{O}(n)$ max-flow
\nodepart{third}
\textit{Randomly round w/o SDP soln}
}
child {
node[punkt] [rectangle split, rectangle split, rectangle split parts=3, text ragged]
{
SDP Size=$c_1 \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: MWUM + $\tilde{O}(n)$ max-flow
\nodepart{third}
\textit{Randomly round w/o SDP soln}
}
edge from parent
node[kant, below, pos=.6] {}
}
child {
node [punkt,rectangle split, rectangle split,
rectangle split parts=3]
{
SDP Size=$(1-c_1)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: MWUM + $\tilde{O}(n)$ max-flow
\nodepart{third}
\textit{Randomly round w/o SDP soln}
}
child {
node [punkt,rectangle split, rectangle split,
rectangle split parts=3]
{
SDP Size=$ (1-c_1)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} c_2 \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: MWUM + $\tilde{O}(n)$ max-flow
\nodepart{third}
\textit{Randomly round w/o SDP soln}
}
edge from parent node[below, kant, pos=.6] {} }
child {
node [punkt, rectangle split, rectangle split parts=3]{ SDP Size=$(1-c_1)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}(1-c_2)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} N$
\nodepart{second}
Solve: MWUM + $\tilde{O}(n)$ max-flow
\nodepart{third}
\textit{Randomly round w/o SDP soln}
}
edge from parent node[kant, above] {}}
edge from parent{ node[kant, above] {}}
};
\end{tikzpicture}
\end{minipage}}
\protect\caption{Divide-\&-Conquer SDP Algorithm 2: MWUM based}
\label{fig:LR2}
\end{figure}
\section{The Leighton-Rao Divide-\&-Conquer Paradigm\label{sec:The-Leighton-Rao-Divide-=000026-Conque}}
Not that, the min-Morse Problem is reduced to min-POP (which is essentially
equivalent to min-FAS with rigid edges). See \autoref{sec:Reduction-of-MMUP}.
Now, min-FAS can be approximated by a divide-\&-conquer procedure
namely the Leighton-Rao method wherein one effectively solves the
min-FAS problem instance by decomposing it into multiple min-directed
balanced cut (min-DBC) approximation problem instances. See \cite{Sh95}
Section 5.1 and Section 5.4. The approximation algorithm for min-DBC
acts as a subroutine that adds up to approximate the min-FAS approximation
instance. In a similar vein, one approximates the min-POP instance
by using approximation algorithm for min-DBCRE (min-directed balanced
cut with rigid edges) as a subroutine in Leighton-Rao applied to min-POP.
This section explains an important algorithmic technique (Leighton-Rao
divide-\&-conquer) used in the design of this approximation algorithm.
But, more importantly it also gives a bird's eye view of a variety
of \emph{other} tools and ideas involved. In a series of breakthrough
results, Leighton and Rao designed an elegant meta-algorithm \cite{LR99}
that uses divide-and-conquer strategy to approximate a wide range
of combinatorial problems with impressive performance guarantees.
Their algorithms also provide approximate max-flow min-cut theorems
for multicommodity flow problems. Given a combinatorial problem $\varPi_{1}$,
the time complexity and approximation ratio of their algorithm(s)
$\mathcal{A}_{1}$ is intimately tied to an \emph{external} state-of-art
c-balanced cut approximation algorithm $\mathcal{A}_{2}$. The algorithm
algorithm $\mathcal{A}_{2}$ is used as a subroutine for algorithm
$\mathcal{A}_{1}$. We will provide a highly simplified overview that
addresses only the essential underlying idea. \cite{Sh95} provides
an excellent survey, which may be of interest to the more inclined
reader. Given a minimization problem $\varPi_{1}$ on an input graph
$G(V,E)$ of size $N$:
\begin{description}
\item [{Step}] 1. We solve the c-balanced cut problem on $G(V,E)$ by application
of the $\alpha-$approximation. The factor $\alpha$ ensures that
a single balanced cut will be at most $\alpha$ times the cost of
the optimal balanced cut at that recursion level.
\item [{Step}] 2. The cut will also divide the vertices of the original
graph into two vertex sets $V_{1}$ and $V_{2}$. The edges that are
not cut can be used to construct induced subgraphs $G(V_{1},E_{1})$
and $G(V_{2},E_{2})$. Let $|V_{1}|=N_{1}$ and $|V_{2}|=N_{2}$.
We now apply \textbf{Step} 1 on $G(V_{1},E_{1})$ and $G(V_{2},E_{2})$.
The recursion stops each time we encounter a solitary vertex.
\end{description}
Please see \autoref{fig:LRDC} for a quick overview. In the remainder
of this section we will sketch the details of this scheme that gives
us an $O(\alpha\log n)$ factor algorithm for problem $\varPi_{1}$.
\cite{LR99,Sh95} offer more details. We follow the treatment delineated
in \cite{Sh95}. Assume that we have an $\alpha$-factor approximation
algorithm for min-DBCRE. Now, observe that every solution of min-POP
is a linear ordering (DMF) corresponding to some partial order(DGVF).
Observe that irrespective of which linear order we choose, we may
obtain a balanced cut from it and the cost of the directed balanced
cut will be an additive part of the objective of min-POP. Please refer
to \autoref{fig:LRDC}.The cost of this linear ordering is at least
the cost of this directed balanced cut and hence the cost of optimal
min-DBCRE is upper bounded by cost of optimal solution of min-POP.
We may apply the same idea recursively on problems of size $cn$ and
$(1-c)n$. Clearly the objective value of divide-\&-conquer algorithm
will satisfy
\begin{equation}
\mathcal{A}(\mathcal{G})\leq\max\left\{ {\mathcal{A}(\mathcal{G}_{1}),\mathcal{A}(\mathcal{G}_{2})}\right\} +\textnormal{DBCRE}{}_{\textnormal{APX}}(\mathcal{G}_{1}|\mathcal{G}_{2})
\end{equation}
Since $\textnormal{DBCRE}{}_{\textnormal{OPT}}\leq\textnormal{minPOP}\textnormal{\ensuremath{_{OPT}}}$,
we also have $\textnormal{DBCRE}{}_{\textnormal{APX}}(\mathcal{G}_{1}|\mathcal{G}_{2})\leq\alpha\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}\left(\textnormal{minPOP}{}_{\textnormal{OPT}}(\mathcal{G})\right)$
where $\alpha$ is the approximation ratio of min-DBCRE approximation
subroutine . Therefore, we can write the above equation as
\begin{equation}
\mathcal{A}(\mathcal{G})\leq\max\left\{ {\mathcal{A}(\mathcal{G}_{1}),\mathcal{A}(\mathcal{G}_{2})}\right\} +\alpha\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}\left(\textnormal{minPOP}{}_{\textnormal{OPT}}(\mathcal{G})\right)
\end{equation}
For each level of recursion, we incur a cost of at most $\alpha\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}\left(\textnormal{minPOP}{}_{\textnormal{OPT}}(\mathcal{G}_{k})\right)$
while noting that $\textnormal{minPOP}{}_{\textnormal{OPT}}(\mathcal{G}_{k})\leq\textnormal{minPOP}{}_{\textnormal{OPT}}(\mathcal{G})$.
Now, since there are $\left\lceil O(\log n)\right\rceil $ levels
of recursion, (where base of the logarithm depends on the ratio $c$
used in the balancing the cut). Applying a basic inductive argument
we conclude that:
\begin{equation}
\mathcal{A}(\mathcal{G})\leq\alpha\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} O(\log n)\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}\left(\textnormal{minPOP}{}_{\textnormal{OPT}}(\mathcal{G})\right)
\end{equation}
Now, we consider the approximation ratio of min-directed balanced
cut with rigid edges (min-DBCRE).
\begin{enumerate}
\item To begin with note that if we are ready to forgo the nearly linear
aspect of time complexity, then we may obtain an approximation ratio
of $O(\sqrt{\log n})$ as proved in \autoref{lem:agre} from \autoref{sub:Extension-to-Agarwal}.
This is a direct consequence of our ability to handle rigid edges
within the ARV framework of rounding as seen in \autoref{sec:Rounding-Based-on},
specifically in \autoref{lem:orrhr} and more importantly in \autoref{lem:agre}
from \autoref{sub:Extension-to-Agarwal}.
\item However, if we wish nearly linear time computation, we may use single
commodity max-flows as subroutines within violation checking oracles.
With this approach one obtains an approximation ratio of $O(\log n)$.
Note that in \autoref{sec:MWUM-for-solving} specifically in \autoref{lem:vcoracle}
we show how to handle rigid edge in MWUM violation checking oracle.
The ability to handle rigid edges within MWUM framework gives us a
nearly linear time algorithm for an $O(\log n)$ approximate solution
for min-DBCRE.
\end{enumerate}
Clearly if we use ARV-based rounding for min-DBCRE as described in
\autoref{lem:agre} in \autoref{sub:Extension-to-Agarwal}, then $\alpha=O(\sqrt{\log n})$,
giving us an approximation ratio of $O(\log^{3/2}n)$ for MMUP. In
contrast, if we use MWUM based solution for min-DBCRE as described
in \autoref{lem:vcoracle} in \autoref{sec:MWUM-for-solving}, we
get $\alpha=O(\log n)$ giving us an approximation ratio of $O(\log^{2}n)$
for MMUP. Please see \autoref{tab:algocomp} for a summary of complexity
implications.
\begin{table}
\protect\caption{Tabular summary of algorithms and complexity}
\begin{tabular}{|c|c|c|}
\hline
& ARV based rounding & MWUM based solution\tabularnewline
\hline
\hline
min-DBCRE-APX ratio & $O(\sqrt{\log n})$ & $O(\log n)$\tabularnewline
\hline
MMUP-APX ratio & $O(\log^{3/2}n)$ & $O(\log^{2}n)$\tabularnewline
\hline
Methodology & Leighton-Rao (Interior Point+ARV) & Leighton-Rao (MWUM)\tabularnewline
\hline
Time Complexity & Polynomial Time & Nearly linear time\tabularnewline
\hline
\end{tabular}
\label{tab:algocomp}
\end{table}
\begin{figure}
\begin{tikzpicture}[shorten >=1pt,node distance=1.5cm,on grid, scale=1]
\node[stateit] (A) {\textsf A};
\node[stateit] (B) [below right=2cm of A] {\textsf B };
\node[stateit] (C) [below left=1.5cm of B] {\textsf C};
\node[stateit] (D) [below right=2.25cm of C] {\textsf D};
\node[stateit] (E) [below left=1.75cm of D] {\textsf E};
\node[stateit] (F) [below right=2.25cm of E] {\textsf F};
\node[stateit] (G) [below left=3cm of F] {\textsf G};
\node[stateit] (H) [below right=2.25cm of G] {\textsf H};
\node[stateit] (I) [below left=1.75cm of H] {\textsf I};
\node[stateit] (J) [below right=2.5cm of I] {\textsf J};
\node[stateit] (K) [below left=1.5cm of J] {\textsf K};
\node[stateit] (L) [below right=1.75cm of K] {\textsf L};
\path (A) edge [bend left=45,cutit] node {} (C)
edge [bend right=45,cutit] node {} (D)
edge [rigidit] node {} (F)
edge [cutit] node {} (H);
\path (B) edge [bend left=45,backit] node {} (A)
edge [bend right=45,rigidit] node {} (E)
edge [cutit] node {} (G)
edge [cutit] node {} (J);
\path (C) edge [cutit] node {} (K)
edge [bend left=45,cutit] node {} (D)
edge [bend right=45,backit] node {} (B)
edge [cutit] node {} (G);
\path (D) edge [cutit,bend left=45] node {} (E)
edge [cutit,bend right=45] node {} (H)
edge [rigidit] node {} (J)
edge [cutit] node {} (I);
\path (E) edge [cutit] node {} (L)
edge [bend left=45,backit] node {} (C)
edge [cutit] node {} (F)
edge [cutit] node {} (H);
\path (F) edge [bend left=45,cutit] node {} (G)
edge [bend right=45,backit] node {} (A)
edge [cutit] node {} (J)
edge [cutit] node {} (I);
\path (G) edge [bend left=45,cutit] node {} (K)
edge [bend right=45,backit] node {} (F)
edge [cutit] node {} (I)
edge [cutit] node {} (H);
\path (H) edge [backit] node {} (C)
edge [bend right=45,cutit] node {} (I)
edge [cutit] node {} (K);
\path (I) edge [bend left=45,backit] node {} (G)
edge [bend right=45,cutit] node {} (F)
edge [rigidit] node {} (K);
\path (J) edge [bend left=45,backit] node {} (H)
edge [backit] node {} (I);
\path (K) edge [cutit] node {} (L)
edge [bend left=45,backit] node {} (H)
edge [bend right=45,backit] node {} (F);
\path (L) edge [bend left=45,backit] node {} (E)
edge [bend right=45,backit] node {} (J)
edge [backit] node {} (I);
\node (M) [below right=1.5cm] at (F)
{
};
\node (N) [above left=1.5cm] at (G)
{
};
\path (M) edge [thickit] node {} (N);
\end{tikzpicture}\begin{tikzpicture}[shorten >=1pt,node distance=1.5cm,on grid, scale=1]
\node[stateit] (A) {\textsf A};
\node[stateit] (B) [below right=2cm of A] {\textsf B };
\node[stateit] (C) [below left=1.5cm of B] {\textsf C};
\node[stateit] (D) [below right=2.25cm of C] {\textsf D};
\node[stateit] (E) [below left=1.75cm of D] {\textsf E};
\node[stateit] (F) [below right=2.25cm of E] {\textsf F};
\node[stateit] (G) [below left=3cm of F] {\textsf G};
\node[stateit] (H) [below right=2.25cm of G] {\textsf H};
\node[stateit] (I) [below left=1.75cm of H] {\textsf I};
\node[stateit] (J) [below right=2.5cm of I] {\textsf J};
\node[stateit] (K) [below left=1.5cm of J] {\textsf K};
\node[stateit] (L) [below right=1.75cm of K] {\textsf L};
\path (A) edge [bend left=45,cutit] node {} (C)
edge [bend right=45,cutit] node {} (D)
edge [rigidit] node {} (F);
\path (B) edge [bend left=45,backit] node {} (A)
edge [bend right=45,rigidit] node {} (E);
\path (C) edge [bend left=45,cutit] node {} (D)
edge [bend right=45,backit] node {} (B);
\path (D) edge [cutit,bend left=45] node {} (E);
\path (E) edge [bend left=45,backit] node {} (C)
edge [cutit] node {} (F);
\path (F) edge [bend right=45,backit] node {} (A);
\path (G) edge [bend left=45,cutit] node {} (K)
edge [cutit] node {} (I)
edge [cutit] node {} (H);
\path (H) edge [bend right=45,cutit] node {} (I)
edge [cutit] node {} (K);
\path (I) edge [bend left=45,backit] node {} (G)
edge [rigidit] node {} (K);
\path (J) edge [bend left=45,backit] node {} (H)
edge [backit] node {} (I);
\path (K) edge [cutit] node {} (L)
edge [bend left=45,backit] node {} (H);
\path (L) edge [bend right=45,backit] node {} (J)
edge [backit] node {} (I);
\node (M) [below left=1.5cm] at (C){};
\node (N) [above right=1.5cm] at (D){};
\path (M) edge [thickit] node {} (N);
\node (P) [below left=1.5cm] at (I){};
\node (Q) [above right=1.5cm] at (J){};
\path (P) edge [thickit] node {} (Q);
\end{tikzpicture}\begin{tikzpicture}[shorten >=1pt,node distance=1.5cm,on grid,scale=1]
\node[stateit] (A) {\textsf A};
\node[stateit] (B) [below right=2cm of A] {\textsf B };
\node[stateit] (C) [below left=1.5cm of B] {\textsf C};
\node[stateit] (D) [below right=2.25cm of C] {\textsf D};
\node[stateit] (E) [below left=1.75cm of D] {\textsf E};
\node[stateit] (F) [below right=2.25cm of E] {\textsf F};
\node[stateit] (G) [below left=3cm of F] {\textsf G};
\node[stateit] (H) [below right=2.25cm of G] {\textsf H};
\node[stateit] (I) [below left=1.75cm of H] {\textsf I};
\node[stateit] (J) [below right=2.5cm of I] {\textsf J};
\node[stateit] (K) [below left=1.5cm of J] {\textsf K};
\node[stateit] (L) [below right=1.75cm of K] {\textsf L};
\path (A) edge [bend left=45,cutit] node {} (C);
\path (B) edge [bend left=45,backit] node {} (A);
\path (C) edge [bend right=45,backit] node {} (B);
\path (D) edge [cutit,bend left=45] node {} (E);
\path (E) edge [cutit] node {} (F);
\path (G) edge [cutit] node {} (I)
edge [cutit] node {} (H);
\path (H) edge [bend right=45,cutit] node {} (I);
\path (I) edge [bend left=45,backit] node {} (G);
\path (K) edge [cutit] node {} (L);
\path (L) edge [bend right=45,backit] node {} (J);
\node (M) [below right=1.5cm] at (B){};
\node (N) [above left=1.5cm] at (C){};
\path (M) edge [thickit] node {} (N);
\node (P) [below left=1.5cm] at (E){};
\node (Q) [above right=1.5cm] at (F){};
\path (P) edge [thickit] node {} (Q);
\node (R) [below right=1.5cm] at (H){};
\node (S) [above left=1.5cm] at (I){};
\path (R) edge [thickit] node {} (S);
\node (U) [below left=1.5cm] at (K){};
\node (V) [above right=1.5cm] at (L){};
\path (U) edge [thickit] node {} (V);
\end{tikzpicture}
\protect\caption{Divide-\&-Conquer: $\quad$black: forward normal edges, red: backward
normal edges, green: rigid edges}
\label{fig:LRDC}
\end{figure}
\section{Reduction of MMUP to min-POP\label{sec:Reduction-of-MMUP}}
\subsection{Definitions: A Garden of Edges}
The procedure involves gadget construction with a flavor reminiscent of the Garey-Johnson book. Given an input graph $\mathcal{H}$ we construct a gadget $\mathcal{H}_{R}$.
\begin{definition}[Rigid edges, Forbidden edges and Normal edges in gadget $\mathcal{H}_{R}$] We have three types of edges, namely:
\begin{itemize}
\item \textbf{Rigid edges (R-edges)} are a set of prespecified oriented edges whose inclusion in every desired output solution is made mandatory.
\item The edges complementary to R-edges are known as \textbf{Forbidden edges (F-edges)} and we enforce the prohibition of forbidden edge orientations in every desired output.
\item \textbf{Normal edges (N-edges)} are edges whose inclusion/exclusion is not enforced. An edge with an orientation complementary to an N-edge is also an N-edge. In the desired solution, we are free to choose either of the two orientations - an N-edge or its complementary N-edge.
\end{itemize}
\end{definition}
\begin{definition}[Paths and Cycles in gadget $\mathcal{H}_{R}$]
A path(or a cycle) in $\mathcal{H}_{R}$ composed entirely of R-edges is known as an \textbf{R-Path(or an R-cycle)}. Analogously we may also define \textbf{N-Paths} and \textbf{N-cycles}. A path(or a cycle) in $\mathcal{H}_{R}$ composed of R-edges as well as N-edges is known as an \textbf{RN-Path(or an RN-cycle)}.
\end{definition}
\begin{shnote}In the context of our problem, we have two type of N-edges: \textbf{down-N-edges} (with default downward orientation) and \textbf{up-N-edges}(associated with matching induced reorientation) and and our typical objective is to either optimize or count the number of up-N-edges. up-N-edges are also denoted as $\overline{N}$-edges and $\underline{N}$-edges respectively. The basic idea is to do the following: Given a Hasse graph, we enlarge the vertex set and edge set of $\mathcal{H}$ to form a gadget $\mathcal{H}_{R}$ to ensure that for every cycle in $\mathcal{H}$, there is a corresponding cycle in $\mathcal{H}_{R}$. We refer to these cycles as \textbf{C-cycles}. More importantly, for every matching in $\mathcal{H}$ there is a corresponding cycle in $\mathcal{H}_{R}$. We refer to these cycles as \textbf{M-cycles}. Rigid edges introduced to break some C-cycles in $\mathcal{H}_{R}$ are known as CR-edges and those introduced to break M-cycles are known as MR-edges. (R-edges=CR-edges+MR=edges). We denote an acyclic orientation of $\mathcal{H}_{R}$ by $\overrightarrow{\mathcal{H}_{R}}$. One can easily retrieve $\overline{\mathcal{H_K}}$ from $\mathcal{H}_R$ where $\overline{\mathcal{H_K}}$ is a directed acyclic matching induced reorientation of original Hasse graph of complex $\mathcal{K}$, namely $\mathcal{H_K}$.
\end{shnote}
We now parsimoniously reduce the problem of finding an acyclic matching on $\mathcal{H}$ to that of finding an acyclic orientation on $\mathcal{H}_{R}$: with the added condition that if $e$ is an R-edge in $\mathcal{H}_{R}$ then $e\in \overrightarrow{\mathcal{H}_{R}}$.
\begin{definition} We define the decision problem \textbf{k-Partially-Oriented Problem} or (k-POP) as follows: Given a digraph $\mathcal{H}(\mathcal{V},\mathcal{E})$, whose edge set is composed of \textbf{R-edges} and \textbf{N-edges}, i.e. $\mathcal{E}=\mathcal{E}_{R}+\mathcal{E}_{N}$, we optimize the number of $\overline{N}-$edges, while ensuring that the gadget formed out of the selected edges remains acyclic. The corresponding optimization versions of the problem are referred to as min-POP and max-POP respectively.
\end{definition}
The objective function is a linear (possibly weighted) function of $\overline{N}$-edges. Essentially, in order to optimize in presence of constraints, we break all cycles in $\mathcal{H_R}$ by rejecting a set of $\overline{N}$-edges and $\underline{N}$-edges, s.t. the number of $\overline{N}$-edges rejected are minimized.
\subsection{The Construction \label{sub:The-Construction}}
\begin{description}
\item [Edge Duplication $H_0\rightarrow H_1$] Given an undirected Hasse graph $H_0$ construct a new graph $H_1$ with same incidence relations as in $H_0$ except that every undirected edge between nodes $A$ and $B$ in $H_0$ becomes two directed edges: $A\rightarrow B$ and $B \rightarrow A$ in the new graph $H_1$.\\ E.g. graph A.0 in \autoref{fig:indiemc} $\rightarrow$ graph C.1 in \autoref{fig:jointmc}
\item [Edge-Pair Isolation \& Cloning of Vertices $H_1 \rightarrow H_2$] We isolate all $\overline{N}$-$\underline{N}$-edge pairs that have two vertices in common. In the new graph $H_1$, for every vertex, say $v_i$ we will have $d(v_i)$in clones - one clone for each $\overline{N}$-$\underline{N}$-edge pair incident on $v_i$. This creates a graph $H_2$ with all edge-pairs disconnected from each other. While every vertex will have $d(v_i)$ clones of itself, every edge from $H_1$ is uniquely represented as one of the edges in the edge pair $H_2$ which gives us a 1-1 correspondence with edges in $H_1$. \\ E.g. Graph A $\rightarrow$ Graph B in \autoref{fig:fft}. Also, Graph A $\rightarrow$ Graph B in \autoref{fig:fft2}.
\item [Addition of CR-edges $H_2 \rightarrow H_3A$] We say that two $\overline{N}$-edges are adjacent if there is an $\underline{N}$-edge joining them. Now we treat each $\overline{N}$-edge in $H_2$ as if it were a vertex in $H_3$ and we join the two $\overline{N}$-edges in $H_3A$ iff they are adjacent in $H_2$. These adjacency edges are essentially $\underline{R}$-edges that form the CR-edges. This construction is reminiscent of that of a line graph except that we are working with oriented graphs and even more specifically with $\overline{N}$-edges as vertices and $\underline{N}$-edges as adjacency arcs. See \autoref{note:matchprop} on why cycles are necessarily "biparite" (i.e. restricted to a single level). \\ E.g. Graph B.1 $\rightarrow$ Graph B.2 in \autoref{fig:indiemc}
\item [Addition of MR-edges/FFT-edges $H_2 \rightarrow H_3B$] We have two ways of enforcing matching constraints using rigid edges: MR-edges and pseudo-FFT edges. The MR-gadget is conceputally simpler of the two, but can be potentially quadratic in the size of the input -- an undesirable bottleneck. The pseudo-FFT gadget is sophisticated but has the advantage of being linear in size. Clearly, if two conflicting edges match, we have an RN-cycle of size 4 with two $\overline{N}$-edges and two $\underline{R}$-edges connecting them (where $\overline{N}$-edges are part of the matching induced reorientation. This basic observation allows us to express matching constraints in form of cycle constraints. Surely, there is a (so-far) uninvestigated Model theory aspect involved here.
\begin{description}
\item[MR-edges $H_2 \rightarrow H_3B.1$ Method 1] For each pair of $\overline{N}$-edges $E_{i}$ and $E_{j}$ in $H_2$ which share a vertex and thus have a matching conflict, we form an $\underline{R}$-edge that joins the top of $E_{i}$ to the bottom of $E_{j}$ and another $\underline{R}$-edge that that top of $E_{j}$ to bottom of $E_{i}$ in graph $H_{3B.1}$. At the end of this we obtain $H_{3B.1}$. \\ E.g. Graph A.1 $\rightarrow$ Graph A.2 in \autoref{fig:indiemc}
\item[pseudo-FFT edges $H_2 \rightarrow H_3B.2$ Method 2] Here we take inspiration from the fast fourier transform (FFT). When one looks at the discrete Fourier transform, we see that each of the ouputs has several inputs. Na\"{i}ve\xspace interpretation leads to $O(N^2)$ complexity. But, a closer examination allows us to exploit the partial order structure by reusing redundant computation. Here too, we can enforce dominance relations using a special gadget that mimics the MR-gadget in linear space and time. \footnote{While the idea for this gadget emerged accidentally while thinking about FFT, the analogy with FFT stays at the metaphorical level, as far as the author can tell. There doesn't seem to be an obvious theoretical link between the two. Hence the prefix \emph{pseudo-}.} \\ E.g. Graph B $\rightarrow$ Graph C in \autoref{fig:fft} (and also Graph B $\rightarrow$ Graph C in \autoref{fig:fft2})
\end{description}
\end{description}
\begin{shnote}[Property of matching-induced bipartite cycles] It is easy to see that in order to specify cycle constraints in graph $H$, it is enough to specify the set of $\overline{N}$-edges in the cycle constraint. For instance, in \autoref{fig:indiemc} to specify that $A-\overline{B}-C-\overline{D}-E-\overline{F}$ forms a cycle, it is enough to specify $\overline{N}$-edges A-C-E as a combination of edges form a cycle. (The argument follows from the fact that since the orientation of the Hasse graph is matching induced, the edges $B, D, E$ will necessarily be $\underline{N}$-edges. Needless to say that edges belong to the same level because of matching property.) This elementary observation motivates the construction of $H^{1}$. Put differently, matching induced orientation allows us to specify cycle constraints merely in terms of $\overline{N}$-edges.
\end{shnote} \label{note:matchprop}
\begin{definition}[MMFEP] min-Morse feedback edge problem: \\
We are now in a position to define the MMFEP problem which is an extension of the MMUP in the following sense: The goal of MMFEP is to find a vector field that minimizes the number of critical cells over the set of all DGVFs that can be defined on a regular cell complex say $\mathcal{K}$, along with certain additional prescriptions involving rigid/forbidden edges.
\end{definition}
\begin{example}[MMFEP] Please refer to \autoref{fig:jointmc} Part D. Graph D.1 shows an MMFEP instance specified as a min-POP instance with $v_{10}\rightarrow v_{12}$ and $v_{13} \rightarrow v_{9}$ specified as rigid edges. Graph D.2 shows a candidate solution. (Note that $v_{12}\rightarrow v_{10}$ and $v_{9} \rightarrow v_{13}$ are depicted as rigid edges and they will necessarily be part of every solution including the candidate solution depicted in Graph D2 of \autoref{fig:jointmc})
\end{example}
\subsection{The $\mathcal{H}_{R}$ gadget = Matching+Cycle Constraints}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (6) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (100.0pt, -112.5pt) {\textcolor{black}{$v_{6}$}};
\node (7) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (162.5pt, -112.5pt) {\textcolor{black}{$v_{7}$}};
\node (8) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (237.5pt, -112.5pt) {\textcolor{black}{$v_{8}$}};
\node (4) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (75.0pt, -50.0pt) {\textcolor{black}{$v_{4}$}};
\node (5) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (125.0pt, -50.0pt) {\textcolor{black}{$v_{5}$}};
\node (1) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (75.0pt, -175.0pt) {\textcolor{black}{$v_{1}$}};
\node (2) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (137.5pt, -175.0pt) {\textcolor{black}{$v_{2}$}};
\node (3) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (187.5pt, -175.0pt) {\textcolor{black}{$v_{3}$}};
\node (11) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (37.5pt, -150.0pt) {\textcolor{black}{$v_{9}$}};
\draw [line width=0.625, color=black] (1) to [in=218, out=99] (6);
\draw [line width=0.625, color=black] (2) to [in=218, out=99] (7);
\draw [line width=0.625, color=black] (3) to [in=202, out=81] (8);
\draw [line width=0.625, color=black] (6) to [in=151, out=270] (2);
\draw [line width=0.625, color=black] (7) to [in=141, out=260] (3);
\draw [line width=0.625, color=black] (8) to [in=67, out=156] (1);
\draw [line width=0.625, color=black] (6) to [in=61, out=180] (11);
\draw [line width=0.625, color=black] (4) to [in=142, out=260] (6);
\draw [line width=0.625, color=black] (5) to [in=38, out=280] (6);
\node at (114.375pt, -144.375pt) [rotate=300] {\textcolor{black}{$B$}};
\node at (162.5pt, -155.625pt) [rotate=299] {\textcolor{black}{$D$}};
\node at (153.75pt, -95.625pt) [rotate=14] {\textcolor{black}{$F$}};
\node at (198.125pt, -129.375pt) [rotate=49] {\textcolor{black}{$E$}};
\node at (133.75pt, -133.75pt) [rotate=66] {\textcolor{black}{$C$}};
\node at (73.125pt, -130.0pt) [rotate=61] {\textcolor{black}{$A$}};
\node at (59.375pt, -114.375pt) [rotate=31] {\textcolor{black}{$G$}};
\node at (130.0pt, -88.125pt) [rotate=68] {\textcolor{black}{$J$}};
\node at (70.0pt, -88.125pt) [rotate=292] {\textcolor{black}{$I$}};
\node at (138.75pt, -214.375pt) [rotate=0] {\textcolor{purple}{\textbf{(A.0) Original graph $H_0$} }};
\end{small}
\end{tikzpicture}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (6) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (100.0pt, -112.5pt) {\textcolor{black}{$v_{6}$}};
\node (4) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (75.0pt, -50.0pt) {\textcolor{black}{$v_{4}$}};
\node (5) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (125.0pt, -50.0pt) {\textcolor{black}{$v_{5}$}};
\node (1) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (75.0pt, -175.0pt) {\textcolor{black}{$v_{1}$}};
\node (2) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (137.5pt, -175.0pt) {\textcolor{black}{$v_{2}$}};
\node (11) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (37.5pt, -150.0pt) {\textcolor{black}{$v_{11}$}}; \draw [line width=0.625, ->, >=latex, color=black] (1) to [in=218, out=99] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=151, out=270] (2);
\draw [line width=0.625, ->, >=latex, color=black] (2) to [in=331, out=90] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=39, out=280] (1);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=61, out=180] (11);
\draw [line width=0.625, ->, >=latex, color=black] (4) to [in=142, out=260] (6);
\draw [line width=0.625, ->, >=latex, color=black] (5) to [in=38, out=280] (6);
\draw [line width=0.625, ->, >=latex, color=black] (11) to [in=241, out=0] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=322, out=81] (4);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=218, out=99] (5);
\node at (73.125pt, -130.0pt) [rotate=61] {\textcolor{black}{$A$}};
\node at (114.375pt, -144.375pt) [rotate=300] {\textcolor{black}{$\overline{B}$}};
\node at (128.75pt, -123.75pt) [rotate=311] {\textcolor{black}{$B$}};
\node at (102.5pt, -160.625pt) [rotate=58] {\textcolor{black}{$\overline{A}$}};
\node at (59.375pt, -114.375pt) [rotate=31] {\textcolor{black}{$\overline{G}$}};
\node at (70.0pt, -88.125pt) [rotate=292] {\textcolor{black}{$\overline{I}$}};
\node at (130.0pt, -88.125pt) [rotate=68] {\textcolor{black}{$\overline{J}$}};
\node at (55.625pt, -141.25pt) [rotate=13] {\textcolor{black}{$G$}};
\node at (85.625pt, -69.375pt) [rotate=306] {\textcolor{black}{$I$}};
\node at (114.375pt, -69.375pt) [rotate=54] {\textcolor{black}{$J$}};
\node at (96.25pt, -195.0pt) [rotate=0] {\textcolor{purple}{\textbf{(A.1) Duplicated graph $H_1$\footnote{For sake of clarity, we depict $H_1$ restricted to edges incident on $v_6$ as opposed to showing the entire $H_1$}} }};
\end{small}
\end{tikzpicture}
\hrule
\vspace*{0.5\baselineskip}
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (26) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (87.5pt, -125.0pt) {\textcolor{black}{$A_{6}$}};
\node (25) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (87.5pt, -200.0pt) {\textcolor{black}{$A_{1}$}};
\node (10) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (150.0pt, -125.0pt) {\textcolor{black}{$B_{6}$}};
\node (4) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (150.0pt, -200.0pt) {\textcolor{black}{$B_{2}$}};
\node (11) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (337.5pt, -125.0pt) {\textcolor{black}{$J_{5}$}};
\node (5) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (337.5pt, -200.0pt) {\textcolor{black}{$J_{6}$}};
\node (28) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (275.0pt, -125.0pt) {\textcolor{black}{$I_{4}$}};
\node (27) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (275.0pt, -200.0pt) {\textcolor{black}{$I_{6}$}};
\node (12) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (212.5pt, -125.0pt) {\textcolor{black}{$G_{6}$}};
\node (9) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (212.5pt, -200.0pt) {\textcolor{black}{$G_{11}$}};
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (4) to [in=241, out=119] (10);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (5) to [in=238, out=119] (11);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (11) to [in=61, out=302] (5);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (25) to [in=241, out=119] (26);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (26) to [in=61, out=299] (25);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (27) to [in=238, out=119] (28);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (28) to [in=61, out=302] (27);
\draw [line width=0.625, ->, double, semitransparent,color=myindigo] (26) to (4);
\draw [line width=0.625, ->, double, semitransparent,color=prussianblue] (10) to (25);
\draw [line width=0.625, ->, double, semitransparent,color=teal] (28) to (5);
\draw [line width=0.625, ->, double, semitransparent,color=mylapuslazuli] (11) to (27);
\draw [line width=0.625, ->, double, semitransparent,color=teal] (28) to (25);
\draw [line width=0.625, ->, double, semitransparent,color=teal] (28) to (4);
\draw [line width=0.625, ->, double, semitransparent,color=mylapuslazuli] (11) to (4);
\draw [line width=0.625, ->, double, semitransparent,color=mylapuslazuli] (11) to (25);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (9) to [in=241, out=119] (12);
\draw [line width=0.625, ->, double, semitransparent,color=myindigo] (26) to (9);
\draw [line width=0.625, ->, double, semitransparent,color=mylapuslazuli] (11) to (9);
\draw [line width=0.625, ->, double, semitransparent,color=redwood] (12) to (25);
\draw [line width=0.625, ->, double, semitransparent,color=redwood] (12) to (4);
\draw [line width=0.625, ->, double, semitransparent,color=redwood] (12) to (27);
\draw [line width=0.625, ->, double, semitransparent,color=redwood] (12) to (5);
\draw [line width=0.625, ->, double, semitransparent,color=myindigo] (26) to (27);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (12) to [in=61, out=299] (9);
\draw [line width=0.625, ->, double, semitransparent,color=teal] (28) to (9);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (10) to [in=61, out=299] (4);
\draw [line width=0.625, ->, double, semitransparent,color=prussianblue] (10) to (5);
\draw [line width=0.625, ->, double, semitransparent,color=myindigo] (26) to (5);
\draw [line width=0.625, ->, double, semitransparent,color=prussianblue] (10) to (27);
\draw [line width=0.625, ->, double, semitransparent,color=prussianblue] (10) to (9);
\node at (144.375pt, -158.125pt) {\textcolor{purple}{$B$}};
\node at (331.875pt, -158.75pt) {\textcolor{purple}{$J$}};
\node at (341.875pt, -158.75pt) {\textcolor{purple}{$\overline{J}$}};
\node at (81.875pt, -158.125pt) {\textcolor{purple}{$A$}};
\node at (91.875pt, -166.875pt) {\textcolor{purple}{$\overline{A}$}};
\node at (268.75pt, -158.75pt) {\textcolor{purple}{$I$}};
\node at (280.0pt, -158.75pt) {\textcolor{purple}{$\overline{I}$}};
\node at (206.25pt, -158.125pt) {\textcolor{purple}{$G$}};
\node at (217.5pt, -166.875pt) {\textcolor{purple}{$\overline{G}$}};
\node at (154.375pt, -166.875pt) {\textcolor{purple}{$\overline{B}$}};
\node at (212.5pt, -220.0pt) [rotate=0] {\textcolor{purple}{\textbf{(A.2) Matching gadget} }};
\end{small}
\end{tikzpicture}
\end{center}
\vspace*{1.0\baselineskip}
.\xrfill{3pt}.
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (6) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (100.0pt, -112.5pt) {\textcolor{black}{$v_{6}$}};
\node (7) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (162.5pt, -112.5pt) {\textcolor{black}{$v_{7}$}};
\node (8) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (237.5pt, -112.5pt) {\textcolor{black}{$v_{8}$}};
\node (1) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (75.0pt, -175.0pt) {\textcolor{black}{$v_{1}$}};
\node (2) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (137.5pt, -175.0pt) {\textcolor{black}{$v_{2}$}};
\node (3) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (187.5pt, -175.0pt) {\textcolor{black}{$v_{3}$}};
\node (11) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (37.5pt, -150.0pt) {\textcolor{black}{$v_{11}$}};
\draw [line width=0.625, ->, >=latex, color=black] (1) to [in=218, out=99] (6);
\draw [line width=0.625, ->, >=latex, color=black] (2) to [in=218, out=99] (7);
\draw [line width=0.625, ->, >=latex, color=black] (3) to [in=202, out=81] (8);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=151, out=270] (2);
\draw [line width=0.625, ->, >=latex, color=black] (7) to [in=141, out=260] (3);
\draw [line width=0.625, ->, >=latex, color=black] (8) to [in=67, out=156] (1);
\draw [line width=0.625, ->, >=latex, color=black] (8) to [in=23, out=260] (3);
\draw [line width=0.625, ->, >=latex, color=black] (7) to [in=39, out=280] (2);
\draw [line width=0.625, ->, >=latex, color=prussianblue] (2) to [in=331, out=90] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=39, out=280] (1);
\draw [line width=0.625, ->, >=latex, color=black] (3) to [in=312, out=90] (7);
\draw [line width=0.625, ->, >=latex, color=black] (1) to [in=260, out=321] (8);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=61, out=180] (11);
\draw [line width=0.625, ->, >=latex, color=prussianblue] (11) to [in=241, out=0] (6);
\node at (73.125pt, -130.0pt) [rotate=61] {\textcolor{black}{$A$}};
\node at (133.75pt, -133.75pt) [rotate=66] {\textcolor{black}{$C$}};
\node at (198.125pt, -129.375pt) [rotate=49] {\textcolor{black}{$E$}};
\node at (114.375pt, -144.375pt) [rotate=300] {\textcolor{black}{$\overline{B}$}};
\node at (162.5pt, -155.625pt) [rotate=299] {\textcolor{black}{$D$}};
\node at (153.75pt, -95.625pt) [rotate=14] {\textcolor{black}{$\overline{F}$}};
\node at (226.25pt, -158.75pt) [rotate=49] {\textcolor{black}{$\overline{E}$}};
\node at (153.75pt, -143.75pt) [rotate=69] {\textcolor{black}{$\overline{C}$}};
\node at (128.75pt, -123.75pt) [rotate=311] {\textcolor{prussianblue}{$B$}};
\node at (102.5pt, -160.625pt) [rotate=58] {\textcolor{black}{$\overline{A}$}};
\node at (172.5pt, -141.25pt) [rotate=295] {\textcolor{black}{$\overline{D}$}};
\node at (168.75pt, -204.375pt) [rotate=15] {\textcolor{black}{$F$}};
\node at (59.375pt, -114.375pt) [rotate=31] {\textcolor{black}{$\overline{G}$}};
\node at (55.625pt, -141.25pt) [rotate=13] {\textcolor{prussianblue}{$G$}};
\node at (138.75pt, -214.375pt) [rotate=0] {\textcolor{purple}{\textbf{(B.1) Duplicated Graph $H_1$ with cycles $A\overline{B} C\overline{D} E\overline{F}$ and $ F \overline{E} D \overline{C} B \overline{A}$\footnote{We do not depict edges incident on vertices $v_4$ and $v_5$ since we are interested in cycles with length $\geq3$}} }};
\end{small}
\end{tikzpicture}
\end{center}
\hrule
\vspace*{-1.0\baselineskip}
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}, scale=0.75]
\begin{small}
\node (12) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (362.5pt, -87.5pt) {\textcolor{black}{$F_{8}$}};
\node (9) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (362.5pt, -162.5pt) {\textcolor{black}{$F_{1}$}};
\node (30) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (300.0pt, -87.5pt) {\textcolor{black}{$E_{8}$}};
\node (29) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (300.0pt, -162.5pt) {\textcolor{black}{$E_{3}$}};
\node (11) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (237.5pt, -87.5pt) {\textcolor{black}{$D_{7}$}};
\node (5) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (237.5pt, -162.5pt) {\textcolor{black}{$D_{3}$}};
\node (28) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (175.0pt, -87.5pt) {\textcolor{black}{$C_{7}$}};
\node (27) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (175.0pt, -162.5pt) {\textcolor{black}{$C_{2}$}};
\node (10) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (112.5pt, -87.5pt) {\textcolor{black}{$B_{6}$}};
\node (4) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (112.5pt, -162.5pt) {\textcolor{black}{$B_{2}$}};
\node (26) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (50.0pt, -87.5pt) {\textcolor{black}{$A_{6}$}};
\node (25) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (50.0pt, -162.5pt) {\textcolor{black}{$A_{1}$}};
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (4) to [in=238, out=119] (10);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (10) to [in=61, out=302] (4);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (5) to [in=238, out=119] (11);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (11) to [in=61, out=302] (5);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (9) to [in=238, out=119] (12);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (12) to [in=61, out=302] (9);
\draw [line width=0.625, ->, double, semitransparent,color=MediumOrchid] (10) to [in=77, out=77] (9);
\draw [line width=0.625, ->, double, semitransparent,color=MediumOrchid] (11) to (4);
\draw [line width=0.625, ->, double, semitransparent,color=MediumOrchid] (12) to (5);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (25) to [in=238, out=119] (26);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (26) to [in=61, out=302] (25);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (27) to [in=238, out=119] (28);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (28) to [in=61, out=302] (27);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (29) to [in=238, out=119] (30);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (30) to [in=61, out=302] (29);
\draw [line width=0.625, ->, double, semitransparent,color=MediumOrchid] (26) to (27);
\draw [line width=0.625, ->, double, semitransparent,color=MediumOrchid] (28) to (29);
\draw [line width=0.625, ->, double, semitransparent,color=MediumOrchid] (30) to [in=77, out=135] (25);
\node at (94.375pt, -129.375pt) {\textcolor{purple}{$B$}};
\node at (132.5pt, -129.375pt) {\textcolor{purple}{$\overline{B}$}};
\node at (113.5pt, -129.375pt) {\textcolor{MediumOrchid}{$\widehat{B}$}};
\node at (219.375pt, -129.375pt) {\textcolor{purple}{$D$}};
\node at (257.5pt, -129.375pt) {\textcolor{purple}{$\overline{D}$}};
\node at (238.5pt, -129.375pt) {\textcolor{MediumOrchid}{$\widehat{D}$}};
\node at (345.0pt, -129.375pt) {\textcolor{purple}{$F$}};
\node at (381.875pt, -129.375pt) {\textcolor{purple}{$\overline{F}$}};
\node at (31.875pt, -129.375pt) {\textcolor{purple}{$A$}};
\node at (70.0pt, -129.375pt) {\textcolor{purple}{$\overline{A}$}};
\node at (157.5pt, -129.375pt) {\textcolor{purple}{$C$}};
\node at (194.375pt, -129.375pt) {\textcolor{purple}{$\overline{C}$}};
\node at (176pt, -129.375pt) {\textcolor{MediumOrchid}{$\widehat{C}$}};
\node at (282.5pt, -129.375pt) {\textcolor{purple}{$E$}};
\node at (319.375pt, -129.375pt) {\textcolor{purple}{$\overline{E}$}};
\node at (301pt, -129.375pt) {\textcolor{MediumOrchid}{$\widehat{E}$}};
\node at (206.25pt, -60pt) {\textcolor{MediumOrchid}{$\widehat{F}$}};
\node at (300pt, -45pt) {\textcolor{MediumOrchid}{$\widehat{A}$}};
\node at (206.25pt, -190pt) [rotate=0] {\textcolor{purple}{\textbf{(B.2) Cycle gadget}}};
\end{small}
\end{tikzpicture}
\end{center}%
\end{minipage}}
\protect\caption{Matching Gadget A.0 $\protect\mapsto$ A.1 $\protect\mapsto$ A.2\qquad{}
Cycle Gadget B.1 $\protect\mapsto$ B.2}
\label{fig:indiemc}
\end{figure}
\begin{example}
Consider the subgraph determined by vertices $v_{1}$ $v_{2}$ $v_{3}$, $v_{4}$, $v_{5}$ $v_{6}$ $v_{7}$ $v_{8}$ and $v_{9}$. Here we depict the gadget formed by this subgraph that take into consideration both matching+cycle constraints.
\end{example}
The $\mathcal{H}_R$ gadget models all matching constraints as well as cycle constraint of original Hasse graph $\mathcal{H_K}$ as strictly cycle constraints only albeit with introduction of the so-called rigid edges.
Note that the $\mathcal{H}_R$ gadget can be described as follows: $H_2\rightarrow H_3A+H_3B.1$. \\
E.g. Graph C.1 $\rightarrow$ Graph C.2 in \autoref{fig:jointmc} constitutes an example of such a reduction. \\
This gadget includes the MR-edges as well as the CR-edges. This gadget still suffers from the drawback that MR-edges may be quadratic in number for certain graphs. This situation can be rectified by using pseudo-FFT gadget in place of matching gadget i.e. $H_3B.2$ instead of $H_3B.1$. The pseudo-FFT gadget is described in \autoref{sub:How-to-linearize:}. .
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (6) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (100.0pt, -112.5pt) {\textcolor{black}{$v_{6}$}};
\node (7) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (162.5pt, -112.5pt) {\textcolor{black}{$v_{7}$}};
\node (8) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (237.5pt, -112.5pt) {\textcolor{black}{$v_{8}$}};
\node (4) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (75.0pt, -50.0pt) {\textcolor{black}{$v_{4}$}};
\node (5) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (125.0pt, -50.0pt) {\textcolor{black}{$v_{5}$}};
\node (1) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (75.0pt, -175.0pt) {\textcolor{black}{$v_{1}$}};
\node (2) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (137.5pt, -175.0pt) {\textcolor{black}{$v_{2}$}};
\node (3) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (187.5pt, -175.0pt) {\textcolor{black}{$v_{3}$}};
\node (11) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (37.5pt, -150.0pt) {\textcolor{black}{$v_{9}$}};
\draw [line width=0.625, ->, >=latex, color=black] (1) to [in=218, out=99] (6);
\draw [line width=0.625, ->, >=latex, color=black] (2) to [in=218, out=99] (7);
\draw [line width=0.625, ->, >=latex, color=black] (3) to [in=202, out=81] (8);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=151, out=270] (2);
\draw [line width=0.625, ->, >=latex, color=black] (7) to [in=141, out=260] (3);
\draw [line width=0.625, ->, >=latex, color=black] (8) to [in=67, out=156] (1);
\draw [line width=0.625, ->, >=latex, color=black] (8) to [in=23, out=260] (3);
\draw [line width=0.625, ->, >=latex, color=black] (7) to [in=39, out=280] (2);
\draw [line width=0.625, ->, >=latex, color=black] (2) to [in=331, out=90] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=39, out=280] (1);
\draw [line width=0.625, ->, >=latex, color=black] (3) to [in=312, out=90] (7);
\draw [line width=0.625, ->, >=latex, color=black] (1) to [in=260, out=321] (8);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=61, out=180] (11);
\draw [line width=0.625, ->, >=latex, color=black] (4) to [in=142, out=260] (6);
\draw [line width=0.625, ->, >=latex, color=black] (5) to [in=38, out=280] (6);
\draw [line width=0.625, ->, >=latex, color=black] (11) to [in=241, out=0] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=322, out=81] (4);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=218, out=99] (5);
\node at (73.125pt, -130.0pt) [rotate=61] {\textcolor{black}{$A$}};
\node at (133.75pt, -133.75pt) [rotate=66] {\textcolor{black}{$C$}};
\node at (198.125pt, -129.375pt) [rotate=49] {\textcolor{black}{$E$}};
\node at (114.375pt, -144.375pt) [rotate=300] {\textcolor{black}{$\overline{B}$}};
\node at (162.5pt, -155.625pt) [rotate=299] {\textcolor{black}{$D$}};
\node at (153.75pt, -95.625pt) [rotate=14] {\textcolor{black}{$\overline{F}$}};
\node at (226.25pt, -158.75pt) [rotate=49] {\textcolor{black}{$\overline{E}$}};
\node at (153.75pt, -143.75pt) [rotate=69] {\textcolor{black}{$\overline{C}$}};
\node at (128.75pt, -123.75pt) [rotate=311] {\textcolor{black}{$B$}};
\node at (102.5pt, -160.625pt) [rotate=58] {\textcolor{black}{$\overline{A}$}};
\node at (172.5pt, -141.25pt) [rotate=295] {\textcolor{black}{$\overline{D}$}};
\node at (168.75pt, -204.375pt) [rotate=15] {\textcolor{black}{$F$}};
\node at (59.375pt, -114.375pt) [rotate=31] {\textcolor{black}{$\overline{G}$}};
\node at (70.0pt, -88.125pt) [rotate=292] {\textcolor{black}{$\overline{I}$}};
\node at (130.0pt, -88.125pt) [rotate=68] {\textcolor{black}{$\overline{J}$}};
\node at (55.625pt, -141.25pt) [rotate=13] {\textcolor{black}{$G$}};
\node at (85.625pt, -69.375pt) [rotate=306] {\textcolor{black}{$I$}};
\node at (114.375pt, -69.375pt) [rotate=54] {\textcolor{black}{$J$}};
\node at (138.75pt, -214.375pt) [rotate=0] {\textcolor{purple}{\textbf{(C.1) Duplicated Graph $H_1$} }};
\end{small}
\end{tikzpicture}
\end{center}
\hrule
\vspace*{1\baselineskip}
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}, scale=0.9]
\begin{small}
\node (24) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (200.0pt, -175.0pt) {\textcolor{black}{$C_{7}$}};
\node (23) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (200.0pt, -250.0pt) {\textcolor{black}{$C_{2}$}};
\node (32) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (262.5pt, -175.0pt) {\textcolor{black}{$D_{7}$}};
\node (31) [circle, minimum size=16.25pt, line width=0.625pt, fill=mypink!60, draw=myred, double] at (262.5pt, -250.0pt) {\textcolor{black}{$D_{3}$}};
\node (30) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (325.0pt, -175.0pt) {\textcolor{black}{$E_{8}$}};
\node (29) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (325.0pt, -250.0pt) {\textcolor{black}{$E_{3}$}};
\node (34) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (387.5pt, -175.0pt) {\textcolor{black}{$F_{8}$}};
\node (33) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (387.5pt, -250.0pt) {\textcolor{black}{$F_{1}$}};
\node (9) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (137.5pt, -250.0pt) {\textcolor{black}{$B_{2}$}};
\node (26) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (12.5pt, -175.0pt) {\textcolor{black}{$G_{6}$}};
\node (25) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (12.5pt, -250.0pt) {\textcolor{black}{$G_{9}$}};
\node (12) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (137.5pt, -175.0pt) {\textcolor{black}{$B_{6}$}};
\node (10) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (75.0pt, -175.0pt) {\textcolor{black}{$A_{6}$}};
\node (4) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (75.0pt, -250.0pt) {\textcolor{black}{$A_{1}$}};
\node (28) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (37.5pt, -50.0pt) {\textcolor{black}{$I_{4}$}};
\node (27) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (37.5pt, -125.0pt) {\textcolor{black}{$I_{6}$}};
\node (5) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (112.5pt, -125.0pt) {\textcolor{black}{$J_{6}$}};
\node (11) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (112.5pt, -50.0pt) {\textcolor{black}{$J_{5}$}};
\draw [line width=1.875, ->, semitransparent,color=purple!90] (4) to [in=241, out=119] (10);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (5) to [in=238, out=119] (11);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (11) to [in=61, out=302] (5);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (25) to [in=241, out=119] (26);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (26) to [in=61, out=299] (25);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (27) to [in=238, out=119] (28);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (28) to [in=61, out=302] (27);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=redwood] (26) to (4);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=violet] (10) to (25);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=teal] (28) to (5);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=mylapuslazuli] (11) to (27);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=teal] (28) to (25);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=teal] (28) to (4);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=mylapuslazuli] (11) to (4);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=mylapuslazuli] (11) to (25);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (9) to [in=241, out=119] (12);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=redwood] (26) to (9);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=mylapuslazuli] (11) to (9);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myrobinegg] (12) to (25);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myrobinegg] (12) to (4);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myrobinegg] (12) to (27);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myrobinegg] (12) to (5);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=redwood] (26) to (27);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (12) to [in=61, out=299] (9);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=teal] (28) to (9);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (10) to [in=61, out=299] (4);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=violet] (10) to (5);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=redwood] (26) to (5);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=violet] (10) to (27);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=violet] (10) to (9);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (23) to [in=241, out=119] (24);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (24) to [in=61, out=299] (23);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (29) to [in=241, out=119] (30);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (30) to [in=61, out=299] (29);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (31) to [in=241, out=119] (32);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (32) to [in=61, out=299] (31);
\draw [line width=1.875, ->, semitransparent,color=purple!90] (33) to [in=241, out=119] (34);
\draw [line width=0.625, ->, semitransparent,color=purple!90] (34) to [in=61, out=299] (33);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myindigo] (10) to (23);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myindigo] (24) to (29);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myindigo] (30) to [in=58, out=156] (4);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myindigo] (12) to (31);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myindigo] (32) to (33);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=myindigo] (34) to [in=77, out=139] (9);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=prussianblue] (24) to (31);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=prussianblue] (32) to (23);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=magenta] (30) to (33);
\draw [line width=0.625, ->, >=latex, double, semitransparent,color=magenta] (34) to (29);
\node at (69.375pt, -208.125pt) {\textcolor{purple}{$A$}};
\node at (105.0pt, -86.25pt) {\textcolor{purple}{$J$}};
\node at (117.5pt, -93.125pt) {\textcolor{purple}{$\overline{J}$}};
\node at (6.875pt, -208.125pt) {\textcolor{purple}{$G$}};
\node at (16.875pt, -216.875pt) {\textcolor{purple}{$\overline{G}$}};
\node at (30.0pt, -83.75pt) {\textcolor{purple}{$I$}};
\node at (45.0pt, -92.5pt) {\textcolor{purple}{$\overline{I}$}};
\node at (131.875pt, -208.125pt) {\textcolor{purple}{$B$}};
\node at (141.875pt, -216.875pt) {\textcolor{purple}{$\overline{B}$}};
\node at (79.375pt, -216.875pt) {\textcolor{purple}{$\overline{A}$}};
\node at (193.75pt, -208.125pt) {\textcolor{purple}{$C$}};
\node at (205.0pt, -216.875pt) {\textcolor{purple}{$\overline{C}$}};
\node at (318.75pt, -208.125pt) {\textcolor{purple}{$E$}};
\node at (330.0pt, -216.875pt) {\textcolor{purple}{$\overline{E}$}};
\node at (256.875pt, -208.125pt) {\textcolor{purple}{$D$}};
\node at (266.875pt, -216.875pt) {\textcolor{purple}{$\overline{D}$}};
\node at (381.25pt, -208.125pt) {\textcolor{purple}{$F$}};
\node at (392.5pt, -216.875pt) {\textcolor{purple}{$\overline{F}$}};
\node at (200.0pt, -270.0pt) [rotate=0] {\textcolor{purple}{\textbf{(C.2) Cycle + Matching gadget}}};
\end{small}
\end{tikzpicture}
\end{center}
\vspace*{1.0\baselineskip}
.\xrfill{3pt}.
\vspace*{1\baselineskip}
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}, scale=0.8]
\begin{small}
\node (10) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (12.5pt, -50.0pt) {\textcolor{black}{$v_{10}$}}; \node (4) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (50.0pt, -37.5pt) {\textcolor{black}{$v_{4}$}}; \node (5) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (100.0pt, -37.5pt) {\textcolor{black}{$v_{5}$}}; \node (9) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (137.5pt, -50.0pt) {\textcolor{black}{$v_{9}$}}; \node (6) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (75.0pt, -112.5pt) {\textcolor{black}{$v_{6}$}}; \node (11) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (25.0pt, -187.5pt) {\textcolor{black}{$v_{11}$}}; \node (1) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (62.5pt, -187.5pt) {\textcolor{black}{$v_{1}$}}; \node (33) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (225.0pt, -187.5pt) {\textcolor{black}{$v_{11}$}}; \node (23) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (262.5pt, -187.5pt) {\textcolor{black}{$v_{1}$}}; \node (24) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (300.0pt, -187.5pt) {\textcolor{black}{$v_{2}$}}; \node (29) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (325.0pt, -125.0pt) {\textcolor{black}{$v_{7}$}}; \node (28) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (275.0pt, -125.0pt) {\textcolor{black}{$v_{6}$}}; \node (32) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (225.0pt, -62.5pt) {\textcolor{black}{$v_{10}$}}; \node (26) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (250.0pt, -62.5pt) {\textcolor{black}{$v_{4}$}}; \node (27) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (300.0pt, -62.5pt) {\textcolor{black}{$v_{5}$}}; \node (31) [regular polygon, regular polygon sides=8, minimum size=16.25pt, fill=lightestimpurple, line width=0.625pt, draw=medimpurple, double] at (325.0pt, -62.5pt) {\textcolor{black}{$v_{9}$}}; \node (25) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (350.0pt, -187.5pt) {\textcolor{black}{$v_{3}$}}; \node (30) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (375.0pt, -125.0pt) {\textcolor{black}{$v_{8}$}}; \node (12) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (12.5pt, -112.5pt) {\textcolor{black}{$v_{12}$}}; \node (8) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (175.0pt, -112.5pt) {\textcolor{black}{$v_{8}$}}; \node (7) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (125.0pt, -112.5pt) {\textcolor{black}{$v_{7}$}}; \node (2) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (100.0pt, -187.5pt) {\textcolor{black}{$v_{2}$}}; \node (3) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (150.0pt, -187.5pt) {\textcolor{black}{$v_{3}$}}; \node (14) [forbidden sign, minimum size=12.5pt, fill=white, line width=0.625pt, draw=myred] at (12.5pt, -75.0pt) {}; \node (15) [forbidden sign, minimum size=12.5pt, fill=white, line width=0.625pt, draw=myred] at (162.5pt, -75.0pt) {}; \node (13) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (200.0pt, -112.5pt) {\textcolor{black}{$v_{13}$}}; \node (17) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (400.0pt, -125.0pt) {\textcolor{black}{$v_{13}$}}; \node (16) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (225.0pt, -125.0pt) {\textcolor{black}{$v_{12}$}}; \draw [line width=0.625, ->, >=latex, color=black] (1) to [in=231, out=113] (6); \draw [line width=0.625, ->, >=latex, color=black] (2) to [in=222, out=103] (7); \draw [line width=0.625, ->, >=latex, color=black] (3) to [in=222, out=103] (8); \draw [line width=0.625, ->, >=latex, color=black] (6) to [in=138, out=260] (2); \draw [line width=0.625, ->, >=latex, color=black] (7) to [in=138, out=260] (3); \draw [line width=0.625, ->, >=latex, color=black] (8) to [in=90, out=157] (1); \draw [line width=0.625, ->, >=latex, color=black] (8) to [in=42, out=280] (3); \draw [line width=0.625, ->, >=latex, color=black] (7) to [in=42, out=280] (2); \draw [line width=0.625, ->, >=latex, color=prussianblue] (2) to [in=318, out=77] (6); \draw [line width=0.625, ->, >=latex, color=black] (6) to [in=51, out=290] (1); \draw [line width=0.625, ->, >=latex, color=black] (3) to [in=309, out=90] (7); \draw [line width=0.625, ->, >=latex, color=black] (1) to [in=270, out=337] (8); \draw [line width=0.625, ->, >=latex, color=black] (6) to [in=86, out=207] (11); \draw [line width=0.625, ->, >=latex, color=black] (10) to [in=166, out=285] (6); \draw [line width=0.625, ->, >=latex, color=black] (4) to [in=139, out=260] (6); \draw [line width=0.625, ->, >=latex, color=black] (9) to [in=14, out=255] (6); \draw [line width=0.625, ->, >=latex, color=black] (5) to [in=41, out=280] (6); \draw [line width=0.625, ->, >=latex, color=prussianblue] (11) to [in=265, out=27] (6); \draw [line width=0.625, ->, >=latex, color=prussianblue] (6) to [in=345, out=105] (10); \draw [line width=0.625, ->, >=latex, color=prussianblue] (6) to [in=195, out=75] (9); \draw [line width=0.625, ->, >=latex, color=prussianblue] (6) to [in=318, out=80] (4); \draw [line width=0.625, ->, >=latex, color=prussianblue] (6) to [in=222, out=100] (5); \draw [line width=0.625, ->, >=latex, color=black] (23) to [in=228, out=108] (28); \draw [line width=0.625, ->, >=latex, color=black] (24) to [in=218, out=99] (29); \draw [line width=0.625, ->, >=latex, color=black] (29) to [in=141, out=260] (25); \draw [line width=0.625, ->, >=latex, color=black] (30) to [in=86, out=156] (23); \draw [line width=0.625, ->, >=latex, color=black] (30) to [in=39, out=280] (25); \draw [line width=0.625, ->, >=latex, color=black] (28) to [in=81, out=202] (33); \draw [line width=0.625, ->, >=latex, color=black] (32) to [in=157, out=280] (28); \draw [line width=0.625, ->, >=latex, color=black] (26) to [in=142, out=260] (28); \draw [line width=0.625, ->, >=latex, color=black] (31) to [in=23, out=260] (28); \draw [line width=0.625, ->, >=latex, color=black] (27) to [in=38, out=280] (28); \draw [line width=0.625, ->, >=latex, color=black] (28) to [in=141, out=260] (24); \draw [line width=0.625, ->, color=teal] (10) to (12); \draw [line width=0.625, ->, color=teal] (13) to (9); \draw [line width=0.625, ->, >=latex, double, color=teal] (16) to (32); \draw [line width=0.625, ->, >=latex, double, color=teal] (31) to (17);
\node at (51.25pt, -141.875pt) [rotate=76] {\textcolor{black}{$A$}};
\node at (104.375pt, -125.0pt) [rotate=55] {\textcolor{black}{$C$}}; \node at (153.75pt, -125.0pt) [rotate=55] {\textcolor{black}{$E$}};
\node at (88.75pt, -168.75pt) [rotate=306] {\textcolor{black}{$\overline{B}$}}; \node at (124.875pt, -171.375pt) [rotate=301] {\textcolor{black}{$\overline{D}$}}; \node at (104.375pt, -107.5pt) [rotate=25] {\textcolor{black}{$\overline{F}$}}; \node at (179.375pt, -158.125pt) [rotate=70] {\textcolor{black}{$\overline{E}$}}; \node at (110.625pt, -166.875pt) [rotate=56] {\textcolor{black}{$\overline{C}$}}; \node at (103.125pt, -136.875pt) [rotate=294] {\textcolor{black}{$B$}}; \node at (81.25pt, -172.5pt) [rotate=63] {\textcolor{black}{$\overline{A}$}}; \node at (137.5pt, -145.0pt) [rotate=291] {\textcolor{black}{$D$}}; \node at (133.75pt, -193.75pt) [rotate=26] {\textcolor{black}{$F$}}; \node at (32.5pt, -137.5pt) [rotate=56] {\textcolor{black}{$\overline{G}$}}; \node at (28.125pt, -96.25pt) [rotate=315] {\textcolor{black}{$\overline{H}$}}; \node at (43.75pt, -81.25pt) [rotate=288] {\textcolor{black}{$\overline{I}$}}; \node at (121.875pt, -96.25pt) [rotate=46] {\textcolor{black}{$\overline{K}$}}; \node at (106.25pt, -81.25pt) [rotate=72] {\textcolor{black}{$\overline{J}$}}; \node at (40.0pt, -170.0pt) [rotate=38] {\textcolor{black}{$G$}}; \node at (35.625pt, -66.875pt) [rotate=329] {\textcolor{black}{$H$}}; \node at (114.375pt, -66.875pt) [rotate=31] {\textcolor{black}{$K$}}; \node at (61.875pt, -60.0pt) [rotate=302] {\textcolor{black}{$I$}}; \node at (88.125pt, -60.0pt) [rotate=58] {\textcolor{black}{$J$}}; \node at (252.5pt, -146.25pt) [rotate=72] {\textcolor{black}{$A$}}; \node at (296.25pt, -146.25pt) [rotate=66] {\textcolor{black}{$C$}}; \node at (323.0pt, -170.125pt) [rotate=299] {\textcolor{black}{$\overline{D}$}}; \node at (308.125pt, -113.75pt) [rotate=20] {\textcolor{black}{$\overline{F}$}}; \node at (378.75pt, -165.0pt) [rotate=65] {\textcolor{black}{$\overline{E}$}}; \node at (234.375pt, -142.5pt) [rotate=51] {\textcolor{black}{$\overline{G}$}}; \node at (231.875pt, -103.125pt) [rotate=305] {\textcolor{black}{$\overline{H}$}}; \node at (242.5pt, -91.875pt) [rotate=283] {\textcolor{black}{$\overline{I}$}}; \node at (322.5pt, -96.875pt) [rotate=60] {\textcolor{black}{$\overline{K}$}}; \node at (293.125pt, -89.375pt) [rotate=75] {\textcolor{black}{$\overline{J}$}}; \node at (283.75pt, -156.25pt) [rotate=291] {\textcolor{black}{$\overline{B}$}};
\node at (200.0pt, -210.0pt) [rotate=0] {\textcolor{purple}{\textbf{(D.1) MMFEP specified as min-POP $\rightarrow$ (D.2) A Solution.}}};
\end{small}
\end{tikzpicture}
\end{center}%
\end{minipage}}
\protect\caption{Cycle + Matching Gadget C.1 $\protect\mapsto$ C.2 \qquad{} MMFEP
$\protect\mapsto$ min-POP}
\label{fig:jointmc}
\end{figure}
\subsection{Correctness and Complexity}
Note that all R-cycles are spurious. If an RE-formulation, say $\mathcal{H}_{R}$, were to have an R-cycle then the formulation itself is incorrect, given the fact that all R-edges need to be included in $\overrightarrow{\mathcal{H}_{R}}$.
\begin{notation}[Isolated edges] The notation $e_{ij}$ represents an $\overline{N}$ edge, whereas $\overline{e_{ij}}$ is an $\underline{N}$ edge. Together $e_{ij}$ and $\overline{e_{ij}}$ form an isolated edge-pair of normal edges. This edge pair is formed at the time of $H_1 \rightarrow H_2$ transformation as described in Step 2 of the reduction procedure described in\autoref{sub:The-Construction}. The head-node of $e_{ij}$ is denoted as $e_{ij}^T$ and the tail node of $e_{ij}$ is denoted as $e_{ij}^B$. Also, $e_{ij}$ (in graphs $H_2$, $H_3B$, $H_3B$) is a representative of the edge(s) between vertices $v_i$ and $v_j$ (in graphs $H_0$ and $H_1$) prior to their isolation. For instance let $i=6, j=1, k=2$ in \autoref{fig:indiemc}. Then the edges $A$ (between nodes $v_1$ and $v_6$) and $B$ (between nodes $v_6$ and $v_2$) from Graph A.0 and Graph A.1 in \autoref{fig:indiemc} become isolated edge pairs $A=e_{16}, \overline{A}=\overline{e_{16}}$ and $B=e_{26}, \overline{B}=\overline{e_{26}}$ respectively in Graph A.2 and Graph B.2 respectively. In Graph A.2 and Graph B.2, $A_6=e_{16}^{T}$, $A_1=e_{16}^{B}$, $B_6=e_{26}^{T}$ and $B_2=e_{26}^{B}$.
\end{notation}
\begin{lemma} The rigid edge formulation for Morse Matching has no R-cycles.
\end{lemma}
\begin{proof} Consider the rigid edge formulation namely $\mathcal{H}_R$. Now delete all the N-edges from this graph to obtain another graph say $\mathcal{H}_1$. In graph $\mathcal{H}_1$, the vertex set consists of $e_{ij}^B$ and $e_{ij}^T$ for every edge $e_{ij}\in \mathcal{G}$. It is easy to see that, in graph $\mathcal{H}_1$, all top nodes namely $e_{ij}^T, e_{ik}^T, \dots$ for edges $e_{ij}, e_{ik}\dots \in \mathcal{G}$ are source nodes and all the bottom nodes namely $e_{ij}^B, e_{ik}^B, \dots$ are sink nodes. Since every node in $\mathcal{H}_1$ is either a source node or a sink node and since we have accounted for all the R-edges of graph $\mathcal{H}_R$ in graph $\mathcal{H}_1$, we conclude that the RE-formulation for Morse Matching has no R-cycles.
\end{proof}
The two set of cycles i.e. (C-cycles and M-cycles) do not interfere. i.e. It can be easily seen that given an Orientaion of $\mathcal{H}_{R}$, say $\overleftrightarrow{\mathcal{H}_{R}}$ (that has cycles), if there exists an RN-cycle in $\overleftrightarrow{\mathcal{H}_{R}}$ which is ``composite" i.e. if this RN-cycle is composed of both the CR-edges as well as MR-edges, then there exists at least one M-cycle in $\overleftrightarrow{\mathcal{H}_{R}}$ such that both the N-edges of this M-cycle belong to the mixed cycle. In other words, every composite cycle will include an \emph{entire} M-cycle.
\begin{theorem} \label{thm:REMain} The rigid edge formulation ensures that there exists an RN-cycle in the rigid edge formulation that corresponds to a matching violation or to a cycle violation.
\end{theorem}
\begin{proof} (a.) Matching Violation $\Rightarrow$ RN-cycle: \\
Suppose there exists a matching violation between edges $e_{ij}$ and $e_{ik}$ which correspond to N-edges $e_{ij}^B\rightarrow e_{ij}^T$ and $e_{ik}^{B} \rightarrow e_{ik}^{T}$. Since we have the rigid edges $e_{ij}^T\rightarrow e_{ik}^B$ and $e_{ik}^T\rightarrow e_{ij}^B$ in our RE-formulation $\mathcal{H}_R$, we have the following RN-cycle $e_{ij}^B\rightarrow e_{ij}^T, e_{ij}^T\rightarrow e_{ik}^B, e_{ik}^{B} \rightarrow e_{ik}^{T}, e_{ik}^T\rightarrow e_{ij}^B$. \\
(b.) Cycle Violation $\Rightarrow$ RN-cycle: \\
Note that if we have $e_{ij}, e_{jk}$ and $e_{kl}$ as edges in original graph $\mathcal{H}_0$, then in our RE-formulation, we have CR-edge going from $e_{ij}^T$ to $e_{kl}^B$. Therefore, w.l.o.g., if we have edges $A-\overline{B}-C-\overline{D}-E-\overline{F}$ forming a path in $\mathcal{H}_0$ (where $A, C$ and $E$ are up-edges) then we have a corresponding RN-path in the RE-formulation since we have CR-edges going from $A^T \rightarrow C^{B}$ and $C^T \rightarrow E^B$. Since every path in $\mathcal{H}_0$ has a corresponding path in the RE-formulation, we can say the same thing about cycles. \\
(c.) RN-cycle $\Rightarrow$ A Matching Violation or a Cycle Violation: \\
The essential idea is the following: If no consecutive pair of $\overline{N}$-edges in an RN-cycle share a vertex, then all edges together form a cycle and we have a cycle violation. Else if we do have a pair of up-N-edges within the RN-cycle that share a vertex then we have a matching violation. Every RN-cycle therefore represents either a cycle or a matching violation. \\
In the RN-formulation, assume that the up-N-edge $e_{ij}$ is selected as part of the output. Edge $e_{ij}$ is the sole incoming edge for the vertex $e_{ij}^T$ whereas it will have more than one outgoing MR-edges each of the form $e_{ij}^T \rightarrow e_{ik}^B$ or $e_{ij}^{T} \rightarrow e_{kj}^B$. Now, either some such adjacent edge $e_{ik}$ or $e_{kj}$ is selected as part of the ouput or not selected. If we assume it is selected then there will be an RN-cycle of the form described in part a. of the proof which corresponds to a matching violation. Suppose we none of the adjacent edges of the form $e_{ik}$ or $e_{kj}$ are selected as part of the output. Now, if we are to have a cycle involving $e_{ij}$ then such a cycle must include the outgoing CR-edges from $e_{ij}^T$. As observed in part b. of the proof, such outgoing edges go from $e_{ij}^T$ to $e_{kl}^B$ everytime we have $e_{ij}, e_{jk}, e_{kl}$ as adjacent edges in graph $\mathcal{H}_0$. So if $e_{kl}$ is selected as part of the output we have $v_i, v_j, v_k \dots$ as a path in both $\mathcal{H}_0$. We inductively argue the same way by observing the outgoing CR-edges and MR-edges of $e_{kl}^T$. If an MR-edge is included as part of the cycle then a matching violation is assured and when we exclude the possibility of involving MR-edges each time, we extend the path using CR-edges only. These CR-edges mimic connectivity in $\mathcal{H}_0$. Therefore when this path involving strictly CR-edges becomes a cycle in $\mathcal{H}_R$, we can infer the existence of a corresponding cycle in $\mathcal{H}_0$
\end{proof}
\subsection{How to linearize: The Pseudo-FFT Gadget \label{sub:How-to-linearize:}}
\begin{definition}[Above \& Below Degrees]\label{def:abd} Consider the original Hasse graph $\mathcal{H}_\mathcal{K}(\mathcal{V},\mathcal{E})$ corresponding to complex $\mathcal{K}$. Given a vertex $v_{k}\in\mathcal{V}$, where vertex $v_{k}$ represents simplex $v_{k}^d\in\mathcal{K}$, the above degree $\mathcal{D_A}_{k}$ of $v_{k}$ equals the number of cofaces incident on simplex $v_{k}^d$ in comple $\mathcal{K}$. Also, the below degree $\mathcal{D_B}_{k}$ represents the dimension of the simplex $v_{k}^d$ in complex $\mathcal{K}$
\end{definition}
This procedure of adding forbidden edges to take care of matching requirements that is mentioned above is disadvantageous because we need to add $^{(\mathcal{D_A}_{k}+\mathcal{D_B}_{k})}C_{2}$ for each vertex $v_{k}^d\in\mathcal{V}$. For a simplicial complex $\mathcal{D_B}_{k}=(d+1)$ where $d$ is the dimension of the simplex corresponding to vertex $v_k^{d}$ in the Hasse graph. In applied contexts, for a simplicial/cubical complex, the maximum dimension is typically less than 4. So, $\mathcal{D_B}_{k}$ is often bounded by a small constant. However, $\mathcal{D_A}_{k}$ can easily be unbounded and can potentially be as large as $O(\mleft|\mathcal{V}\mright|)$. This would mean that in constructing the matching gadget, we may end up with quadratic number of forbidden edges (w.r.t. $\mleft|\mathcal{E}\mright|$) by adding $^{(\mathcal{D_A}_{k}+\mathcal{D_B}_{k})}C_{2}$ edges for each vertex $v_k$. Worst case complexity of matching gadget is therefore quadratic. The complexity bound $O(\mathcal{D_A}_{k}+\mathcal{D_B}_{k})$ i.e. linear number of edges per vertex $v_{k}\in\mathcal{V}$ is achievable if we use the pseudo-FFT-gadget as a replacement for the matching gadget.
Please see \autoref{fig:fft} and \autoref{fig:fft2} for illustrative examples. \autoref{fig:pseudorationale} provides further intuition. Essentially, we replace \emph{rigid edges} with \emph{rigid paths}. Superficially, this seems like an additional layer of complexity. The point to note however is that unlike rigid edges, these rigid paths are not \emph{exclusive} to the any two pair of vertices. If we look at \autoref{fig:pseudorationale}, what we really need to model is a domination relation from $X_1$ to $Y_2$ and from $Y_1$ to $X_2$. In matching gadget, we modeled these domination relations using the most elementary tool available to us: namely rigid edges. In case of pseudo-FFT, gadget we use \emph{non-exclusive} rigid paths to model domination relations from $X_1$ to $Y_2$ and from $Y_1$ to $X_2$ respectively. The non-exclusivity of these paths leads to a drastic reduction in complexity.
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (26) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (87.5pt, -125.0pt) {\textcolor{black}{$X_{1}$}};
\node (25) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (87.5pt, -200.0pt) {\textcolor{black}{$X_{2}$}};
\node (10) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (150.0pt, -125.0pt) {\textcolor{black}{$Y_{1}$}};
\node (4) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (150.0pt, -200.0pt) {\textcolor{black}{$Y_{2}$}};
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (4) to [in=241, out=119] (10);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (25) to [in=241, out=119] (26);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (26) to [in=61, out=299] (25);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (10) to [in=61, out=299] (4);
\draw [line width=0.625, ->, double, semitransparent,color=myindigo] (26) to (4);
\draw [line width=0.625, ->, double, semitransparent,color=prussianblue] (10) to (25);
\node at (144.375pt, -158.125pt) {\textcolor{purple}{$Y$}};
\node at (81.875pt, -158.125pt) {\textcolor{purple}{$X$}};
\node at (91.875pt, -166.875pt) {\textcolor{purple}{$\overline{X}$}};
\node at (154.375pt, -166.875pt) {\textcolor{purple}{$\overline{Y}$}};
\node at (119.5pt, -220.0pt) [rotate=0] {\textcolor{purple}{\textbf{Matching gadget} }};
\node (16) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (240pt, -125.0pt) {\textcolor{black}{$X_{1}$}};
\node (15) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (240pt, -200.0pt) {\textcolor{black}{$X_{2}$}};
\node (20) [regular polygon, regular polygon sides=6, minimum size=16.25pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (320pt, -125.0pt) {\textcolor{black}{$Y_{1}$}};
\node (24) [circle, minimum size=16.25pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (320pt, -200.0pt) {\textcolor{black}{$Y_{2}$}};
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (24) to [in=241, out=119] (20);
\draw [line width=1.875, ->, >=latex, semitransparent,color=purple!90] (15) to [in=241, out=119] (16);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (16) to [in=61, out=299] (15);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple!90] (20) to [in=61, out=299] (24);
\draw [line width=0.625, ->, decorate,decoration=zigzag, semitransparent,color=prussianblue] (16) to (24);
\draw [line width=0.625, ->, decorate,decoration=zigzag, semitransparent,color=prussianblue] (20) to (15);
\node at (314.375pt, -158.125pt) {\textcolor{purple}{$Y$}};
\node at (231.875pt, -158.125pt) {\textcolor{purple}{$X$}};
\node at (241.875pt, -166.875pt) {\textcolor{purple}{$\overline{X}$}};
\node at (324.375pt, -166.875pt) {\textcolor{purple}{$\overline{Y}$}};
\node at (280pt, -220.0pt) [rotate=0] {\textcolor{purple}{\textbf{ pseudo-FFT gadget} }};
\end{small}
\end{tikzpicture}
\end{center}
\vspace*{1\baselineskip}
\hrule
$\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad$\begin{tikzpicture}[thick]
\draw[line width=0.625, ->, double, semitransparent,color=myindigo] (3,2.5) -- (5.3,2.5);
\draw[line width=0.625, ->, decorate,decoration=zigzag, semitransparent,color=prussianblue] (3,2) -- (5.3,2);
\node at (6.5, 2.5) {\textcolor{myindigo}{Rigid Edge}};
\node at (6.5, 2) {\textcolor{prussianblue}{Rigid Path}};
\end{tikzpicture}%
\end{minipage}}
\protect\caption{pseudo-FFT rationale}
\label{fig:pseudorationale}
\end{figure}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{center}
\usetikzlibrary{shapes.geometric} \begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (1) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (131.25pt, -125.0pt) {\textcolor{black}{1}};
\node (3) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (81.25pt, -193.75pt) {\textcolor{black}{3}};
\node (2) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (112.5pt, -193.75pt) {\textcolor{black}{2}};
\node (4) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (150.0pt, -193.75pt) {\textcolor{black}{4}};
\node (5) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (181.25pt, -193.75pt) {\textcolor{black}{5}};
\node (6) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -75.0pt) {\textcolor{black}{6}};
\node (8) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (62.5pt, -75.0pt) {\textcolor{black}{8}};
\node (7) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (37.5pt, -75.0pt) {\textcolor{black}{7}};
\node (9) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (87.5pt, -75.0pt) {\textcolor{black}{9}};
\node (10) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (112.5pt, -75.0pt) {\textcolor{black}{10}};
\draw [line width=0.625, ->, color=black] (1) to [in=107, out=225] (2);
\draw [line width=0.625, ->, color=black] (2) to [in=287, out=45] (1);
\draw [line width=0.625, ->, color=black] (1) to [in=276, out=217] (6);
\draw [line width=0.625, ->, color=black] (1) to [in=294, out=191] (7);
\draw [line width=0.625, ->, color=black] (1) to [in=281, out=191] (8);
\draw [line width=0.625, ->, color=black] (1) to [in=281, out=162] (9);
\draw [line width=0.625, ->, color=black] (1) to [in=281, out=119] (10);
\draw [line width=0.625, ->, color=black] (6) to [in=197, out=299] (1);
\draw [line width=0.625, ->, color=black] (7) to [in=212, out=270] (1);
\draw [line width=0.625, ->, color=black] (8) to [in=198, out=270] (1);
\draw [line width=0.625, color=black] (9) to [in=162, out=281] (1);
\draw [line width=0.625, ->, color=black] (9) to [in=135, out=307] (1);
\draw [line width=0.625, ->, color=black] (10) to [in=101, out=299] (1);
\draw [line width=0.625, ->, color=black] (3) to [in=191, out=101] (1);
\draw [line width=0.625, ->, color=black] (1) to [in=29, out=259] (3);
\draw [line width=0.625, ->, color=black] (4) to [in=264, out=127] (1);
\draw [line width=0.625, ->, color=black] (5) to [in=287, out=143] (1);
\draw [line width=0.625, ->, color=black] (1) to [in=73, out=315] (4);
\draw [line width=0.625, ->, color=black] (1) to [in=101, out=331] (5);
\node at (116.25pt, -165.0pt) {\textcolor{black}{B}};
\node at (37.5pt, -136.875pt) {\textcolor{black}{E}};
\node at (43.125pt, -102.5pt) {\textcolor{black}{F}};
\node at (75.0pt, -91.25pt) {\textcolor{black}{G}};
\node at (115.0pt, -95.625pt) {\textcolor{black}{I}};
\node at (98.125pt, -102.5pt) {\textcolor{black}{H}};
\node at (88.75pt, -165.625pt) {\textcolor{black}{A}};
\node at (141.25pt, -163.75pt) {\textcolor{black}{C}};
\node at (163.75pt, -167.5pt) {\textcolor{black}{D}};
\node at (131.5pt, -210.0pt) {\textcolor{black}{\textbf{A. Hasse graph $H_1$ with duplicated edges} }};
\end{small}
\end{tikzpicture}
\end{center}
\vspace*{1\baselineskip}
\hrule
\vspace*{1\baselineskip}
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (1) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -18.75pt) {\textcolor{black}{$A_1$}}; \node (2) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -56.25pt) {\textcolor{black}{$A_3$}}; \node (3) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (37.5pt, -18.75pt) {\textcolor{black}{$B_1$}}; \node (4) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (37.5pt, -56.25pt) {\textcolor{black}{$B_2$}}; \node (5) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (62.5pt, -18.75pt) {\textcolor{black}{$C_1$}}; \node (6) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (62.5pt, -56.25pt) {\textcolor{black}{$C_4$}}; \node (7) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (87.5pt, -18.75pt) {\textcolor{black}{$D_1$}}; \node (8) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (87.5pt, -56.25pt) {\textcolor{black}{$D_5$}};
\node (9) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (112.5pt, -18.75pt) {\textcolor{black}{$E_6$}};
\node (10) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (112.5pt, -56.25pt) {\textcolor{black}{$E_1$}}; \node (11) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (137.5pt, -18.75pt) {\textcolor{black}{$F_7$}};
\node (12) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (137.5pt, -56.25pt) {\textcolor{black}{$F_1$}}; \node (13) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (162.5pt, -18.75pt) {\textcolor{black}{$G_8$}};
\node (14) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (162.5pt, -56.25pt) {\textcolor{black}{$G_1$}}; \node (15) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (187.5pt, -18.75pt) {\textcolor{black}{$H_9$}};
\node (16) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (187.5pt, -56.25pt) {\textcolor{black}{$H_1$}}; \node (17) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (212.5pt, -18.75pt) {\textcolor{black}{$I_{10}$}};
\node (18) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (212.5pt, -56.25pt) {\textcolor{black}{$I_1$}};
\draw [line width=0.625, ->, >=latex, color=black] (1) to [in=119, out=241] (2);
\draw [line width=0.625, ->, >=latex, color=black] (2) to [in=299, out=61] (1);
\draw [line width=0.625, ->, >=latex, color=black] (3) to [in=119, out=241] (4);
\draw [line width=0.625, ->, >=latex, color=black] (4) to [in=299, out=61] (3);
\draw [line width=0.625, ->, >=latex, color=black] (5) to [in=119, out=241] (6);
\draw [line width=0.625, ->, >=latex, color=black] (6) to [in=299, out=61] (5);
\draw [line width=0.625, ->, >=latex, color=black] (7) to [in=119, out=241] (8);
\draw [line width=0.625, ->, >=latex, color=black] (8) to [in=299, out=61] (7);
\draw [line width=0.625, ->, >=latex, color=black] (9) to [in=119, out=241] (10);
\draw [line width=0.625, ->, >=latex, color=black] (10) to [in=299, out=61] (9);
\draw [line width=0.625, ->, >=latex, color=black] (11) to [in=119, out=241] (12);
\draw [line width=0.625, ->, >=latex, color=black] (12) to [in=299, out=61] (11);
\draw [line width=0.625, ->, >=latex, color=black] (13) to [in=119, out=241] (14);
\draw [line width=0.625, ->, >=latex, color=black] (14) to [in=299, out=61] (13);
\draw [line width=0.625, ->, >=latex, color=black] (15) to [in=119, out=241] (16);
\draw [line width=0.625, ->, >=latex, color=black] (16) to [in=299, out=61] (15);
\draw [line width=0.625, ->, >=latex, color=black] (17) to [in=119, out=241] (18);
\draw [line width=0.625, ->, >=latex, color=black] (18) to [in=299, out=61] (17);
\node at (137.5pt, -70.0pt) {\textcolor{black}{\textbf{B. Hasse graph $H_2$ after edge pair isolation and cloning of vertices } }};
\end{small}
\end{tikzpicture}
\end{center}
\vspace*{1\baselineskip}
\hrule
\begin{center}
\begin{tikzpicture} [every node/.style={inner sep=0pt}]
\begin{small}
\node (1) [circle, minimum size=50.0pt, fill=violet!30, line width=0.625pt, draw=violet, double] at (162.5pt, -112.5pt) {\textcolor{black}{${I}_{+}$}};
\node (2) [circle, minimum size=50.0pt, fill=violet!30, line width=0.625pt, draw=violet, double] at (162.5pt, -212.5pt) {\textcolor{black}{${ABCDEFGH}_{+}$}};
\node (3) [circle, minimum size=50.0pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (237.5pt, -112.5pt) {\textcolor{black}{${ABCDEFGH}_{-}$}};
\node (4) [circle, minimum size=50.0pt, fill=lightmayanblue!50, line width=0.625pt, draw=myrobinegg, double] at (237.5pt, -212.5pt) {\textcolor{black}{${I}_{-}$}};
\node (5) [circle, minimum size=37.5pt, fill=PaleGreen, line width=0.625pt, draw=teal, double] at (100.0pt, -112.5pt) {\textcolor{black}{${I}_{+}$}};
\node (17) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (50.0pt, -112.5pt) {\textcolor{black}{${I}_{+}$}};
\node (27) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -112.5pt) {\textcolor{black}{$I_{10}$}};
\node (6) [circle, minimum size=37.5pt, fill=PaleGreen, line width=0.625pt, draw=teal, double] at (100.0pt, -187.5pt) {\textcolor{black}{${ABCD}_{+}$}};
\node (7) [circle, minimum size=37.5pt, fill=PaleGreen, line width=0.625pt, draw=teal, double] at (100.0pt, -250.0pt) {\textcolor{black}{${EFGH}_{+}$}};
\node (8) [circle, minimum size=37.5pt, fill=PaleGreen, line width=0.625pt, draw=teal, double] at (300.0pt, -62.5pt) {\textcolor{black}{${ABCD}_{-}$}};
\node (9) [circle, minimum size=37.5pt, fill=PaleGreen, line width=0.625pt, draw=teal, double] at (300.0pt, -187.5pt) {\textcolor{black}{${I}_{-}$}};
\node (10) [circle, minimum size=37.5pt, fill=PaleGreen, line width=0.625pt, draw=teal, double] at (300.0pt, -125.0pt) {\textcolor{black}{${EFGH}_{-}$}};
\node (18) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (50.0pt, -175.0pt) {\textcolor{black}{${AB}_{+}$}};
\node (19) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (50.0pt, -212.5pt) {\textcolor{black}{${CD}_{+}$}};
\node (20) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (50.0pt, -250.0pt) {\textcolor{black}{${EF}_{+}$}};
\node (21) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (50.0pt, -287.5pt) {\textcolor{black}{${GH}_{+}$}};
\node (22) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (350.0pt, -25.0pt) {\textcolor{black}{${AB}_{-}$}};
\node (23) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (350.0pt, -62.5pt) {\textcolor{black}{${CD}_{-}$}};
\node (24) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (350.0pt, -100.0pt) {\textcolor{black}{${EF}_{-}$}};
\node (25) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (350.0pt, -137.5pt) {\textcolor{black}{${GH}_{-}$}};
\node (26) [circle, minimum size=25.0pt, fill=mypink!60, line width=0.625pt, draw=myred, double] at (350.0pt, -175.0pt) {\textcolor{black}{${I}_{-}$}};
\node (38) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -168.75pt) {\textcolor{black}{$A_1$}};
\node (39) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -187.5pt) {\textcolor{black}{$B_1$}};
\node (40) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -206.25pt) {\textcolor{black}{$C_1$}};
\node (41) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -225.0pt) {\textcolor{black}{$D_1$}};
\node (42) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -243.75pt) {\textcolor{black}{$E_6$}};
\node (43) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -262.5pt) {\textcolor{black}{$F_7$}};
\node (44) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -281.25pt) {\textcolor{black}{$G_8$}};
\node (45) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (12.5pt, -300.0pt) {\textcolor{black}{$H_9$}};
\node (46) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -18.75pt) {\textcolor{black}{$A_3$}};
\node (47) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -37.5pt) {\textcolor{black}{$B_2$}};
\node (48) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -56.25pt) {\textcolor{black}{$C_4$}};
\node (49) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -75.0pt) {\textcolor{black}{$D_5$}};
\node (50) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -93.75pt) {\textcolor{black}{$E_1$}};
\node (51) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -112.5pt) {\textcolor{black}{$F_1$}};
\node (52) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -131.25pt) {\textcolor{black}{$G_1$}};
\node (53) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -150.0pt) {\textcolor{black}{$H_1$}};
\node (54) [circle, minimum size=12.5pt, fill=myorange!30, line width=0.625pt, draw=persianorange, double] at (387.5pt, -168.75pt) {\textcolor{black}{$I_1$}};
\node at (169.5pt, -330.0pt) {\textcolor{black}{\textbf{C. Hasse graph $H_3B.2$ on addition of pseudo-FFT rigid edges} }};
\node at (169.5pt, -340.0pt) {\textcolor{black}{\textbf{'+' suffix-nodes are from-nodes whereas '-' suffix-nodes are to-nodes} }};
\draw [line width=0.625, ->, >=latex, semitransparent,color=violet] (1) to (3);
\draw [line width=0.625, ->, >=latex, semitransparent,color=violet] (2) to (4);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (5) to (1);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (6) to (2);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (7) to (2);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (6) to (10);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (7) to [in=194, out=73] (8);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myrobinegg] (3) to (8);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myrobinegg] (3) to (10);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myrobinegg] (4) to (9);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (17) to (5);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (18) to (6);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (19) to (6);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (20) to (7);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (21) to (7);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (8) to (22);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (8) to (23);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (10) to (24);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (10) to (25);
\draw [line width=0.625, ->, >=latex, semitransparent,color=teal] (9) to (26);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (18) to [in=225, out=354] (23);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (19) to [in=162, out=81] (22);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (20) to [in=186, out=37] (25);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (21) to [in=261, out=342] (24);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (38) to (18);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (39) to (18);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (40) to (19);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (41) to (19);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (42) to (20);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (43) to (20);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (44) to (21);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (45) to (21);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (22) to (46);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (22) to (47);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (23) to (48);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (23) to (49);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (24) to (50);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (24) to (51);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (25) to (52);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (25) to (53);
\draw [line width=0.625, ->, >=latex, semitransparent,color=myred] (26) to (54);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (27) to [in=197, out=294] (17);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (38) to (47);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (39) to [in=169, out=61] (46);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (40) to [in=156, out=66] (49);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (41) to [in=163, out=66] (48);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (42) to [in=246, out=336] (51);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (43) to [in=246, out=336] (50);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (44) to [in=180, out=37] (53);
\draw [line width=0.625, ->, >=latex, semitransparent,color=purple] (45) to [in=246, out=343] (52);
\end{small}
\end{tikzpicture}
\end{center}%
\end{minipage}}
\protect\caption{The pseudo-FFT gadget for vertex $v_{2}$ (labelled '1' in this figure)}
\label{fig:fft}
\end{figure}
\begin{algorithm}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{ntabbing}
123\=123\=\kill
\prc \tbf{createNodes}($\mathscr{A},N$) \\
\tabinput Set of arcs $\mathscr{A}$ \\
\taboutput $\log n$ levels of hierarchically constructed ‘from-nodes’ $\mathcal{F}$ and ‘to-nodes’ $\mathcal{T}$\\
\> $\varcopyright\triangleright$ \alghalfindent \{ The $1^{\tnn{st}}$ level From-nodes $\mathcal{F}_{i}^1$ are the same as $\mathscr{A}_{i}^{T}$. \} \label{}\\
\> $\varcopyright\triangleright$ \alghalfindent \{ The $1^{\tnn{st}}$ level To-nodes $\mathcal{T}_{i}^1$ are the same as $\mathscr{A}_{i}^{B}$. \} \label{}\\
\> Assign labels $\llbracket i_{+} \rrbracket $ to the from-nodes $\mathcal{F}_{i}^1$. Assign labels $\llbracket i_{-} \rrbracket$ to the to-nodes $\mathcal{T}_{i}^1$. \label{}\\
\> Subsequently, $\forall \ell \textsf{ s.t. }\mleft\lceil N\mright\rceil /2^{\ell} \geq 1$ create $\mleft\lceil N\mright\rceil /2^{\ell}$ nodes for $\ell^{\tnn{th}}$ levels $\mathcal{F}^\ell$ and $\mathcal{T}^\ell$. \label{}\\
\> \treturn $\mathcal{N}_{\mathcal{L}}$ (i.e. the number of levels created including the $1^{\tnn{st}}$ level). \label{}\\
\> $\varcopyright\triangleright$ \alghalfindent \{ Label assignment for levels $\mathcal{F}^\ell$ and $\mathcal{T}^\ell$ for levels $\ell>1$ is deferred. \} \label{}\\
\\
\reset
\NoIndentAfterCmd \\
\prc $\bm{\mathcal{T}^{\ell}}\tbf{edges}$($\mathcal{T}^p,\mathcal{T}^q, M$) \\
\tabinput To-nodes of levels $\mathcal{T}^p$ and $\mathcal{T}^q$ \\
\NoIndentAfterCmd \\
\taboutput Arcs joining nodes of $\mathcal{T}^p$ to-nodes of $\mathcal{T}^q$. Labels for nodes of level $\mathcal{T}^{p}$.\\
\>Assign $i\longleftarrow 1, j \longleftarrow 1$.\label{}\\
\>\tdo \label{todo}\\
\>\> \tif $j\leq M-1$ \tthen \label{}\\
\>\>\algindent Create arcs $\mathcal{T}_{i}^p\rightarrow\mathcal{T}_{j}^q$ \tand $\mathcal{T}_{i}^p\rightarrow\mathcal{T}_{j+1}^q$. \label{todoublea}\\
\>\> \algindent Assign label $\mathscr{L}(\mathcal{T}_{i}^p)\longleftarrow \llbracket \mathscr{L}(\mathcal{T}_{j}^q) \Join \mathscr{L}(\mathcal{T}_{j+1}^q)\rrbracket; j \longleftarrow j+2$. \label{todoubleb}\\
\>\> \algindent $\varcopyright\triangleright$ \alghalfindent \{ The operator $\llbracket \mathpzc{x} \Join \mathpzc{y} \rrbracket$ appends string $\mathpzc{x}$ to string $\mathpzc{y}$.\} \label{}\\
\>\> \telse \label{}\\.
\>\>\algindent Create arcs $\mathcal{T}_{i}^p\rightarrow\mathcal{T}_{j}^q$; Assign label $\mathscr{L}(\mathcal{T}_{i}^p)\longleftarrow \llbracket \mathscr{L}(\mathcal{T}_{j}^q)_{-}\rrbracket; j \longleftarrow j + 1.$ \label{tosingle}\\
\>\> $i \longleftarrow i + 1$. \label{}\\
\>\twhile $j<M$ \label{todoends}\\
\\
\reset
\NoIndentAfterCmd \\
\prc $\bm{\mathcal{F}^{\ell}}\tbf{edges}$($\mathcal{F}^p,\mathcal{F}^q, \mathcal{T}^r, M$) \\
\tabinput from-nodes of levels $\mathcal{F}^p,\mathcal{F}^q$ and to-nodes of $\mathcal{T}^q$ \\
\NoIndentAfterCmd \\
\taboutput Arcs joining nodes of $\mathcal{F}^p$ to nodes of $\mathcal{F}^q$ \& $\mathcal{T}^r$. Labels for nodes of level $\mathcal{F}^{q})$.\\
\>Assign $i\longleftarrow 1, j \longleftarrow 1$.\label{}\\
\>\tdo \label{fordo}\\
\>\> \tif $i\leq M-1$ \tthen \label{}\\
\>\>\algindent Create arcs $\mathcal{F}_{i}^p\rightarrow\mathcal{F}_{j}^q$, $\mathcal{F}_{i+1}^p\rightarrow\mathcal{F}_{j}^q$ , $\mathcal{F}_{i}^p\rightarrow\mathcal{T}_{i+1}^r$ \tand $\mathcal{F}_{i+1}^p\rightarrow\mathcal{T}_{i}^r$. \label{fromdoublea}\\
\>\> \algindent Assign label $\mathscr{L}(\mathcal{F}_{j}^q)\longleftarrow \llbracket \mathscr{L}(\mathcal{F}_{i}^p) \Join \mathscr{L}(\mathcal{F}_{i+1}^p) \rrbracket; i \longleftarrow i+2$. \label{fromdoubleb}\\
\>\> \telse \label{}\\.
\>\>\algindent Create arcs $\mathcal{F}_{i}^p\rightarrow\mathcal{F}_{j}^q$; Assign label $\mathscr{L}(\mathcal{F}_{j}^q)\longleftarrow\llbracket \mathscr{L}(\mathcal{F}_{i}^p)_{-}\rrbracket; i \longleftarrow i + 1.$ \label{fromsingle}\\
\>\> $j \longleftarrow j + 1$. \label{}\\
\>\twhile $i<M$ \label{fordoends}\\
\\
\reset
\NoIndentAfterCmd \\
\prc \tbf{pseudo-FFT} ($\mathscr{A},N$) \\
\tabinput $\mathscr{A}$ is a set of N-edge pairs with cardinality $|\mathscr{A}|=N$ \\
\NoIndentAfterCmd \\
\taboutput The pseudo-FFT gadget that linearizes matching constraints\\
\> $\mathcal{N}_{\mathcal{L}}\longleftarrow$\tbf{createNodes}($\mathscr{A},N$) \label{}\\
\> $\forall \ell\in[1,\mathcal{N}_{\mathcal{L}}-1], \bm{\mathcal{T}^{\ell}}\tbf{edges}(\mathcal{T}^{\ell},\mathcal{T}^{\ell-1}, \mleft | \mathcal{T}^{\ell-1}\mright |)$ \label{allto}\\
\> $\forall \ell\in[1,\mathcal{N}_{\mathcal{L}}-1], \bm{\mathcal{F}^{\ell}}\tbf{edges}(\mathcal{F}^{\ell},\mathcal{F}^{\ell+1}, \mathcal{T}^{\ell}, \mleft | \mathcal{F}{^\ell}\mright |)$ \label{allfrom}\\
\> Create arcs $\mathcal{F}_1^{\mathcal{N}_{\mathcal{L}}}\rightarrow \mathcal{T}_2^{\mathcal{N}_{\mathcal{L}}}$ \& $\mathcal{F}_2^{\mathcal{N}_{\mathcal{L}}}\rightarrow \mathcal{T}_1^{\mathcal{N}_{\mathcal{L}}}$ \label{lastlevel}\\
\end{ntabbing}%
\end{minipage}}
\label{alg:pseudo-FFT}\protect\caption{Pseudo-FFT Construction}
\end{algorithm}
\begin{definition}[Truncated Binary Tree] We define a \textbf{truncated binary tree} to be any binary tree with the root node (and the edges incident on it) removed.
\end{definition}
\begin{definition}[Node Levels in Truncated Binary Trees] Given a binary tree $\mathscr{T}$ of depth $\delta$, let the \textbf{distances} of a node $\sigma\in\mathscr{T}$ from the root $\rho$ be denoted as $\mathpzc{d}(\sigma)$. We define the \textbf{rank} of a node in the tree as $\mathpzc{r}(\sigma)=\mathpzc{d}(\sigma)+1$. Finally, we define the \textbf{level} of a node $\sigma$ in the binary tree as $\mathpzc{l}(\sigma)=\delta +1-\mathpzc{r}(\sigma)$. Analogously, we may define the distance, the rank and the level of truncated binary trees by imagining a fictitious root vertex $\rho_{f}$ at the top of the truncated tree.
\end{definition}
\begin{definition}[up-directed, down-directed Binary Trees] A binary tree with directed edges is known as a directed binary tree. An \textbf{down-directed binary tree} is a binary tree with directed edges from the root to its children on the next level and in general from any parent node (in the sense of a binary tree) to a child node on the next level such that the root is the source. Also, all leaf nodes are sinks of directed paths that start from the root node. An \textbf{up-directed binary tree} is a tree that is similar to a down-directed binary tree, except for the fact that all its edge orientations are inverted. In this case, the root node is the unique sink of all directed paths and all leaf nodes are sources.
\end{definition}
\textbf{Truncated up-directed} (and \textbf{truncated down-directed})
binary trees are up-directed (and respectively down-directed trees)
with corresponding root nodes removed.
\begin{definition}[From-nodes and To-nodes] Given a set of paired normal edges $\{\mathscr{A}_i\}$, all nodes in the truncated up-directed tree with $\{\mathscr{A}_i^{T}\}$ as their leaf nodes are known as \textbf{from-nodes} whereas all nodes in the truncated down-directed tree with $\{\mathscr{A}_i^{B}\}$ as their leaf nodes are known as \textbf{to-nodes}. We refer to the corresponding trees as the \textbf{up-directed from-tree} and the \textbf{down-directed to-tree} respectively.
\end{definition}
\begin{notation}[Number of Levels $\mathcal{N}_\mathcal{L}$]
Define $\mathcal{N}_\mathcal{L}=$ total number of levels of the truncated from-tree $=$ total number of levels of truncated to-tree.
\end{notation}
\begin{definition}[Label Inheritance Property] We say that node $\mathpzc{v}$ has \textbf{s-label inheritance} if there exists a node $u$ s.t. $\mathpzc{u}\rightarrow \mathpzc{v}\implies \mathscr{L}(\mathpzc{u})\subseteq \mathscr{L}(\mathpzc{v})$. Analogously, we say that node $\mathpzc{u}$ has \textbf{d-label inheritance property} if there exists a node $\mathpzc{v}$ s.t. $\mathpzc{u}\rightarrow \mathpzc{v} \implies \mathscr{L}(\mathpzc{v})\subseteq \mathscr{L}(\mathpzc{u})$.
\end{definition}
\begin{definition}[Label Merging] We say that two labels $\mathscr{L}(\mathpzc{x})$ and $\mathscr{L}(\mathpzc{y})$ merge at a from-level $i$ if $\mathscr{L}(\mathpzc{x})\cap\mathscr{L}(\mathpzc{y})=\emptyset$ and if there exists a label $\mathscr{L}(\mathpzc{x})\cup\mathscr{L}(\mathpzc{y})$ at level $i+1$. A similar definition holds for labels at to-levels $j$ and $j+1$.
\end{definition}
\begin{definition}[Label Direct Inheritance] If there exists from-labels $\mathscr{L}(\mathpzc{x})$ and $\mathscr{L}(\mathpzc{y})$ at levels $i$ and $i+1$ respectively such that $\mathscr{L}(\mathpzc{x})=\mathscr{L}(\mathpzc{y})$ then we say that $\mathpzc{y}$ has \textbf{directly inherited} (i.e. without merging) its label from $\mathpzc{x}$. A similar notion of direct inheritance holds for nodes of to-levels $j$ and $j+1$.
\end{definition}
\begin{definition}[Binary Complements] Binary complements are defined as follows:
\begin{enumerate}
\item For level $p$ s.t. $1\leq p \leq \mathcal{N}_\mathcal{L}-1$, if we have three from-nodes say $\mathpzc{u}=\mathcal{F}_{i}^p, \mathpzc{v}=\mathcal{F}_{j}^p$ and $\mathpzc{w}=\mathcal{F}_{k}^{p+1}$ s.t. $\mathpzc{u}\rightarrow \mathpzc{w}$ and $\mathpzc{v}\rightarrow \mathpzc{w}$ then $\mathpzc{u}$ and $\mathpzc{v}$ are known as \textbf{binary from-complements}. In this case, $\mathscr{L}(\mathpzc{u})+\mathscr{L}(\mathpzc{v})=\mathscr{L}(\mathpzc{w})$.
\item The ultimate from-level $\mathcal{L}$ has at most two from-nodes. If the are precisely two in number, then they are considered to be binary from-complements.
\item Similarly, for level $p$ s.t. $1\leq p \leq \mathcal{N}_\mathcal{L}-1$, if we have three to-nodes $\mathpzc{x}=\mathcal{T}_{i}^p, \mathpzc{y}=\mathcal{T}_{j}^{p+1}$ and $\mathpzc{z}=\mathcal{T}_{k}^{p+1}$ s.t. $\mathpzc{x}\rightarrow \mathpzc{y}$ and $\mathpzc{x}\rightarrow \mathpzc{z}$ then $\mathpzc{y}$ and $\mathpzc{z}$ are known as \textbf{binary to-complements}. Here we have $\mathscr{L}(\mathpzc{y})+\mathscr{L}(\mathpzc{z})=\mathscr{L}(\mathpzc{x})$.
\item The ultimate to-level $\mathcal{L}$ has at most two from-nodes. If the are precisely two in number, they are considered to be binary to-complements.
\item If $FCC_{i}^p$ and $\mathcal{F}_{j}^p$ make binary form-complements and $TCC_{i}^p$ and $\mathcal{T}_{j}^p$ are respective binary to-complements at level $p$, then the pair of nodes $\{\mathcal{F}_{i}^p, \mathcal{T}_{j}^p\}$ (as well as the pair $\{\mathcal{T}_{i}^p, \mathcal{F}_{i}^p\}$) form \textbf{mirror-binary complements}.
\end{enumerate}
\end{definition}
If the number of nodes at a particular level are odd, then all nodes
except the last node of that level has binary complements.
\begin{definition}[Label Domination] If we have an edge $\mathpzc{x}\rightarrow\mathpzc{y}$ between two vertices $\mathpzc{x}$ and $\mathpzc{y}$ then we say that $\mathpzc{x}$ dominates $\mathpzc{y}$. Also, the relation of domination is transitive. i.e. if we have $\mathpzc{x}\rightarrow\mathpzc{y}$ and $\mathpzc{y}\rightarrow\mathpzc{z}$ then we say that $\mathpzc{x}$ dominates $\mathpzc{z}$.
\end{definition}
Therefore this gadget introduces three types of edges $\mathpzc{x}\rightarrow\mathpzc{y}$,
namely:
\begin{enumerate}
\item the from-node edges $\mathpzc{x}\rightarrow\mathpzc{y}$ s.t. $\mathscr{L}(\mathpzc{x})\subseteq\mathscr{L}(\mathpzc{y})$
(via s-label inheritance)
\item the to-node edges $\mathpzc{x}\rightarrow\mathpzc{y}$ s.t. $\mathscr{L}(\mathpzc{y})\subseteq\mathscr{L}(\mathpzc{x})$
(via d-label inheritance) and
\item the from-to edges $\mathpzc{x}\rightarrow\mathpzc{y}$ s.t the from-node $\mathpzc{x}$
dominates its mirror complement to-node $\mathpzc{y}$.
\end{enumerate}
\begin{lemma} \label{lem:inherit} All from-nodes of the FFT gadget have s-label inheritance whereas all to-nodes have d-label inheritance.
\end{lemma}
\begin{proof} For the proof of this lemma we would referring to the construction as delineated in Algorithm 2. We begin with observing $\toproc$. Let $\mathpzc{u}=\mathcal{T}_{i}^p, \mathpzc{v}=\mathcal{T}_{j}^q$ and $\mathpzc{w}=\mathcal{T}_{j+1}^q$ Line~\ref{todoublea} of $\toproc$ creates edges $\mathpzc{u}\rightarrow \mathpzc{v}$ and $\mathpzc{u}\rightarrow \mathpzc{w}$ whereas Line~\ref{todoubleb} of $\toproc$ is equivalent to saying $\mathscr{L}(\mathpzc{v})\subseteq \mathscr{L}(\mathpzc{u})$ and $\mathscr{L}(\mathpzc{w})\subseteq \mathscr{L}(\mathpzc{u})$. For the alternative case on Line~ \ref{tosingle}, we have $\mathpzc{u}\rightarrow \mathpzc{v}$ followed by $\mathscr{L}(\mathpzc{v})\subseteq \mathscr{L}(\mathpzc{u})$. Since all to-nodes labels are created in $\toproc$, we conclude from the two cases above that to-nodes have the d-label inheritance property.
Now we turn our attention to $\fromproc$. Let $\mathpzc{t}=\mathcal{F}_{i}^p, \mathpzc{u}=\mathcal{F}_{i+1}^p, \mathpzc{v}=\mathcal{F}_{j}^q$. Line~\ref{fromdoublea} of $\fromproc$ creates edges $\mathpzc{t}\rightarrow \mathpzc{v}$ and $\mathpzc{u}\rightarrow \mathpzc{v}$. Moreover, Line~\ref{fromdoubleb} assigns label to $\mathpzc{v}$ in a manner such that $\mathscr{L}(\mathpzc{t})\subseteq \mathscr{L}(\mathpzc{v})$ and $\mathscr{L}(\mathpzc{u})\subseteq \mathscr{L}(\mathpzc{v})$. Finally, for the alternative case in Line~\ref{fromsingle}, $\mathpzc{t}\rightarrow \mathpzc{v}$ followed by $\mathscr{L}(\mathpzc{t})\subseteq \mathscr{L}(\mathpzc{v})$. All from-node labels are created in $\fromproc$. Hence, from-nodes have s-label inheritance property.
\end{proof}
\begin{lemma}[Existence and Uniqueness of Labels for Every From-Level] \label{lem:labeleveryfromlevel} Suppose we are given a set of N-edge pairs $\mathscr{A}$ with cardinality $|\mathscr{A}|=N$. Then, $\forall i\in[1\dots N]$, the labels $\mathscr{L}(\mathcal{F}_{i}^1)$ where $\mathcal{F}_{i}^1=\mathscr{A}_{i}^{T}$ are s.t. there exists a set of indices $\{i_p\}$ that satisfy the \textbf{Label Property} namely:
\[\mathscr{L}(\mathcal{F}_{i}^1)\subseteq\mathscr{L}(\mathcal{F}_{i_p}^p)\quad \forall \:2\leq p \leq \mathcal{N}_\mathcal{L}\] Moreover, such an index $i_p$ that satisfies the label property for a given label $\mathscr{L}(\mathcal{F}_{i}^1)$ is unique for every level $p$.
\end{lemma}
\begin{proof} To begin with, note that, the subroutine $\bm{\mathcal{F}^{\ell}}\tbf{edges}$ is called from Line~\ref{allfrom} of subroutine \tbf{pseudo-FFT}. For each $\ell$, edges joining nodes of $\mathcal{F}^{\ell}$ to nodes of $\mathcal{F}^{\ell+1}$ are formed. Denote the total number of nodes of level $\mathcal{F}^{\ell}$ (namely $|\mathcal{F}^{\ell}|$) by $M$. Also, let $W=\nicefrac{(M-1)}{2}+1$. In the while loop of Lines \ref{fordo}--\ref{fordoends} of subroutine $\bm{\mathcal{F}^{\ell}}\tbf{edges}$, if $M$ is even then, $\forall i$, s.t. $1\leq i \leq \nicefrac{M}{2}$, we merge labels of pairs of nodes $(\mathpzc{x}_i,\mathpzc{x}_{i+1})$ of level $\mathcal{F}^{\ell}$ by creating edges $\mathpzc{x}_i\rightarrow\mathpzc{y}_j$ and $\mathpzc{x}_{i+1}\rightarrow\mathpzc{y}_j$ between such a pair with a unique vertex $\mathpzc{y}_j$ of level $\mathcal{F}^{\ell+1}$. For each level $\ell$, there will be $\nicefrac{M}{2}$ such pairs of level $\mathcal{F}^{\ell}$ that form edges with $\nicefrac{M}{2}$ nodes of level $\mathcal{F}^{\ell+1}$. Else if the number $M=|\mathcal{F}^{\ell}|$ is odd, then we merge labels of $\nicefrac{(M-1)}{2}$ pairs of nodes of level $\mathcal{F}^{\ell}$ namely $(\mathpzc{x}_i,\mathpzc{x}_{i+1})$ by creating edges between each pair of nodes of $\mathcal{F}^{\ell}$ with a unique node $\mathpzc{y}_j$ of level $\mathcal{F}^{\ell+1}$ (as before) whereas the final node of $\mathcal{F}^{\ell+1}$ namely $\mathpzc{y}_{W}$ direct inherits its label from final node $\mathpzc{x}_M$ of level $\mathcal{F}^{\ell}$ via edge $\mathpzc{x}_M\rightarrow\mathpzc{y}_{W}$. In both cases, whether even or odd, for every node $\mathpzc{x}$ in $\mathcal{F}^{\ell}$, we have exactly one node $\mathpzc{y}$ in level $\mathcal{F}^{\ell+1}$ such that $\mathpzc{x}\rightarrow\mathpzc{y}$. This gives us \emph{per level} uniqueness property. Now, since the subroutine $\bm{\mathcal{F}^{\ell}}\tbf{edges}$ is invoked for every consecutive pair of levels $(\ell,\ell+1)$ within subroutine \tbf{pseudo-FFT}, we conclude that by:
\begin{enumerate*}[label=(\textbf{\alph{enumi})}]
\item transitivity of dominance relations formed by directed edges across nodes of successive levels
\item and owing to \emph{per level} uniqueness property as discussed above,
\end{enumerate*}
the label property as described in the statement of the theorem is satisfied.
\end{proof}
\begin{lemma}[Existence and Uniqueness of Labels for Every To-Level] \label{lem:labeleverytolevel} Suppose we are given a set of N-edge pairs $\mathscr{A}$ with cardinality $|\mathscr{A}|=N$. Then, $\forall j\in[1\dots N]$, the labels $\mathscr{L}(\mathcal{T}_{j}^1)$ where $\mathcal{T}_{j}^1=\mathscr{A}_{j}^{B}$ are s.t. there exists a set of indices $\{j_p\}$ that satisfy the \textbf{Label Property} namely:
\[\mathscr{L}(\mathcal{T}_{j}^1)\subseteq\mathscr{L}(\mathcal{T}_{j_p}^p)\quad \forall \:2\leq p \leq \mathcal{N}_\mathcal{L}\] Moreover, such an index $j_p$ that satisfies the label property for a given label $\mathscr{L}(\mathcal{T}_{j}^1)$ is unique for every level $p$.
\end{lemma}
\begin{proof}
The subroutine $\bm{\mathcal{T}^{\ell}}\tbf{edges}$ is invoked in Line~\ref{allto} of subroutine \tbf{pseudo-FFT}. For each $\ell$, edges joining nodes of $\mathcal{T}^{\ell}$ to nodes of $\mathcal{T}^{\ell-1}$ are formed. Denote the total number of nodes of level $\mathcal{F}^{\ell}$ (namely $|\mathcal{F}^{\ell}|$) by $M$. Also, let $W=\nicefrac{(M-1)}{2}+1$. In the while loop of Lines \ref{todo}--\ref{todoends} of subroutine $\bm{\mathcal{T}^{\ell}}\tbf{edges}$, if $M$ is even then, $\forall i$, s.t. $1\leq i \leq \nicefrac{M}{2}$, we merge labels of pairs of nodes $(\mathpzc{x}_i,\mathpzc{x}_{i+1})$ of level $\mathcal{T}^{\ell-1}$ by creating edges $\mathpzc{y}_j\rightarrow\mathpzc{x}_i$ and $\mathpzc{y}_j\rightarrow\mathpzc{x}_{i+1}$ between such a pair with a unique vertex $\mathpzc{y}_j$ of level $\mathcal{T}^{\ell}$. For each level $\ell$, we have $\nicefrac{M}{2}$ such pair of nodes from level $\mathcal{T}^{\ell-1}$ that form edges with $\nicefrac{M}{2}$ nodes of level $\mathcal{T}^{\ell}$. Else if the number $M=|\mathcal{T}^{\ell-1}|$ is odd, then we merge labels of $\nicefrac{(M-1)}{2}$ pairs of nodes of level $\mathcal{T}^{\ell-1}$ namely $(\mathpzc{x}_i,\mathpzc{x}_{i+1})$ by creating edges $\mathpzc{y}_j\rightarrow\mathpzc{x}_i$ and $\mathpzc{y}_j\rightarrow\mathpzc{x}_{i+1}$ between each pair of nodes of $\mathcal{T}^{\ell-1}$ with a unique node $\mathpzc{y}_j$ of level $\mathcal{T}^{\ell}$ (as before) whereas the final node of $\mathcal{T}^{\ell-1}$ namely $\mathpzc{y}_{W}$ direct inherits its label from final node $\mathpzc{x}_M$ of level $\mathcal{T}^{\ell}$ via the edge $\mathpzc{y}_{W}\rightarrow\mathpzc{x}_M$. In both cases, whether even or odd, for every node $\mathpzc{x}$ in $\mathcal{T}^{\ell-1}$, we have exactly one node $\mathpzc{y}$ in level $\mathcal{T}^{\ell}$ such that $\mathpzc{y}\rightarrow\mathpzc{x}$. This gives us \emph{per level} uniqueness property. Now, (symmetric to the argument made in \autoref{lem:labeleveryfromlevel}) since the subroutine $\bm{\mathcal{T}^{\ell}}\tbf{edges}$ is invoked for every consecutive pair of levels $(\ell,\ell+1)$ within subroutine \tbf{pseudo-FFT}, we conclude that by:
\begin{enumerate*}[label=(\textbf{\alph{enumi})}]
\item transitivity of dominance relations formed by directed edges across nodes of successive levels
\item and owing to \emph{per level} uniqueness property as described above,
\end{enumerate*}
the label property as described in the statement of the theorem is satisfied
\end{proof}
\begin{theorem} \label{thm:matchmimic} Given a set of edges $\mathscr{A}$ we can ensure that every node of the type $\mathscr{A}_i^{T}$ dominates every node of the type $\mathscr{A}_j^{B}$ (for $i\neq j$) using rigid edges that are part of the pseudo-FFT-gadget.
\end{theorem}
\begin{proof} Consider two edges $\mathscr{A}_i (\mathscr{A}_i^{T},\mathscr{A}_i^{B})$ and $\mathscr{A}_j (\mathscr{A}_j^{T},\mathscr{A}_j^{B})$. We will show that there exists a directed path from $\mathscr{A}_i^{T}$ to $\mathscr{A}_j^{B}$ and another directed path from $\mathscr{A}_j^{T}$ to $\mathscr{A}_i^{B}$. Let $\mathcal{F}_{j}^1=\mathscr{A}_j^{T}$ and $\mathcal{F}_{i}^1=\mathscr{A}_i^{T}$. Also, let $\mathcal{T}_{j}^1=\mathscr{A}_j^{B}$ and $\mathcal{T}_{i}^1=\mathscr{A}_i^{B}$. We know from \autoref{lem:labeleveryfromlevel}, that the label property will be satisfied for all subsequent from-levels $2\leq p \leq \mathcal{N}_\mathcal{L}$. Let $\mathcal{F}_{i_p}^p$ be such that $\mathscr{L}(\mathcal{F}_{i}^1)\subseteq\mathscr{L}(\mathcal{F}_{i_p}^p)$ and $\mathcal{F}_{j_p}^p$ be such that $\mathscr{L}(\mathcal{F}_{j}^1)\subseteq\mathscr{L}(\mathcal{F}_{j_p}^p)$ We consider two cases here: \\
\textbf{Case 1:} The labels $\mathcal{F}_{j_q}^q$ and $\mathcal{F}_{i_q}^q$ are binary complements for some from-level $1\leq q \leq \mathcal{N}_\mathcal{L}-1$. In this case, from Line~\ref{fromdoublea} of subroutine $\bm{\mathcal{F}^{\ell}}\tbf{edges}$ we concur that edges are formed between the mirror-binary complements $\mathcal{F}_{i_q}^q\rightarrow\mathcal{T}_{j_q}^q$ and $\mathcal{F}_{j_q}^q\rightarrow\mathcal{T}_{i_q}^q$. However, using s-label inheritance property along with label property for from-nodes, we know that $\mathcal{F}_{i}^1=\mathscr{A}_i^{T}$ dominates $\mathcal{F}_{i_q}^q$ (also, $\mathcal{F}_{j}^1=\mathscr{A}_j^{T}$ dominates $\mathcal{F}_{j_q}^q$) and by using d-label inheritance property along with label property for to-nodes, we know that $\mathcal{T}_{j_q}^q$ dominates $\mathcal{T}_{j}^1=\mathscr{A}_j^{B}$ (also, $\mathcal{T}_{i_q}^q$ dominates $\mathcal{T}_{i}^1=\mathscr{A}_i^{B}$). Thus for this particular case the statement of the above theorem holds.\\
\textbf{Case 2:} Now consider the case when $\mathcal{F}_{j_q}^q$ and $\mathcal{F}_{i_q}^q$ are not binary complements for any of the from-levels $1\leq q \leq \mathcal{N}_\mathcal{L}-1$. Since there are only two from-nodes for level $L=\mathcal{N}_\mathcal{L}$, one of the nodes is $\mathcal{F}_{i_L}^L$ and the other is $\mathcal{F}_{j_L}^L$. By definition, the two nodes of the final level are always binary complements and from Line~\ref{lastlevel} of subroutine \textbf{pseudo-FFT} we have edges between mirror-binary complements. Once again, (as before in Case 1) by applying s-level inheritance property and label property for from-nodes for $\mathcal{F}_{i_L}^L$ and d-level inheritance property for $\mathcal{T}_{j_L}^L$ along with label property for to-nodes, we conclude that the statement of the theorem above holds for final level mirror binary complements
\\ The two cases exhaust all possibilities (i.e. two given labels $\mathscr{A}_i^{T}$ and $\mathscr{A}_j^{T}$ have labels $\mathcal{F}_{i_q}^q$ and $\mathcal{T}_{j_q}^q$ containing them that form binary complements at exactly one level $1\leq q \leq \mathcal{N}_\mathcal{L}$ and at that particular level we have edges joining mirror binary complements which ensure that domination holds through s-label and d-label inheritance properties.) Hence proved.
\end{proof}
\begin{notation} We use the notation $a\rightsquigarrow b$ to indicate a path joining nodes $a$ and $b$ and the notation $a\rightarrow b$ to indicate an edge joining nodes $a$ and $b$.
\end{notation}
\begin{theorem} The pseudo-FFT gadget and the matching gadget are equivalent.
\end{theorem}
\begin{proof} \textbf{Part I} Matching $\Rightarrow$ pseudo-FFT \\
Suppose there exists a rigid cycle involving normal edges $\mathscr{A}_i^{B}\rightarrow\mathscr{A}_i^{T}$, $\mathscr{A}_j^{B}\rightarrow\mathscr{A}_j^{T}$ and rigid edges belonging to matching gadgets $\mathscr{A}_i^{T}\rightarrow\mathscr{A}_j^{B}$ and $\mathscr{A}_j^{T}\rightarrow\mathscr{A}_i^{B}$. This would mean that there is a matching conflict involving edges $\mathscr{A}_i$ and $\mathscr{A}_j$. Using \autoref{thm:matchmimic}, in turn, implies that there will be corresponding rigid paths from $\mathscr{A}_i^{T}\rightsquigarrow\mathscr{A}_j^{B}$ and $\mathscr{A}_j^{T}\rightsquigarrow\mathscr{A}_i^{B}$ respectively for the pseudo-FFT gadget. The presence of normal edges $\mathscr{A}_i^{B}\rightarrow\mathscr{A}_i^{T}$ and $\mathscr{A}_j^{B}\rightarrow\mathscr{A}_j^{T}$ will therefore result into a cycle for the pseudo-FFT gadget (just as it would if we were to use the ematching gadget instead).\\
\textbf{Part II} pseudo-FFT $\Rightarrow$ Matching \\
To begin with, observe that rigid edges belonging to the pseudo-FFT gadget either join \begin{enumerate*}[label=(\textbf{\alph{enumi})}]
\item a lower level from-node to a higher-level from-node
\item or a higher level to-node to a lower level to-node
\item or a from-node to a to-node of the same level,
\end{enumerate*}
thus maintaining a hierarchy. The from-nodes and to-nodes of level $p$ such that $2\leq p \leq \mathcal{N}_\mathcal{L}$ do not have any normal edges through them. Consider an RN-cycle $\mathscr{C}$ involving pseudo-FFT edges of the form $\mathscr{A}_1^{B}\rightarrow\mathscr{A}_1^{T}$, $\mathscr{A}_1^{T} \rightsquigarrow\mathscr{A}_2^{B}$, $\mathscr{A}_2^{B}\rightarrow\mathscr{A}_2^{T}$, $\dots,$ $\mathscr{A}_n^{B}\rightarrow\mathscr{A}_n^{T},$ $\mathscr{A}_n^{T}\rightsquigarrow\mathscr{A}_1^{B}$. Suppose that the FFT-rigid path $\mathscr{A}_i^{T}\rightsquigarrow \mathscr{A}_j^{B}$ is part of RN-cycle $\mathscr{C}$. Then the corresponding rigid path $\mathscr{A}_j^{T}\rightsquigarrow \mathscr{A}_i^{B}$ is also a part of the FFT-gadget (These rigid paths mirror the rigid edges $\mathscr{A}_i^{T}\rightarrow\mathscr{A}_j^{B}$ and $\mathscr{A}_j^{T}\rightarrow\mathscr{A}_i^{B}$ of the matching gadget.). Also, if $\mathscr{A}_i^{T}\rightsquigarrow \mathscr{A}_j^{B}$ is part of the RN-cycle then clearly $\overline{N}$-edges $\mathscr{A}_i^B\rightarrow\mathscr{A}_i^T$ and $\mathscr{A}_j^B\rightarrow\mathscr{A}_j^T$ are also a part of the RN-cycle $\mathscr{C}$. Now, this ensures that a matching conflict between edges $\mathscr{A}_i^B\rightarrow\mathscr{A}_i^T$ and $\mathscr{A}_j^B\rightarrow\mathscr{A}_j^T$ occurs because of presence of the sequence of edges and paths: $\mathscr{A}_i^B\rightarrow\mathscr{A}_i^T$, $\mathscr{A}_i^{T}\rightsquigarrow\mathscr{A}_j^{B}$, $\mathscr{A}_j^B\rightarrow\mathscr{A}_j^T$, and $\mathscr{A}_j^{T}\rightsquigarrow\mathscr{A}_i^{B}$. So if there is a cycle violation in pseudo-FFT gadget there will be at least one cycle violation in the corresponding matching gadget. In fact, in the above case, every pair $\mathscr{A}_i, \mathscr{A}_j$ where $i,j<n$ and $\mathscr{A}_i^{T}\rightsquigarrow \mathscr{A}_j^{B}$ is an FFT-rigid path results into an RN-cycle indicative of a matching conflict. Hence proved.
\end{proof}
\begin{theorem} \label{thm:REMain2} Suppose that we replace matching gadgets by the pseudo-FFT gadgets for matching conflicts at every vertex in the Hasse graph, then the rigid edge formulation ensures that there exists an RN-cycle in the rigid edge formulation that corresponds to a matching violation or to a cycle violation.
\end{theorem}
\begin{proof} The proof of this theorem mirrors the proof of \autoref{thm:REMain} if we replace every rigid edge $\mathscr{A}^{T}\rightarrow \mathscr{B}^{B}$ of the matching gadget with rigid path $\mathscr{A}^{T}\rightsquigarrow \mathscr{B}^{B}$ of the pseudo-FFT-gadget.
\end{proof}
\subsection{Objective: From Gradient Pairs to Critical Cells }
To begin with, we now need to translate the count of number of non-matching
edges incident per vertex to the count of critical cells. Let $\mathcal{C}(v_{i})$
be a function which is $0$ when $v_{i}$ is matched and $1$ when
it is unmatched. Let $\overline{\mathcal{E}}_{i}$ be the set of $\underline{N}$-edges
incident on vertex $v_{i}\in\mathcal{H}$ where $\mathcal{H}$ is the original Hasse
graph. Consider the term: ${\normalcolor \Upsilon(v_{i})=\sum\limits _{\overline{e_{ij}}\in\overline{\mathcal{E}}_{i}}\overline{e_{ij}}-\mleft(d(v_{i})-1\mright)}$.
By definition, the maximum number of incident $\underline{N}$-edges
$\left|\overline{\mathcal{E}}_{i}\right|=d\left(v_{i}\right).$ If all of
them are unmatched then $\left|\overline{\mathcal{E}}_{i}\right|=d\left(v_{i}\right)$
whereas if one of them is unmatched then $\left|\overline{\mathcal{E}}_{i}\right|=d\left(v_{i}\right)-1$.
If $v_{i}$ is matched with some vertex $v_{j}$, then by definition,
$\underline{N}$-edge $\overline{e_{ij}}=0$ whereas $\overline{e_{ij}}=1$
for all $j$ such that $v_{i},v_{j}$ are not matched. So, when $v_{i}$
is matched we have, $\Upsilon(v_{i})=\left(d\left(v_{i}\right)-1\right)-\left(d\left(v_{i}\right)-1\right)=0$,
whereas when $v_{i}$ is unmatched we have, $\Upsilon(v_{i})=\left(d\left(v_{i}\right)\right)-\left(d\left(v_{i}\right)-1\right)=1$.
Let $\mathcal{C}(v_{i})$ be an indicator variable for criticality of vertex
$v_{i}$. Therefore, $\mathcal{C}(v_{i})=\Upsilon(v_{i})$.
We formulate the objective thus:
{\footnotesize{}\begin{equation}
\begin{aligned}
\Upsilon(v_i) & = \sum\limits_{\overline{e_{ij}}\in\overline{\mathcal{E}}_{i}} \overline{e_{ij}} - \mleft (d(v_i) -1\mright)\\
\Upsilon(\mathcal{H}) & = \sum\limits_{v_i\in\mathcal{V}}\mleft[\sum\limits_{\overline{e_{ij}}\in\overline{\mathcal{E}}_{i}} \overline{e_{ij}} - \mleft (d(v_i) -1\mright)\mright] \\
& = \sum\limits_{v_i\in\mathcal{V}} \mathcal{C}(v_i)
\label{eq:MMUPobj}
\end{aligned}
\end{equation}}{\footnotesize \par}
Recall that, the reason MMUP approximates the number of criticalities in terms of aggregate Morse number whereas MMMP approximates number of gradient pairs (Morse matchings) in terms optimal number of gradient pairs. As per the SMOC condition described in , if we were to assume that the optimal aggregate Morse number closely models the aggregate Betti number then, the approximation of MMUP is equivalently an approximation of the Betti numbers. Thus under SMOC, MMUP as a problem-model offers significant advantage over MMMP since it allows us to approximate the aggregate Betti numbers using a significantly smaller complex.
\begin{theorem}\label{thm:minred}
There is an approximation preserving reduction from MMUP/MMFEP to min-POP.
\end{theorem}
\begin{proof}
The proof is relatively straightforward. From \autoref{eq:MMUPobj} along with \autoref{thm:REMain} (for min-POP using cycle+matching gadgets) or \autoref{thm:REMain2} (for cycle+pseudo-FFT gadgets), we conclude that the values of the objective functions itself are, in fact, \emph{equal} for MMUP and the corresponding min-POP problem. This means that Condition 2 of \emph{approximation preserving reductions} as specified in \autoref{eq:apcond2} is satisfied. Since objective values are equal for all candidate solutions, they are also equal for the optimal solution. Therefore, Condition 1 of \emph{approximation preserving reductions} as specified in \autoref{eq:apcond1} is satisfied. The reduction procedures are clearly polynomial time ( quadratic if cycle+matching gadgets are used and linear if cycle+pseudo-FFT gadgets are used). The procedure to construct the solution for MMUP given a solution for min-POP is also linear. All that is required to obtain the MMUP solution is to count the number of $\underline{N}$ from the min-POP solution. Once we have a count of $\underline{N}$ edges, from \autoref{eq:MMUPobj} we get the number of criticalities (unmatched nodes) and we are done. Identical reasoning holds for MMFEP to min-POP reduction.
\end{proof}
\begin{theorem} \label{thm:linred} The approximation preserving reduction from MMUP/MMFEP to min-POP is linear in space and time.
\end{theorem}
\begin{proof} Note that during the edge isolation procedure the number of vertices increase. Specifically, we start with $2 \times |\mathcal{E}|$ vertices as opposed to starting with $|\mathcal{V}|$ vertices. The pseudo-FFT gadget adds linear number of nodes and edges which eventually makes the entire graph reduction linear. The argument goes as follows:
\begin{itemize}
\item For every vertex $v$ in the Hasse graph, we need to construct a pseudo-FFT gadget. Level 1 has $\mathcal{D}_v$ nodes (where $\mathcal{D}_v$ is the sum of above degree $\mathcal{D_A}_v$ and below degree $\mathcal{D_B}_v$ of vertex $v$ in the original Hasse graph). Level k has $\left\lceil \nicefrac{\mathcal{D}_v}{2^k} \right\rceil$ new nodes.
\item The total number of levels are bounded by $\log \mathcal{D}_v$.
\item From a simple power series calculation we can deduce that the total number of newly introduced from-nodes (from label merging) in the pseudo-FFT gadget for some vertex $v$ is upper bounded by $\mathcal{D}_v$. Also, since there are $\log \mathcal{D}_v$ number of from-levels, the number of from-nodes created via direct label inheritance are upper bounded by $\log \mathcal{D}_v$.
\item Similar counting argument can be made for total number of to-nodes. Hence, there are in all $2 \times \mleft(\mathcal{D}_v + \log \mathcal{D}_v\mright)$ number of newly introduced nodes.
\item Total number of nodes (original + newly introduced) will be $3\times \mathcal{D}_v + 2 \times \log \mathcal{D}_v$.
\item Each node in pseudo-FFT gadget has at most two outgoing rigid edges. Therefore total number of rigid edges introduced per vertex $v$ in the original Hasse graph will be bounded by $6\times \mathcal{D}_v + 4 \times \log \mathcal{D}_v$. For a sufficiently large $\mathcal{D}_v$, this estimate is upper bounded by $7 \times \mathcal{D}_v$
\item Each edge is counted once as part of above degree of some vertex and once as the below degree. When summed across all vertices in the Hasse graph, we get an upper bound of $14 \times \mathcal{E}$ on the number of newly introduced, taking into account every node in the Hasse graph.
\item The cycle gadget introduces $|\mathcal{E}|$ new rigid edges. Therefore, number of rigid edges introduced (cycle+pseudo-FFT) will be upper bounded by $15 \times \mathcal{E}=O(\mathcal{E})$.
\item Also, we assume that the dimension of the input simplicial complex is a small constant. So, we have $O(|\mathcal{E}|)=O(|\mathcal{V}|)$. So, total number of nodes and edges added to the graph is bouned by $O(\mathcal{V})$.
\end{itemize}
Since the number of new nodes and edges introduced are linear, the space and time complexity of the reduction procedure is also linear.
\end{proof}
\begin{example}
Please see \autoref{fig:fft} and \autoref{fig:fft2}. In \autoref{fig:fft2}, $v_{1}$ has dimension $\mathcal{D_B}_{v_{1}}=4$ and incidence $\mathcal{D_A}_{v_{1}}=12$ whereas in \autoref{fig:fft}, the node $v_2$ has dimension $\mathcal{D_B}_{v_{2}}=4$ and incidence $\mathcal{D_A}_{v_{2}}=5$. If we were to use the matching gadget, we would need $^{16}C_{2}$ for \autoref{fig:fft2} and $^{9}C_{2}$ for \autoref{fig:fft} specifications of rigid edges which is huge. Now we shall demonstrate a new graph gadget that achieves the effect of forbidden edges for matching using significantly fewer edges. It should be noted that every edge depicted in examples of pseudo-FFT gadget is a rigid edge. The nodes in red, blue, purple and green are the newly introduced nodes whereas peach color nodes are introduced at the time of edge isolation. It can be easily verified from the construction that the number of newly introduced nodes that are specific to the pseudo-FFT gadget is always smaller than $2\times\mleft(\mathcal{D_A}_{v}+\mathcal{D_B}_{v} +\mleft( \mathcal{D_A}_{v}+\mathcal{D_B}_{v} \mright)\mright)$ by a simple power series calculation (whereas nodes created at the time of edge isolation are bounded by $\mleft(\mathcal{D_A}_{v}+\mathcal{D_B}_{v}\mright)$). Therefore, total number of nodes in our examples are asymptotically bounded by $4\times\mleft(\mathcal{D_A}_{v}+\mathcal{D_B}_{v}\mright)$. All the the peach nodes on the left, all the red, blue and green nodes have two going rigid edges going out from them, all the purple nodes have one outgoing rigid edge whereas all peach nodes on the right have none. So, total number of new rigid edges introduced (for all $4\times\mleft(\mathcal{D_A}_{v_{1}}+\mathcal{D_B}_{v_{1}}\mright)$ nodes counted once) is asymptotically upper bounded by $8\times \mleft(\mathcal{D_A}_{v_{1}}+\mathcal{D_B}_{v_{1}}\mright)$. Therefore, if we apply this bound on all vertices of the original Hasse graph and then sum over all the vertices, we get $8 \times \mleft(|\mathcal{E}|+|\mathcal{V}| \times |\mathcal{D_B}_{max}| \mright)$ edges (when summed over pseudo-FFT gadget of every vertex in the orginal Hasse graph). We consider the number of edges to be linear in number of vertices and the dimension of the simplicial complex to be a small constant. Therefore, under these assumptions, we achieve linearity.
\\ In the example depicted in \autoref{fig:fft2} we see the pseudo-FFT gadget for a vertex $v_1$ which has degree $\mleft(\mathcal{D_A}_{v_{1}}+\mathcal{D_B}_{v_{1}}\mright)=16$ that happens to be a power or 2. This example happens to be quite symmmetric since all newly formed vertices obtain their labels through \emph{merging}.
\\ In contrast, in the exampe depicted in \autoref{fig:fft}, we see a gadget for a vertex $v_2$ which has degree $\mleft(\mathcal{D_A}_{v_{2}}+\mathcal{D_B}_{v_{2}}\mright)=9$ which isn't a power of 2. Hence, we see a mix of \emph{merging} and \emph{direct label inheritance}.
\end{example}
\section{Lower Bound on min-POP }
\begin{theorem}
It is NP-hard to approximate min-POP with a factor better than $O(\log(n))$
\end{theorem}
\begin{proof}
The proof follows via an approximation preserving reduction from min-Set-Cover to min-POP. The construction and proof are relatively straightforward. The curious connection it establishes between two classical problems is surprising. Construction and proof deferred.
\end{proof}
\section{SDP Formulation for Directed c-Balanced Separator}
From \autoref{thm:linred}, we conclude that, in order to design an
efficient algorithm for MMUP we need to design an efficient algorithm
for min-POP (which is essentially min-FAS with rigid edges). To solve
min-FAS with rigid edges we have two strategies: LP-based approximation
and SDP-based approximation. In this work, we focus on the SDP based
solution.
\begin{definition}[Directed Semimetric on a Graph $\mathcal{G}$]
Let $\mathcal{G}$ = ($\mathcal{V}$,$\mathcal{E}$) be a directed graph. A directed semimetric on a graph $\mathcal{G}$ satisfies the following conditions:
\begin{enumerate}
\item $\forall i \in \mathcal{V}, \mathcal{D}(v_i,v_i) = 0. $
\item $\forall v_i,v_j,v_k \in \mathcal{V}, \mathcal{D}(v_i,v_j) + \mathcal{D}(v_j,v_k)\geq \mathcal{D}(v_i,v_k)$
\end{enumerate}
\end{definition}
Agarwal et al. ~\cite{ACMM05} use the directed semimetric $\mathcal{D}(v_i,v_j) = \mleft|v_i-v_j\mright|^{2}-\mleft|v_0-v_i\mright|^{2}+\mleft|v_0-v_j\mright|^{2}$
for a fixed reference vector $v_{0}$ and vectors $v_{i},v_{j}$ s,t,
$i,j\neq0$ etc. are in 1-1 correspondence with vertices $v_{i},v_{j}\in\mathcal{V}$\footnote{We use the symbol $v_i$ to denote a vector that represents vertex $v_i\in\mathcal{G}$. Whether we mean to represent the vertex or whether we mean to discuss the vector that represents it, each time we use the symbol $v_i$ is sufficiently clear from the context.}.
Also, the quantities $\mathcal{D}(v_{i},v_{j})$ are in 1-1 correspondence
with the edges $v_{i}\rightarrow v_{j}\in\mathcal{E}$ .
\begin{definition}[minimum directed c-balanced separator]
Let $\mathcal{G}$ = ($\mathcal{V}$,$\mathcal{E}$) be a directed graph. Then, the directed edge expansion of a cut ($S$,$\bar{S}$) is $\frac{\delta^{out}(S)} {min(\mleft|S\mright| ,\mleft|\bar{S}\mright| )} $. A cut ($S$,$\bar{S}$) that satisfies $\mleft|S\mright| \geq c \mleft|\mathcal{V}\mright|$ and $\mleft|\bar{S}\mright| \geq c \mleft|\mathcal{V}\mright|$ is a c-balanced cut. A c-balanced cut with minimum directed edge expansion is a minimum directed c-balanced separator. Here, $\delta^{out}(S)$ denotes the set of outgoing edges from $S$.
\end{definition}
Recall that min-POP requires dealing with pre-specified edge orientations
i.e. dealing with rigid edges and forbidden edges respectively. We
achieve this effect with a very simple strategy that we shall outline
below: Let $\mathcal{H}_{1}(\mathcal{V},\mathcal{E})$ be the edge
duplicated Hasse digraph from the construction described in \autoref{sub:The-Construction}.
Let $(u_{i},v_{i})_{i\in\left\{ {1\cdots\mathcal{F}}\right\} }\in\mathcal{E_{\oslash}}\subseteq\mathcal{E_{\star}}$
be the set of \textit{forbidden edges} and $(v_{i},u_{i})_{i\in{1\cdots\mathcal{R}}}\in\mathcal{E_{\dagger}}\subseteq\mathcal{E}_{\star}$
be the set of \textit{rigid edges }where $\mathcal{E_{\star}}$ is
the \textit{extended set of edges} of the graph $\mathcal{H}_{\star}=(\mathcal{V}_{\star},\mathcal{E}_{\star})$
obtained at the end of the reduction procedure. Note that $\mathcal{E_{N}}\cup\mathcal{E_{\dagger}}\cup\mathcal{E_{\oslash}}=\mathcal{E_{\star}}$
and $\left(\left|\mathcal{E_{\dagger}}\right|=\mathcal{R}\right)=\left(\left|\mathcal{E_{\oslash}}\right|=\mathcal{F}\right)$.
Note that the set of normal edges $\mathcal{E_{N}}$ is in 1-1 correspondence
with the edges $\mathcal{E}$ of the Hasse graph. Assume that the
condition since $\mathcal{D}(u_{i},v_{i})+\mathcal{D}(v_{i},u_{i})=1$
has already been enforced for every $\mathcal{D}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}},\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$.
Therefore, we need to implement either $\mathcal{D}(u_{i},v_{i})=0$
(enforcing a forbidden edge) or $\mathcal{D}(v_{i},u_{i})=1$ (enforcing
a rigid edge) noting that the conditions are equivalent. When using
a divide-\&-conquer procedure with min-directed c-balanced cut as
a subroutine, it is easier to implement $\mathcal{D}(u_{i},v_{i})=0$
(i.e. to specify that we do not cut edge $(u_{i},v_{i})$ at \emph{this}
(or any) recursion level) rather than implementing $\mathcal{D}(v_{i},u_{i})=1$
(i.e. to specify that we cut the rigid edge $(v_{i},u_{i})$ at \emph{this}
recursion level). Ensuring $\mathcal{D}(u_{i},v_{i})=0$ at each recursion
level ensures we have $\mathcal{D}(v_{i},u_{i})=1$ at the end of
divide-\&-conquer recursion procedure.
\begin{definition}[Orthogonality Constraint for edge $(u,v)$] \label{def:orthocs}
We define orthogonality constraint for edge $(u,v)$ as follows: \\
$\mathcal{D}(u,v)=\mleft|v_{0}-v\mright|^{2}-\mleft|v_{0}-u\mright|^{2}+\mleft|v-u\mright|^{2}=0$ \\
which is equivalent to: $\mleft|v_{0}-v\mright|^{2} + \mleft|v-u\mright|^{2} = \mleft|v_{0}-u\mright|^{2}$
\end{definition}
In this case, we will proceed with the the most na\"{i}ve\xspace strategy.
i.e. We shall use the directed metric from Agarwal et.al. and to equate
it to $0$ for all forbidden edges. The vector programming formulation
for min-POP is basically an extension of Agarwal et.al.'s \cite{ACMM05}
formulation for \emph{minimum Directed c-Balanced Separator Problem
(Conditions I, II, III)} with an extra\textbf{ }\emph{IV}$^{th}$
\emph{condition}\textbf{\emph{ }}for handling the forbidden edges.
\textit{Vector programming formulation\textit{\footnote{Once the vector programming formulation is written down it is merely a triviality to write down the corresponding SDP formulation. See \cite{Va03,WS11}. Going forward we shall use the terms vector program and SDP interchangeably unless otherwise specified. The meaning is of course extremely obvious from the context..} }
for (Divide \& Conquer Subroutine of) min-POP:}
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{(min-POPs)}]
\lpobj{min}{\frac{1}{8}\sum\limits _{(\langle v_i, v_j \rangle \in\mathcal{E}}\mathcal{D}(v_{i},v_{j})}
\lpeq[forbidden1]{\sum\limits _{\langle u_i, v_i \rangle \in \mathcal{F}}\mathcal{D}(u_{i},v_{i})=0}{}
\lpeq[c-balanced1]{\sum_{i < j} |v_{i}-v_{j}|^{2} \geq 4c(1-c)n^2}{}
\lpeq[triangle1]{|v_{i}-v_{j}|^2 + |v_{j}-v_{k}|^2 \geq |v_{k}-v_{i}|^{2}}{v_i,v_j,v_k \in \mathcal{V}\cup\{v_0\}}
\lpeq[unit1]{|v_{i}|^{2}=1}{v_i \in \mathcal{V} \cup\{v_0\}}
\lplabel{lp:form}
\end{lpformulation}%
\end{minipage}}
Consider the SDP formulation described in \autoref{lp:form}. Note
that, if we were to remove the \autoref{c-balanced1}, this would
actually be an SDP relaxation for min-POP (See \autoref{thm:formdc}).
However, we would not know of a rounding procedure to such a relaxation.
Now, \autoref{c-balanced1} ensures that if a cut based rounding is
used then the cut will be a c-balanced cut(a necessity for Leighton-Rato
divide-\&-conquer approximation schemes). Hence we must add \autoref{c-balanced1}
so that we can use a balanced-cut based rounding procedure that gets
applied as part of a hierarchical divide-\&-conquer scheme.
This SDP is indeed a relaxation since every assignment of boolean
variables corresponds to a feasible set of vectors:
\begin{itemize}
\item $v_{i}=v_{0}$, if $x_{i}=1$.
\item $v_{i}=-v_{0}$, if $x_{i}=0$.
\end{itemize}
We now discuss the equations in the formulation for the min-POP subroutine.
\begin{itemize}
\item \autoref{unit1} ensures that all vertices lie on the unit
sphere.
\item \autoref{triangle1} is the familiar triangle inequality.
\item \autoref{forbidden1} is the forbidden edges condition. It
is the new condition that we add to the formulation (to incorporate
rigid edges). It is essentially the sum of all the forbidden edge
constraints described by \autoref{def:orthocs}. Note that since all
the $\mathcal{D}(u_{i},v_{i})$ are non-negative, equating their sum to $0$
essentially means that all of them have to be $0$.
\end{itemize}
Agarwal et.al.\cite{ACMM05} give a $\frac{c}{2}-$balanced cut that
approximates the minimum directed c-balanced separator within a factor
of $O(\sqrt{\log(n)})$. Their work is based on the the work of Arora
et al. ~\cite{ARV09}, more specifically on the powerful ARV Structure
Theorem, that is briefly stated in \autoref{thm:arvst} in \autoref{sec:Rounding-Based-on}.
We modify the Agarwal et al.'s algorithm to provide an $O(\sqrt{\log(n)})$
factor approximation for every subroutine of the divide-\&-conquer
approximation algorithm for min-POP.
Now, we consider a formulation for feedback-arc-set on some graph
$\mathcal{G}(\mathcal{V},\mathcal{E})$ by Grötschel, Jünger and Reinelt in \cite{GJR85}.
This special formulation specifies the acyclicity constraints for
given edge set $\mathcal{E}$ in form of ordering constraints on the complete
graph with vertex set $\mathcal{V}$. They show that destroying cycles that
are specific to edge set $\mathcal{E}$ is equivalent to destroying all length
2 and length 3 cycles in the complete graph. The information about
the edge set $\mathcal{E}$ is accommodated only in the objective function
as shown in the formulation \autoref{lp:GJR}.
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{(min-FAS(GJR))}]
\lpobj{min} { \sum_{ e_{i,j} \in \mathcal{E} } d_{ij} }{}
\lpeq[0cond] { d_{ij} \geq 0 }{}
\lpeq[antisymm]{d_{ij} + d_{ji} =1}{}
\lpeq[triineq]{d_{ij} + d_{jk} \geq d_{ik}}{}
\lplabel{lp:GJR}
\end{lpformulation}%
\end{minipage}}
\begin{theorem} \label{thm:formdc}
The formulation described in \autoref{lp:form} is a correct formulation for a subroutine within the SDP-based Divide \& Conquer Algorithm for min-POP
\end{theorem}
\begin{proof}
To begin with we check if the directed semi-metric $\mathcal{D}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}, \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ in SDP formulation min-POP satisfies the conditions prescribed for $d_{ij}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ in the LP formulation \autoref{lp:GJR}.
\begin{itemize}
\item Since $\mathcal{D}(u,v)=\mleft|v_{0}-v\mright|^{2}-\mleft|v_{0}-u\mright|^{2}+\mleft|v-u\mright|^{2}=0$ is sum of positive real numbers, clearly, $\mathcal{D}(u,v)\geq 0$.
\item Secondly, note that, by simple application of definition, $\mathcal{D}(u,v)+\mathcal{D}(v,w)-\mathcal{D}(u,w)=\mleft| v-u \mright|^{2} + \mleft| w-v \mright|^{2} - \mleft|w-u\mright|^{2}$. However since we work with an $l_2^{2}$ metric in the ARV algorithm, we have $\mleft| v-u \mright|^{2} + \mleft| w-v \mright|^{2} \geq \mleft|w-u\mright|^{2}$ which gives us $\mathcal{D}(u,v)+\mathcal{D}(v,w)\geq \mathcal{D}(u,w)$.
\item Finally, the rounding algorithm of Agarwal et al. employs a directed cut hierarchically until every directed edge (or its counterpart) is cut at some recursion level. Therefore, clearly we satisfy $\mathcal{D}(u,v)+\mathcal{D}(v,u)=1$ for every directed edge $\mathcal{D}(v,w)$ on the complete graph.
\end{itemize}
It follows from the Grötschel-Jünger-Reinelt formulation of min-feedback arc set given in the LP formulation \autoref{lp:GJR}, that the graph obtained at the end of the Divide-\&-Conquer based SDP algorithm will not have any cycles. Also, we have the forbidden edge constraint $\sum\limits _{\langle u_i, v_i \rangle \in \mathcal{F}}\mathcal{D}(u_{i},v_{i})=0$.\footnote{In \autoref{sub:Extension-to-Agarwal}, \autoref{lem:agre}, we prove that the Agarwal et al. rounding procedure preserves the rigid edges. In other words, the rounding procedure ensures that none of the forbidden edges are cut.} Now, we have already proved in \autoref{thm:REMain} and \autoref{thm:REMain} that if
\begin{enumerate*}[label=(\textbf{\alph{enumi})}]
\item the rigid edge formulation does not have any cycles
\item and if the rigid edges are preserved,
\end{enumerate*}
then a corresponding orientation of the original Hasse graph will not have any matching or cycle constraint violations. Hence, if we were to merely consider \autoref{forbidden1}, \autoref{triangle1} and \autoref{unit1} then we would get an SDP formulation for min-POP. But, since (for purpose of appropriate rounding), we need to approximate a c-balanced cut at each recursion level, we need \autoref{c-balanced1} in addition to the above three equations in order to formulate an SDP meant for the min-POP subroutine.
\end{proof}
\section{An $\ensuremath{O({\log^{\nicefrac{3}{2}}(\mathcal{N})})}$ factor Approximation
Algorithm based on ARV-rounding \label{sec:Rounding-Based-on}}
\subsection{Adding the Orthogonality Constraint to the ARV formulation}
\subsubsection{Sparsest Cut and the ARV Structure Theorem}
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{Sparsest Cut }]
\lpobj{min}{\sum_{i, j \in E} c_{ij} \| v_i - v_j \|^2}{}
\lpeq[unitspherecond] { \sum_{i,j} \|v_i - v_j\|^2 = 1 }{}
\lpeq[lppcond]{\|v_i - v_j\|^2 + \|v_j - v_k\|^2 \geq \|v_i - v_k\|^2}{}
\lpeq[wellsepcond]{\sum_{i,j} \|v_i - v_j\|^2 = \Omega(n^2)}{}
\lplabel{lp:SPCP}
\end{lpformulation}%
\end{minipage}}
The original ARV algorithm was designed for the sparsest cut problem
that is specified in \autoref{lp:SPCP}. The ARV algorithm crucially
depends on the ARV structure theorem to prove the approximability
result is stated below. See \autoref{thm:arvst}.
\begin{definition}[$l_2^2$ metric space] \label{def:l22} $l_2^2$ representation of a graph G is an assignment of a point vector to each node $v_i \in \mathbb{R}^k$ for each i. s.t. \[ |v_i - v_j|_2^2 + |v_j - v_k|_2^2 \geq |v_i - v_k|_2^2 \] It is a unit $l_2^2$ representation if on unit sphere i.e. $|v_i| = 1, \forall i$. \end{definition}
\begin{theorem}[ARV Structure Theorem] \label{thm:arvst}
For a set of points $v_1, v_2, \dots v_n$ if the following conditions are satisfied:
\begin{enumerate}
\item All points lie on the unit ball in $\mathbb{R}^n$.
\item The points form an $l_2^2$ metric w.r.t. $d_{ij}$ where $d_{ij} = \|v_i - v_j\|_2^2$,
\item All points are well-separated i.e. $\sum_{i,j} d_{ij}/n^2 \geq \delta = \Omega(1)$.
\end{enumerate}
Then $\exists S, T$ disjoint subsets of V s.t. $|S|, |T| \geq \Omega(n)$ which satisfy the following property: \[ min_{i \in S, j \in T} d_{ij} \geq \Omega(1/\sqrt{\log{n}})\].
\end{theorem}
\begin{algorithm}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{tikzpicture}[every node/.style={inner sep=0pt}]
\fill[color=myred,fill=mypink!60,fill opacity=0.5] (3.6,-2.58) -- (2.58,-3.58) -- (-3.72,2.72) -- (-2.72,3.72) -- cycle;
\fill [color=shadecolor, opacity=0.5](0,0) circle (3cm);
\draw [color=shadestcolor] (0,0) circle (3cm);
\draw [color=black] (3.6,-2.58)-- (2.58,-3.58);
\draw [color=black] (2.58,-3.58)-- (-3.72,2.72);
\draw [color=black] (-3.72,2.72)-- (-2.72,3.72);
\draw [color=black] (-2.72,3.72)-- (3.58,-2.58);
\draw (3.08,-3.08)-- (-3.22,3.22);
\fill [color=blue!70] (-0.6,1.9) circle (1.5pt);
\fill [color=blue!70] (1.94,-0.48) circle (1.5pt);
\fill [color=blue!70] (2.66,-0.8) circle (1.5pt);
\fill [color=myteal] (-2.4,0.38) circle (1.5pt);
\fill [color=myteal] (0.86,-1.96) circle (1.5pt);
\fill [color=myteal] (0.64,-2.58) circle (1.5pt);
\fill [color=black] (0,0) circle (2.0pt);
\draw[color=black] (0.16,0.28) node {$O$};
\fill [color=black] (0,3) circle (2.0pt);
\draw[color=black] (0.2,3.28) node {$v_0$};
\fill [color=red!70] (-1.74,1.04) circle (1.5pt);
\fill [color=myteal] (-1.8,-0.1) circle (1.5pt);
\fill [color=blue!70] (0.9,2.28) circle (1.5pt);
\fill [color=blue!70] (-1.07,2.8) circle (1.5pt);
\fill [color=red!70] (-0.56,1.06) circle (1.5pt);
\fill [color=blue!70] (0,2) circle (1.5pt);
\fill [color=teal] (0.18,-1.38) circle (1.5pt);
\fill [color=red!70] (1.58,-1.48) circle (1.5pt);
\fill [color=blue!70] (2.72,0.66) circle (1.5pt);
\fill [color=blue!70] (1.6,0.38) circle (1.5pt);
\fill [color=myteal] (-0.84,-1.22) circle (1.5pt);
\fill [color=blue!70] (0.48,0.88) circle (1.5pt);
\fill [color=myteal] (-2.4,-0.42) circle (1.5pt);
\fill [color=black] (0,-3) circle (2.0pt);
\draw[color=myteal] (0.22,-2.72) node {$\overline{v_0}$};
\fill [color=myteal] (-2.2,-0.86) circle (1.5pt);
\fill [color=uququq] (3.08,-3.08) circle (1.5pt);
\fill [color=uququq] (-3.22,3.22) circle (1.5pt);
\fill [color=blue!70] (1.38,1.28) circle (1.5pt);
\fill [color=blue!70] (2.16,0.6) circle (1.5pt);
\fill [color=myteal] (-1.5,-1.78) circle (1.5pt);
\fill [color=myteal] (-1.62,-1.24) circle (1.5pt);
\fill [color=myteal] (-0.84,-2.3) circle (1.5pt);
\fill [color=blue!70] (1.22,0.76) circle (1.5pt);
\draw [color=blue!70] (-1.15,2.4) node {$A$};
\draw [color=blue!70] (-0.95, 2.4) circle (2.0pt);
\draw[color=red!70] (-0.15, 1.00) node {$B$};
\draw[color=red!70] (-0.35, 1.00) circle (2.0pt);
\draw [color=blue!70] (1.9,0.9) node{$C$};
\draw [color=blue!70] (2.10,0.9) circle (2.0pt);
\draw[color=red!70] (-1.25,1.40) node {$D$};
\draw[color=red!70] (-1.45,1.40) circle (2.0pt);
\draw[color=myteal] (-0.14,-2.08) node {$E$};
\draw [color=myteal] (-0.34,-2.08) circle (2.0pt);
\draw[color=myteal] (-0.98,-0.7) node {$F$};
\draw [color=myteal] (-1.18,-0.7) circle (2.0pt);
\fill [color=blue!70] (2.16,1.56) circle (1.5pt);
\fill [color=blue!70] (1.46,2.04) circle (1.5pt);
\fill [color=blue!70] (1,1.86) circle (1.5pt);
\end{tikzpicture}\captionof{figure} {ARV rounding}%
\end{minipage}}
\begin{algorithmic}[1]
\Require A unit-$l_2^{2}$ representation of $m$ vectors $\mathcal{V}= \mleft \{ v_{1}\dots v_{m} \mright \}$
\Ensure There exist constants $\alpha$ and $\beta$ s.t. the Algorithm returns a partition of the above set of vectors $\Omega$ into sets $\mathcal{X}$ and $\mathcal{Y}$ both of which are $\alpha$-large and are $\Delta$-separated where $\Delta=\frac{\beta}{\sqrt{\log(n)}}.$
\LState Pick a random line $r$ through the origin. (Determines the fat random hyperplane normal to it.)
\LState Let $\mathcal{U} =\mleft\{ v_{i}:v_{i}.r\geq\frac{1}{\sqrt{\sigma}}\mright\} $
\LState Let $\mathcal{V} =\mleft\{ v_{i}:v_{i}.r\leq\frac{1}{\sqrt{\sigma}}\mright\} $
\LineComment Discard pairs of points from $v_i\in \mathcal{U}$ and $v_j\in \mathcal{V}$ such that $\mleft|v_{i}-v_{j}\mright|^{2}\leq\Delta$ i.e. Start with $\mathcal{S}=\mathcal{U},\mathcal{T}=\mathcal{V}$ and sequentially delete pair of points violating the `$\Delta$-condition'
\ForAll{$v_{i}\in \mathcal{S}, v_{j} \in \mathcal{T}$ such that $\mleft|v_{i}-v_{j}\mright|^{2}\ \leq\Delta$}
\LState {$\mathcal{S}\leftarrow \mathcal{U} -\mleft\{v_i \mright \}$}
\LState {$\mathcal{T}\leftarrow \mathcal{V} - \mleft\{v_j \mright \}$}
\EndFor
\LineComment{\textcolor{myindigo}{Handling of rigid edges $\mathscr{R}_{v_i}$ with $v_i$ as the destination node. }}
\LState{Choose random $ \Delta \geq 0$} \label{lst:line:restart}
\LState{Choose $\sigma$ randomly s.t. $0\leq \sigma\leq\Delta$.}
\LState{$\mathcal{X}\leftarrow S$ and let $Z=V_{\diamond}-\mathcal{A}$.}
\ForAll {$v_i$ such that $\mleft|v_{i}-x\mright|^{2}\leq \Delta $ for $x \in S$ }
\If {$ \nexists v_{k}\in Z$ such that $\mleft\{ v_{k}\rightarrow v_{i}\mright\} \in \mathcal{E_{\dagger}}$}
\LState $\mathcal{X}\leftarrow\mathcal{X} + \mleft\{ v_{i} \mright\}$
\EndIf
\EndFor
\LState{$\mathcal{Y}=\Omega-\mathcal{X}$.}
\LState{ \textbf{return } $\mleft (\mathcal{X}, \mathcal{Y} \mright )$.} \label{lst:line:reend}
\end{algorithmic}%
\end{minipage}}
\protect\caption{ARV-Algorithm}
In the figure accompanying \autoref{alg:ARV}, a hypersphere is shown projected as a disc. \emph{In this particular figure}, the line seen passing though the origin depicts the random hyperplane. In the later sections, we use slightly different visual conventions. The strip in pink is the fat hyperplane around the random hyperplane. All points lie \emph{on} the sphere. This 2-dim. picturization is a universally followed convention. In our case however, apart from the visual role, some of the correctness proof depend on the 2-dim projections and the related the projective and affine geometry.
\label{alg:ARV}
\end{algorithm}
\begin{shnote}[Fat hyperplanes and rigid edges] \label{note:fatandrigid}
In \autoref{alg:ARV}, we handle the rigid edges in Lines~\ref{lst:line:restart}-\ref{lst:line:reend}. This variation to the ARV algorithm is justified because the ARV structure theorem guarantees us that $S$ and $T$ which are of size $\Omega(n)$ are well-separated. Following this, The addition of new points to $S$ and $T$ to form sets $\mathcal{X}$ and $\mathcal{Y}$ is done in a manner that respects the rigid edge specifications. Also, $S\subseteq\mathcal{X}$ and $T\subseteq\mathcal{Y}$. So, clearly $\mathcal{X}$ and $\mathcal{Y}$ are both of size $\Omega(n)$. So, if ARV structure theorem was applicable to $(S,T)$ then it is also applicable to $(\mathcal{X},\mathcal{Y})$. Finally, we have $\mathcal{X}\cup\mathcal{Y}=\Omega$. So, the $(\mathcal{X},\mathcal{Y})$ partition of vectors produces a graph cut.
\end{shnote}
\subsubsection{Randomized hyperplanes and 2-plane projections}
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\includegraphics[scale=0.475]{figure4.png}%
\end{minipage}}
\protect\caption{Sections and Axes of Hyperplanes}
\label{fig:multiple2sections}
\end{figure}
Note that the SDP solution constitutes of distribution of vectors
on an n-dimensional sphere. In \autoref{sub:Handling-Forbidden-Edges:}
we introduce a new type of constraint called the orthogonality constraint.
In order to \emph{see} how and why the orthogonality constraint helps
us model rigid edges, we need to consider the projection of the n-dimensional
SDP solution on a 2-dimensional plane. Apart from the visual cues
and the intuition, 2-projections provides a more convenient language
for proof and reasoning behind the mechanism of the orthogonality
constraint. To be sure, we\emph{ do not} project the n-dimensional
SDP solution on the 2-dimensional plane \emph{as part of the algorithm},
but only \emph{as part of the analysis}.
\begin{definition}[Section of the Hyperplane] Given a hypersphere and an intersecting hyperplane that passes through its center, the section of a hyperplane is obtained by taking all points of intersection between the hypersphere and the hyperplane.
\end{definition}
\begin{definition}[2-Section of a Hyperplane] Given a hypersphere, a hyperplane passing through the center of the hypersphere and a 2-plane, we obtain a 2-section of a hyperplane by projecting a section of a hyperplane on a 2-plane.
\end{definition}
We disregard the measure zero case when the 2-plane is parallel to
the hyperplane.
\begin{definition}[Major Axis of a 2-Section of a Hyperplane] The hypersphere when projected on a 2-plane gives us a circle. Major axis of a 2-section of a hyperplane is the maximal length straight line joining any two points on the 2-section of a hyperplane.
\end{definition}
It is easy to check that the length of the major axis of a 2-section
of a hyperplane is always equal to the diameter of the hypersphere.
We refer to the major axis of a 2-section of a hyperplane simply as
'major axis of a hyperplane'.
\begin{definition}[Hypercircle] Here, we use the non-standard term hypercircle to mean a cross-section of the hypersphere obtained by taking all points on the hypersphere that are at a given distance $\Phi$ from a fixed point $v_0$. Different values of $\Phi$ give us different hypercircles.
\end{definition}
In \autoref{fig:multiple2sections}, we see that two different hyperplanes
may share the same major axes. The set of hyperplanes that share a
major axis forms an equivalence class. Given a projection on a 2-plane,
although a hyperplane is better (and yet not uniquely) represented
by a 2-section, it is sufficient to limit our interests strictly to
the major axis of the hyperplane because the behavior or the hyperplane
can be entirely gauged from the role its major axis plays in separating
various points on a 2-plane. In \autoref{fig:2plane5case}, we can
visualize the five different cases concerning 2-plane projections.
To follow a convention, we will say that a point is \emph{accepted}
if it lies \emph{in the same hemisphere as $v_{0}$ (i.e. above the
major axis)} and \emph{rejected} if it \emph{lies in the same hemisphere
as $\bar{v_{0}}$ (i.e. below the major axis)}.
\begin{enumerate}
\item b lies on the major axis of the hyperplane and a lies under it ($b$
may or may not be accepted, $a$ is rejected).
\item a lies on the major axis of the hyperplane and b lies above it ($b$
is accepted, $a$ may or may not be rejected).
\item Major axis of the hyperplane lies between $b$ and $a$. $b$ is above
and $a$ is below. ($b$ is accepted and $a$ is rejected.)
\item Both $a$ and $b$ lie below the major axis. (Both $a$ and $b$ are
rejected.)
\item Both $a$ and $b$ lie above the major axis. (Both $a$ and $b$ are
accepted.)
\end{enumerate}
Now, given a rigid edge $b\rightarrow a$, we put two requirements
on the 2-plane in order to determine the fate of $b\rightarrow a$:
\begin{enumerate}
\item The 2-section of the hyperplane coincides with its major axis. This
happens when the hyperplane is perpendicular to the 2-plane of projection.
\item Having made the hyperplane perpendicular to the 2-plane, we still
have an additional degree of freedom: we may rotate it along the line
perpendicular to the hyperplane and we keep doing that till $b$ appears
\emph{on} \emph{the circle}(which is same as saying that the 2-plane
$\ensuremath{\boxplus_{Obv_{0}}}$ (determined by $O$, $b$ and $v_{0}$)
is identical to the 2-plane of projection.
\end{enumerate}
Going forward we will not distinguish between the 2-section and the
major axis of the hyperplane since we have chosen our plane of projection
to be perpendicular to the line of vision. Also, henceforth, the major
axis will be considered to be the projection of the hyperplane.
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\includegraphics[scale=0.07]{figure5.png}%
\end{minipage}}
\protect\caption{2-plane Projection: 5 Cases}
\label{fig:2plane5case}
\end{figure}
The following lemma demonstrates how one may use edge-by-edge 2-projections
to infer whether or not a particular edge is cut by the hyperplane
in n-dimensions.
\begin{lemma}[From 2-dimensions to n-dimensions] \label{lem:2all}
Given two points $x$ and $y$, if their 2-projections say $x^{\prime}$ and $y^{\prime}$ lie on the opposite side of the major axis, then then $x$ and $y$ lie on the opposite side of the hyperplane in n-dimensions.
\end{lemma}
\begin{proof} All these conclusions follow from a single elementary observation: Suppose we are given an n-dimensional sphere and suppose that a given hyperplane passes through its center. Irrespective of which 2-plane you project it on, the major axis will always be of the size of the diameter and will divide the projection into two equal semicircular arcs. In \autoref{fig:2plane5case}, apart from $b$, the lengths of the segments $v_0-P$ and $v_0-Q$ are also preserved. Also, if two points say $x^{\prime}$ and $y^{\prime}$ lie on the opposite side of the major axis, then within the 2-projection, $x^{\prime}-y^{\prime}$ will intersect with the major axis at a point say $z^{\prime}$. Now, clearly, in the n-dimensions, there exists a point say $z$ that lies \emph{on the hyperplane} and also on the segment $x-y$ s.t. $x^{\prime}$ and $y^{\prime}$ are projections of $x$ and $y$. Therefore, if given any two points $x$ and $y$ if their projections $x^{\prime}$ and $y^{\prime}$ lie on opposite sides of the major axis, then then $x$ and $y$ lie on the opposite sides of the hyperplane.
\end{proof}
\begin{example}
In \autoref{fig:2plane5case}, we observe that for case 1, $a$ and $v_0$ lie on opposite sides, for case 2, $b$ and $\bar{v_0}$ lie on opposite sides, for case 3, $\{b,v_0\}$ and $a$ are on opposite sides, for case 4, $\{b,a\}$ and $v_0$ lie on opposite sides and for case 5, $\{a,b\}$ and $v_0$ lie on opposite sides. These cases are mutually exclusive and collectively exhaustive and in each of the cases, it is ensured that the edge $a\rightarrow b$ isn't cut.
\end{example}
\subsubsection{Handling Forbidden Edges: ARV Rounding\label{sub:Handling-Forbidden-Edges:}}
The purpose of this subsection is merely to illustrate that the purpose
of this section is to merely establish the fact that one can accommodate
the orthogonality constraint within the randomized hyperplane rounding
employed within the ARV framework. In \autoref{sub:Extension-to-Agarwal},
we use the ARV structure theorem for directed balanced cut rounding.
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\includegraphics[scale=0.5]{figure6.png}%
\end{minipage}}
\protect\caption{Random Hyperplane: Orthogonality Constraint}
\label{fig:RHOC}
\end{figure}
\begin{lemma}[Orthogonality and Random Hyperplane Rounding] \label{lem:orrhr}
If the Orthogonality condition for edge $(a,b)$ is satisfied (i.e. $\mathcal{D}(a,b)=\mleft|v_{0}-b\mright|^{2}-\mleft|v_{0}-a\mright|^{2}+\mleft|b-a\mright|^{2}=0$) and if we cut the unit hypersphere $\mathcal{S}$ with a random hyperplane $\mathcal{P}$, then only one of the two possibilities can occur:
\begin{enumerate}
\item $(b,a)$ is cut thus ensuring that $(a,b)$ will never be cut.
\item Neither $(a,b)$ nor $(b,a)$ is cut.
\end{enumerate}
\end{lemma}
\begin{proof}
Suppose that we have solved the SDP and all vectors obtained from Cholesky decomposition lie in their optimal configuration on the unit hypersphere $\mathcal{S}$. To understand the effect of separation of vectors on application of rounding by some hyperplane $\mathcal{P}$, project the hypersphere $\mathcal{S}$ and $\mathcal{P}$ with optimal configuration of all points on the 2-plane $\boxplus_{Obv_0}$ which is determined by $v_0$, $b$ and the origin $O$. Note that since $\boxplus_{Obv_0}$ contains both $b$ and $v_0$, the projections of $b$ and $v_0$ will lie on the edge of the projection of $\mathcal{S}$ on $\boxplus_{Obv_0}$(which will be a disc). Hence, $b$ and $v_0$ will lie on the bounding circle that will contain projections of all the vectors on the 2-plane $\boxplus_{Obv_0}$. The projection of $\mathcal{P}$ on $\boxplus_{Obv_0}$ will be a line (as long as we disregard the possibility of the event where the random hyperplane $\mathcal{P}$ is parallel to the plane of projection $\boxplus_{Obv_0}$ the probability measure of which is $0$.) Since the line $v_0-O-\overline{v_0}$ is parallel to plane of projection $\boxplus_{Obv_0}$, the projection of $\overline{v_0}$ will lie on the circle and the projection of the line $v_0-O-\overline{v_0}$ will determine a diameter of this circle. Thus $\overline{v_0}$ and $v_0$ will subtend a right angle at b. i.e. the two lines $b-\overline{v_0}$ and $b-v_0$ will remain perpendicular before and after projection since both these lines are parallel to the plane of projection $\boxplus_{Obv_0}$. But the line $\overline{v_0}-b$ is one of the lines contained in the hyperplane (say $H_0$) perpendicular to the line ${v_0}-b$. Hence the projection of $H_0$ will be $\overline{v_0}-b$. The position of point $a$ after projection will be somewhere along the segment $\overline{v_0}-b$ since the segment $a-b$ lies in the hyperplane perpendicular to the line ${v_0}-b$. It may very well happen that $a$ and $b$ get projected on the same point in 2-plane $\boxplus_{Obv_0}$. But, this happens when segment $a-b$ is perpendicular to the 2-plane $\boxplus_{Obv_0}$. If we consider all degrees of freedom of $a-b$, then the case of being precisely perpendicular to the 2-plane $\boxplus_{Obv_0}$ is a measure zero event. So, we consider all possible positions of $a$ on segment $\overline{v_0}-b$ except point $b$.
In \autoref{fig:RHOC} we observe five different positions of $a$(shown as $a^1_1$, $a^2_1$ and $a^3_1$ and $a^1_2$, $a^2_2$ and $a^3_2$ respectively for two different positions of $b$, i.e. $b_1$ and $b_2$ . Let us assume a position for $b$ (without loss of generality) and proceed with projecting the hypersphere $\mathcal{S}$ on the 2-plane $\boxplus_{Obv_0}$ as described above. Then, any random hyperplane (denoted as a line here after projection on the 2-plane) will lie between lines $O-AB$ and $O-DA$ as shown in the figures below. So now we shall have five cases
\begin{itemize}
\item \textbf{Case I:} The projection of the hyperplane lies between between lines $O-AB$ and $O-BC$. $\mathcal{D}(b,a)=0$ and $\mathcal{D}(a,b)=0$ (Since both $a$ and $b$ lie in the bottom hemisphere (w.r.t $\mathcal{P}$)\footnote{We always mean the hemisphere above or below the random hyperplane $\mathcal{P}$ unless specified otherwise.} while $v_{0}$ lies in the top hemisphere and neither $ab$ nor $ba$ is cut.)
\item \textbf{Case II:} The projection of the hyperplane lies between between lines $O-BC$ and $O-CD$. $\mathcal{D}(b,a)=1$ and $\mathcal{D}(a,b)=0$ (Since $b$ lies in the top hemisphere along with $v_{0}$ and $a$ lies in the bottom hemisphere and $ba$ is cut but $ab$ is not .)
\item \textbf{Case III:} The projection of the hyperplane lies between between lines $O-CD$ and $O-DA$. $\mathcal{D}(b,a)=0$ and $\mathcal{D}(a,b)=0$ (Since all three points, $v_{0}$, $a$ and $b$ lie in the top hemisphere and neither $ab$ nor $ba$ is cut.)
\item \textbf{Case IV:} $b$ lies on the projection of the hyperplane. So the point $v_0$ and the segment $\bar{v_0}-b$ lies on the opposite side of the projected hyperplane. We have seen before that the projection of $a$ lies on $\bar{v_0}-b$. Therefore, $a$ and $v_0$ lies on the opposite side of the hyperplane. Therefore using the conclusions of \autoref{lem:2all}, we conclude that $a$ and $v_0$ lie on opposite sides of the hyperplane in n-dimensions. Finally, $b$ lies in the fat hyperplane. So, it may go to either of the two sides, while $a$ stays opposite of $v_0$. Hence $ab$ is never cut. However, $ba$ may be cut (if $b$ is assigned to the same set as $v_0$) or neither of the two will be cut (if $b$ is assigned to the opposite side of $v_0$).
\item \textbf{Case V:} $a$ lies on the projection of the hyperplane. Since $v_0-a$ is the hypotenuse (even in the projected space), $b$ and $\bar{v_0}$ will lie on opposite sides of the projected hyperplane (since we know that projection of $a$ lies on the segment $v_0-b$. Therefore $b$ and $v_0$ are on the same side while $a$ lies in the fat hyperplane. So, if $a$ is assigned to the same set as $\{b,v_0\}$, then neither of the edges $ba$ and $ab$ will be cut, whereas if $a$ is assigned to the same set as $\bar{v_0}$, then edge $ba$ is cut. Note that we implicitly use the results of \autoref{lem:2all} to generalize our conclusions from 2-dimensional projection to n-dimensions.
\end{itemize}
Since from our point of view angles vary only from $0$ to $\pi$, once we reach $O-DA$ starting from $)-AB$ we encounter repetition. So, with the above five cases in mind, we are done. In other words, for Case II, $\mathcal{D}(a,b)=1$ and $\mathcal{D}(b,a)=0$ \emph{in this iteration.} For Case I and Case III, the decision is deferred to later iterations of divide and conquer. (i.e. while in this iteration neither $(a,b)$ nor $(b,a)$ is cut, eventually in one of the iterations, one of the two has to be cut for the algorithm to terminate.) Since our choice of $a$ and $b$ was arbitrary, this analysis can be seen to hold for all the specifications of the type $\mathcal{D}(a,b)=0$
\end{proof}
Apart from the measure zero event when $a$ and $b$ coincide to a single point when projected on the 2-plane $\boxplus_{Obv_0}$, we can always switch from the 2-dimensional reasoning to the n-dimensional reasoning. To see this, let $\mathcal{P}_2$ be the projection of the hyperplane on the 2-plane. Once again, ignoring the measure zero event where the random hyperplane is parallel to the 2-plane, the projection $\mathcal{P}_2$ will be seen as a line through the origin.
We never really use the \autoref{lem:orrhr} directly. So, whatever
is discussed in this section is mostly to make the following section
more readable.
\subsection{Extension to Agarwal et.al's Rounding: Why forbidden edges are never
cut?\label{sub:Extension-to-Agarwal}}
To begin with, note that Agarwal et al's algorithm is a pseudo-approximation
algorithm. i.e. Although we are supposed to find the approximately
minimal directed c-balanced cut, what Agarwal et al.'s algorithm instead
offers is a $c^{\prime}$-balanced cut where $c^{\prime}\geq ac$
for \emph{any fixed} $a<1$. For the purpose of divide-\&-conquer
the $c^{\prime}$-balanced cut serve equally well and the analysis
for Leighton-Rao algorithm is actually identical. So we shall use
the terms pseudo-approximation algorithm and approximation-algorithm
(in context of Agarwal et al.'s algorithm) interchangeably.
\begin{algorithm}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{tikzpicture}[every node/.style={inner sep=0pt}, scale=0.8]
\begin{small}
\fill[color=zzttqq,fill=zzttqq,fill opacity=0.1] (3.58,-2.58) -- (2.58,-3.58) -- (-3.72,2.72) -- (-2.72,3.72) -- cycle; \draw(0,0) circle (3cm);
\draw [color=zzttqq] (3.6,-2.58)-- (2.58,-3.58);
\draw [color=zzttqq] (2.58,-3.58)-- (-3.72,2.72);
\draw [color=zzttqq] (-3.72,2.72)-- (-2.72,3.72);
\draw [color=zzttqq] (-2.72,3.72)-- (3.58,-2.58);
\draw (3.08,-3.08)-- (-3.22,3.22);
\draw [line width=1.2pt,dash pattern=on 5pt off 5pt,color=cccccc] (-2.79,1.1)-- (2.79,1.1);
\draw (0,0)-- (0,3);
\draw [color=teal,dash pattern=on 5pt off 5pt](0,3)-- (2.79,1.1);
\draw [color=teal,dash pattern=on 5pt off 5pt](0,3)-- (-2.79,1.1);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt] (0,3)-- (2.6,1.49);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt] (2.6,1.49)-- (0,-3);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt] (0,3)-- (2.56,-1.57);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt] (2.56,-1.57)-- (0,-3);
\fill [color=violet] (2.6,1.49) circle (1.5pt);
\draw[color=violet] (2.9,1.78) node {$b$};
\fill [color=violet] (2.4,1.14) circle (1.5pt);
\draw[color=violet] (2.2,1) node {${a}^1$};
\fill [color=violet] (1.22,-0.89) circle (1.5pt);
\draw[color=violet] (1.4,-0.62) node {${a}^2$};
\fill [color=violet] (0.79,-1.64) circle (1.5pt);
\draw[color=violet] (0.92,-1.36) node {${a}^3$};
\fill [color=violet] (2.56,-1.57) circle (1.5pt);
\draw[color=violet] (2.84,-1.3) node {$B$};
\fill [color=violet] (0.77,-2.57) circle (1.5pt);
\draw[color=violet] (1.06,-2.3) node {$A^1$};
\fill [color=violet] (1.31,-2.27) circle (1.5pt);
\draw[color=violet] (1.62,-1.98) node {$A^2$};
\fill [color=violet] (2.02,-1.87) circle (1.5pt);
\draw[color=violet] (2.32,-1.58) node {$A^3$};
\fill [color=uququq] (0,0) circle (1.5pt);
\draw[color=uququq] (0.16,0.28) node {$\mathcal{O}$};
\fill [color=black] (0,3) circle (1.5pt);
\draw[color=black] (0.2,3.28) node {$v_0$};
\fill [color=red!70] (-1.74,1.04) circle (1.5pt); \fill [color=black] (-1.8,-0.12) circle (1.5pt); \fill [color=black] (0.9,2.28) circle (1.5pt); \draw[color=black] (1.14,2.56) node {${\mathcal{U}_{\frac{1}{2}}}^{+}$}; \fill [color=xdxdff] (-1.07,2.8) circle (1.5pt); \fill [color=black] (-0.6,1.9) circle (1.5pt); \fill [color=red!70] (-0.56,1.06) circle (1.5pt); \fill [color=black] (0,2) circle (1.5pt); \fill [color=black] (0.18,-1.38) circle (1.5pt); \fill [color=black] (0.86,-1.96) circle (1.5pt); \fill [color=red!70] (1.58,-1.48) circle (1.5pt); \fill [color=black] (1.94,-0.48) circle (1.5pt); \fill [color=black] (2.72,0.66) circle (1.5pt); \fill [color=black] (1.6,0.38) circle (1.5pt); \fill [color=black] (2.66,-0.8) circle (1.5pt); \fill [color=black] (-0.34,-2.08) circle (1.5pt); \fill [color=black] (-1.18,-0.7) circle (1.5pt); \fill [color=black] (-0.84,-1.22) circle (1.5pt); \draw[color=black] (-0.62,-0.94) node {$\mathcal{V}^{-}$}; \fill [color=black] (0.48,0.88) circle (1.5pt); \fill [color=black] (0.64,-2.58) circle (1.5pt); \fill [color=black] (-2.4,0.38) circle (1.5pt); \fill [color=black] (-2.4,-0.42) circle (1.5pt); \fill [color=black] (0,-3) circle (1.5pt); \draw[color=black] (0.22,-2.72) node {$\overline{v_0}$}; \fill [color=black] (-2.2,-0.86) circle (1.5pt); \fill [color=uququq] (3.08,-3.08) circle (1.5pt); \fill [color=uququq] (-3.22,3.22) circle (1.5pt); \fill [color=black] (1.38,1.28) circle (1.5pt); \fill [color=black] (2.16,0.6) circle (1.5pt); \fill [color=black] (-1.5,-1.78) circle (1.5pt); \fill [color=black] (-1.62,-1.24) circle (1.5pt); \fill [color=black] (-0.84,-2.3) circle (1.5pt); \fill [color=black] (1.22,0.76) circle (1.5pt); \fill [color=black] (2.16,1.56) circle (1.5pt); \fill [color=black] (1.46,2.04) circle (1.5pt); \fill [color=black] (1,1.86) circle (1.5pt); \fill [color=black] (2.52,-0.12) circle (1.5pt); \draw[color=black] (2.9,0.16) node {${\mathcal{U}_{\frac{1}{2}}}^{-}$}; \fill [color=black] (-2.52,1.4) circle (1.5pt);
\draw[color=black] (-2.4,1.2) node {$\mathcal{V}^{+}$};
\end{small}
\end{tikzpicture}$\qquad$\begin{tikzpicture}[every node/.style={inner sep=0pt}, scale=0.8]
\begin{small}
\fill[color=zzttqq,fill=zzttqq,fill opacity=0.1] (2.15,-3.85) -- (0.83,-4.33) -- (-2.22,4.04) -- (-0.89,4.52) -- cycle;
\draw(0,0) circle (3cm);
\draw (0,3)-- (0,0);
\draw [color=zzttqq] (2.15,-3.85)-- (0.83,-4.33);
\draw [color=zzttqq] (0.83,-4.33)-- (-2.22,4.04);
\draw [color=zzttqq] (-2.22,4.04)-- (-0.89,4.52);
\draw [color=zzttqq] (-0.9,4.52)-- (2.15,-3.85);
\draw (1.49,-4.09)-- (-1.56,4.28);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt](0,3)-- (-2.43,1.76);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt](-2.43,1.76)-- (0,-3);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt](0,3)-- (-2.34,-1.88);
\draw [color=zzzzff,dash pattern=on 5pt off 5pt](-2.34,-1.88)-- (0,-3);
\draw [line width=1.2pt,dash pattern=on 5pt off 5pt,color=cccccc] (-2.9,0.7)-- (2.9,0.7);
\draw [color=teal,dash pattern=on 5pt off 5pt](0,3)-- (2.9,0.7);
\draw [color=teal,dash pattern=on 5pt off 5pt](0,3)-- (-2.9,0.7);
\draw[color=violet] (-2.12,2.04) node {$b$};
\fill [color=violet] (-2.43,1.76) circle (1.5pt);
\draw[color=violet] (-2.04,-1.6) node {$B$};
\fill [color=violet] (-2.28,1.48) circle (1.5pt);
\draw[color=violet] (-2.14,1.76) node {$a^1$};
\fill [color=violet] (-1.73,0.4) circle (1.5pt);
\draw[color=violet] (-1.46,0.68) node {$a^2$};
\fill [color=violet] (-0.6,-1.81) circle (1.5pt);
\draw[color=violet] (-0.3,-1.54) node {$a^3$};
\fill [color=violet] (-1.76,-2.16) circle (1.5pt);
\draw[color=violet] (-1.46,-1.88) node {$A^1$};
\fill [color=violet] (-1.27,-2.39) circle (1.5pt);
\draw[color=violet] (-0.96,-2.12) node {$A^2$};
\fill [color=violet] (-0.7,-2.66) circle (1.5pt);
\draw[color=violet] (-0.42,-2.38) node {$A^3$};
\fill [color=uququq] (0,0) circle (1.5pt); \draw[color=uququq] (0.16,0.28) node {$\mathcal{O}$}; \fill [color=black] (0,3) circle (1.5pt); \draw[color=black] (0.2,3.28) node {$v_0$}; \fill [color=black] (0,-3) circle (1.5pt); \draw[color=black] (0.22,-2.72) node {$\overline{v_0}$}; \fill [color=uququq] (1.49,-4.09) circle (1.5pt); \fill [color=uququq] (-1.56,4.28) circle (1.5pt); \fill [color=black] (-2.58,1.3) circle (1.5pt); \fill [color=black] (-2.02,0.92) circle (1.5pt); \draw[color=black] (-1.72,1.2) node {$\mathcal{V}^{+}$}; \fill [color=black] (-1.86,1.64) circle (1.5pt); \fill [color=black] (-1.36,0.96) circle (1.5pt); \fill [color=black] (-2.06,1.3) circle (1.5pt); \fill [color=black] (-2.42,1.06) circle (1.5pt); \fill [color=black] (-2.14,1.7) circle (1.5pt); \fill [color=black] (-1.8,2.12) circle (1.5pt); \fill [color=black] (-1.56,1.5) circle (1.5pt); \fill [color=black] (-2.26,-0.04) circle (1.5pt); \fill [color=black] (-1.44,-0.2) circle (1.5pt); \fill [color=black] (-1.44,-1.44) circle (1.5pt); \draw[color=black] (-1.14,-1.16) node {$\mathcal{V}^{-}$}; \fill [color=black] (-0.64,-1.8) circle (1.5pt); \fill [color=black] (-2.28,-0.84) circle (1.5pt); \fill [color=black] (0.5,2.28) circle (1.5pt); \fill [color=black] (2.04,1.64) circle (1.5pt); \fill [color=black] (0.98,1.32) circle (1.5pt); \fill [color=black] (1.08,2.32) circle (1.5pt); \fill [color=black] (1.9,1.06) circle (1.5pt); \fill [color=black] (1.38,1.68) circle (1.5pt); \draw[color=black] (1.84,1.96) node {${\mathcal{U}_{\frac{1}{2}}}^{+}$}; \fill [color=black] (0.84,1.76) circle (1.5pt); \fill [color=black] (1.56,-0.08) circle (1.5pt); \fill [color=black] (2.46,-0.06) circle (1.5pt); \fill [color=black] (2.04,-0.8) circle (1.5pt); \draw[color=black] (2.5,-0.52) node {${\mathcal{U}_{\frac{1}{2}}}^{-}$}; \fill [color=black] (1.84,-1.3) circle (1.5pt); \fill [color=black] (1,0.22) circle (1.5pt); \fill [color=black] (1.96,0.22) circle (1.5pt); \fill [color=black] (2.58,-0.86) circle (1.5pt); \fill [color=black] (-1.32,-2.1) circle (1.5pt); \fill [color=red!70] (-1.18,2.12) circle (1.5pt); \fill [color=red!70] (-0.16,2.02) circle (1.5pt); \fill [color=red!70] (-0.28,1.3) circle (1.5pt); \fill [color=red!70] (-0.7,2.54) circle (1.5pt); \fill [color=red!70] (0.32,0.32) circle (1.5pt);
\draw[color=uququq] (3.22,0.98) node {$radius=r$};
\end{small}
\end{tikzpicture} %
\end{minipage}}
\begin{algorithmic}[1]
\Require SDP Formulation of min-POP
\Ensure A $\frac{\alpha}{2}$-Balanced Cut $\mleft(\mathcal{A}, \overline{\mathcal{A}} \mright)$ that approximates a single subroutine of min-POP within a factor of $O(\sqrt{\log (n)})$
\LState Solve the SDP relaxation for min-POP to obtain a unit-$l_2^{2}$ representation of $m$ vectors $V_{\diamond}= \mleft \{ v_{1}\dots v_{m} \mright \}$
\LState Apply the ARV Algorithm on the set of vectors $V_{\diamond}$ to find $\alpha$-large, $\Delta$-separated sets $\mathcal{U}$ and $\mathcal{V}$
\LState Find radius $\Phi$ such that at least half of the vectors corresponding to vertices from $\mathcal{U}$ lie inside the ball of radius $\Phi$ with center at the point $v_0$, and at least half of the vectors lie outside the ball (with boundary points counted inside as well as outside).
\LState Let ${\mathcal{U}_{\frac{1}{2}}}^+$ = $\mleft\{i \in \mathcal{U}: \mleft|v_0 - v_i \mright|^2 \leq \Phi^{2} \mright\}$.
Let ${\mathcal{U}_{\frac{1}{2}}}^-$ = $\mleft\{i \in \mathcal{U}: \mleft|v_0 - v_i \mright|^2 > \Phi^{2} \mright\}$.
\LState Let $\mathcal{V}^+$ = $\mleft\{i \in \mathcal{V}: \mleft|v_0 - v_i \mright|^2 \leq \Phi^{2} \mright\}$.
Let $\mathcal{V}^-$ = $\mleft\{i \in \mathcal{V}: \mleft|v_0 - v_i \mright|^2 > \Phi^{2} \mright\}$.
\If{ $\mleft| \mathcal{V}^+ \mright| \geq \mleft| \mathcal{V}^- \mright|$ }
\LState $\mathcal{A}=\mathcal{V}^+$, $\mathcal{B}={\mathcal{U}_{\frac{1}{2}}}^-$
\Else
\LState $\mathcal{A}={\mathcal{U}_{\frac{1}{2}}}^+$,
$\mathcal{B}=\mathcal{V}^-$
\EndIf
\LState{Choose random $ \Delta \geq 0$}
\LState Choose $\sigma$ randomly s.t. $0\leq \sigma\leq\Delta$, and let $Z=V_{\diamond}-\mathcal{A}$
\LineComment{\textcolor{myindigo}{Handling of rigid edges $\mathscr{R}_{v_i}$ with $v_i$ as the destination node. }}
\ForAll {$a\in \mathcal{A}$ such that $\mleft\{v_{i}:\mleft|v_{i}-a\mright|^{2}\leq \sigma \mright\}$ } \label{lst:line:restart2}
\If {$ \nexists v_{k}\in Z$ such that $\mleft\{ v_{k}\rightarrow v_{i}\mright\} \in \mathcal{E_{\dagger}}$}
\LState $\mathcal{A}\leftarrow\mathcal{A} + \mleft\{ v_{i} \mright\}$
\EndIf
\EndFor
\LState $\overline{\mathcal{A}}=V_{\diamond}-\mathcal{A}$
\Return $\mleft( \mathcal{A},\overline{\mathcal{A}} \mright)$ \label{lst:line:reend2}
\end{algorithmic}
\protect\caption{Rounding the min-DBCRE SDP (Modification of Algorithm by Agarwal et
al.)}
\label{alg:AgarwalMod}
\end{algorithm}
The original algorithm by Agarwal et al. and its analysis can be found
in Algorithm 4 pg.7 of \cite{ACMM05}. Finally in the the lemma that
follows, we establish the fact that the rounding procedure described
in \autoref{alg:AgarwalMod} (i.e. the identification of sets $\mathcal{A}$
and $\mathcal{B}$
\begin{lemma} \label{lem:agre}
The Agarwal et.al. style of rounding (i.e. ARV rounding modified for directed balanced cuts), preserves forbidden edges.
\end{lemma}
\begin{proof}
As shown in the figures below, consider the line joining $v_0$ and the origin $\mathcal{O}$. The set of points equidistant from $v_0$ lying at a distance $\Phi$ on hypersphere $\mathcal{S}$ will lie on a hypercircle\footnote{Here, we use the non-standard term hypercircle to mean a section of the hypersphere obtained by taking all points on the hypersphere that are distance $\Phi$ from $v_0$. Each time we change the distance $\Phi$ we get a different hypercircle corresponding to it}. The projection of the hypercircle on a non-parallel 2-plane will be a segment. In the projection, this segment will always be perpendicular to line joining $v_0$ and the origin $\mathcal{O}$. The dotted segments in green denotes distance or radius $\Phi$ from $v_0$, while the dotted segment in gray is the projection of the hypercircle on the 2-plane which as we see is perpendicular to line $v_0 - \mathcal{O}$.
Consider forbidden edge constraints $\mathcal{D}(a,b)=0$ and $\mathcal{D}(A,B)=0$.
We have shown two cases of the \autoref{alg:AgarwalMod} namely when $\mleft|\mathcal{V}^{-}\mright| > \mleft|\mathcal{V}^{+}\mright|$ in which case $\mathcal{A} = {\mathcal{U}_{\frac{1}{2}}}^{+} $
and when $\mleft|\mathcal{V}^{+}\mright| > \mleft|\mathcal{V}^{-}\mright| $ in which case $\mathcal{A} = \mleft|\mathcal{V}^{+}\mright|$.
In both cases, consider the gray dotted line (projection of hypercircle) to be the Hyperplane.\\
Assume without loss of generality, that in both cases, $b$ is above the Hyperplane i.e. assume that $\mathcal{D}(v_0,b)\leq{\Phi}^{2}$ and assume that $\mathcal{D}(v_0,B)>{\Phi}^{2}$. Therefore, in both cases, $b$ that lies in $\mathcal{A}$ and $B$ that lies in $\overline{\mathcal{A}}$. We make our argument using the visual aid of the figure alongside \autoref{alg:AgarwalMod} where we consider various positions for and b. These positions considered represent all possible scenarios and there is no loss of generality in restricting our attention to the specific positions in the diagram.
Also, in both cases, consider without loss of generality three possible positions for $a$ i.e. $a^1$ lying in $\mathcal{A}$ while $a^2$ and $a^3$ lie in $\overline{\mathcal{A}}$. \\
As before the 2-plane of projection is the plane determined by points $\mathcal{O}$, $v_0$ and $b$.(We disregard the measure zero case where segment $a-b$ is perpendicular to the 2-plane determined by $\mathcal{O}$, $v_0$ and $b$.) Hence we see $b$ on the edge of the circle. But, by construction, edge $(v_0,a)$ is the hypotenuse of a right angled triangle i.e. $\mathcal{D}(v_0,a) \geq \mathcal{D}(v_0,b)$. Hence, if $b$ is in $\mathcal{A}$ then $a$ may or may not be in $\mathcal{A}$ e.g. in position $a^1$ it is seen to be in $\mathcal{A}$ while in position $a^2$ and $a^3$ it is seen to be in $\overline{\mathcal{A}}$. Hence if $b\in \mathcal{A}$, we are done. \\
Coming to the more important case, consider the 2-plane of projection is plane determined by points $\mathcal{O}$, $v_0$ and $B$ with $B$ on the edge of the circle \footnote{We disregard the measure zero case where segment $A-B$ is perpendicular to the 2-plane determined by $\mathcal{O}$, $v_0$ and $B$.}. Observe that, $B\in \overline{\mathcal{A}}$. However, once again by construction, edge $(v_0,A)$ is the hypotenuse of a right angled triangle i.e. $\mathcal{D}(v_0,A)\geq\mathcal{D}(v_0,B)$. So irrespective of position of $A$, i.e. $A^1$, $A^2$, $A^3$ (without loss of generality) or any other position, clearly the point $A \in \overline{\mathcal{A}}$. Hence we are done. i.e. the directed balanced cut rounding preserves the forbidden edge condition. \\
Finally, the only modification we make to Agarwal et al,'s rounding procedure in \autoref{alg:AgarwalMod} is the handling of rigid edges in context of \emph{fat hyperplanes} in Lines~\ref{lst:line:restart2}-\ref{lst:line:reend2}. The justification for this modification is identical to the one provided in \autoref{note:fatandrigid}.
\end{proof}
\begin{shnote}[Handling Forbidden edges via Sum of Orthogonality constraints:] \label{note:orthoARV}
From \autoref{lem:agre} it is clear that, if for all $i\in\{1,\dots,\mathcal{F}\}$ we simply add the constraint: \\
\framebox{\begin{minipage}[t]{1\columnwidth}
$\sum\limits_{i\in\{1,\dots,\mathcal{F}\}}\mathcal{D}(u_{i},v_{i})=\sum\limits_{i\in\{1,\dots,\mathcal{F}\}}(\mleft|v_{0}-v_{i}\mright|^{2}-\mleft|v_{0}-u_{i}\mright|^{2}+\mleft|v_{i}-u_{i}\mright|^{2})=0$
\end{minipage}}
to the \textbf{minimum Directed c-Balanced cut formulation} as applied on the extended Hasse Graph $\mathcal{H}_{\star}=(\mathcal{V}_{\star},\mathcal{E}_{\star})$, we can ensure that the rounding will either
\begin{enumerate}
\item choose a rigid edges (by cutting it) i.e. avoid the complimentary forbidden edges or
\item defer the decision to some later iteration
\end{enumerate}
and give us an $O(\sqrt{\log(n)}$-factor pseudo-approximation at each iteration of the modified balanced cut subroutine.
\end{shnote}
\section{An MWUM based $O(\log^{2}(\mathcal{N}))$ factor Nearly Linear Time
Algorithm \label{sec:MWUM-for-solving}}
\subsection{Arora-Kale's Primal Dual Matrix MWUM}
The solution to \textsf{MMUP} using \textsf{MWUM} has \emph{special}
consequences across a wide range of problems in computational topology.
We touch upon the most important of its applications here namely computation
of ordinary homology, persistence homology and scalar field topology
described in \ref{sec:Applications}.
\subsubsection{Multiplicative Weights Update Method}
The authors of multiplicative weights update method (MWUM), Arora,
Hazan and Kale ~ \cite{AHK05} have following to say about its general
place in computer science algorithms:
\begin{quote}
``$\ldots$We feel that this meta-algorithm and its analysis are
simple and useful enough that they should be viewed as a basic tool
taught to all algorithms students together with divide-and-conquer,
dynamic programming, random sampling, and the like$\ldots$''
\end{quote}
While MWUM as an algorithmic flavor has been there around for a while,
the recent work ~\cite{AHK05}, ~\cite{AK07} provides a more general
framework encompassing several variants over different fields developed
independently over a long period of time. See \autoref{tab:mwum}
for an analogy between the algorithmic primitives and an experts'
framework.
\begin{table}
\begin{tabular}{|c|c|}
\hline
Algorithm & Analogy\tabularnewline
\hline
\hline
The algorithm & A decision maker\tabularnewline
\hline
Weights on n variables & Weights on n experts\tabularnewline
\hline
Iteration of weight updates & Perception shift on expert supplied value\tabularnewline
\hline
Maximize objective & High total payoff in long run\tabularnewline
\hline
\end{tabular}
\protect\caption{MWUM: Experts Model}
\label{tab:mwum}
\end{table}
While obviously the best decision is not known beforehand, we do know
the objective function and there is a way to calculate the payoff
in each round. This makes it possible to to eventually arrive at optimal
decision-making by associating weights to each expert-advice, and
choosing a prediction each time based on weighted majority of experts'
prediction. Depending on whether or not the prediction is accurate,
the weights associated to experts are updated, in each round. Eventually
the weights converge leading to higher payoff decisions in successive
rounds.
The simplest approach would be to attach a 0/1 outcome to the loss
suffered by an expert at the end of round $k$. The next leap is to
generalize the setting by allowing losses suffered by the
experts to be real numbers in $[0,1]$ instead of binary values. The
penalty suffered by the $i^{th}$ expert in round $k$ is denoted
by $\mathpzc{l}_{i}^{(k)}\in[0,1]$. The multiplicative weights update method
is therefore a probabilistic experts algorithm with the following
steps:
\begin{enumerate}
\item Let $w_{i}$ be the weight assigned to the $i^{th}$ expert. Initialize $w_{i}=1$.
\item The prediction will weigh in the opinions of all the experts with probability of choosing that specific expert being proportional to $w_{i}$. In other words. the probability of choosing the $i^{th}$ expert will be:
$\frac{w(i)}{\mathcal{W}}$ where $\mathcal{W}$ is the sum of all weights.
\item Finally, update all weights at the end of the round by setting $w_{i} \leftarrow w_{i}(1-\epsilon)^{\mathpzc{l}_{i}^{(k)}}$ for all experts (assuming $\mathpzc{l}_{i}^{(k)}\geq 0$). The generalized case where $\mathpzc{l}_{i}^{(k)}$ is allowed to be smaller than 0 is treated in \cite{AHK12}.
\end{enumerate}
\pagebreak{}
\subsubsection{Primal Dual Matrix Multiplicative Weights Update Method}
\begin{notation} For the remainder of the section, we use small letters for scalars, small bold letter for vectors and capital bold letters for matrices.
\end{notation}
We shall start with the description of the combinatorial-primal dual
matrix MWUM as outlined in Arora et.al. \cite{AK07}. Accordingly,
consider the standard formulation of primal min-SDP and its dual max-SDP:
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{minSDP}]
\lpobj{min}{\bm{{C\bullet X}}}
\lpeq[minsdpbasicineq1]{\bm{A_i} \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} \bm{X} \geq b_i}{i\in[1 \dots N]}
\lpeq[minsdpbasicineq2]{\bm{X}\succeq 0}{}
\lplabel{lp:minSDPbasic}
\end{lpformulation}
\begin{lpformulation}[\textsf{maxSDP}]
\lpobj{max}{\bm{b} \raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}} \bm{y}}
\lpeq[maxbasic1]{\sum\limits_{i=1}^{n}y_i \bm{A_i}\preceq \bm{C}}{}
\lpeq[maxbasic2]{\bm{y}\geq 0}{}
\lplabel{lp:maxSDPbasic}
\end{lpformulation}%
\end{minipage}}
Here $\bm{y}$ and $\bm{b}$ are vectors of the form $\mathbf{y}=\left\{ y_{1,}y_{2,}\ldots,\, y_{n}\right\} $
and $\mathbf{b}=\left\{ b_{1,}b_{2,}\ldots,\, b_{n}\right\} $
\begin{algorithm}
\begin{algorithmic}[1]
\Require {(i.)The primal min-SDP problem instance. (ii.)Candidate objective value $\alpha$ (iii.)Accuracy parameter $\delta$ }
\Ensure {$\delta$-Feasible dual solution $\mathbf{y}$ with dual objective value $\geq \mleft( 1- \delta \mright)\alpha$ and $\delta$-Feasible primal solution $\mathbf{X}$ with primal objective value $\leq \mleft( 1+ \delta \mright)\alpha$ }
\LState {Set $\mathbf{X}^{1}=\mathbf{I}$. Let $\epsilon=\nicefrac{\delta\alpha}{2\rho n}$. Let $\varepsilon=-\ln(1+\epsilon)$. }
\ForAll{$i=1$ to $\nicefrac{8\rho^{2}n^{2}\ln(n)}{\delta^{2}\alpha^{2}}$}
\If {$\mathbf{y^i} \leftarrow$ \textcolor{myindigo}{Mod-Violation-Checking-Oracle}($\mathbf{X^i}$) fails}
\Return $\mleft(\mathbf{y^{i}},\mathbf{X^{i}}\mright)$
\Else
\LState $\mathbf{M}^{i}=\nicefrac{\mleft(\sum\limits _{j=1}^{m}\mathbf{A}_{j}y_{j}^{i}-\mathbf{C}+\rho\mathbf{I}\mright)}{2\rho}$
\LState $\mathbf{W}^{i+1}=\mleft(1+\epsilon\mright)^{\sum\limits _{k=1}^{i}\mathbf{M}^{k}}=\exp\mleft(-\varepsilon\mleft(\sum\limits _{k=1}^{i}\mathbf{M}^{k}\mright)\mright)$
\LState $\mathbf{X}^{i+1}=\nicefrac{n\mathbf{W}^{i+1}}{\mathbf{Tr}\mleft(\mathbf{W}^{i+1}\mright)}$
\EndIf
\EndFor
\Return $\mleft(\mathbf{y^{i}},\mathbf{X^{i}}\mright)$
\end{algorithmic}
\protect\caption{Arora-Kale Primal Dual Matrix MWUM for SDP}
\end{algorithm}
By using binary search, the optimization problem is reduced to a feasibility
problem. Let $\alpha$ be the binary search parameter that is passed
as an input to the primal dual matrix MWUM Algorithm below. A subroutine
called violation checking Oracle certifies the validity of the current
iterate $\mathbf{X^{i}}$ to declare whether it is primal feasible
and has objective value $\geq\alpha$. It should be noted that, the
violation-checking Oracle need not point to a single violating constraint
but may return a convex combination of constraints as an evidence
of violation. The dual vector $\mathbf{y}^{i}$ that is returned by
the Oracle plays the role of choosing a convex combination of violated
constraints. If these conditions are satisfied then the binary search
parameter $\alpha$ is updated. If not, the violation checking Oracle
generates a feedback vector $\mathbf{y^{i}}$ which is used to update
the ``multiplicative weights'' $\mathbf{W^{i+1}}$ and consequently
generate the next primal iterate $\mathbf{X^{i+1}}$ . In the next
iteration, $\mathbf{X^{i+1}}$ will be used to generate a new dual
$\mathbf{y^{i+1}}$. This interdependence of primal and dual iterates
makes this algorithm primal-dual.
\begin{shnote}[The Primal Dual Ideology] The rounding algorithm (in our case the ARV Rounding with (Agarwal et.al. + Forbidden edge) modifications) is performed directly on the Cholesky decomposition of the $\delta$-feasible solution. If the rounding algorithm succeeds, then it produces the approximately integral solution. If the rounding algorithm fails, the failure is usually dramatic enough to apply rapid corrections to candidate solution by enforcing feedback through dual solution $\mathbf{y}$.
\end{shnote}
The convergence analysis of the algorithm uses $\mathbf{Tr}\left(\mathbf{W}^{i+1}\right)$
as a potential function which is used as a normalizing factor in computing
the next primal iterate $\mathbf{X^{i+1}}$ . The number of iterations
required for determining whether $\alpha$ (as an optimal objective
value) is a good guess or not depends on the so called 'width parameter'
$\rho$. The idea is to ensure that the SDP formulation is such that
we can find the smallest real number $\rho$ which satisfies the condition
$\left\Vert \mathbf{A}_{i}y_{i}-\mathbf{C}\right\Vert \leq\rho$.
We shall assume that $\mathbf{Tr}(\mathbf{X})=n$. Also, there are
subtleties involved in fast computations involving the matrix exponential
which we won't be discussing. For these and other details please refer
to Arora et.al \cite{AK07}.
\subsubsection{The Forbidden Edges Implementation within the Violation-Checking-Oracle}
For adapting the min-POP problem to MWUM methods \footnote{ We have tried to cover maximum background possible while keeping the length of the exposition in mind. That said, for a reader wishes a more indepth understanding of the material at hand, a reading of the reference source \cite{AK07} is highly recommended. Appendix \textbf{A.2} on pg.19 which constructs Violation-Checking-Oracle is even more directly relevant. To keep things simple, for the reader, we shall use the same notation as they do wherever possible. }
we need to reformulate some of the equations of the vector program
\autoref{lp:form}\footnote{The SDP works as a relaxation of both the min-POP as well as the min-DBCRE. However, any rounding of the SDP will approximate only the min-DBCRE.}.
Following the approach delineated in \cite{AK07}, the triangle inequalities
in \autoref{triangle1} from the original formulation are replaced
with path inequalities in \autoref{triangle2} and the c-balanced
conditions \autoref{c-balanced1} are replaced by the spreading constraints
\autoref{c-balanced2}(where $a=4c(1-c)-\epsilon$ and $\overline{{\epsilon}}=1-\epsilon$).
The path inequalities are implied by the triangle inequalities and
the the spreading constraints are implied by the c-balanced conditions.
Also, the formulation \autoref{lp:form2} introduces a new condition
in \autoref{extrawidth2} that wasn't a part of the \autoref{lp:form}
formulation. As discussed in \cite{AK07}, these conditions along
with the additional n+1 variables therein are included to keep the
width bounded.
We are now in a position to formulate the SDP and its dual. The formulation
\autoref{lp:form2} as an \emph{SDP }and its dual in \autoref{lp:form3}
and \autoref{lp:form4}. Accordingly, let $\mathbf{v_{i}}\cdot\mathbf{v_{j}}=\bm{{X_{ij}}}$
i.e. $\mathbf{v_{i}}$ are a set of vectors obtained from the Cholesky
decomposition of $\bm{{X}}$. Let $c_{ij}$ be the coefficient of
$\left|\mathbf{v_{i}}-\mathbf{v_{j}}\right|^{2}$ in the objective
of the min-POP formulation above. We can now rewrite the objective
as $\min\bm{{C\bullet X}}$. Also, let $\mathbf{D}_{ij}$ be the matrix
representative for directed metric $\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})$.
The SDP and its dual formulation are described in \autoref{lp:form3}
and \autoref{lp:form4}.
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{minPOPs}]
\lpobj{min}{\frac{1}{8}\sum\limits_{(\langle \mathbf{v_i}, \mathbf{v_j} \rangle \in\mathcal{E}}\mleft|\mathbf{v_i}-\mathbf{v_j}\mright|^{2}-\mleft|\mathbf{v_0}-\mathbf{v_i}\mright|^{2}+\mleft|\mathbf{v_0}-\mathbf{v_j}\mright|^{2})}
\lpeq[forbidden2]{\sum\limits _{\langle \mathbf{u_i}, \mathbf{v_i} \rangle \in \mathcal{F}}\!\!\!\!\! \mleft|u_i-\mathbf{v_i}\mright|^{2}-\mleft|\mathbf{v_0}-\mathbf{u_i}\mright|^{2}+\mleft|\mathbf{v_0}-\mathbf{v_i}\mright|^{2} =0}{}
\lpeq[c-balanced2]{\sum_{i,j\in\mathcal{S}} |\mathbf{v_i}-\mathbf{v_j}|^{2} \geq a n^2}{\mathcal{S} \textsf{ s.t. } |\mathcal{S}|\geq \overline{\epsilon}n}
\lpeq[triangle2]{\sum\limits_{j=1}^{k-1}\mleft|\mathbf{v_{i_{j}}}-\mathbf{v_{i_{j+1}}}\mright|^{2}\geq\mleft||\mathbf{v_{i_{1}}}-\mathbf{v_{i_{k}}}\mright|^{2}}{\textsf{paths p }}
\lpeq[extrawidth2] {\mleft|\mathbf{v_{n+i}}-\mathbf{v_{n+j}}\mright|^2=0}{i,j\in[1\dots (n+1)]}
\lpeq[unit2]{|\mathbf{v_{i}}|^{2}=1}{i \in [1 \dots 2(n+1)]}
\lplabel{lp:form2}
\end{lpformulation}%
\end{minipage}}
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{minPOPs2}]
\lpobj{min}{\bm{{C\bullet X}}}
\lpeq[forbidden3]{\sum\limits _{k=1}^{\kappa} \mathbf{D}_{i_{k}j_{k}} \bullet\mathbf{X} =0}{}
\lpeq[c-balanced3]{\sum_{i,j\in \mathcal{S}} \bm{K_\mathcal{S}\bullet X} \geq a n^2 }{\mathcal{S} \textsf{ s.t. } |\mathcal{S}|\geq \overline{\epsilon}n}
\lpeq[triangle3]{\bm{T_p \bullet X} \geq 0}{\textsf{paths p }}
\lpeq[extrawidth3]{\bm{E_{ij}\bullet X} =0}{i,j\in [1 \dots (n+1)]}
\lpeq[unit3]{ \bm{X_{ii}}=1}{i \in [1 \dots 2(n+1)]}
\lpeq[sdpcond3]{\bm{X}\succeq 0}{}
\lplabel{lp:form3}
\end{lpformulation}%
\end{minipage}}
\fbox{\begin{minipage}[t]{1\columnwidth}%
\begin{lpformulation}[\textsf{maxPOPs}]
\lpobj{max}{\sum_{i}x_{i}+an^{2}\sum_{\mathcal{S}}z_{\mathcal{S}}}
\lpeq[ineqmax]{\!\! \boxbslash_{\bm{{x}}}+\!\!\!\!\!\sum_{i,j\in[1\dots n]} \!\!\!\!\!\!\! q_{ij}\bm{{E_{ij}}}+\!\! \sum_{p} f_{p}\bm{{T_{p}}}+\!\! \sum_{\mathcal{S}} \!\! z_{\mathcal{S}}\bm{{K_{\mathcal{S}}}}+\!\! \sum\limits _{k=1}^{\kappa}\mathbf{D}_{i_{k}j_{k}}y_{k}\preceq\bm{{C}} }{}
\lpeq[unitsmax1]{z_{\mathcal{S}}\geq 0}{\mathcal{S} \textsf{ s.t. }|\mathcal{S}|\geq \overline{\epsilon}n}
\lpeq[unitsmax2]{y_{k}\geq 0}{k\in[1\dots\kappa]}
\lpeq[unitsmax3]{f_{p}\geq 0}{\textsf{paths p }}
\lplabel{lp:form4}
\end{lpformulation}%
\end{minipage}}
\begin{lemma}\label{lem:telemetric} If the triangle inequalities (alternatively, path inequalities) are satisfied, then we have $\sum\limits _{j=1}^{k-1}\mathcal{D}(\mathbf{v_{i_{j}}},\mathbf{v_{i_{j+1}}}) \geq \mathcal{D}\mleft(\mathbf{v_{i_{1}}},\mathbf{v_{i_{k}}}\mright)$
\end{lemma}
\begin{proof}
Consider the term $\sum\limits_{j=1}^{k-1}\mathcal{D}(\mathbf{v_{i_{j}}},\mathbf{v_{i_{j+1}}})$. Using the definition of directed semi-metric along with a telescopic inequality we get,
\begin{eqnarray}
\sum\limits _{j=1}^{k-1}\mathcal{D}(\mathbf{v_{i_{j}}},\mathbf{v_{i_{j+1}}}) & = & \sum\limits _{j=1}^{k-1}\mleft(\mleft|\mathbf{v_{0}}-\mathbf{v_{i_{j+1}}}\mright|^{2}+\mleft|\mathbf{v_{i_{j}}}-\mathbf{v_{i_{j+1}}}\mright|^{2}-\mleft|\mathbf{v_{0}}-\mathbf{v_{i_{j}}}\mright|^{2}\mright)\label{eq:c1}\\
& = & \sum\limits _{j=1}^{k-1}\mleft(\mleft|\mathbf{v_{i_{j}}}-\mathbf{v_{i_{j+1}}}\mright|^{2}\mright)-\mleft|\mathbf{v_{0}}-\mathbf{v_{i_{1}}}\mright|^{2}+\mleft|\mathbf{v_{0}}-\mathbf{v_{i_{k}}}\mright|^{2}\label{eq:c2}\\
& \geq & \mleft|\mathbf{v_{i_{1}}}-\mathbf{v_{i_{k}}}\mright|^{2}-\mleft|\mathbf{v_{0}}-\mathbf{v_{i_{1}}}\mright|^{2}+\mleft|\mathbf{v_{0}}-\mathbf{v_{i_{k}}}\mright|^{2}\label{eq:c3}\\
& = & \mathcal{D}\mleft(\mathbf{v_{i_{1}}},\mathbf{v_{i_{k}}}\mright)\label{eq:c4}
\end{eqnarray}
\end{proof}
Note that the objective of the vector program \hyperlink{hdbcr2}{DBCRE-PATH(VP)}
is written as an inequality constraint as follows:
\begin{equation}
\sum_{\langle\mathbf{v_{i}},\mathbf{v_{j}}\rangle\in\mathcal{E}_{\underline{N}}}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})\leq\alpha\label{eq:objcon}
\end{equation}
So, instead of an optimization problem we have a feasibility problem.
Each iteration is a binary search on parameter $\alpha$. Every binary
search iteration involves solving a feasibility problem with constraints
specified in \autoref{eq:objcon}, \autoref{forbidden2} \autoref{c-balanced2},
\autoref{triangle2}, \autoref{extrawidth2} and \autoref{unit2}.
\begin{lemma}\label{lem:flipflop} If we can find a multicommodity flow such that:
\begin{enumerate}
\item (Degree constraint) For any node, the total flow on all paths starting from that node is at most $d=\tilde{O}(\nicefrac{\gamma}{n})$,
\item (Capacity constraints)For any edge, the total flow on all paths using the edge is at most $1$,
\item (Total flow constraint) Let $f_{ij}$ denote the total flow on all paths from $i$ to $j$, then $\sum\limits_{ij}f_{ij}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})>\gamma$
\end{enumerate}
Then, either the objective constraint (specified in \autoref{eq:objcon}) is violated or the triangle inequalities set of constraints (alternatively the set of path constraints in Lemma~\ref{lem:telemetric}) is violated.
\end{lemma}
\begin{proof}
We deduce \autoref{eq:a1} from Lemma~\ref{lem:telemetric}. \autoref{eq:a2} is merely a change of variables. Finally \autoref{eq:a3} follows from Condition 3 of the statement of Lemma~\ref{lem:flipflop} (i.e. total flow constraint).
\begin{eqnarray}
\sum\limits_{\mathcal{P}(i,j)}\sum\limits_{\langle\mathbf{v_{k}},\mathbf{v_{l}}\rangle\in\mathcal{P}(i,j)}f_{\mathcal{P}(i,j)}\mathcal{D}(\mathbf{v_{k}},\mathbf{v_{l}}) & \geq & \sum\limits _{\mathcal{\mathcal{P}}(i,j)}f_{\mathcal{P}(i,j)}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}}) \label{eq:a1}\\
& = & \sum\limits _{ij}f_{ij}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})\label{eq:a2}\\
& > & \gamma\label{eq:a3}
\end{eqnarray}
\autoref{eq:b1} is merely a rearrangement of summations. \autoref{eq:b2} is obtained from Condition 2 of the statement of Lemma~\ref{lem:flipflop} (i.e. capacity constraints). Finally \autoref{eq:b3} follows from assuming that the objective constraint in \autoref{eq:objcon} holds.
\begin{eqnarray}
\sum\limits_{\mathcal{P}(i,j)}\sum\limits_{\langle\mathbf{v_{k}},\mathbf{v_{l}}\rangle\in\mathcal{P}(i,j)}f_{\mathcal{P}(i,j)}\mathcal{D}(\mathbf{v_{k}},\mathbf{v_{l}}) & = & \sum\limits _{\langle v_{i}v_{j}\rangle\in\mathcal{E}}\sum\limits _{\mathcal{P}(i,j)}f_{\mathcal{P}(i,j)}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})\label{eq:b1}\\
& \leq & \sum\limits _{\langle v_{i}v_{j}\rangle\in\mathcal{E}}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})\label{eq:b2}\\
& \leq & \gamma\label{eq:b3}
\end{eqnarray}
\end{proof}
\begin{lemma}\label{lem:aklemma} Let $\mathcal{S}\subseteq\mathcal{V}$ be a set of nodes of size $\Omega(n)$. Suppose we are given for all $i \in \mathcal{S}$, vectors $v_i, w_i$ of length $O(1)$, s.t. $\forall i,j, \left\Vert w_i - w_j \right\Vert^{2}\leq o(1)$ and $\sum\limits_{i,j\in\mathcal{S}}\left\Vert \mathbf{v_{i}}-\mathbf{v_{j}}\right\Vert ^{2}\geq\Omega(n^{2})$. For any given $\alpha$,
\begin{enumerate}
\item There is an algorithm which using a single max-flow computation, either outputs a $c^{\prime}$-balanced cut of expansion $O(\log (n) \frac{\alpha}{n}$ or a valid O($\frac{\log (n)\alpha}{n}$)-regular directed flow $f_{ij}$ s.t. $\sum\limits_{i,j} f_{ij}\mathcal{D}(i,j)\geq \alpha$.
\item There is an algorithm which using $O(\log n)$-max flow computations outputs either:
\begin{enumerate}
\item $\Omega(\frac{n}{\sqrt{\log n}})$ vertex disjoint paths s.t. the path inequality along these paths is violated by $\Omega(1)$ or
\item a $c^{\prime}$-balanced cut of expansion $O\sqrt{\log n}\frac{\alpha}{n}$ or
\item a valid $O(\frac{\alpha}{n})$-regular direced flow $f_p$ flow s.t. $\sum\limits_{i,j}f_{ij}\mathcal{D}(i,j)\geq\alpha$.
\end{enumerate}
\end{enumerate}
\end{lemma}
\begin{proof} Please see Lemma 3, Section 4.3 of \cite{AK07}. The proof is provided in Section A.2 pg.21 of \cite{AK07}.
\end{proof}
We need to find a procedure to verify the following forbidden edge
constraints:
\begin{equation}
\boldsymbol{D}_{i_{k}j_{k}}\bullet\boldsymbol{X}\leq\sigma\ensuremath{\quad\forall k\in[1,\kappa]}\label{eq:forbidobjmain}
\end{equation}
Instead we check for the single constraint \autoref{eq:forbidobj} which
if satisfied implies all constraints in \autoref{eq:forbidobjmain}.
\begin{equation}
\sum_{k=1}^{\kappa}\boldsymbol{D}_{i_{k}j_{k}}\bullet\boldsymbol{X}\leq\sigma\ensuremath{}\label{eq:forbidobj}
\end{equation}
We need to do this in a manner that the width of the formulation stays
bounded.
In particular, our Oracle needs to find a feedback matrix $\mathbf{Q}$
that satisfies the following two equations:
\begin{equation}
\boxbslash_{\boldsymbol{x}}\bullet\boldsymbol{X}-\mathbf{Q}\bullet\boldsymbol{X} \leq 0
\label{eq:vcwide1}
\end{equation}
\begin{equation}
\mleft\Vert \boxbslash_{\boldsymbol{x}}-\mathbf{Q}\mright\Vert \leq\rho
\label{eq:vcwide2}
\end{equation}
$\boxbslash_{\boldsymbol{x}}$ is a diagonal matrix with vector $\mathbf{x}$
on its diagonal .\autoref{eq:vcwide1} needs to be an evidence of violation
of a convex combination of forbidden edge constraints in \autoref{eq:forbidobjmain}.
\autoref{eq:vcwide2} is the width constraint.
Our VC-Oracle can be described by the following steps:
\begin{enumerate}
\item Assume that the Oracle has already verified the constraints sepecified
in \autoref{c-balanced3}, \autoref{extrawidth3} and \autoref{unit3}.
If any of these constraints are violated then a feedback matrix is
generated in a manner identical to the Oracle for min-directed c-balanced
cut as specified in Appendix A.2 of \cite{AK07}.
\item Owing to step 1, the conditions of Lemma 10, Appendix B. of \cite{AK07}
are therefore satisfied by vectors $\mathbf{v_{i}}$. We pick a random
unit vector $\mathbf{u}$. Analogous to the ARV-Algorithm, we let $\mathcal{L}=\left\{ i:\mathbf{v}_{\mathbf{i}}\cdot\mathbf{u}<-\sigma\right\} $
and let $\mathcal{R}=\left\{ i:\mathbf{v_{i}}\cdot\mathbf{u}>+\sigma\right\} $,
for $\sigma=\varTheta\left(\frac{1}{\log n}\right)$. Therefore, the
sets obtained $\mathcal{L}$ and $\mathcal{R}$ obtained are of size
$\Omega(n)$ with constant probability. Subsequently, connect all
nodes in $\mathcal{L}$ to a single source with edges of capacity=degree$(d)$,
and all nodes in $\mathcal{R}$ to a single sink, once again with
edges of capacity $d$. Run a single-commodity max-flow algorithm
along with the dummy source and sink. To begin with, we attempt to
find a $O\left(\frac{\log(n)\alpha}{n}\right)$-regular directed flow
$f_{ij}$ s.t. $\sum\limits _{i,j}f_{ij}\mathcal{D}(i,j)\geq\alpha$
(which is shown to be equivalent to the condition of max-flow value
exceeding $O(\alpha\log n)$). Ignoring the flow on source and sink
edges we obtain the max-flow gives the required multicommodity flow
$f_{ij}$ prescribed by Lemma~\ref{lem:aklemma} .
\begin{enumerate}
\item If we are successful in finding such a flow, then using Lemma~\ref{lem:flipflop}(substituting
$\gamma=\alpha$), we infer that either the objective constraints
specified in \autoref{eq:objcon} or the set of path constraints specified
in Lemma~\ref{lem:telemetric} are violated. Moreover, because of Condition
1 of Lemma~\ref{lem:flipflop}(i.e. Degree constraints), the width of
the formulation is bounded. Assuming that the condition \autoref{c-balanced3}
is satisifed, we choose any set $\mathcal{S}$, s.t. $\sum_{i,j\in\mathcal{S}}\bm{K_{\mathcal{S}}\bullet X}\geq an^{2}$
and $\mathcal{S}\textsf{ s.t. }|\mathcal{S}|\geq\overline{\epsilon}n$\}. Now, a
Laplacian $D$ of the complete weighted graph is constructed s.t.
for ever edge $e_{ij}$ s.t. $i\in\mathcal{S}$ and $j\in\mathcal{T}$,
we associate weight $f_{ij}$. For all other edges we assign weight
$0$. It is shown in Thm. 5 of \cite{AK07}, that using $\boxbslash_{\boldsymbol{x}}-D$
is the requisite feedback matrix. Because the flow is d-regular, the
width stays bounded.
\item It is proved in \cite{AK07} that if this max-flow has value less
than $O(\alpha\log n)$, then the min-cut found in the process is
an $O(\log n)$ approximation to the min directed c-balanced separator.
\end{enumerate}
\item If we find the requisite max-flow in step 2, then we obtain a feedback
matrix and we are done. Else we have found a min-cut from step 2,
which happens to be an $O(\log n)$ approximation to min-directed
c-balanced separator. Let us denote this cut as $\mathcal{C}_{1}$.
Now, we apply the same procedure as step 2, except that now we try
to find a flow that satisfies the path inequalities constraints from
Lemma~\ref{lem:telemetric} along with \autoref{eq:forbidobj}. Here
we intend to apply Lemma~\ref{lem:flipflop} with \autoref{eq:forbidobj}
as a replacement for objective constraint \autoref{eq:objcon}.
\begin{enumerate}
\item Now, owing to the outcome of step 1, we already know that the path
inequalities are satisified. So, if we are able to find the requisite
flow, it must surely be an indicator of the fact that the forbidden-edge
constraints are violated. Here we are using Lemma~\ref{lem:flipflop}
by substituting $\gamma=\sigma$. As in case of 2.(a), we construct
a new Laplacian, say $D_{2}$ of complete weighted graph. Note that,
$D_{2}\bullet X=\sum\limits_{ij}f_{ij}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})\geq\sigma$.
(However, in this case the degree of the flow will be $O(\frac{\sigma}{n})$.)
\item If we are unable to find such a flow, then we output cut $\mathcal{C}_{1}$
as our required cut which approximates min-directed balanced cut (with
rigid edges) up to a factor of $O(\log n)$.
\end{enumerate}
\end{enumerate}
\begin{lemma}[VC-Oracle] The Violation Checking Oracle from \cite{AK07} can be extended to include a check on the forbidden edges condition violation. This check can be done in linear time in a manner that respects the width condition of the primal-dual formulation.
\end{lemma} \label{lem:vcoracle}
\begin{proof}
The handling of equations \autoref{c-balanced3}, \autoref{unit3} and \autoref{extrawidth3} in step 1 and handling of \autoref{eq:objcon}+\autoref{triangle3} in step 2 is identical to handling of corresponding equations in the directed balanced cut formulation of Section A.2 in \cite{AK07}. So, we need to prove correctness of step 3. \\
Suppose that we are able to find the requisite max-flow in step 3, s.t.
\begin{enumerate}
\item Total flow on all paths starting from that node is at most $d=\tilde{O}(\nicefrac{\sigma}{n})$,
\item For any edge, the total flow on all paths using the edge is at most $1$,
\item (Let $f_{ij}$ denote the total flow on all paths from $i$ to $j$, then $\sum\limits_{ij}f_{ij}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})>\sigma$
\end{enumerate}
then the three conditions of Lemma~\ref{lem:flipflop} are satisfied and we have evidence that either the path inequalities of Lemma~\ref{lem:telemetric} are violated or the constraint specified in \autoref{eq:forbidobj} is violated. But, from step 2, we already know that path inequalities are satisifed. So, it must be that \autoref{eq:forbidobj} is violated.
Let $\boxbslash_{\boldsymbol{x}}$ be the diagonal matrix with the vector $\boldsymbol{x}$ on the diagonal. In order to satisfy \autoref{eq:vcwide1}, we Choose $x_{i}=\nicefrac{\sigma}{n}$ and $\mathbf{Q}=D_2$. Thus, we obtain $\left(\frac{\sigma}{n}I-D_{2}\right)\bullet X\leq\sigma-\sigma=0$ since $D_{2}\bullet X=\sum\limits_{ij}f_{ij}\mathcal{D}(\mathbf{v_{i}},\mathbf{v_{j}})\geq\sigma$. Hence, $\boxbslash_{\boldsymbol{x}}-D_2$ is our feedback matrix. Also, from Thm. 5 of \cite{AK07}, we know that for a d-regular flow, we have $0\preceq D_2 \preceq 2dI$. In our case, $d=\nicefrac{\sigma}{n}$. Also, $\boxbslash_{\boldsymbol{x}}=\frac{\sigma}{n}I$. Therefore, clearly, $\mleft\Vert \boxbslash_{\boldsymbol{x}}-\mathbf{Q}\mright\Vert \leq\tilde{O}\left(\nicefrac{\sigma}{n}\right)$. But since $\sigma$ can be chosen to be a small number. For our purposes, it suffices to let $\sigma=\varTheta(\nicefrac{1}{\log n})$. Also, clearly, our minimum for MMUP must be always at least 1, irrespective of the complex under consideration. So we always have $\alpha\geq 1$. In other words $\alpha \geq \sigma $. Therefore, $\mleft\Vert \boxbslash_{\boldsymbol{x}}-\mathbf{Q}\mright\Vert \leq\tilde{O}\left(\nicefrac{\alpha}{n}\right)=\rho$, satisfying \autoref{eq:vcwide2}. \\
Now, if we are unable to find the max-flow that guarantees a forbidden-edge violation, then clearly the current configuration of vectors is such that almost all the forbidden edge constraints are \emph{nearly} satisfied. We say nearly because ideally we would have preferred $\boldsymbol{D}_{i_{k}j_{k}}\bullet\boldsymbol{X}=0, \forall k\in[1,\kappa]$. $\sigma$ is made so small that we can be sure that for all practical purposes, for any forbidden edge $\mathcal{D}(\mathbf{v_i},\mathbf{v_j})$, the segments $v_{i}-v_{0}$ and $v_{j}-v_{i}$ are almost orthogonal. So for \emph{any} cut, nearly all forbidden edge constraints will be obeyed. Using first part of Lemma~\ref{lem:aklemma} we get an approximation of $O(\log n)$ for min directed c-balanced cut with (nearly satisfied) forbidden edge constraints. Now suppose it so happens that this cut violates some of the forbidden edge constraints. For each of the forbidden edges, we have $\mathcal{D}(\mathbf{v_i},\mathbf{v_j})<\sigma$. Since, $\sigma=\varTheta(\nicefrac{1}{\log n})$, we can ensure that if a pair violates forbidden edges inspite of obeying the relaxed forbidden edge constraint in \autoref{eq:forbidobjmain}, then at least one of the two points $\mathbf{v_i},\mathbf{v_j}$ must lie \emph{inside the fat hyperplane}. Note that the approximation of $O(\log n)$ for min-directed c-balanced cut is achieved through points that lie outside the fat hyperplane and redistribution of points that lie inside the fat hyperplane to either of the two sets $\mathcal{L}$ or $\mathcal{R}$ does not affect approximation ratio. Therefore, the points in the fat hyperplane are redistributed in a manner analogous to the Agarwal et al. rounding. So, we can use cut $\mathcal{C}_1$ from step 2 and we are done.
\end{proof}
The authors of \cite{AK07} point out a striking analogy when they
say that their matrix multiplicative weights update method is the
SDP analog of Young's ``Randomized rounding without solving the LP''\cite{Yo95}.
Now, apart from the orthogonality constraint, all other conditions
for the violation checking Oracle are treated in exactly the same
manner as directed balanced separator explained in Section 4.3 pg
14 of \cite{AK07}. The treatment of triangle inequalities (that are
rewritten as path inequalities in the above formulation) deserve a
special mention. In effort to refute these inequalities one comes
face-to-face with the ``round-without-solve'' character of MWUM.
For this reason, we briefly discuss how single commodity flow (alternatively
multicommodity flow) may be used to
\begin{enumerate}
\item find a combination of constraints (triangle inequalities and/or objective-function
constraint) that are violated
\item directly obtain an approximate solution of the SDP!
\end{enumerate}
\subsubsection{From Triangle Inequalities to Multicommodity Flow \label{sub:From-Triangle-Inequalities}}
Note that, the objective function of the SDP is modeled as a constraint:
\begin{equation}
C\bullet X\leq\alpha\label{eq:objcon2}
\end{equation}
where $\alpha$ is a binary search parameter thereby reducing the
optimization problem into a a sequence of feasibility problems dictated
binary search. For every iteration, the violation checking Oracle
must check if \autoref{eq:objcon2}is violated. With a certain ingenuity,
\cite{AK07} reduce the problem of checking the first constraint and
the triangle inequality constraints into a multicommodity flow problem
which is specified in the following manner (pg 8. of \cite{AK07}).
\begin{enumerate}
\item For any node, the total flow on all paths starting from the node is
constrained by at most $d=\tilde{O}(\nicefrac{\alpha}{n})$ (degree
constraints)
\item For any edge $e_{ij}$, the total flow on all paths using the edge
is at most $c_{ij}$ (capacity constraints) (where $\sum_{i,j\in E}c_{ij}\|v_{i}-v_{j}\|^{2}$
is the objective function)
\item If $f_{ij}$ is the total flow on all paths from $i$ to $j$, then
$\sum\frac{1}{4}f_{ij}\left\Vert v_{i}-v_{j}\right\Vert ^{2}>\alpha$
\end{enumerate}
Lemma 1 on pg 8. of \cite{AK07} proves that there exists an algorithm
which either outputs a multicommodity flow prescribed above or outputs
a $c^{\prime}$-balanced cut of expansion $O(\sqrt{\log n}\frac{\alpha}{n})$.
If we find the prescribed multicommodity flow, then we have found
a violation. Else we have found the desired cut of certain expansion
that eventually helps us arrive at an $O(\sqrt{\log n})$ ratio. The
multicommodity flow problem has certain subtleties concerning the
degree regularity condition. Once these subtleties are taken care
of, one may solve the multicommodity flow using algorithms devised
in \cite{Fl00,GK98}. We are however interested in the more efficient
max-flow solution described in \autoref{sub:Max-flows-in-MWUM}.
\subsubsection{Max-flows in MWUM\label{sub:Max-flows-in-MWUM}}
The description of the max-flow subroutine is provided in proof of
Lemma 1 on pg.12 of \cite{AK07}. To begin with one assumes that all
other conditions except for the condition specified in \autoref{eq:objcon2}
and the triangle inequalities are satisfied in this iteration. Consider
a random vector $\mathbf{u}$. It is proved in Lemma 10, Appendix
B of \cite{AK07} that it is possible to find two sets $L$ and $R$
or size $\Omega(n)$ s.t.:
\begin{equation}
\forall i,j\quad(\mathbf{v_{i}}-\mathbf{v_{j}})\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}\mathbf{u}\geq\nicefrac{\sigma}{\sqrt{n}}
\end{equation}
for some constant $\sigma$. The two sets $L$ and $R$ are extracted
using projections on random vector $\mathbf{u}$. Here the procedure
is analogous to one found in \autoref{alg:ARV} in \autoref{sec:Rounding-Based-on}.
The proofs of correctness are similar to those employed in Lemma 10
of \cite{AK07}.
The second part of Lemma 1 on pg 8. of \cite{AK07} concerns the use
of max-flows in MWUM computations. If the max-flow has value less
than $O(\log n\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}}\alpha)$, then the min-cut found in the
process is an $O(\log(n))$ approximation to the minimum c-balanced
separator. Otherwise the max-flow gives the required multicommodity
flow that establishes violation of a combination of constraints. Therefore
we obtain an $O(\log n)$ approximation to
\subsubsection{Discussion on Time Complexity}
We have already observed in \autoref{thm:REMain2}, that the reduction
procedure (from MMUP to min-POP) takes linear time. Now we need to
establish the nearly linear time complexity of the approximation algorithm
for min-POP. To begin with recall that we use Leighton-Rao style divide-\&-conquer
by recursively solving min-directed balanced cut (with rigid edges)
(min-DBCRE) in order to solve min-POP. We have modified the matrix
MWUM from \cite{AK07} to solve the directed balanced cut (min-DBC)
problem (with rigid edges) in \autoref{sec:MWUM-for-solving}. In
\cite{AK07}, pg. 3 Table 1, the run times for directed balanced separator
are listed. Also, Theorem 8 on pg. 14, Section 4.3 along with Appendix
A.2 of \cite{AK07} gives us details of how one may use polylogarithmic
number of single commodity max-flow computations to obtain an $O(\log n)$
pseudo-approximation. Moreover, the computational bottleneck of this
MWUM algorithm for min-DBC is the runtime of the max-flow subroutine.
At the time when \cite{AK07} was written the (asymptotically) best
algorithm for max-flow was $O(m^{1.5})$ time. which allowed them
to obtain an $O(\log n)$ time approximation in time $\tilde{O}(m^{1.5})$
time. But since then, $\tilde{O}(m)$ time single commodity max-flow
algorithms have been independently developed in \cite{Ke14} and \cite{Sh13}.
If we use these new state-of-art max-flow algorithms as subroutines
in the violation checking oracles of MWUM, we obtain an $\tilde{O}(m)$
time algorithm for min DBC and hence also for the min-DBCRE given
the fact that we do not add any extra superlinear computational cost
to handle rigid edges in min-DBCRE. Now, since we use Leighton-Rao
style divide-\&-conquer, we can use an elementary computation to see
that inspite of the recursive application of min-DBCRE (on smaller
subproblems), the asymptotic computational cost of min-POP remains
bounded by $\tilde{O}(m)$.
\section{Applications\label{sec:Applications}}
\subsection{Homology}
Computing the homology groups has several applications, particularly,
in material sciences, imaging, pattern classification and CAPD (computer
assisted proofs in dynamics). Once effective implementations for homology
computation like CHomP (by Mischaikow's group) and RedHom (by Mrozek's
group) became publicly available, problems in dynamics were routinely
translated to problems about computing homology on cubical sets. More
recently, homology is appraised to be a more widely applicable computational
invariant of topological spaces arising from practical data sets of
interest to a rapidly growing community of computer scientists and
applied mathematicians \cite{Ca09}. Ordinary homology groups are
algebraic objects that encode geometric features (n-dimensional holes).
Given a point cloud sample, the goal of persistent homology (as opposed
to ordinary homology) is the study of multiscale features of that
space. The notion of filtration allows us to study the multiscale
representation of that space.
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\centering
\begin{tikzpicture}
\fill [color=blue!30] (-1.7,0)--(1,-0.49)--(0,1.7);
\draw [color=red!50] (-1.7,0)--(1,-0.5)--(0,1.7)--(-1.7,0);
\draw [color=red!50] (1,-0.5)--(2.5,0.4)--(0,1.7);
\node [below] at (-1.7,0) {$v_0$};
\node [below] at (1,-0.5) {$v_1$};
\node [above] at (0,1.7) {$v_2$};
\node [above] at (2.5,0.4) {$v_3$};
\end{tikzpicture}
\begin{displaymath}
\partial_1=
\bordermatrix{ ~ & v_0v_1&v_0v_2&v_1v_2&v_1v_3&v_2v_3
\cr v_0&-1&-1&0&0&0
\cr v_1& 1&0&-1&-1&0
\cr v_2& 0&1&1&0&-1
\cr v_3& 0&0&0&1&1
\cr },
\end{displaymath} \begin{displaymath}
\partial_2=
\bordermatrix{ ~&v_0v_1v_2\cr v_0v_1&1\cr v_0v_2&-1\cr v_1v_2&1\cr v_1v_4&0\cr v_2v_3&0\cr }.
\end{displaymath}%
\end{minipage}}'
\protect\caption{Simplicial Complex $\mathcal{K}$ and its boundary operator }
\label{fig:complex_x}
\end{figure}
\begin{algorithm}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\protect\caption{MMUP-based Homology Computation}
\begin{algorithmic}[1]
\LineComment{ Let $\mathcal{K}$ denote the input simplicial complex and $\mathcal{H}$ denote its Hasse graph Representation.}
\LState {$\mathcal{H} \colonequals \textbf{HasseDiagramRepresentation}(\mathcal{K})$}
\LState {$\mathcal{D}\colonequals\textbf{MMUP-DGVF}(\mathcal{H})$}
\LState {$\mathcal{F} \colonequals \textbf{TopologicalSort}(\mathcal{D}$)}
\LState {$\Delta \colonequals\textbf{DesignBoundaryOperator}(\mathcal{F},\mathcal{D})$}
\LState {$\Delta \colonequals\textbf{CriticalCellCancellaton}(\Delta)$}
\LState {$H(K,A) \colonequals \textbf{SmithNormalForm}(\Delta_{\varLambda},\varLambda,A$)}
\end{algorithmic}
\begin{shnote}
Any total order (compatible with the partial order specified by the DGVF) on cells gives an implicit discrete Morse function. To get such a function, one simply assigns the index of the cell in the sorted list as the value of the Morse function. This explains the use of topological sort in the above Algorithm.
\end{shnote}%
\end{minipage}}
\end{algorithm}
Let $\varLambda=\sum\limits _{i=0}^{2}m_{i}$ and let $\varUpsilon=\sum\limits _{i=0}^{2}\beta_{i}$.
\begin{theorem}[Complexity of computing boundary operator]
The boundary operator can be computed in $O(\varLambda \times \mathcal{N})$ time.
\end{theorem}
We refer the reader to \autoref{sub:Boundary-Operator-Computation}
for the Algorithm~\ref{alg:bdryOp}. \autoref{thm:bdryOpCorr} establishes
the correctness of this algorithm.
Morse theory allows cancellation of pairs of critical cells using
gradient reversals. One then computes a new boundary operator upon
cancellation. The criterion for cancellation of critical cells is
easily identifiable from the boundary operator. In principle (to get
a further reduction in number of critical cells), such a cancellation
routine precedes the homology subroutine and succeeds the boundary
operator design subroutine. To keep the discussion simple we do not
discuss critical cell cancellation here. A discussion of the newly
non-smooth cancellations can be found found in \autoref{sub:Smooth-and-Non-smooth}.
It is easy to see that under the weak Morse optimality condition (WMOC)
discussed in \autoref{sub:Problem-Definition}, we can compute homology
groups in nearly linear time by using the MMUP algorithm to reduce
the complex size. combinatorial Hodge theory has been used in past
to compute homology (instead of the standard Smith normal form)\cite{Fri96}.
We show that using Hodge theory \cite{Ra14f}, we can obtain drastic
reductions in runtime of the algorithm (under the WMOC)\cite{Ra14f}.
\subsection{Persistent Homology }
We will describe how the most na\"{i}ve\xspace application of MMUP-APX algorithm
to compute persistent homology leads to substantial gains in time
complexity. A filtration on a simplicial complex $\mathcal{K}$ is a collection
of subcomplexes $\{\mathcal{K}(t)\ |\ t\in\mathbb{R}\}$ of $\mathcal{K}$ s.t.
$\mathcal{K}(t_{i})\subset\mathcal{K}(t_{j})$ when $t_{i}\leq t_{j}$. Typically,
In \emph{persistent homology}, one studies the topological invariants
of the nested family of complexes described by such a filtration.
Alternatively one studies sublevel sets of an explicitly provided
scalar function i.e., one studies topological features of a space
induced either by a function or a filtration. Its algebraic framework
lends it generality whereas its geometric proximity to Morse theory
gives it stability with regard to noise. Persistence gives us a natural
grading of features at multiple resolutions and sieves out noise from
features. This has led to several applications in shape analysis,
image analysis, and data analysis\cite{Ca09,EH08,EH10}. As we can
see from the \autoref{fig:cechcomplex} and \autoref{fig:vietorisripscomplex},
the very notion of 'holes' is fuzzy for a point cloud dataset, in
the sense that it is dependent on the scale at which we construct
the complex. An increase in scale leads to gradual thickening of the
space. Clearly, some holes last longer (i.e. they are more persistent)
than others. The underlying intuition behind using persistent homology
for a variety of applications is that noise is likely to have low
persistence whereas features that are truly reflective of underlying
topology will be more persistent.
\begin{definition}[Nerve of a Complex] Suppose that we are given a finite collection of nonempty convex sets $\mathcal{S}=\mleft\{U_j,\ j \in J\mright\}$. Then \emph{nerve} of $\mathcal{S}$ is given by
\begin{displaymath}
\mathscr{N}(\mathcal{S})=\mleft\{ \sigma \subseteq I \mid \cap_{j \in \sigma}U_j\neq \varnothing \mright\}.
\end{displaymath}
\end{definition}
\begin{definition}[\v{C}ech complex]
Suppose that we are given a $(\mathbb{R}^n,d)$ a metric space, $\varXi$ a finite set of points in $\mathbb{R}^n$, and $\kappa>0$. Then \v{C}ech complex for $\varXi$ and $\kappa$ is isomorphic to the nerve of the collection of $n-$balls of radius $r$ with centers in $\varXi$. \begin{displaymath} \mathscr{C}_{\kappa}(\varXi)=\mleft\{\sigma \subseteq \varXi \mid \cap_{j \in \sigma}B_{v_j}(\kappa)\neq \varnothing \mright\}. \end{displaymath}
\end{definition}
\begin{definition}[Vietoris-Rips complex] Suppose that we are given $(Q,d)$, a metric space, $\varXi$ a finite set of points in $Q$, and $\kappa>0$. The Vietoris-Rips complex of parameter $\kappa$ of $\varXi$, denoted by $\mathcal{R}_{\kappa}(\varXi)$, is the abstract simplicial complex whose $n$-simplices correspond to unordered $(n+1)$-tuples of vertices in $\varXi$ that are within pairwise distance less than $2\kappa$ of each other. \end{definition}
Consider a simplicial complex $\mathcal{K}$ of size $n$. A \emph{filtration} of a simplicial complex is an ordering relation on its simplices which respects inclusion. Consider a function $\theta : \mathcal{S} \rightarrow {\rm I\!R}$. Here, $\mathcal{S}$ denotes the set of non-empty subsets of $\mathcal{K}$. The filtration parameter $\theta$ is monotonic in the sense that, for any two simplices $\alpha \subseteq \gamma$ in $\mathcal{K}$, $\theta$ satisfies $\theta(\alpha) \leq \theta(\gamma)$ (analogous to discrete Morse theory). We will call $\theta(\sigma)$ the filtration value of the simplex $\sigma$. Topologically speaking, the monotonicity condition is a semantic necessity as it is required to ensure that sublevel sets of $\mathcal{K} (q) = \theta^{-1}(-\infty,q]$ are subcomplexes of $\mathcal{K}$, for every $q \in {\rm I\!R}$. Consequently, we arrive at the following sequence of $n+1$ subcomplexes: \[ \emptyset = \mathcal{K}_0 \subseteq \mathcal{K}_1 \subseteq \cdots \subseteq \mathcal{K}_n = \mathcal{K} \textnormal{ s.t. } \mathcal{K}_i = \theta^{-1}(-\infty,\theta_i] \] The filtration induces a sequence of homomorphisms in the homology and cohomology groups which can be written as:
\begin{eqnarray} 0 = H_p(\mathcal{K}_0) \rightarrow H_p(\mathcal{K}_1) \rightarrow \cdots \rightarrow H_p(\mathcal{K}_{n-1}) \rightarrow H_p(\mathcal{K}_n) = H_p(\mathcal{K}) \label{hom}\\ 0 = H^p(\mathcal{K}_0) \leftarrow H^p(\mathcal{K}_1) \leftarrow \cdots \leftarrow H^p(\mathcal{K}_{n-1}) \leftarrow H^p(\mathcal{K}_n) = H^p(\mathcal{K}) \label{cohom}
\end{eqnarray}
New simplices are introduced one-by-one in the order prescribed by the filtration. Note that the filtration order (as required in persistent homology) is a total order as opposed to partial orders prescribed by discrete Morse functions. It can be formally proved (but can also be intuitively seen) that when simplex $\alpha^d$ is introduced at time T in the sequence, it either leads to an increase in Betti number $\beta_d$ by 1 or a decrease in $\beta_{d-1}$ by 1. Accordingly $\alpha^d$ is called a positive or a negative simplex. Therefore, computing persistent homology of a filtration is akin to pairing each simplex that creates a homology feature (positive simplex) with the one that destroys it (negative simplex). This is analogous to cancellation of critical cells as observed in discrete Morse theory. While typically one outputs a \emph{persistence diagram}, which is a plot of the points $(\theta(\alpha),\theta(\gamma))$ for each persistent pair $(\alpha,\gamma)$ at times the underlying algebraic structures (namely the persistent homology modules) are the more applicable part of the output. In summary, the computational goal of persistent homology is to detect when a new homology class is born and when an existing class dies as one proceeds in the order prescribed by the filtration.
\begin{figure}
\begin{center}
\newcommand{\gbullet}[1]
\fill[color=red!70!black,semitransparent] #1 circle (2pt);
}
\newcommand{\covset}[2]
\fill[color=blue!30,semitransparent] #2 circle (#1);
}
\newcommand{\covering}[1]
\covset{#1}{(90:1)}
\covset{#1}{(210:1)}
\covset{#1}{(330:1)}
}
\begin{tabular}{c@{\qquad}c@{\qquad}c}
\begin{tikzpicture}[baseline={(0,0)}]
\covering{0.4}
\gbullet{(90:1)}
\gbullet{(210:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\begin{tikzpicture}[baseline={(0,0)}]
\covering{0.92}
\draw (90:1) -- (210:1);
\draw (210:1) -- (330:1);
\draw (330:1) -- (90:1);
\gbullet{(90:1)}
\gbullet{(210:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\begin{tikzpicture}[baseline={(0,0)}]
\covering{1.1}
\fill[color=red!50,semitransparent] (90:1) -- (210:1) -- (330:1) -- cycle;
\begin{scope}[color=red]
\draw (90:1) -- (210:1);
\draw (210:1) -- (330:1);
\draw (330:1) -- (90:1);
\end{scope}
\gbullet{(90:1)}
\gbullet{(210:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\\[1em]
\end{tabular}
\end{center}
\caption{ \v Cech Complex with Increasing Radii(parameter)}
\label{fig:cechcomplex}
\end{figure}
\begin{figure}
\begin{center}
\newcommand{\gbullet}[1]
\fill[color=red!70!black,semitransparent] #1 circle (2pt);
}
\newcommand{\covset}[2]
\fill[color=blue!30,semitransparent] #2 circle (#1);
}
\newcommand{\covering}[1]
\covset{#1}{(60:1)}
\covset{#1}{(180:1)}
\covset{#1}{(330:1)}
}
\begin{tabular}{c@{\qquad}c@{\qquad}c} \begin{tikzpicture}[baseline={(0,0)}]
\covering{0.4}
\gbullet{(60:1)}
\gbullet{(180:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\begin{tikzpicture}[baseline={(0,0)}]
\covering{0.8}
\draw (330:1) -- (60:1);
\gbullet{(60:1)}
\gbullet{(180:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\begin{tikzpicture}[baseline={(0,0)}]
\covering{0.92}
\draw (60:1) -- (180:1);
\draw (330:1) -- (60:1);
\gbullet{(60:1)}
\gbullet{(180:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\begin{tikzpicture}[baseline={(0,0)}]
\covering{1.1}
\fill[color=red!50,semitransparent] (60:1) -- (180:1) -- (330:1) -- cycle;
\begin{scope}[color=red]
\draw (60:1) -- (180:1);
\draw (180:1) -- (330:1);
\draw (330:1) -- (60:1);
\end{scope}
\gbullet{(60:1)}
\gbullet{(180:1)}
\gbullet{(330:1)}
\end{tikzpicture}
\\[1em]
\end{tabular}
\end{center}
\caption{ Vietoris Rips with Increasing Radii(parameter)}
\label{fig:vietorisripscomplex}
\end{figure}
Although, so far. we have discussed persistence only in relation to
simplicial complexes. the same set of ideas hold while computing persistence
of cell complexes.
\begin{notation}[$H_{+}^{+}(\cdot), H_{-}^{-}(\cdot)$]
If $\alpha$ is a positive simplex in the filtration, we denote it by $H_{+}^{+}(\alpha)$. If $\alpha$ is a negative simplex in the filtration, we denote it by $H_{-}^{-}(\alpha)$.
\end{notation}
We denote the boundary operator of the original simplicial complex
by $\eth$ and the boundary operator of the discrete Morse complex
by $\Delta$. We process simplices one-by-one in the order prescribed
by the persistence filtration. At times we denote the boundary operator
associated to the \textsf{DGVF} $\mathcal{V}_{i}$ by $\Delta_{\mathcal{V}_{i}}$
and the \textsf{DGVF} associated to the boundary operator $\Delta_{i}$
by $\mathcal{\mathcal{V}}_{\Delta_{i}}$.
\autoref{thm:Morsecell} is one of the deep theorems of discrete Morse
theory. It tells us that the chain complex expressed in \autoref{thm:morsehom}
encodes the boundary operator of a \emph{new} cell complex $\mathcal{Q}$
which is homotopy equivalent to the original simplicial complex $\mathcal{K}$.
The formula for the boundary operator is given by \autoref{thm:boundop}.
We exploit this fact to do matrix operations on the boundary matrices
of the new cell complex $\mathcal{Q}$ instead of doing them on the
boundary matrix of the original complex $\mathcal{K}$. Assuming that
the number of critical cells is substantially small when compared
to the size of the complex, this gives us a new family of persistence
homology computation algorithms that operate directly on the cell
complex boundary instead of operating on the the simplicial boundary
matrix. The fact that we use the \textsf{MMUP-APX} algorithm for obtaining
a nearly optimal \textsf{DGVF} makes it merely an instance in a family
of (discrete Morse theory based) persistence homology algorithms that
can directly exploit the cell complex boundary matrix.
Note that Forman gives a criterion for critical pair cancellation
as expressed in the Theorem below.
\begin{theorem}[Cancellation of critical pairs Forman[Fo1998, Fo2002]] \label{thm:cancelgrad} Let $f$ be a discrete Morse function on a simplicial complex $\mathcal{L}$ such that $\sigma^{(p+1)}$ and $\tau^{p}$ are critical. Let there be a unique gradient path from $\partial\sigma$ to $\tau$. Then there is another Morse function $g$ on $\mathcal{L}$ with the same set of critical simplices except $\tau$ and $\sigma$. Also, the gradient vector field associated to $g$ is equal to the gradient vector field associated to $f$except along the unique gradient path from $\partial\sigma$ to $\tau$. \end{theorem}
If a pair of critical cells $\langle\tau,\sigma\rangle$ satisfy critical
pair cancellation criterion , we denote it as $\cancel{\langle\tau,\sigma\rangle}$.
The precise details of critical cell cancellations used in the algorithm
sketched in \autoref{fig:perHom} is described in \cite{RPN14}.
\begin{figure}
\shadowbox{\begin{minipage}[t]{1\columnwidth}%
\begin{tikzpicture}[
grow=down,
level 1/.style={sibling distance=6cm,level distance=3cm},
level 2/.style={sibling distance=6cm, level distance=3cm},
edge from parent/.style={very thick,draw=blue!40!black!60, shorten >=5pt, shorten <=5pt},
edge from parent path={(\tikzparentnode.south) -- (\tikzchildnode.north)}, kant/.style={text width=2cm, text centered, sloped}, every node/.style={text ragged, inner sep=2mm,font=\scriptsize},
punkt/.style={rectangle, rounded corners, shade, top color=white, bottom color=blue!50!black!20, draw=blue!40!black!60, very thick }
]
\node[punkt] [rectangle split, rectangle split, rectangle split parts=3, text ragged]
{
\textbf{A.} Compute $\Delta(\sigma)$ for newly arrived simplex $\sigma^d$ via Dyn.Prog. Complexity: $\tilde{O}(\varLambda_{\mathcal{K}_{i}})$
\nodepart{second}
\textsf{GOAL:} Determine whether $H_{+}^{+}(\sigma)$ or $H_{-}^{-}(\sigma)$.
\nodepart{third}
Inspect $\Delta(\sigma)$ Find $\tau^{d-1}$ s.t. $\cancel{\langle\tau,\sigma\rangle}$. Takes nearly constant time.
}
child {
node[punkt] [rectangle split, rectangle split, rectangle split parts=3, text ragged]
{
\textbf{B.} $\exists\tau$ s.t. $\cancel{\langle\tau,\sigma\rangle}\Rightarrow$ Reverse gradient path.
\nodepart{second}
Recompute boundary operator $\Delta$.
\nodepart{third}
Complexity: $O(\mathcal{N}) + \Delta(\varLambda\mathcal{N})$
}
edge from parent
node[kant, below, pos=.6] {}
}
child {
node [punkt,rectangle split, rectangle split,
rectangle split parts=3]
{
\textbf{C.} $\cancel{\exists}\tau$ s.t. $\cancel{\langle\tau,\sigma\rangle}\Rightarrow$use $\Delta_{\mathcal{K}_{i}+\sigma}$ instead of using
\nodepart{second}
$\eth_{\mathcal{K}_{i}+\sigma}$ to perform elementary column ops.
\nodepart{third}
Complexity is $\tilde{O}(\varLambda^{2})$.
}
child {
node [punkt,rectangle split, rectangle split,
rectangle split parts=3]
{
\textbf{D.} Case: $H_{-}^{-}(\sigma)$: Use MMUP-APX to compute $\mathcal{V}_{i+1}$
\nodepart{second}
Compute $\Delta_{\mathcal{V}_{i+1}}$ from $\mathcal{V}_{i+1}$ using Dyn. Prog.
\nodepart{third}
Complexity: $\tilde{O}(\varLambda\mathcal{N}) + \tilde{O}(\mathcal{N})$.
}
edge from parent node[below, kant, pos=.6] {} }
child {
node [punkt, rectangle split, rectangle split parts=3]{
\textbf{E.} Case $H_{+}^{+}(\sigma)$: $\Delta_{i+1} \leftarrow \Delta_{\mathcal{K}_{i}+\sigma}$
\nodepart{second}
$\mathcal{V}_{\Delta_{i+1}}$ stays $\tilde{O}(\log^2(N))$-optimal
\nodepart{third}
No computation required.
}
edge from parent node[kant, above] {}}
edge from parent{ node[kant, above] {}}
};
\end{tikzpicture}%
\end{minipage}}
\protect\caption{Bird eye view: MMUP-APX based persistence homology algorithm}
\label{fig:perHom}
\end{figure}
\begin{theorem} The algorithm sketched in \autoref{fig:perHom} correctly computes persistent homology
\end{theorem}
\begin{proof}[sketch]
The correctness of the algorithm follows from \autoref{thm:MorsecellNone}, \autoref{thm:Morsecell}, \autoref{thm:morsehom}, \autoref{thm:boundop} and \autoref{thm:cancelgrad}. We sketch the proof below.
\begin{description}
\item{Step A}: To begin with, we justify the use of the Morse boundary operator over the use of the simplicial boundary operator owing to \autoref{thm:morsehom} and \autoref{thm:boundop}. The procedure for boundary operator construction is described in \autoref{sub:Boundary-Operator-Computation}. This justifies Step A.
\item{Step B}: Since the cancellation of critical cells pairs up an existing unpaired criticality ($d-1$ dimensional) with a new one ($d$ dimensional), clearly there is a reduction in $\beta_{d-1}$. Hence, in step B, the advent of $\sigma^{d}$ s.t. there exists an existing criticality $\tau$ that can be paired with it, establishes $\sigma$ as a negative simplex. The recomputation of boundary operator upon cancellation ensures that it is up to date for next iteration. The optimality is preserved because OPT(i+1) = OPT(i) - 1. In this step, we observe maximum conceptual similarity and overlap between discrete Morse theory and persistent homology.
\item{Step C}: Clearly we may perform column operations on this operator instead of the simplicial boundary operator because both should give the same Betti numbers.
\item{Step D}: The boundary operator is recomputed upon reapplication of MMUP. So it is up to date for the next iteration. The reapplication of MMUP ensures the $\log^{2}(\mathcal{N})$ factor optimality of MMUP.
\item{Step E}: Since none of the prior simplices is destroyed, the boundary operator stays up to date with a single computation. The boundary operator correctness and the $\log^{2}(\mathcal{N})$ factor optimality arguments follow from induction. Also, if vector field was $\log^{2}(\mathcal{N})$ factor optimal for $i$, then obviously it will be $\log^{2}(\mathcal{N})$-factor optimal for $i+1$ because in this particular case, OPT(i+1)=OPT(i)+1.
\end{description}
\end{proof}
\begin{theorem}The complexity of the MMUP algorithm is $\tilde{O}(\mathcal{N}^{2}) assuming the WMOC condition$
\end{theorem}
\begin{proof}The MMUP-APX algorithm which runs in time complexity $\tilde{O}(\mathcal{N})$ and the dynamic programming algorithm with pseudolinear complexity $\tilde{O}(\varLambda\times\mathcal{N})$ to compute the boundary operator described in \autoref{thm:boundop} is one of the the key computational components that drive down the time complexity.
\begin{enumerate}
\item The first step is to determine the existence of $\cancel{\langle\tau,\sigma\rangle}$ by inspecting $\Delta(\sigma)$. Since the no. of critical cells of $\mathcal{K}_{i}$ in vector field $\mathcal{V}_{i}$ are within polylogarithmic bound of the optimal, the number of non-zero elements in $\Delta(\sigma)$ is $\mathcal{\tilde{O}}(\varLambda)$. The inspection step involves going through this list to find if the coefficients of one of the entries is either $+1$ or $-1$ which takes $\mathcal{\tilde{O}}(\varLambda)$. time.
\item Depending on existence of $\cancel{\langle\tau,\sigma\rangle}$, we have two cases:
\begin{enumerate}
\item When a critical pair (in the sense of discrete Morse theory) is found, the gradient is reversed and boundary operator $\Delta$ is recomputed using a dynamic update data structure. The time complexity of this operation is $\tilde{O}(\varLambda\mathcal{N})$.
\item Elementary column operations performed on $\Delta$ which has complexity $\tilde{O}(\varLambda^{2})$. \end{enumerate}
\item Note that every time we encounter a negative simplex, we employ the na\"{i}ve\xspace approach of recomputing $\mathcal{V}_{i+1}$ from scratch using the MMUP-APX algorithm. Consequently we also need to compute the boundary operator of the entire subcomplex. The time complexity of MMUP-APX is $\tilde{O}(\mathcal{N})$. The time complexity of boundary operator computation is $\tilde{O}(\varLambda\mathcal{N})$. Finally, under the WMOC hypothesis, $\tilde{O}(\varLambda\mathcal{N})=\tilde{O}(\mathcal{N})$ when MMUP-APX is employed given the polylogarithmic approximation ratio in bounding the number of criticalities.
\end{enumerate}
\begin{description}
\item[Worst Case Complexity] The total complexity is therefore $\tilde{O}(\varLambda\mathcal{N}+\varLambda^{2})$. In pathological cases, when the WMOC is not satisfied, we have $\tilde{O}(\varLambda)\thickapprox\tilde{O}(\mathcal{N})$ , the complexity cost we incur per step is no better than the worst case which is known to be $O(\mathcal{N}^{2})$ which makes the overall worst case complexity $O(\mathcal{N}^{3})$.
\item[Complexity under the WMOC condition] But, under the reasonable assumption of WMOC, i.e. when $\varLambda\lll\mathcal{N}$, the complexity of the processing time of one simplex in the filtration sequence can be brought down to $\tilde{O}(\mathcal{N})$ thereby bringing down the complexity of persistence homology to $\tilde{O}(\mathcal{N}^{2})$ (a sizeable gain over $O(\mathcal{N}^{3})$ worst case methods).
\end{description}
\end{proof}
The MMUP-APX algorithm which runs in time complexity $\tilde{O}(\mathcal{N})$
and the dynamic programming algorithm with pseudolinear complexity
$\tilde{O}(\varLambda\times\mathcal{N})$ to compute the boundary
operator described in \autoref{thm:boundop} are two of the the key
computational components that drive down the time complexity of persistent
homology to $\mbox{\ensuremath{\widetilde{O}(\mathcal{N}^{2})}}$.
\subsection{Witten Morse functions \& Scalar Field Design: Rigor, Quality and
Efficiency.}
We develop the idea of compatibility between input scalar field and
the set of all Morse-Witten gradient vector fields. MMFEP provides
a powerful and sufficiently general modeling tool for obtaining approximation
guarantees for closely related problems in discrete Morse theory.
To begin with, we rigorously define what we mean by ‘compatibility’
between an input scalar field and the corresponding discrete Morse
function. To the best of our knowledge there hasn't been much of an
attempt in prior research literature to rigorously formulate the notion
of \emph{compatibility}. We then optimize over the set of all possible
compatible fields so as to minimize the number of ‘spurious’ critical
cells. Using the approximation algorithm for min-POP as developed
in this paper, we obtain a nearly optimal (i.e. logarithmic multiple
of the optimum) discrete Witten Morse function given an input scalar
field. Moreover, the runtime of the algorithm is nearly linear. In
summary, our work hopes to lend rigor and computational speed to the
task of finding a nearly optimal discrete flat Witten Morse function
compatible with an input scalar field.
\begin{figure}
\centering
\begin{scriptsize}
\begin{tikzpicture}
\fill [color=red!70] (-1.7,0) circle (1.5pt);
\fill [color=red!70] (1,-0.5) circle (1.5pt);
\fill [color=red!70] (0,1.7) circle (1.5pt);
\fill [color=red!70] (2.5,0.4) circle (1.5pt);
\fill [color=blue!50,semitransparent,] (-1.7,0)--(1,-0.5)--(0,1.7);
\fill [color=blue!60,semitransparent,] (-1.7,0)--(1,-0.5)--(2.5,0.4);
\fill [color=blue!70,semitransparent,] (0,1.7)--(1,-0.5)--(2.5,0.4);
\draw [color=red!50] (-1.7,0)--(1,-0.5)--(0,1.7)--(-1.7,0);
\draw [color=red!50] (1,-0.5)--(2.5,0.4)--(0,1.7);
\draw [color=red!50,dotted] (2.5,0.4)--(-1.7,0);
\node [below] at (-1.7,0) {$\mathcal{F}(v_0)=5$};
\node [below] at (1,-0.5) {$\mathcal{F}(v_1)=8$};
\node [above] at (0,1.7) {$\mathcal{F}(v_2)=12$};
\node [above] at (2.5,0.4) {$\mathcal{F}(v_3)=3$};
\end{tikzpicture}
\end{scriptsize}
\protect\caption{Simplicial Complex $\mathcal{K}$ }
\label{fig:complex_x2}
\end{figure}
We now define the set of all compatible $\varepsilon$-discrete Witten
Morse functions (DWMF) $\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$, given an input scalar
field $\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$. The definition below isn't constructive
because for all cells of dimension $k>0$, it allows a certain range
of values for the DWMF given a scalar field rather than fixing them
up.
\begin{definition}[Compatible $\epsilon$-DWMFs $\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ for scalar field $\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$] \label{def:compfield} We say that a a discrete Witten Morse function $\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ is compatible with a given scalar field $\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ if:
\begin{enumerate}
\item For every dimension $k>1$, let $\varepsilon^{k}\in\mathbb{R}$ be the $k$-dim. perturbation error s.t. $\varepsilon^{k}>\varepsilon^{j}$ for every $k>j$. Also, $\forall i,j,k$, we must have $\varepsilon^{k}<|\mathcal{F}(\sigma_{i}^{0})-\mathcal{F}(\sigma_{j}^{0})|$ where $i,j$ range across all 0-dimensional simplices and $k$ ranges across all dimensions.
\item To begin with, we equate the DMF $\mathpzc{f}(\sigma^{0})$ for every 0-dimensional simplex (i.e. a vertex) with the corresponding input scalar function $\mathcal{F}(\sigma^{0})$
\item Now, inductively we let $\mathpzc{f}(\tau_{j}^{k})=\left\{\max\limits_{\sigma_{i}^{0}\prec\!\prec\tau^{k}}\mathcal{F}(\sigma_{i}^{0})\right\}\pm \varepsilon^{k}$ where the symbol $\prec\!\prec$ indicates incidence of 0-dimensional cell $\sigma_{i}^{0}$ on k-dimensional cell $\tau_{j}^{k}$.
\end{enumerate}
\end{definition}
Let $\sigma_{Y}^{0}$ be the 0-dim. simplex of $\tau_{j}^{k}$ for
which the maximum $\max\limits _{\sigma_{i}^{0}\prec\!\prec\tau^{k}}\mathcal{F}(\sigma_{i}^{0})$
is attained. In this case we say that $\tau_{j}^{k}$ \emph{inherits}
its Morse function value from $\sigma_{Y}^{0}$ .
We denote compatibility by symbol $\circeq$. i.e. if a scalar function
$\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ and a discrete Morse function $\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$
are compatible then we write it as $\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})\circeq\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$.
\begin{example} \label{example:scalar} Consider the simplex with a scalar field $\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ as shown in \autoref{fig:complex_x2}. The scalar field is given as: $\mathcal{F}(v_0)=5$, $\mathcal{F}(v_1)=8$, $\mathcal{F}(v_2)=12$, $\mathcal{F}(v_3)=3$. Let $\varepsilon^{1},\varepsilon^{2}, \varepsilon^{3} $ be arbitrary real numbers satisfying conditions specified in the definition above. In this case we will have $\mathpzc{f}(v_{0} v_{1})=8\pm\varepsilon^{1}$, $\mathpzc{f}(v_{0} v_{2})=12\pm\varepsilon^{1}$, $\mathpzc{f}(v_{0} v_{3})=5\pm\varepsilon^{1}$, $\mathpzc{f}(v_{1} v_{2})=12\pm\varepsilon^{1}$, $\mathpzc{f}(v_{1} v_{3})=8\pm\varepsilon^{1}$, $\mathpzc{f}(v_{2} v_{3})=12\pm\varepsilon^{1}$. Similarly, we will have $\mathpzc{f}(v_{0} v_{1} v_{2})=12\pm\varepsilon^{2}$, $\mathpzc{f}(v_{0} v_{1} v_{3})=8\pm\varepsilon^{2}$, $\mathpzc{f}(v_{1} v_{2} v_{3})=12\pm\varepsilon^{2}$ and $\mathpzc{f}(v_{0} v_{2} v_{3})=12\pm\varepsilon^{2}$. Finally $\mathpzc{f}(v_{0} v_{1} v_{2} v_{3})=12\pm\varepsilon^{3}$
\end{example}
Possibly the most important issue in designing a discrete Morse function
that corresponds to an input scalar field is minimization of \emph{spurious}
critical cells. This is where MMUP approximation algorithm comes in
picture. So, now our objective is to find a $\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ s.t.
$\min\limits _{\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})\circeq\mathcal{F}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})}\sum\limits _{0\leq i\leq D}\mathpzc{m}_{i}$
where $m_{i}$ are the Morse numbers and $D$ is the dimension of
the complex.
It is easy to see that Definition~\ref{def:compfield} is equivalent
to saying that each simplex $\gamma^{k}$ will inherit a function
value $\mathpzc{f}(\raisebox{-0.25ex}{\scalebox{1.2}{$\cdot$}})$ where $\mathpzc{f}(\varpi^{k})=\max\limits _{\vartheta_{j}^{k-1}\prec\sigma^{k}}\mathpzc{f}(\nu_{j}^{k-1})\pm|\varepsilon^{k}-\varepsilon^{k-1}|$
.
Consider $\varpi^{k}$ s.t. $\ensuremath{\mathpzc{f}(\varpi{}^{k})=\left\{ \max\limits _{\sigma_{i}^{0}\prec\!\prec\tau^{k}}\mathcal{F}(\sigma_{i}^{0})\right\} \pm\varepsilon^{k}}$.
Then let $\varkappa^{0}$ be the vertex incident on $\varpi^{k}$
for which the maximum is attained. Now, given the structure of a simplicial
complex, it is easy to see that $\varkappa^{0}$ is a vertex that
is incident on all the $(k-1)$-dimensional faces $\nu_{j}^{k-1}$
of $\varpi$ with exactly one exception (say $\vartheta_{X}^{k-1}$)
each time (irrespective of dimension $k$). For instance, in Example~\ref{example:scalar},
if we look at the function value of faces of $v_{1}v_{2}v_{3}$, $v_{1}v_{3}$
is the exception. In case of $v_{0}v_{1}v_{3}$, $v_{0}v_{3}$ is
the exceptional face and so on. This means that except for $\vartheta_{X}^{k-1}$,
all other faces of $\varpi$ inherit their discrete Morse function
value from $\varkappa^{0}$. Consequently, we may \emph{potentially}
form a gradient pair $\langle\vartheta_{j}^{k-1},\varpi\rangle$ for
any of the faces $\vartheta_{j}^{k-1}$ of $\varpi^{k}$ (except $\vartheta_{X}^{k-1}$)
merely by perturbation of at the most the value $\varepsilon^{k}$
for $\varpi^{k}$ and the value $\varepsilon^{k-1}$ for $\vartheta_{j}^{k-1}$
thereby ensuring that the gradient pair is compatible with Morse function
value. This allows us to use the following approach.
\begin{enumerate}
\item For dimension $d=1$: We start with one dimensional simplices (say
$\varpi_{i}^{1}$). Of the two incident 0-dimensional simplices, we
allow matchings will one of the 0-dimensional cells while \emph{prohibiting}
matching with the other 0-dimensional simplex. We also record the
0-dim. simplex from which it inherits its function value.
\item For dimensions $d>1$: For every simplex $\varpi_{i}^{k}$ we allow
matchings with all but one $(k-1)$-dimensional cell (say $\vartheta_{X}^{k-1}$)
s.t. $\varpi_{i}^{k}$ and $\vartheta_{X}^{k-1}$ inherit their function
values from a 0-dimensional simplex other than $\varkappa^{0}$. We
therefore prohibit the matching $\langle\vartheta_{X}^{k-1},\varpi_{i}^{k}\rangle$
while allowing matchings of $\varpi_{i}^{k}$ with all of its other
$(k-1)$ dimensional faces.
\end{enumerate}
For every dimension $k>1$, the prohibition of specific matchings
for each simplex $\varpi_{i}^{k}$ where $i$ ranges across all simplices
of dim. $k$, we use the notion of \emph{forbidden edges} (in the
corresponding oriented Hasse graph). The complement of these edges
will be \emph{rigid edges}. Note that, these rigid/forbidden edges
are \emph{additional} to those that are introduced in the MMUP approximation
algorithm design. In other words, these rigid edges are the additional
rigid edges introduced (apart from MR and CR rigid edges of the MMUP-APX
problem). This makes the scalar field topology problem an instance
of the MMFEP problem described above. This problem can (in a manner
analogous to MMUP) be reduced to min-POP. This reduction followed
by application of divide-\&-conquer SDP-based approximation algorithm
gives us an $\tilde{O}(N)$ complexity algorithm. The solution will
minimize the number of spurious critical cells (up to polylogarithmic
factor $\tilde{O}(\log^{2}(N))$).
\begin{shnote} This formulation is equivalent to saying that we optimize over $\epsilon$-perturbations of all compatible discrete Witten Morse functions to obtain a nearly optimal discrete Witten Morse function in nearly linear time. We may say that the output is not a perturbation of but a precise a discrete Witten Morse function because the output of MMUP-APX is actually a gradient vector field and it is always possible to choose values of a discrete Morse function so as to reflect the additional criteria satisfied by a discrete Witten Morse function.
\end{shnote}
An indepth treatment that includes \textbf{stability }and\textbf{
robustness }properties of $\varepsilon$-discrete Morse functions
will be treated in a separate paper.
The use of MMUP in context of simplicial maps and Conley index computations
is deferred.
\section{Improvements \& further directions.}
\begin{table}
\begin{center}
\begin{tabular}{|>{\centering}m{2.2cm}||c||>{\centering}m{3cm}||>{\centering}m{3cm}||>{\centering}m{1cm}|}
\hline
How & When & \multicolumn{1}{c||}{Why} & What & Ref.\tabularnewline
\hline
\hline
Expansions, Thickening, Deformations. & Pre/Post & Zeeman, Andrew-Curtis, Poincare Conjectures & ZC: $\mathcal{K}^{2}$ Contractible $\implies$$\mathcal{K}^{2}\times I$
$\ssearrow\raisebox{-0.25ex}{\scalebox{1.2}{$\ldotp$}}$ & \cite{HMS93,Mat07}\tabularnewline
\hline
\hline
Subdivisions & Pre/Post & 3-dim. Topology & Results by Chillingworth & \cite{Ch67,Ch80}\tabularnewline
\hline
\hline
Iterations. & Post & CHomP-RedHom & More gains in fewer iterations. & \cite{HMMNWJD10,MN13}\tabularnewline
\hline
\hline
Cancel $\cancel{\langle\tau,\sigma\rangle}$ & Post & $\widetilde{O}(n)$ assuming WMOC & $\widetilde{O}(n)$-post MMUP-APX.
Else $\Omega(n^{2})$, & \cite{Fo98a,RPN14}\tabularnewline
\hline
\hline
Bdry. Pruning & Pre & Full removal. $O(n)$-time. & $\triangleq$ Strong Homotopy/LC reductions. & \cite{Ba08,RPN14} \tabularnewline
\hline
\end{tabular}
\end{center}
\protect\caption{MMUP: Improvements, Extensions, Further Directions}
\end{table}
\subsection{From Morse to Poincare: Zeeman's conjecture and Forman's Theory}
MMUP-APX algorithm can be used to obtain computational certificates
in favor or Zeeman's conjecture, which is an important open problem
in pure mathematics. In fact, the Poincare conjecture is implied by
Zeeman's conjecture. A more careful analysis of implications of discrete
Morse theory, especially the use of MMUP-APX based algorithms in relation
to the Poincar\'{e}\xspace conjecture needs to be done.
\subsection{Subdivisions and 3-dimensional Topology}
For a contractible yet non-collapsible complex an \emph{increase}
in collapsibility can potentially be observed under simplicial subdivisions.
While heuristic algorithms are not in a position to harness this increase
in collapsibility, our algorithm is well-placed to exploit it. An
indepth description of this application is deferred.
\section{Concluding Remarks }
As we move into the world of modern massive datasets, we are faced
with a familiar, and yet in context of growing sizes, a far more contingent
challenge in context of computational topology namely to minimize
the time to compute topological features. One of the most promising
approaches is to `efficiently' reduce the problem instance to another
problem instance of a significantly smaller size (and yet with the
same topology)\cite{HMMNWJD10,MM95,MN13}. Incidentally, discrete
Morse theory happens to be the best reduction method in the literature.
Now since the problem of computing the optimal Morse function is NP-hard,
approximation algorithms is our best bet for designing provably efficient
algorithms. In this work, we have harnessed tools from modern algorithmics
to design a nearly optimal Morse function. More significantly, we
do this in nearly linear time. In context of large datasets, a fast
reduction algorithm can be the very linchpin in driving down the complexity
of a wide range of problems in computational topology. In fact, we
believe that an efficient implementation of the MMUP-APX algorithm
(MWUM-based) can possibly constitute the core understructure of topological
libraries.
\section*{Acknowledgments }
The debt the author owes Prof. Natarajan knows no depth. His inputs,
at each stage, have been invaluable in shaping this work. The author
is deeply grateful to Michael Lesnick for several helpful discussions.
He would like to thank Prof. Carlsson, Prof. Edelsbrunner, Prof. Mischaikow,
and Prof. Mrozek for providing encouragement (during early stages).
|
\section{Introduction}
\subsection*{Next-Generation Sequencing Analysis}
The rapid increase in the utilisation of Next Generation Sequencing (NGS) technologies amongst disparate fields of research has resulted in a new set of challenges for scientific researchers. Whilst technology selection, sequencing costs and efficacy of sample preparation are still considerable concerns, pragmatic approaches in decision-making and implementation will result in satisfactory sequence generation. However, potentially vast quantity of data generated by sequencing efforts suggests that the true bottleneck is the computational analysis of the sequence data\cite{scholz12, lin07}. Indeed, without considered planning of the bioinformatics analysis, data processing time and costs can far exceed those of the actual sequencing itself \cite{angiuoli11}, diminishing the benefits attributed to high-throughput technologies.\par
Whilst the utilisation of computing resources and data analysis tools is generally accessible to the scientific and research communities, the ability to harness their full potential is often limited to specialists. Faced with the increasing ubiquity of bioinformatics tools for post-sequencing analysis and the realisation that decentralised research is becoming more prevalent, bioinformaticians are tasked with creating applications that are reliable, scalable, and user-friendly.\par
The prevalence of bioinformatics workflows and pipelines such as Galaxy\cite{goecks10}, QIIME core analysis\cite{caporaso10} and Taverna\cite{hull06}, and the increasing collaborative efforts through shared infrastructures\cite{ragan-kelley13} suggests greater uptake by non-specialists will be forthcoming. As NGS technology develops, the challenges faced by both bioinformaticians and users relate specifically to the competency of the software tools and the performance of the hardware to handle increasingly larger and more complex datasets. \par
The lack of appropriate hardware infrastructure is the greatest contributing factor to the bioinformatics bottleneck and the rise in virtual environments, parallelised code and super-computing facilities is testament to an understanding of the need for continual development and innovation in NGS data handling and management\cite{schatz10}. However, these structural and programmatic facilitators are not without their drawbacks. For example, cloud computing facilities such as Amazon Web Services' Elastic Compute Cloud \cite{aws13} offer flexible and scalable environments for performing a wide range of bioinformatics, but issues around data security and file transfer rates coupled with per hour usage costs make resource planning an integral requirement for any data analysis project.
\subsection{Understanding capacity and performance}
The capacity for processing and analysing NGS data accurately is dependent on identifying the most suitable software and hardware for the task. Incorrect selection or mismatch between data requirements and architecture will inevitably lead to suboptimal performance and, potentially, poor or erroneous results. With a wide range of bioinformatics tools being utilised by both specialists and non-specialists, there is a need for greater transparency in their deployment to facilitate effective and efficient analysis. As time and cost are often constraining factors in research and corporate environments, the ability to assign resources with \textit{a priori} knowledge of the performance and run time is of great benefit. For example, \textit{Cunningham} was developed to provide accurate runtime estimates for BLAST analysis of large shotgun sequence datasets \cite{white11} .\par
A recent study using the 16S gene for estimating bacterial diversity has shown the quantity and size of sequence clusters affects accuracy in non-parametric diversity calculations, whilst also determining which methods to employ\cite{barriuso11}.\par
Parallelisation of bioinformatics algorithms aimed at dramatically decreasing their processing time by exploiting multiple core processors\cite{galvez10} or GPU capabilities\cite{liu12} has alleviated some of the analysis bottleneck. Code optimisation can also make significant performance gains in highly parallel applications\cite{tan06}, but often requires expertise in coding that is not always accessible or practical for the end-user.\par
Parallel speedup and efficiency are key performance metrics that can be used to assess the most effective use of multiple CPU cores or nodes in a Cluster, Grid or Cloud environment. Whilst it may be intuitive that splitting large computational jobs amongst a greater number of processors should lead to increasing reductions in processing time, the presence of code that must be run serially (\textit{ser}) in most algorithms means that only a fraction of the work benefits from parallel (\textit{par}) speedup. Amdahl's Law\cite{amdahl67} describes the speedup of a process across multiple cores ($P$) given an amount of work ($N$) as:
\begin{equation} \label{eq:speed}
S(N,P) = \frac{t(N, P=1)}{t(N,P)}
\end{equation}\par
In ideal parallel processes, speedup is therefore equal to $1/P$, but with a fraction of code being serial, this equation becomes:
\begin{equation} \label{eq:speed_ser}
S(N,P) = \frac{t(N)_{ser} + t(N)_{par}}{t(N)_{ser} + \frac{t(N)_{par}}{P}}
\end{equation}\par
Knowing the speedup of parallelisation, then the efficiency may also be calculated simply by:
\begin{equation} \label{eq:eff}
E = \frac{S(N,P)}{P}
\end{equation}\par
For algorithms that are parallelisable, it is useful to perform these calculations to get an understanding of the scalability of the processes on any given architecture. This will help users to more appropriately assign resources and avoid problems with latency or parallel overheads.
\subsection{Computational transparency for targeted analysis}
Recent advances in sequencing technology have brought about unprecedented resolution in identification and classification of bacterial species. Sequencing of the highly conserved 16S rRNA gene is popular as it allows for comparative studies of microbial communities, their diversity and structure. The ubiquity of this approach in amplicon-based metagenomics coupled with the dwindling cost of high-throughput sequencing has put emphasis on development of tools and hardware infrastructure that can handle increasingly data rich sequence analysis.\par
A discretised pipeline was developed to model the relationship between sequence data size and complexity, and computational resource. An overview of the pipeline components is shown in Fig.~\ref{fig:PP} (Appendix~\ref{suppmat}), comprising several typical processing and analysis protocols available in the QIIME software. The performance was measured using the time taken to complete each process step in real and CPU metrics. The clock or \textit{real} time is necessary to determine actual resource cost, but can be skewed on systems where other extraneous processes are running, adding load to the shared resource. CPU time, characterised as the sum of \textit{user} and \textit{system} time, is reflective of the actual work done by the process being monitored.
\section{Methods}
\subsection{System architecture}
The system used for evaluation was a 64-bit 2 x 6-Core (Intel Xeon 2.66 GHz CPU) Apple MacPro with 32GB RAM running OS X 10.8.2.\par
For the performance testing 16 logical cores (8 physical cores with hyper-threading) were used for the parallelisable components of the pipeline (e.g. Denoising) and single CPU otherwise. All analysis results are specific to this architecture and configuration. All cited instances of QIIME relate to its OSX compilation, MacQIIME Version 1.6.0.
\subsection{Training and validation datasets}
\subsubsection{Training data}
Three datasets containing 16S microbial rRNA gene fragments were used to develop the performance models. The first training dataset (\textit{MFC}) was taken from an acetate fed microbial fuel cell reactor inoculated with arctic soil sourced from Arctic soil (Ny-\r{A}lesund, Spitsbergen, Svalbard), and operated at 26.5\textdegree C \cite{heidrich12}. The second dataset was generated from a sample taken from a small eutrophic lake in the English Lake District (\textit{Priest Pot}) in 2008 \cite{quince09}. The third dataset (\textit{Arctic}) were sequences derived from DNA extracted from Arctic mineral soil samples collected from the Svalbard region [Unpublished].\par
The Priest Pot sequences were generated using standard 454 GS-FLX chemistry and targeted the V5 hyper-variable region, whilst the MFC and Arctic samples were sequenced with the more recent GS-FLX Titanium chemistry, which gives longer read lengths and targeted the V4-V5 regions, as at the time of sequencing (2011), they provided the highest classification accuracy with lowest amplification bias.\par
In the case of the MFC and Arctic data, samples were originally pooled using barcodes to provide a multiplexed dataset. However, only one sample was selected and processed through the pipeline to avoid the effects of redundancy when processing multiple samples as a single batch.
\subsubsection{Validation data}
The first validation dataset (\textit{Mix}) were sequences taken from a laboratory scale batch reactor sample, which had been used to study anaerobic digestion of domestic wastewater at 15\textdegree C. The sample was sequenced in January 2013 using the same method employed for the {\textit{Arctic}} data. The second validation dataset (\textit{Artificial}) consisted of an artificial community from 90 clones that was pyrosequenced over the V5 region of the 16S rRNA gene with a 454 GS-FLX sequencer\cite{quince09}.
Table~\ref{tab:datasets} summarises each of the datasets used in this analysis and includes the number of reads, average read length, number of putative OTUs and $\alpha$ diversity (equitability) of the raw, unfiltered sequences. Because of inherent sequencing errors, the OTU and equitability values are likely to be over- and underestimated, respectively. As can be seen, the samples sequenced with the older 454 GS-FLX chemistry are much shorter than those sequenced with the Titanium chemistry. After trimming to remove primer and barcode, the mean read lengths are approximately 200 bp for GS-FLX and 400 bp for Titanium.
\noindent
\begin{center}
\begin{table*}[ht]
\caption{Test and Validation Datasets: Number of Sequences (Seqs); Average Read Length (bp); Number of OTUs at 97\% Cutoff (OTU); Equitability Estimate \cite{krebs89} of Total Samples ($\alpha$)}
\label{tab:datasets}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
Source & Dataset & Seqs & bp & OTU & $\alpha$ \\ \hline\hline
MFC & Test & 72003 & 411 $\pm$ 45 & 828 & 0.481\\ \hline
Priest Pot & Test & 28361 & 244 $\pm$ 52 & 1146 & 0.613\\ \hline
Arctic & Test & 21576 & 426 $\pm$ 53 & 2267 & 0.807\\ \hline
Mixed sediment & Validation & 19718 & 423 $\pm$ 49 & 2390 & 0.801\\ \hline
Artificial & Validation & 46341 & 260 $\pm$ 38 & 177 & 0.536\\ \hline
\end{tabular}
\end{table*}
\end{center}
\subsection{Analysis steps}
An initial subsampling of the test datasets was performed using the QIIME \textit{subsample\_fasta.py} script to randomly split the raw Fasta file into subsamples of 5\% to 95\% fractions at intervals of 5\%. Six repeats were generated at each interval for the GS-FLX sequenced Artificial dataset and two repeats for the Titanium sequenced Arctic dataset (due to the greatly increased computational time required to denoise these sequences). Thus, for the denoising steps, 76 samples were used for training and 133 samples for validation. \par
A simple bash script was written to simplify and automate the performance testing by looping through the subsamples, processing according to each step in the performance pipeline (Fig.~\ref{fig:PP}) and passing the Real and CPU time for execution to a separate output file.\par
The subsampled Fasta files are pre-processed using the QIIME \textit{split\_libraries.py} script, which applies some basic quality filtering to the sequences for read length trimming, ambiguous base checking and primer and barcode removal.\par
The subsamples are then ready for denoising to correct for errors generated in the PCR and sequencing steps. QIIME Denoiser \cite{reeder10} is a heuristic algorithm that uses a greedy alignment scheme before clustering flowgrams in descending order of abundance. Erroneous reads are filtered from the cluster to produce the final denoised sequences. Denoiser can use pre-filtered Fasta files and matches the IDs of the remaining reads with those present in the text translation of the raw SFF file, to avoid denoising of poor quality sequences. Chimera checking is an optional but often important step that is performed independently from denoising using the ChimeraSlayer\cite{haas11} tool via QIIME's \textit{identify\_chimeric\_seqs.py} script. \par
For denoising using AmpliconNoise, the Standard Flowgram File (SFF) associated with the data was split into 19 subsamples with sizes corresponding to the set used for QIIME denoising using SFF Workbench\cite{heracle12} via a Wine\cite{wine13} translation of the Windows API. The individual SFF files were then converted to text translations using the QIIME \textit{process\_sff.py} script. There is no need for demultiplexing of the data as only a single sample is used in the pipeline test. The AmpliconNoise software\cite{quince09} uses Bayesian theory to generate an approximate likelihood from empirical error distribution data to infer true read identity given sequencing error (PyroNoise) and PCR error (SeqNoise). An additional chimera checking step using Perseus\cite{quince11} is performed after error removal. The software was run using the QIIME wrapper script \textit{ampliconnoise.py} rather than via the stand-alone package to maintain a consolidated workflow.\par
The denoised reads from QIIME denoiser were used for downstream analysis using the following steps:
\begin{itemize}
\item \textbf{De Novo OTU picking:} Clustering of sequences with 97\% similarity threshold using the \textit{uclust} method\cite{edgar10}, before picking representative Operational Taxonomic Units (OTUs) from each cluster by the sequence first assigned to a cluster.
\item \textbf{Assign Taxonomy:} Assignment of taxonomic identities to the OTUs using the curated Greengenes 16S rRNA gene database\cite{mcdonald12} with a na\"{i}ve Bayesian classifier, \textit{RDP}\cite{wang07}.
\item \textbf{Alignment:} Alignment of the sequences is necessary for comparative analysis, such as $\beta$ diversity, in which phylogenetic distances are used to understand differences in community composition from distinct samples. Alignment was performed using the \textit{PyNAST} method and pairwise clustering with \textit{uclust}\cite{edgar10} against a Greengenes database template. Gap only columns and highly variable regions within the alignment files are removed using a filtering step that makes use of Greengenes compatible Lanemask file for excluding these positions. Phylogenic relatedness of organisms within a sample may also be of interest and the creation of a phylogenetic tree using the \textit{FastTree 2.1.3}\cite{price10} method is performed after alignment.
\item \textbf{Diversity analysis:} Diversity metrics are key outputs from the QIIME pipeline and are used to gain an quantitative understanding of the distribution and relatedness of organisms within a sample ($\alpha$ diversity) or between different samples ($\beta$ diversity). $\alpha$ diversity uses the abundance data stored in the OTU table generated from the OTU picking and taxonomy assignment steps to calculate a range of metrics provided by the user, such as Chao1, Shannon and Phylogenetic Distance. Rarefaction plots are generated for each metric based on random subsampling (using a pseudo random number generator) of the OTU table between a given range of sequences per sample and at a given step size. $\beta$ diversity uses both the information stored in the OTU table and a phylogenetic tree, if the phylogenetic metrics are calculated using Unifrac. In this analysis, both quantitative (weighted) and qualitative (unweighted) Unifrac metrics\cite{lozupone05} were calculated and Principal Coordinate Analysis plots generated to display the results.
\end{itemize}\par
Details of the parameters used for each analysis step are provided in Table~\ref{tab:pipeparam} (Appendix~\ref{suppmat}).
\subsection{Performance measures}
Both the wall clock and CPU (Usr + Sys) were recorded using the GNU time command, which allows for formatting of the output in Mac OSX, and stored as text files. Whilst the wall clock time indicates the amount of real time between execution and completion of a process, CPU time is more indicative of the computing effort required to run the process. However, when considering running large datasets through a bioinformatics pipeline, time to completion is the measure by which costs can be assessed.\par
File input size was determined based on the process step being analysed as shown in Table~\ref{tab:seqtype} (Appendix~\ref{suppmat}). The number of reads and mean read length per Fasta file were determined using the QIIME \textit{count\_seqs.py} script. Equitability was calculated using the QIIME $\alpha$ diversity metric script on the raw input fasta files supplied to the denoising algorithms. \par
The time and predictor (diversity, number of reads and read length) data was imported into Matlab\cite{matlab13} and a stepwise Multiple Linear Regression (MLR) was applied to fit a model between the explanatory variables and clock time. The MLR model takes the following generalised form:
\begin{equation} \label{eq:mlr}
y_{i} = \beta_{0} + \beta_{1}x_{i,1} + \beta_{2}x_{i,2} + ... + \beta_{p}x_{i,p} + \epsilon_{i}
\end{equation}
\noindent where $y_{i}$ are the predictands, $x_{i,p}$ the predictors, $\beta_{p}$ the regression coefficients, and $\epsilon_{i}$ is the error term. The regression coefficients are the solution to the least squares estimation:
\begin{equation} \label{eq:mlrcoeff}
\beta = [X^{T}X]^{-1}X^{T}y_{i}
\end{equation}
\noindent where $X$ is the matrix of regressor variables. The Matlab function \textit{LinearModel}, from the Statistics Toolbox (v8.2), was used to perform the stepwise regression using the polynomial form described in eq.~\ref{eq:mlr}. The starting model included the intercept, linear terms, interactions and power terms with interactions up to a factor of 4 for each explanatory variable. The algorithm uses forward and backward regression with the Sum of Squared Error (SSE) to add and remove terms from the model based on the p-values of the F-statistic with and without a potential term. The thresholds for adding and removing terms were p $<$ 0.05 and p $>$ 0.1, respectively. \par
A conventional method for assessing goodness-of-fit for linear regression models is to calculate the $R^{2}$ value or coefficient of determination. As the complexity of the polynomial increases (by adding more variables), the $R^{2}$ value will increase, which may result in a skewed confidence in the reliability of the model. The adjusted-$R^{2}$ value is used to address this issue:
\begin{equation} \label{eq:rsq}
R^{2}_{adj} = 1 - \frac{(\frac{\sum{(\hat{y_{i}} - \bar{y})^2}}{\sum{(y_{i} - \bar{y})^2}})(n - 1)}{n - p -1}
\end{equation}
\noindent where $y_{i}$, $\bar{y}$ and $\hat{y_{i}}$ are the observed data, mean of the observed data and modelled predictands, respectively, $n$ is the number of observations and $p$ the number of regressors. \par
Parallel speedup and efficiency were calculated according to Amdahl's Law for AmpliconNoise based denoising of the MFC sample, as this step is observed to be the most computationally intensive part of the pipeline and is compiled to run in parallel. Denoising was performed using 1, 2, 4, 8, 16, 32, 64 and 128 physical cores, with 48GB RAM per node (six cores per node) on the DIAG resource \cite{diag13} for two sub-samples of the total sequences (20000 and 40000 reads).
\section{Results and discussion}
\subsection{Model development}
Performance modelling was initiated on the two test datasets using the standard QIIME pipeline tools for processing and analysis of 16S rRNA sequencing data. The wall clock (Real) and CPU time were recorded for each pipeline step and a model fit between these variables and input read number was made. For the denoising algorithms (QIIME denoiser and AmpliconNoise) a Multiple Linear Regression (MLR) model was deemed necessary as visualisation of the read number versus clock time highlighted that more than one explanatory variable was influencing the denoising time. Community diversity within a sample, or more specifically taxonomic rank-abundance, is known to influence computational effort of denoising \cite{reeder10}. \par
The equitability, or evenness, defines the homogeneity of species within a community, with higher values of the index indicating a highly even or homogeneous distribution of species. This diversity metric was used as a second explanatory variable in the MLR model as a candidate for determining clustering speed during denoising. Despite diversity calculation being a component of the QIIME pipeline, the parameter values used for the model were generated by calculating equitability from the raw input fasta files, prior to quality checking and noise removal. Although the value is not a true measure of diversity due to the presence of errors, it is assumed the error profile across all samples is equivalent and this will not affect the model.
\subsection{Modelling sequence denoising strategies}
\subsubsection{Training}
Denoising of high-throughput sequencing data with the most commonly used algorithms (QIIME denoiser and AmpliconNoise) is clearly the major performance bottleneck in the analysis pipeline, but also one of the most critical in terms of determination of more accurate OTU numbers and subsequent classification.\par
A MLR model was developed with two explanatory variables (number of reads ($\lambda$) and sample equitability ($\alpha$)) as predictors and wall clock time ($y$) as the response variable.\par
By simply observing the relationship between the explanatory and response variables, it is evident that a non-linear implementation of the MLR model is necessary. The introduction of power terms in the model is intended to reflect the curvlinear nature of the underlying dependencies.\par
Stepwise MLR models were developed using the three training datasets for the QIIME denoiser (Eq.~\ref{eq:mlrqiime}) and AmpliconNoise algorithms (Eq.~\ref{eq:mlran}). The models take the form given by equation~\ref{eq:mlr} with non-linear power terms and cross-products between the two predictor variables.
\begin{multline} \label{eq:mlrqiime}
y_{qd} = \beta_{1}\alpha + \beta_{2}\alpha\lambda + \beta_{3}\lambda^{2} + \beta_{4}\alpha^{2}\lambda + \beta_{5}\alpha\lambda^{2} + \beta_{6}\lambda^{3} + \beta_{7}\alpha^{2}\lambda^{2} + \beta_{8}\alpha\lambda^{3}
\end{multline}
The regression coefficients ($\beta$) are shown in Table~\ref{tab:modelcoeff}. The results from the QIIME denoiser model suggest a conformity between the two explanatory variables selected and the resulting predictand. Fig.~\ref{fig:calmodel_q} shows excellent prediction (Adjusted $R^{2} >$ 0.9) for all training data, which is confirmed by performing an ANOVA on the full model (F-statistic = 7.38$\times 10^{3}$, p-value = 2.54$\times 10^{-95}$) indicating that the non-linear model is highly significant. All plots are shown in relation to equitability for ease of visualisation, however, an example 3D plot (embedded in the uppermost plot in Fig.~\ref{fig:calmodel_q}) for the MFC data shows the excellent fit against both explanatory variables.\par
\begin{figure}[t!]
\centering
\includegraphics[trim=0cm 0cm 0cm 0cm,clip=true,width=4.4in]{wade1-equitmodelfit.jpg}
\caption{QIIME denoiser model performance for the three training datasets; fit plotted along the equitability axis. Two parameter fit is shown in the 3D embedded figure}
\label{fig:calmodel_q}
\end{figure}
For AmpliconNoise, an initial two parameter training model produced good fits for two datasets (Arctic and MFC), but could not fit the Priest Pot data. It was surmised that read length could be an additional factor in determining processing time during the sequencing error removal (\textit{Seqnoise}) step, given the importance of sequence size in influencing error rate distribution \cite{gilles11}. Including mean read length per sample ($\rho$) as a third parameter in the model decreased the Root Mean Square Error of Calibration (RMSEC) from 5750 to 129 and the improvement can also be seen in Fig.~\ref{fig:calmodel_a}. Although prediction is not as convincing as with the QIIME denoiser data, the model, shown in equation~\ref{eq:mlran}, is still highly significant (F-statistic = 1.60$\times 10^{3}$, p-value = 2.60$\times 10^{-51}$).
\begin{multline} \label{eq:mlran}
y_{an} = \beta_{1}\rho + \beta_{2}\lambda + \beta_{3}\alpha\lambda + \beta_{4}\lambda^{2} + \beta_{5}\alpha\rho + \beta_{6}\rho\lambda + \beta_{7}\rho^{2}\\ + \beta_{8}\alpha\lambda^{2} + \beta_{9}\lambda^{3} + \beta_{10}\alpha\lambda^{3}
\end{multline}
\begin{figure}[t!]
\centering
\includegraphics[trim=0cm 0cm 0cm 0cm,clip=true,width=4.4in]{wade2-equitmodelfitan.jpg}
\caption{AmpliconNoise model performance for the three training datasets; fit plotted along the equitability axis}
\label{fig:calmodel_a}
\end{figure}
\subsubsection{Validation}
Validation of regression based models is critical to ascertain their ability to be used with independent data. In complex, highly parameterised models, there is a risk that overfitting of the data may occur, in which the model tends to fit to the training data, but lacks the predictive capacity when fitting to validation or real-world datasets.
The models developed for both denoising algorithms were tested with the independent validation datasets to assess their suitability for prediction of processing time. For both denoising algorithms the models fit the Mixed sediment dataset well, with adjusted $R^{2}$ values of 0.97 and 0.72 for QIIME denoiser and AmpliconNoise, respectively. However, the models do not predict the Artificial data (See Fig.~\ref{fig:valmodel_q}), which suggests that there is some underlying property of the artificially generated sequence communities that is not captured during training. The dataset was constructed to represent a community with log-normal distribution analogous to true community distributions found in the environment\cite{quince09}, and it is possible that this artificial construct has presented some feature in the data that has a significant impact on denoising performance. As discussed, the Priest Pot data was acquired from the same sequencing technology and the inclusion of the mean read length in the AmpliconNoise model had significant impact on training performance. \par
\begin{figure}[!t]
\centering
\includegraphics[trim=0cm 0cm 0cm 0cm,clip=true,width=4.4in]{wade3-equitmodelfitvalidation.jpg}
\caption{QIIME denoiser model prediction for the two validation datasets; fit plotted along the equitability axis}
\label{fig:valmodel_q}
\end{figure}
A main effects plot, shown in Fig.~\ref{fig:maineff} (Appendix~\ref{suppmat}), was generated to look at the contributions from each independent model variable on the clock time. For both the two and three variable models, the number of reads was the largest contributing factor to the output variability. It can also be seen that, although mean read length has a small impact on the AmpliconNoise model, justifying its inclusion, it is not the critical factor in explaining the poor performance with the Artificial dataset.
\subsection{General pipeline model}
The pipeline algorithms deployed sequentially without parallelism (i.e. on a single CPU core) generally contribute insignificant burden to the overall processing time compared to the denoising step. However, based on system memory availability and CPU processor speed, scripts related to sequence alignment, OTU picking, taxonomy assignment and diversity calculation, may become cumbersome, especially for large sequencing runs. QIIME includes several parallel commands to handle such conditions, but in this study the test environment and dataset sizes were such that single CPU processing was sufficient. A MLR model was developed using the total wall clock time measured for all analysis steps independent of denoising, as shown in Table~\ref{tab:seqtype}. Although non-continuous, as quality filtering occurs prior to denoising, whereas all other steps are downstream, the intention is to indicate a relationship between processing time and predictor variables that will aid resource allocation. Additionally, it should be noted that the predictor variables will undergo changes during the pipeline as reads are removed, trimmed and truncated, particularly at the upstream end of the pipeline. However, monitoring changes in read number and diversity is impractical, especially with automated pipelines and given the underlying aim of assessing resource requirements \textit{a priori}. It is assumed that the underlying correlation between predictors and response hold true for the agglomerated model as the relative changes in predictor variables are expected to be uniform across all training samples for any given pipeline. \par
Observation of individual pipeline step results indicated generally linear relationships between number of reads and clock time but, as with the denoising algorithms, additional confounding factors appeared to have a role in determining response across datasets. An initial model was developed using number of reads and equitability, which gave satisfactory training results but produced poor fitting for the Artificial validation data (Adjusted $R^{2}$ of 0.333). A three factor model that included mean read length was investigated and greatly improved prediction of the articitial validation dataset (Adjusted $R^{2}$ of 0.704), but with a slight decrease in the model fit for the Mixed dataset (Adjusted $R^{2}$ reduced from 0.914 to 0.841). The final pipeline MLR model (F-statistic = 1.27$\times 10^{3}$ , p-value = 5.43$\times 10^{-53}$) is shown in equation~\ref{eq:mlrpipe} and the regression coefficients in Table~\ref{tab:modelcoeff}, with the fitted training and prediction curves presented in Figs.~\ref{fig:calmodel_p} and~\ref{fig:valmodel_p}, respectively.
\begin{multline} \label{eq:mlrpipe}
y_{pipe} = \beta_{0} + \beta_{1}\lambda + \beta_{2}\alpha + \beta_{3}\rho + \beta_{4}\alpha\lambda + \beta_{5}\lambda^{2} + \beta_{6}\alpha\rho + \beta_{7}\lambda\rho + \beta_{8}\rho^{2}\\ + \beta_{9}\alpha\lambda^{2}
\end{multline}
\begin{figure}[!t]
\centering
\includegraphics[trim=0cm 0cm 0cm 0cm,clip=true,width=4.4in]{wade4-pipemodelfittest.jpg}
\caption{General pipeline model performance for the three training datasets; fit plotted along the read number axis}
\label{fig:calmodel_p}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[trim=0cm 0cm 0cm 0cm,clip=true,width=4.4in]{wade5-pipemodelfitval.jpg}
\caption{General pipeline model prediction for the two validation datasets; fit plotted along the read number axis}
\label{fig:valmodel_p}
\end{figure}
\subsection{Speedup and Efficiency}
Due to the unfeasibly long processing time for AmpliconNoise and, to a lesser extent, QIIME denoiser, there may be a tendency to invest heavily in high-performance computing solutions to dramatically reduce the run time. Whilst there is some guidance on memory requirements for running the algorithms (1GB for FLX sequences with QIIME denoiser \cite{qiime13}, $>$ 8GB for large datasets when running AmpliconNoise \cite{anoise13}), there may be a tendency for employing as large a resource as economically and logistically feasible. Aside from the costs involved, acquisition of hardware for long-term use may result in redundancy unless demand is significant. Deployment on decentralised systems may also result in conflict if non-essential capacity is being utilised for the denoising task. \par
The results of the speedup and efficiency tests performed with AmpliconNoise on the decentralised DIAG \cite{diag13} resource are shown in Fig.~\ref{fig:speedup}. The speedup plot shows that actual performance improvement is far from ideal when utilising more processors, reaching a threshold of approximately 7.5 times speedup with 128 cores, which corresponds to 5\% efficiency shown in the second plot. AmpliconNoise is clearly not a massively parallel algorithm, with many serial components contributing to the dramatic reduction in efficiency with greater parallel resources.\par
Based on the analysis and considering the tradeoff between time to completion and resource utilisation/expenditure, between 8 and 16 (shown as a vertical bar in Fig.~\ref{fig:speedup}) cores appear reasonable for this architecture. Using more than 16 cores does not deliver significant increases in speedup, whilst efficiency drops below 30\%.
\begin{table}[!t]
\caption{Regression Coefficients for the MLR Models of the QIIME Denoiser and AmpliconNoise Algorithms, and General QIIME Pipeline Steps}
\label{tab:modelcoeff}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Regress. coeff. & QIIME & Ampliconnoise & Pipeline\\ \hline\hline
$\beta_{0}$ & 0 & 0 & 2820.200 \\ \hline
$\beta_{1}$ & 1.718 & -47.687 & -0.012 \\ \hline
$\beta_{2}$ & -5.360 & 1.709 & 556.420 \\ \hline
$\beta_{3}$ & 1.1$\times 10^{-4}$ & -7.098 & -19.001 \\ \hline
$\beta_{4}$ & 4.096 & 2.2$\times 10^{-4}$ & 0.045 \\ \hline
$\beta_{5} $ & -4.6$\times 10^{-4}$ & 127.790 & 2.8$\times 10^{-7}$ \\ \hline
$\beta_{6} $ & 6.2$\times 10^{-10}$ & 0.017 & -1.640 \\ \hline
$\beta_{7} $ & 4.7$\times 10^{-4}$ & -0.184 & -2.1$\times 10^{-5}$ \\ \hline
$\beta_{8} $ & -1.4$\times 10^{-9}$ & -5.2$\times 10^{-4}$ & 0.032 \\ \hline
$\beta_{9} $ & ... & -1.6$\times 10^{-8}$ & -5.7$\times 10^{-7}$ \\ \hline
$\beta_{10} $ & ... & 3.0$\times 10^{-8}$ & ... \\ \hline
\end{tabular}
\end{table}
\begin{figure}[!h]
\centering
\includegraphics[trim=0cm 0cm 0cm 0cm,clip=true,width=4.4in]{wade6-speedup.jpg}
\caption{Speedup and Efficiency metrics for subsampled MFC data (20000 and 40000 reads) across multiple cores on the DIAG resource using AmpliconNoise}
\label{fig:speedup}
\end{figure}
\section{Conclusion}
The QIIME software package is widely employed for microbial community analysis using data derived from a range of NGS technologies. The protocols defined by \cite{kuczynski11} offer a standard methodology for processing sequencing reads and transforming that data into interpretable information such a sample diversity and phylogenetic distances. The benefits of using QIIME for downstream analysis of NGS data is its ability to pipeline a range of bioinformatics steps in a consistent and reproducible manner, several parallelised scripts for data intensive processing and its portability across a range of high-throughput and high-performance environments such as Amazon Cloud, Virtual Box/CloVR and Grid services such as DIAG.\par
When considering what resources to utilise for post-sequencing analysis, it is important to have an understanding of computational requirements for the tools employed. This is vital when considering deploying algorithms on resources that provide a service at a cost related to time utilised. However, it may also be important for users requiring rapid turnaround from sequencing to information, those wishing to optimise a pipeline or invest in additional computational resources, and in cases where the resource is utilised by multiple analysts. \par
The work presented demonstrates that there exists a significant relationship between characteristics of the input sequencing data and the computational time required to process that data. Although the models developed are characteristic of the system that they were calibrated against, the MLR modelling technique can be applied to generate models for any architecture. \par
Two and three variable models developed for sequence denoising algorithms were successful in predicting time for completion, but suffered with the Artificial data. This was potentially due to the nature of the dataset construction, which was not captured in the model development, which consisted entirely of samples from naturally distributed communities. However, a three variable general pipeline model predicted total time for completion of the standard QIIME analyses with 6\% and 7\% error for the Mixed and Artificial datasets, respectively. \par
There is often a temptation when working with computationally intensive algorithms to allocate the maximum amount of resource to the problem without considering if this will be optimal. Parallel speedup and efficiency analysis for the AmpliconNoise algorithm revealed that denoising with increasing number of CPUs is far from ideal. The efficiency of the computation decreases rapidly beyond a single core processor, whilst speedup is not significantly increased beyond 32 cores and the analysis suggests the use of between 8 and 16 processors is sufficient under the test architecture. \par
Accurate modelling of the relationship between input sequence data and time to completion has shown to be a viable method for supporting data analysts in decision-making related to resource allocation. Although restricted to a single architecture and set of tasks, the methodology can and should be applied in a broader context to understand if the approach is transferable across a range of applications (e.g. read mapping, genome assembly). A standardised model may be idealistic, but the modelling effort is minimal when contrasted with the expected data throughput that is beginning to emerge in sequencing today. As sequencing technology is evolving at a remarkable rate, it will be useful to look at assessing the methodology presented here against more diverse data sets, in terms of size and source environment. This should be coupled with attempts to identify how generic the models are across different system architectures and platforms. \par
It is suggested that by developing techniques and tools for modelling computational requirements for bioinformatics analysis, coupled with methods for estimating sequencing effort required to generate the necessary depth of information (see \cite{quince08}), then analysts can be armed with the capabilities to comprehensively plan their research efforts, assess resource requirements and funding needs.
\nomenclature{MFC}{Microbial Fuel Cell}
\nomenclature{MLR}{Multiple Linear Regression}
\nomenclature{OTU}{Operational Taxonomic Unit}
\nomenclature{PCR}{Polymerase Chain Reaction}
\nomenclature{RMSEC}{Root Mean Square Error of Calibration}
\printnomenclature
|
\section{Introduction}
Blokh \cite{Blokh} proved that a transitive map of a connected topological graph is either non-invertible and has the relative specification property, or is an irrational rotation of the circle. In the former case, a transitive map has dense set of periodic points and positive topological entropy. There are other connections between transitivity and topological entropy for one-dimensional maps (see \cite{AKLS}).
This suggests a question: what topological conditions force an analogous result? A natural candidate for the family of continua with similar properties is the class of dendrites. Dendrites generalize topological trees and appear in dynamics (for example as Julia sets for some complex polynomials \cite[Thm. 4.2]{CG}). It is also known that maps on dendrites share some dynamical properties with graph maps (see, for example, \cite{AHNO, K, KKM, Mai, N, S}).
In particular, if a dendrite contains a free arc then a transitive map necessarily has positive topological entropy (see \cite{DSS}).
In \cite{HM}, Hoehn and Mouron introduced a family of self-maps on the Wa\.{z}ewski universal dendrite $D_\omega$ and showed that among these maps there are topologically weakly mixing but not mixing ones. Somewhat imprecisely, we say that a map is a \emph{Hoehn-Mouron map} if it is defined following a similar procedure as in \cite{HM}.
This note contains a proof of the following theorem (here we present only the main part of it, for the precise statement see p. \pageref{Main}):
\begin{thm-main}[short version]
There exists a weakly mixing, not mixing, proximal Hoehn-Mouron map $f\colon D_\omega\to D_\omega$ with zero topological entropy.
\end{thm-main}
Our result gives an affirmative answer to the question of Baldwin \cite{Baldwin}, whether there is
a transitive dendrite map with zero entropy. This also provides a new example of a transitive map of a dendrite which
does not have dense periodic points (because proximality excludes existence of more than one minimal point and every periodic point is minimal). This property of the Hoehn-Mouron example is also noted in \cite{AHNO}.
Observe that dendrites have the fixed point
property, hence every continuous dendrite map has at least one fixed point. An example of a transitive dendrite map with
only one fixed point and no other periodic points is given in \cite{S}, but the map constructed there has positive topological entropy.
\section{Basic definitions}
We assume the reader is familiar with elementary theory of dynamical systems.
A \emph{dendrite} is a compact, pathwise connected, locally connected, non-separating, and nowhere dense subset of the plane, or equivalently a locally connected continuum that contains no simple closed curves. The \emph{universal dendrite} (the \emph{Wa\.zewski dendrite}, see \cite[\S 10.37, p. 181-185]{Nadler}) is a dendrite $D_\omega$ such that each ramification point of $D_\omega$ is of infinite order and for each arc $A\subset D_\omega$, the set of ramification points of $D_\omega$ contained in $A$ is dense in $A$. Let $f\colon D_\omega\to D_\omega$ be a continuous map.
Given non-empty open sets $U,V\subset D_\omega$, define $N(U,V)=\{n\in\N:f^{-n}(U)\cap V\neq\emptyset\}$.
We recall that $f$ is \emph{transitive} if $N(U,V)\neq\emptyset$ for every non-empty open sets $U,V\subset D_\omega$.
A map $f$ is \emph{weakly mixing} if the product map $f\times f$ is transitive.
We say that $f$ is mixing if the set $N(U,V)\neq\emptyset$ has finite complement in $\N$ for every non-empty open sets $U,V\subset D_\omega$.
Equivalently, $f$ is weakly mixing if $N(U,V)$ is thick for all nonempty open sets $U,V\subset X$. Recall that a set $A\subset \N$ is
\emph{thick}, if it contains arbitrarily long blocks of consecutive integers, that is,
for every $n>0$ there is $k\in\N$ such that $\{k,k+1,\ldots,k+n-1\}\subset A$. An $f$-invariant closed nonempty set $M\subset D_\omega$ is \emph{minimal} if an orbit of every point $x\in M$ is dense in $M$. A point $x\in D_\omega$ is a \emph{minimal point} if the closure of its orbit is a minimal set. We say that $f$ is \emph{proximal} if it has a unique minimal point. Note that if $x\in D_\omega$ is a unique minimal point for $f$, then $x$ is fixed by $f$.
\section{The main theorem}
\begin{thm-main} \label{Main} Let $D_\omega$ denote the Wa\.{z}ewski universal dendrite. There exists a set $Z\subset\N$ such that the Hoehn-Mouron map $f_Z\colon D_\omega\to D_\omega$ has the following properties:
\begin{enumerate}
\item $f_Z$ is weakly mixing, but not mixing;
\item all transitive points of $f_Z$ are end points of $D_\omega$;
\item $f_Z$ has a unique fixed point $o$, which is, in fact, a unique minimal point of $f_Z$ (it follows that $f_Z$ is proximal);
\item $f_Z$ is uniquely ergodic, with the only $f_Z$-invariant Borel probability measure being the Dirac measure concentrated on $o$;
\item $f_Z$ is Li-Yorke chaotic, but not DC$_2$-chaotic;
\item $f_Z$ has zero topological entropy.
\end{enumerate}
\end{thm-main}
\begin{figure}
\includegraphics{dendrite}
\caption{An approximation of the Wa\.{z}ewski universal dendrite and subdendrites $E_0,\ldots, E_5$.}
\label{rys}
\end{figure}
\begin{proof}
We follow the notation and terminology introduced in \cite{HM}. We write
\[
D_\omega=\bigcup_{j=0}^\infty E_j, \text{ where }E_i \cap E_j=\{o\} \text{ for }i\neq j,
\]
and each $E_j$ is a homeomorphic copy of $D_\omega$. The subdendrites $E_j$ are defined on page 7 of \cite{HM}.
Let $U_j=E_j\setminus\{o\}$, thus $U_j$ is an open subset of $D_\omega$ for $j=0,1,\ldots$.
Example 17 of \cite{HM} proves that given a set $Z\subset \N$
fulfilling certain assumptions there is a weakly mixing, but not mixing map $f_Z\colon D_\omega\to D_\omega$ satisfying
\begin{align}
f_Z(E_j)&=E_{j-1}, \quad\text{for }j\ge 1,\label{claim1}\\
f_Z^{n}(E_0)&\subset \bigcup_{n+j-1\in Z}E_j, \quad\text{for }n=1,2,\ldots.\label{claims1734}
\end{align}
It remains to find a set $Z$ such that the map $f_Z$ exists and has all the properties listed in the main theorem.
We construct the set $Z$ inductively.
Let $N_0=0$ and $M_0=1$. Assume that we have defined $N_K$ and $M_K>N_K$ for some $K\in\N$. We set
$N_{K+1}=(2^{M_K}+1)\cdot M_K$ and we take $M_{K+1}$ large enough so that
\textsc{Claims} 17.2 and 17.5 of \cite{HM}
are satisfied.\footnote{In the notation of \cite{HM}, it means that if $Z\subset\N$ is a thick set such that $[N_{K+1},M_{K+1}]\cap\N\subset Z$, then
for each $0\le i \le K+1$ there are some $m,n\in\N$ such that $\alpha(s)$ is defined for each
$s\in T_{s_i,\,K+1,\,m}$ and $\alpha^n(T_{s_i,\,K+1,\,m})=T_{\emptyset,\,K+1,\,0}$.}
Take
\[
Z=\bigcup_{K=1}^\infty [N_K,M_K]\cap\N.
\]
Then $f_Z$ as defined in \cite{HM} is weakly mixing but not mixing.
By \eqref{claims1734} for each $n\in\N$ we have
\[
f_Z^{n}(E_0)\subset \bigcup_{n+j-1\in Z}E_j, \quad\text{for }n=1,2,\ldots.
\]
It follows that
\[
f_Z^{n}(U_0)\subset \bigcup_{n+j-1\in Z}U_j\cup\{o\}, \quad\text{for }n=1,2,\ldots.
\]
But with our $Z$ we know that if $N=M_K+1$ for some $K\in\N$, then
\[
f_Z^{N}(U_0)\subset \bigcup_{j=2^{M_K}\cdot M_K}^\infty U_j\cup\{o\},
\]
hence
\[
f_Z^{N+l}(U_0)\subset \bigcup_{j=2^{M_K}\cdot M_K-l}^\infty U_j\cup\{o\}, \quad\text{for }l=1,2,\ldots,2^{M_K}\cdot M_K-1.
\]
Therefore for each $K\in\N$ we have
\begin{equation}\label{eq:disjont}
f_Z^j(U_0) \cap U_0=\emptyset\quad\text{for }j=M_K+1,\ldots,\left(2^{M_K}+1\right)\cdot M_K.
\end{equation}
Let $x\in U_0$ and $N(x,U_0)=\{n\in\N : f_Z^n(x)\in U_0\}$.
It follows from \eqref{eq:disjont} that
\begin{equation}\label{eq:lower_density}
\liminf_{n\to\infty}\frac{|N(x,U_0)\cap\{0,1,\ldots,n-1\}|}{n}=0.
\end{equation}
If $\bar\mu(U_0)>0$ for some $f_Z$-invariant Borel probability measure $\bar\mu$, then without loss of generality we can assume that $\bar\mu$ is ergodic and then \eqref{eq:lower_density} would contradict the pointwise ergodic theorem.
It follows that $\mu(U_0)=0$ for every $f_Z$-invariant Borel probability measure $\mu$. Moreover, we also have $\mu(U_j)=0$ for every $j\in\N$ since
\[
U_j\subset f_Z^{-j}(U_0), \quad\text{for }j=1,2,\ldots.
\]
(this follows easily from \eqref{claim1}).
Hence the Dirac measure concentrated on $o$ is the only $f_Z$ invariant probability measure. By the Variational Principle \cite{Walters}, $f_Z$ has topological entropy zero.
The statement concerning Li-Yorke chaos follows from $f_Z$ being weakly mixing (see \cite{Iwanik}). The statement about distributional chaos is a consequence of the fact that for each $x,y\in D_\omega$ the pair $(x,y)$ is proximal with upper Banach density one (because the Dirac measure concentrated on $o$ is the only $f_Z$-invariant probability measure, see \cite{LT}). We refer the reader to \cite{Iwanik, LT} for the details and definitions of Li-Yorke and DC$_2$-chaos.
\end{proof}
\section*{Acknowledgements}
We are greatly indebted to Vladim{\'{\i}}r {\v{S}}pitalsk{\'y} for his remarks and fruitful discussion. We also gratefully thank Logan Hoehn and Piotr Oprocha for several helpful comments concerning the subject of the paper. The research of Jakub Byszewski and Dominik Kwietniak was supported by the National Science Centre (NCN) under grant no. DEC-2012/07/E/ST1/00185.
|
\section{\textbf{Introduction}}The local Gan-Gross-Prasad conjecture deals with certain restriction problems between $p$-adic groups. In this paper, we shall investigate it for some non-generic case not yet treated before.
Let $E/F$ be a quadratic extension of number fields and $G=U(3)$ be the quasi-split unitary group of rank 3 relative to $E/F$. Then $H=U(2) \times U(1)$ is the unique elliptic endoscopic group for $G$.
In \cite{Ro}, Rogawski has defined a certain enlarged class of $L$-packets, or $A$-packets, of $G$ using endoscopic transfer of one-dimensional characters of $H$ to $G$. In more detail, let $\varrho= \otimes_{v} \varrho_v$ be a one-dimensional automorphic character of $H$. The $A$-packet $\Pi(\varrho) \simeq \otimes \Pi(\varrho_v)$ is the transfer of $\varrho$ with respect to functoriality for an embedding of $L$-groups $\xi : ^LH \to ^LG$. Then for all places $v$ of $F$, $\Pi(\varrho_v)$ contains a certain non-tempered representation $\pi^{n}(\varrho_v)$ and it contains an additional supercuspidal representaton $\pi^{s}(\varrho_v)$ precisely when $v$ remains prime in $E$. Gelbart and Rogawski \cite{Ge1} showed that the representations in this $A$-packet arise in the Weil representation of $G$. Our goal is to study the branching rule of the representations in this $A$-packet.
For the branching problem, there is a fascinating conjecture, the so-called Gan-Gross-Prasad (GGP) conjecture, which was first proposed by Gross and Prasad \cite{ggp} for orthogonal group and later they, together with Gan, extended it to all classical group in \cite{Gan2}. Since our main theorem has to do with it, we shall give a brief review on the GGP conjecture, especially for unitary group.
Let $E/F$ be a quadratic extension of non-archimedean local fields of characteristic zero. Let $V_{n+1}$ be a Hermitian space of dimension $n+1$ over $E$ and $W_n$ a skew-Hermitian space of dimension $n$ over $E$.
Let $V_n \subset V_{n+1}$ be a nondegenerate subspace of codimension $1$, so that if we set
\[ G_n = \U(V_n) \times \U(V_{n+1}) \quad \text{or} \quad \U(W_n) \times \U(W_n) \]
and
\[ H_n = \U(V_n) \quad \text{or} \quad \U(W_n), \]
then we have a diagonal embedding
\[ \Delta: H_n \hookrightarrow G_n. \]
Let $\pi$ be an irreducible smooth representation of $G_n$. In the Hermitian case, one is interested in computing
\[ \dim_\CC \Hom_{\Delta H_n} ( \pi, \CC). \]
We shall call this the \emph{Bessel} case (B) of the GGP conjecture.
For the GGP conjecture in the skew-Hermitian case, we need to introduce a certain Weil representation $\omega_{\psi, \chi, W_n}$ of $H_n$, where $\psi$ is a nontrivial additive character of $F$ and $\chi$ is a character of $E^{\times}$ whose restriction to $F^{\times}$ is the quadratic character $\omega_{E/F}$ associated to $E/F$ by local class field theory. (For the exact definition of $\omega_{\psi, \chi, W_n}$, please refer to Section. \ref{SS}.) In this case, one is interested in computing
\[ \dim_\CC \Hom_{\Delta H_n} ( \pi, \omega_{\psi,\chi, W_n}). \]
We shall call this the \emph{Fourier--Jacobi} case (FJ) of the GGP conjecture. To treat both cases using one notation, we shall let $\nu = \CC$ or $\omega_{\psi,\chi, W_n}$ in the respective cases.
By the results of \cite{agrs} \cite{sun}, it is known that \[ \dim_\CC \Hom_{\Delta H_n} ( \pi, \nu ) \le 1 \] and so the next step is to specify irreducible smooth representations $\pi$ such that $$\Hom_{\Delta H_n} ( \pi, \nu ) = 1. $$(A non-zero element of $\Hom_{\Delta H_n} ( \pi, \nu )$ is called a \emph{Bessel} (\emph{Fourier-Jacobi}) model of $\pi$ in the (skew) hermitian case.)
In \cite{Gan2}, Gan, Gross, Prasad has brought this problem into a more general setting using the notion of relevant pure inner forms of $G_n$ and Vogan $L$-packets. A pure inner form of $G_n$ is a group of the form
\[ G_n' = \U(V_n') \times \U(V'_{n+1}) \quad \text{or} \quad \U(W'_n) \times \U(W''_n) \]
where $V_{n}' \subset V_{n+1}'$ are $n$ and $n+1$ dimensional hermitian spaces over $E$ and $W_n', W_n''$ are $n$-dimensional skew hermitian spaces over $E$.
\noindent Furthermore, if \[ \quad V_{n+1}'/V_{n}' \cong V_{n+1}/V_{n} \quad \text{or} \quad W_n'=W_n'', \] we say that $G_n'$ is relevant pure inner form.\\ (Indeed, there are four pure inner forms of $G_n$ and among them, only two are relevant.)
If $G_n'$ is relevant, we set
\[ H'_n = \U(V'_n) \quad \text{or} \quad \U(W'_n), \]
so that we have a diagonal embedding
\[ \Delta: H'_n \hookrightarrow G'_n. \]
Now suppose that $\phi$ is an $L$-parameter for the group $G_n$. Then the (relevant) Vogan $L$-packet $\Pi_{\phi}$ associated to $\phi$ consists of certain irreducible smooth representations of $G_n$ and its (relevant) pure inner forms $G_n'$ whose $L$-parameter is $\phi$. We denote the relevant Vogan $L$-packet of $\phi$ by $\Pi^R_{\phi}$.
With these notions, we can loosely state the result of Beuzart-Plessis (\cite{bp1},\cite{bp2},\cite{bp3}) for \emph{Bessel} case and Gan-Ichino (\cite{iw}) for \emph{Fourier-Jacobi} case as follows :
\begin{thm}
For a tempered $L$-parameter $\phi$ of $G_n$, the followings hold:
\begin{enumerate}
\item $\sum_{\pi' \in \Pi^R_{\phi}}\dim_{\CC}\Hom_{\Delta H'_n} ( \pi', \nu )=1.$
\item Using the local Langlands correspondence for unitary group, we can pinpoint the unique $\pi' \in \Pi^R_{\phi}$ such that $$\dim_{\CC}\Hom_{\Delta H'_n} ( \pi', \nu )=1.$$
\end{enumerate}
\end{thm}
To emphasize its dependence on the number $n$, we denote the $Bessel$ and $\emph{Fourier-Jacobi}$ case of Theorem 1.1 as (B)$_n$ and (FJ)$_n$ respectively, and later we shall elaborate more on this notation. \\The GGP conjecture predicts that this theorem also holds for a generic $L$-parameter $\phi$ of $G_n$.
Our main theorem is to investigate (FJ)$_3$ for some $L$-parameter of $G_3$ involving a non-generic $L$-parameter of $U(W_3)$. More precisely, we have
\begin{main}For an irreducible smooth representation $\pi_2$ of $U(W_3)$, let $\pi=\pi^{n}(\varrho) \otimes \pi_2$ as a representation of $G_3$. Then
\begin{enumerate}
\item $\Hom_{\Delta H_3}(\pi,\omega_{\psi,\chi,W_3})=0$ if $\pi_2$ is not a theta lift from $U(V_2)$.
\item Assume that $\pi_2$ is the theta lift from $U(V_1')$ and let $\phi=\phi^{n}\otimes \phi_2$ be the $L$-parameter of $\pi$. Then $$\sum_{\pi' \in \Pi^R_{\phi}}\dim_{\CC}\Hom_{\Delta H'_3} ( \pi', \omega_{\psi,\chi,W_3} )=1.$$
\item Using the local Langlands correspondence for unitary groups, we can explicitly describe the representation $\pi' \in \Pi^R_{\phi}$ appearing in \text{(ii)} such that $$\label{p}\dim_{\CC}\Hom_{\Delta H'_3} ( \pi', \omega_{\psi,\chi,W_3} )=1.$$
\end{enumerate}
\end{main}
\begin{rem} As we shall see in Theorem \ref{Te}, the $L$-parameter of $\pi^{n}(\varrho_v)$ is not only non-tempered but also non-generic. Thus if we choose the $L$-parameter of $\phi_2$ in $\phi$ apart from those obtained by the theta lift from $U(V_2)$ to $U(W_3)$, then the first part of the Main Theorem tells us that the GGP conjecture may not be true for non-generic $L$-parameter of $G_n$.
\end{rem}The proof of Main Theorem is based on the following see-saw diagram :
\[
\xymatrix{
\U(W_3) \times \U(W_3) \ar@{-}[dr] \ar@{-}[d] & \U(V_{2}) \ar@{-}[d] \\
\U(W_3) \ar@{-}[ur] & \U(V_1) \times \U(V_{1})}.
\]
Since all elements in the $A$-packet $\Pi(\varrho)$ can be obtained by theta lift from $U(V_1)$, we can write $\pi^n(\varrho)=\Theta_{\psi,\chi,W_3,V_1}(\sigma)$ where $\sigma$ is an irreducible smooth character of $U(V_1)$ and $\psi,\chi$ are some characters, which are needed to fix a relevant Weil representation. Then by the see-saw identity, we have
$$\Hom_{U(W_3)}(\Theta_{\psi,\chi,W_3,V_1}(\sigma)\otimes \omega^{\vee}_{\psi,\chi,W_3},\pi^{\vee}_2) \simeq \Hom_{U(V_1)}(\Theta_{\psi,\chi,V_2,W_3}(\pi^{\vee}_2),\sigma).$$ From this, we see that for having $\Hom_{U(W_3)}(\Theta_{\psi,\chi,W_3,V_1}(\sigma)\otimes \omega^{\vee}_{\psi,\chi,W_3},\pi^{\vee}_2) \ne0$, it should be preceded $\Theta_{\psi,\chi,V_2,W_3}(\pi^{\vee}_2)\ne 0$. This accounts for (i) in the \textbf{Main Theorem} because $$\Hom_{U(W_3)}(\Theta_{\psi,\chi,W_3,V_1}(\sigma)\otimes \omega^{\vee}_{\psi,\chi,W_3},\pi^{\vee}_2)\simeq \Hom_{U(W_3)}(\Theta_{\psi,\chi,W_3,V_1}(\sigma)\otimes \pi_2, \omega_{\psi,\chi,W_3}).$$
If $\Theta_{\psi,\chi,V_2,W_3}(\pi^{\vee}_2)\ne 0$, then by the local theta correspondence, $\pi^{\vee}_2$ should be $\Theta_{\psi,\chi,W_3,V_2}(\pi_0)$, where $\pi_0$ is an irreducible representation of $U(V_2)$. By applying (B)$_1$, we can pinpoint $\pi_0$ and $\sigma$ in the framework of local Langlands correspondence such that $\Hom_{U(V_1)}(\pi_0,\sigma)\ne0$. Next we shall use the precise local theta correspondences for $\big(U(V_1),U(W_3)\big)$ and $\big(U(V_1),U(W_1)\big)$ in order to transfer recipe for (B)$_1$ to (FJ)$_3$.
The rest of the paper is organized as follows; In Section 2, we shall give a brief sketch of the local Langlands correspondence for unitary group. In Section 3, we collect some results on the local theta correspondence for unitary group which we will use in the proof of our main results. In Section 4, we shall prove our Main Theorem.
\subsection*{Acknowledgements} The author would like to express deep gratitude to professor Haseo Ki for introducing him this research area. The influence of the work of Gan and Ichino \cite{iw} on this paper is obvious to the reader. We are grateful to professor Atsushi Ichino for his kind suggestion to pursue the non-tempered aspects of the GGP conjecture. The author expresses his deepest thanks to the referee for his many invaluable comments and bringing our attention to the work of H.Atobe and Gan \cite{ag} containing Theorem 4.1 and Proposition 5.4, both of which are crucial to prove our main theorem. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIP)(ASARC, NRF-2007-0056093).
\subsection{Notations}\label{not}We fix some notations we shall use throughout this paper:
\begin{itemize}
\item $E/F$ is a quadratic extension of non-archimedean local fields of characteristic zero.
\item $c$ is the non-trivial element of Gal$(E/F)$.
\item $\text{Fr}_E$ is a Frobenius element of Gal$(\bar{E}/E)$.
\item Denote by $\Tr_{E/F}$ and $\N_{E/F}$ the trace and norm maps from $E$ to $F$.
\item $\delta$ is an element of $E$ such that $\Tr_{E/F}(\delta)=0$.
\item Let $\psi$ be an additive character of $F$ and define $$\psi^E(x) := \psi(\frac{1}{2}\Tr_{E/F}(\delta x)) \quad \text{and} \quad \psi^E_2(x) := \psi(\Tr_{E/F}(\delta x)).$$
\item Let $\chi$ be a character of $E^{\times}$ whose resriction to $F^{\times}$ is $\omega_{E/F}$, which is the quadratic character assosiated to $E/F$ by local class field theory.
\item For a linear algebraic group $G$, its $F$-points will be denoted by $G(F)$ or simply by $G$.
\end{itemize}
\section{\textbf{Local Langlands correspondence for unitary group}}
By the recent work of Mok \cite{Mok} and Kaletha-M\'inguez-Shin-White \cite{kmsw}, the local Langlands correpondence is now known for unitary group conditional on the stabilization of the twisted trace formula and weighted fundamental lemma. The twisted trace formula has now been stabilized by Waldspurger \cite{Stab1}--\cite{Stab9} and Moeglin-Waldspurger \cite{Stab6}--\cite{Stab10} and the proof of the weighted fundamental lemma is an ongoing project of Chaudouard and Laumon. Since our main results are expressed using the local Langlands correspondence, we shall assume the local Langlands correspondence for unitary group. In this section, we list some of its properties which are used in this paper. Indeed, much of this section are excerpts from Section.2 in \cite{iw}.
\subsection{Hermitian and skew-Hermitian spaces}
For $\varepsilon=\pm1$, let $V$ be a finite $n$-dimensional vector space over $E$ equipped with a nondegenerate $\varepsilon$-hermitian $c$-sesquilinear form $\langle \cdot, \cdot \rangle_V : V \times V \rightarrow E$.
That means for $v, w \in V$ and $a, b \in E$, \[
\langle a v, b w \rangle_V = a b^c \langle v, w \rangle_V, \qquad
\langle w, v \rangle_V = \varepsilon \cdot \langle v, w \rangle_V^c.
\]
\noindent We define $\disc V = (-1)^{(n-1)n/2} \cdot \det V$, so that
\[
\disc V \in
\begin{cases}
F^{\times} / \mathrm{N}_{E/F}(E^{\times})
& \text{if $\varepsilon = +1$;} \\
\delta^n \cdot F^{\times} / \mathrm{N}_{E/F}(E^{\times})
& \text{if $\varepsilon = -1$}
\end{cases}
\]
and we can define $\epsilon(V) = \pm 1$ by
\begin{equation}\label{sign}
\epsilon(V) =
\begin{cases}
\omega_{E/F}(\disc V) & \text{if $\varepsilon = +1$;} \\
\omega_{E/F}(\delta^{-n} \cdot \disc V) & \text{if $\varepsilon = -1$.}
\end{cases}
\end{equation}
By a theorem of Landherr, for a given positive integer $n$, there are exactly two isomorphism classes of $\varepsilon$-hermitian spaces of dimension $n$ and they are distinguished from each other by $\epsilon(V)$.\\
Let $\U(V)$ be the unitary group of $V$ defined by
\[
\U(V) = \{ g \in \GL(V) \, | \,
\text{$\langle g v, g w \rangle_V = \langle v, w \rangle_V$ for $v, w \in V$}
\}.
\]
Then $U(V)$ turns out to be connected reductive algebraic group defined over $F$.
\subsection{$L$-parameters and component groups}
Let $I_F$ be the inertia subgroup of $\text{Gal}(\bar{F}/F)$. Let $W_F=I_F \rtimes \langle \text{Fr}_F \rangle $ be the Weil group of $F$ and $\WD_F = W_F \times \SL_2(\CC)$ the Weil-Deligne group of $F$.
For a homomorphism $\phi: \WD_F \rightarrow \GL_n(\CC)$, we say that it is a representation of $\WD_F$ if
\begin{enumerate}
\item $\phi$ is continuous and $\phi(\text{Fr}_F)$ is semisimple,
\item the restriction of $\phi$ to $\SL_2(\CC)$ is induced by a morphism of algebraic groups $\SL_2 \to \GL_n$
\end{enumerate}
If moreover the image of $W_F$ is bounded then we say that $\phi$ is tempered. Define $\phi^{\vee}$ by $\phi^\vee(w) = {}^t\phi(w)^{-1}$ and call this the contragredient representation of $\phi$. If $E/F$ is a quadratic extension of local fields and $\phi$ is a representation of $\WD_E$, fix $s \in W_F \smallsetminus W_E$ and define a representation $\phi^c$ of $\WD_E$ by $\phi^c(w) = \phi(sws^{-1})$.
The equivalence class of $\phi^c$ is independent of the choice of $s$. We say that $\phi$ is conjugate self-dual if there is an isomorphism $b: \phi \mapsto (\phi^{\vee})^{c}$. Note that there is a natural isomorphism $(((\phi^{\vee})^c)^{\vee})^c\simeq \phi$ so that $(b^{\vee})^c$ can be considered as an isomorphism from $\phi$ onto $(\phi^{\vee})^c$. For $\varepsilon=\pm 1$, we say that $\phi$ is conjugate self-dual with sign $\varepsilon$ if there exists such a $b$ satisfying the extra condition $(b^{\vee})^c=\varepsilon \cdot b$.
Define $\text{As}(\phi):WD_F \to \GL_{n^2}(\mathbb{C})$ by tensor induction of $\phi$ as follows;
$$\text{As}(\phi)(w)=\begin{cases}\phi(w)\otimes \phi(s^{-1}ws) \quad \quad \quad \quad\text{ if } w\in WD_E \\ \iota \circ (\phi(s^{-1}w) \otimes \phi(ws)) \quad \quad \text{ if } w \in WD_F \smallsetminus WD_E\end{cases}$$ where $\iota$ is the linear isomorphism of $\mathbb{C}^n\otimes \mathbb{C}^n$ given by $\iota(x \otimes y )=y \otimes x$.
Then the equivalence class of $\text{As}(\phi)$ is also independent of the choice of $s$. We set $\text{As}^+(\phi)=\text{As}(\phi)$ and $\text{As}^{-}(\phi)=\text{As}(\chi \otimes \phi)$.
Let $V$ be a $n$-dimensional $\varepsilon$-hermitian space over $E$. An $L$-parameter for the unitary group $\U(V)$ is a $GL_n(\mathbb{C})$-conjugacy class of admissible homomorphisms $$\phi:\WD_F \longrightarrow ^LU(V)=GL_{n}(\CC) \rtimes \text{Gal}(E/F)$$ such that the composite of $\phi$ with the projection onto Gal$(E/F)$ is the natural projection of $\WD_F$ to Gal$(E/F)$.
The following proposition from \cite[\S 8]{Gan2} enables us to removes the cumbersome Gal$(E/F)$-factor in the definition of $L$-parameter of $U(V)$.
\begin{prop}Restriction to $\WD_E$ gives a bijection between the set of $L$-parameters for $U(V)$ and the set of equivalence classes of conjugate self-dual representations $$\phi:\WD_E \longrightarrow GL_n(\CC)$$
of sign $(-1)^{n-1}$.
\end{prop}
With this proposition, by an $L$-parameter for $U(V)$, we shall mean a $n$-dimensional conjugate self-dual representation $\phi$ of $\WD_E$ of sign $(-1)^{n-1}$.
Given a $L$-parameter $\phi$ of $U(V)$, we say that $\phi$ is generic if its adjoint $L$-function $L(s,\text{Ad} \circ \phi)=L(s,\text{As}^{(-1)^{n-1}} (\phi))$ is holomorphic at $s=1$.
Write $\phi$ as a direct sum
\[ \phi = \bigoplus_i m_i \phi_i \]
of pairwise inequivalent irreducible representations $\phi_i$ of $\WD_E$ with multiplicities $m_i$.
We say that $\phi$ is square-integrable if it has no multiplicity (i.e. $m_i=1$ for all $i$) and $\phi_i$ is conjugate self-dual with sign $(-1)^{n-1}$ for all $i$. Furthermore, we can associate its component group $S_{\phi}$ to $\phi$.
As explained in \cite[\S 8]{Gan2}, $S_{\phi}$ is a finite 2-abelian group which can be described as
\[ S_{\phi} = \prod_j (\ZZ / 2\ZZ) a_j \]
with a canonical basis $\{ a_j \}$, where the product ranges over all $j$ such that $\phi_j$ is conjugate self-dual with sign $(-1)^{n-1}$.
If we denote the image of $-1 \in \GL_n(\CC)$ in $S_\phi$ by $z_\phi$, we have\[
z_{\phi} = (m_j a_j) \in \prod_j (\ZZ / 2\ZZ) a_j.
\]
\subsection{Local Langlands correspondence for unitary group}
The aim of the local Langlands correspondence for unitary groups is to classify the irreducible smooth representations of unitary groups. To state it, we first introduce some notations.
\begin{itemize}
\item Let $V^+$ and $V^-$ be the $n$-dimensional $\varepsilon$-Hermitian spaces with $\epsilon(V^+) = +1$, $\epsilon(V^-) = -1$ respectively.
\item Let $\Irr(\U(V^{\pm}))$ be the set of irreducible smooth representations of $\U(V^{\pm})$.
\end{itemize}
Then the local Langlands correspondence a form enhenced by Vogan \cite{v}, says that for a $L$-parameter $\phi$ of $U(V^{\pm})$, there is the so-called Vogan $L$-packet $\Pi_{\phi}$, a finite set consisting of irreducible smooth representations of $U(V^{\pm})$, such that
\[ \Irr(\U(V^+)) \sqcup \Irr(\U(V^-)) = \bigsqcup_{\phi} \Pi_{\phi}, \]
where $\phi$ on the right-hand side runs over all equivalence classes of $L$-parameters for $\U(V^\pm)$.
\noindent Then under the local Langlands correspondence, we may also decompose $\Pi_{\phi}$ as
\[ \Pi_{\phi} = \Pi_{\phi}^+\sqcup \Pi_{\phi}^-, \]
where for $\epsilon = \pm 1$, $\Pi_{\phi}^{\epsilon}$ consists of the representations of $\U(V^{\epsilon})$ in $\Pi_{\phi}$.
Furthermore, as explained in \cite[\S 12]{Gan2}, there is a bijection $$J^{\psi}(\phi):\Pi_{\phi} \to \Irr( S_{\phi})$$ which is canonical when $n$ is odd and depends on the choice of an additive character of $\psi:F^{\times} \to \CC$ when $n$ is even. More precisely, such bijection is determined by the $\N_{E/F}(E^{\times})$-orbit of nontrivial additive characters
\[
\begin{cases}
\psi^E:E/F \rightarrow \CC^{\times} & \text{if $\varepsilon = +1$;} \\
\psi:F \rightarrow \CC^{\times} & \text{if $\varepsilon = -1$.}
\end{cases}
\]
According to this choice, when $n$ is even, we write
\[J^{\psi}=
\begin{cases}
J_{\psi^E} & \text{if $\varepsilon = +1$;} \\
J_{\psi} & \text{if $\varepsilon = -1$},
\end{cases}
\] and even when $n$ is odd, we retain the same notation $J^{\psi}(\phi)$ for the canonical bijection.\\
Hereafter, if a nontrivial additive chracter $\psi:F \rightarrow \CC^{\times}$ is fixed, we define $\psi^E:E/F \rightarrow \CC^{\times} $ by
\[ \psi^E(x) : = \psi(\tfrac{1}{2} \Tr_{E/F}(\delta x)) \] and using these two characters, we fix once and for all a bijection $$J^{\psi}(\phi):\Pi_{\phi} \to \Irr( S_{\phi})$$ as above.
With these fixed bijections, we can label all irreducible smooth representations of $\U(V^{\pm})$ as $\pi(\phi,\eta)$ for some unique $L$-parameter $\phi$ of $U(V^{\pm})$ and $\eta \in \Irr( S_{\phi})$.
\subsection{Properties of the local Langlands correspondence}
\label{SS:llc}
We briefly list some properties of the local Langlands correspondence for unitary group, which we will use in this paper:
\begin{itemize}
\item
$\pi(\phi,\eta)$ is a representation of $\U(V^{\epsilon})$ if and only if $\eta(z_\phi) = \epsilon$.
\item $\pi(\phi,\eta)$ is tempered if and only if $\phi$ is tempered.
\item $\pi(\phi,\eta)$ is square-integrable if and only if $\phi$ is square-integrable.
\item The component groups $S_{\phi}$ and $S_{\phi^{\vee}}$ are canonically identified. Under this canonical identification, if $\pi=\pi(\phi,\eta)$, then its contragradient representation $\pi^{\vee}$ is $\pi(\phi^{\vee},\eta\cdot \nu)$ where \[
\nu(a_j) = \begin{cases}
\omega_{E/F}(-1)^{\dim \phi_j} & \text{if $\dim_{\CC} \phi$ is even;} \\
1 & \text{if $\dim_{\CC} \phi$ is odd.} \end{cases} \]
(This property follows from a result of Kaletha \cite[Theorem 4.9]{kel}.)
\end{itemize}
\section{\textbf{Local theta correspondence}}
In this section, we state the local theta correspondence of unitary groups for two low rank cases. From now on, for $\epsilon = \pm 1$, we shall denote by $V_n^\epsilon$ the $n$-dimensional Hermitian space with $\epsilon(V_n^\epsilon) = \epsilon$ and by $W_n^\epsilon$ the $n$-dimensional skew-Hermitian space with $\epsilon(W_n^\epsilon) = \epsilon$, so that $W_n^\epsilon = \delta \cdot V_n^\epsilon$.
\subsection{The Weil representation for Unitary groups}
\label{SS}In this subsection, we introduce the Weil representation.
Let $E/F$ be a quadratic extension of local fields and $(V_m,\langle,\rangle_{V_m})$ be a $m$-dimensional Hermitian space and $(W_n,\langle,\rangle_{W_n})$ a $n$-dimensional skew-Hermitian space over $E$. Define the symplectic space
$$
\mathbb{W}_{V_m,W_n} := \operatorname{Res}_{E/F} V_m \otimes_E W_n
$$
with the symplectic form $$
\langle v \otimes w,v' \otimes w' \rangle_{\mathbb{W}_{V_m,W_n}} := \frac{1}{2}\operatorname{tr}_{E/F}\left(\langle v,v'\rangle_{V_m}{\langle w,w' \rangle}_{W_n}\right).
$$
We also consider the associated symplectic group $Sp(\mathbb{W}_{V_m,W_n})$ preserving $\langle \cdot,\cdot \rangle_{\mathbb{W}_{V_m,W_n}}$ and the metaplectic group $\widetilde{Sp}(\mathbb{W}_{V_m,W_n})$ which sits in a short exact sequence :
$$1\to \CC^\times\to\widetilde{Sp}(\mathbb{W}_{V_m,W_n})\to Sp(\mathbb{W}_{V_m,W_m})\to 1.$$\\
Let $\mathbb{X}_{V_m,W_n}$ be a Lagrangian subspace of $\mathbb{W}_{V_m,W_n}$ and fix an additive character $\psi:F\to\CC^\times$. Then we have a Schr\"odinger model of the Weil Representation $\omega_\psi$ of $\widetilde{Sp}(\mathbb{W})$ on $\mathcal{S}(\mathbb{X}_{V_m,W_n})$, where $\mathcal{S}$ is the Schwartz-Bruhat function space.
If we set \[ \chi_{V_m} :=\chi^m \quad \text{ and } \quad \chi_{W_n}:=\chi^n, \]
where $\chi$ is a character of $E^{\times}$ whose resriction to $F^{\times}$ is $\omega_{E/F}$, which is the quadratic character assosiated to $E/F$ by local class field theory, then $(\chi_{V_m}, \chi_{W_n})$ gives a splitting homomorphism $$
\iota_{\chi_{V_m}, \chi_{W_n}}:U(V_m)\times U(W_n)\to \widetilde{Sp}(\mathbb{W}_{V_m,W_n})
$$ and so by composing this to $\omega_\psi$, we have a Weil representation $\omega_\psi\circ \iota_{\chi_{V_m},\chi_{W_n}}$ of $U(V_m) \times U(W_n)$ on $\mathbb{S}(\mathbb{X}_{V_m,W_n})$.
When the choice of $\psi$ and $(\chi_{V_m}, \chi_{W_n})$ is fixed as above, we simply write $$\omega_{\psi,W_n,V_m}:= \omega_\psi\circ \iota_{\chi_{V_m},\chi_{W_n}}.$$ Throughout the rest of the paper, when it comes to a Weil representations of $U(V_m) \times U(W_n)$, we shall denote it by $\omega_{\psi,W_n,V_m}$ with understanding the choices of characters $(\chi_{V_m},\chi_{W_n})$ as above.
\begin{rem}\label{con} When $m=1$, the image of $U(V_1)$ in $\widetilde{Sp}(\mathbb{W}_{V_1,W_n})$ coincides with the image of the center of $U(W_n)$, and so we regard the Weil representation of $U(V_1) \times U(W_n)$ as a representation of $U(W_n)$. In this case, we denote the Weil representation of $U(W_n)$ as $\omega_{\psi,W_n}$. Furthermore, we can also use $\chi_{V_1}=\chi^{-1}$ for the choice of splitting homomorphism $\iota_{\chi_{V_1}, \chi_{W_n}}$ instead of $\chi_{V_1}=\chi$. In this case, the Weil representation of $U(W_n)$ is $\omega^{\vee}_{\psi,W_n}$.
\end{rem}
\subsection{Local theta correspondence}
Given a Weil representation $\omega_{\psi,W_n,V_m}$ of $U(V_m) \times U(W_n)$ and an irreducible smooth representation $\pi$ of $\U(W_n)$, the maximal $\pi$-isotypic quotient of $\omega_{\psi,V_m,W_n}$ is of the form
\[
\Theta_{\psi,V_m,W_n}(\pi) \boxtimes \pi
\]
for some smooth representation $\Theta_{\psi,V_m,W_n}(\pi)$ of $\U(V_m)$ of finite length.
By the Howe duality\footnote{It was first proved by Waldspurger \cite{w} for all residual characteristic except $p= 2$. Recently, Gan and Takeda \cite{gt1}, \cite{gt2} have made it available for all residual characteristics.}, the maximal semisimple quotient $\theta_{\psi,V_m,W_n}(\pi)$ of $\Theta_{\psi,V_m,W_n}(\pi)$ is either zero or irreducible.
In this paper, we consider two kinds of theta correspondences for $(\U(1) \times \U(3))$ and $(\U(2) \times \U(3))$ :
\subsection{Case (i)} Now we shall consider the theta correspondence for $\U(V_{1}^\epsilon) \times \U(W_3^{\epsilon'})$. The following is a compound of Theorem 3.4 and Theorem 3.5 in \cite{Haan}.
\begin{thm}\label{Te}Let $\phi$ be a $L$-parameter of $U(V_1^{\pm})$. Then we have:
\begin{enumerate}
\item For any $\epsilon,\epsilon'=\pm1$ and any $\pi \in \Pi_{\phi}^{\epsilon'}$, $\Theta_{\psi, W_3^{\epsilon},V_1^{\epsilon'}}(\pi)$ is nonzero and irreducible.
\item $\Theta_{\psi, W_3^{\epsilon},V_1^{\epsilon'}}(\pi)$ $=\begin{cases} \text{a non-tempered representation} & \text{if $\epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=\epsilon \cdot \epsilon'$} \\ \text{a supercuspidal representation } & \text{if $\epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=-\epsilon \cdot \epsilon'$} \end{cases} $ , \\ where $$\psi_2^{E}(x)=\psi (\Tr_{E/F}(\delta x)).$$
\item The $L$-parameter $\theta(\phi)$ of $\Theta_{\psi, W_3^{\epsilon},V_1^{\epsilon'}}(\pi)$ has the following form; \begin{equation}\label{non}\theta(\phi)=\begin{cases} \theta^n(\phi)=\chi |\cdot|_{E}^{\frac{1}{2}} \oplus \phi \cdot \chi^{-2} \oplus \chi |\cdot|_{E}^{-\frac{1}{2}}& \text{if $\epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=\epsilon \cdot \epsilon'$} \\ \theta^s(\phi)= \phi \cdot \chi^{-2} \oplus \chi \boxtimes \textbf{S}_2 & \text{if $\epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=-\epsilon \cdot \epsilon'$} \end{cases},\end{equation} where $\textbf{S}_2$ is the standard 2-dimensional representation of $SL_2(\mathbb{C}).$
\item For $\epsilon, \epsilon'$ such that $\epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=\epsilon \cdot \epsilon'$, the theta correspondence $\pi \mapsto \theta_{\psi, W_{3}^{\epsilon}, V_1^{\epsilon'}}(\pi)$ gives a bijection
\[ \Pi_{\phi} \longleftrightarrow \Pi_{\theta^n(\phi)}. \]
\item For $\epsilon, \epsilon'$ such that $\epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=-\epsilon \cdot \epsilon'$, the theta correspondence $\pi \mapsto \theta_{\psi, W_{3}^{\epsilon}, V_1^{\epsilon'}}(\pi)$ gives an injection
\[ \Pi_{\phi} \hookrightarrow \Pi_{\theta^s(\phi)}. \]
\end{enumerate}
Write \begin{itemize} \item $S_{\phi}\ = \ (\ZZ/2\ZZ)a_1$
\item $S_{\theta^n(\phi)}=(\ZZ/2\ZZ)a_1 \quad \ \quad\quad\quad\quad\quad \text{ if }\ \epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=\epsilon \cdot \epsilon'$
\item $S_{\theta^s(\phi)}=(\ZZ/2\ZZ)a_1 \times (\ZZ/2\ZZ)a_2
\quad \text{ if }\ \epsilon(\frac{1}{2},\phi \otimes \chi^{-3}, \psi_2^{E})=-\epsilon \cdot \epsilon' $\end{itemize}
where $$\psi_2^{E}(x)=\psi (\Tr_{E/F}(\delta x)).$$
(Note that $\theta^s(\phi)$ is a square-integrable $L$-parameter of $U(W_3^{\epsilon})$ and the summand $(\ZZ/2\ZZ)a_2$ of $S_{\theta^s(\phi)}$ arises from the summand $\chi \boxtimes \textbf{S}_2$ in $\theta^s(\phi)$.)
Since we are only dealing with odd dimensional spaces, there are three canonical bijections
\begin{itemize}
\item $J^{\psi}(\phi): \Pi_{\phi} \longleftrightarrow \Irr(S_{\phi}) $
\item $J^{\psi}(\theta^n(\phi)): \Pi_{\theta^n(\phi)} \longleftrightarrow \Irr(S_{\theta^n(\phi)}) $
\item $J^{\psi}(\theta^s(\phi)): \Pi_{\theta^s(\phi)} \longleftrightarrow \Irr(S_{\theta^s(\phi)}).$
\end{itemize}
Using these maps, the following bijection and inclusion
\begin{align*}
\Irr(S_{\phi}) & \longleftrightarrow \Irr(S_{\theta^n(\phi)}) \\
\eta & \mapsto \theta^n(\eta),
\end{align*}
\begin{align*}
\Irr(S_{\phi}) & \hookrightarrow \Irr(S_{\theta^s(\phi)}) \\
\eta & \mapsto \theta^s(\eta)
\end{align*}induced by the theta correspondence can be explicated as follows:
\begin{equation}\label{theta1} \theta^{n}(\eta)(a_1)=\eta(a_1) \cdot \epsilon(\frac{1}{2},\phi \otimes \chi^{-3},\psi_{2}^E ),\end{equation}
\begin{equation}\label{theta2}\theta^{s}(\eta)(a_1)=\eta(a_1) \cdot \epsilon(\frac{1}{2},\phi \otimes \chi^{-3},\psi_{2}^E ) , \\ \quad \quad \theta_{2}(\eta)(a_2)=-1.\end{equation}
\end{thm}
\begin{rem}Note that $\theta^n(\phi)$ is a non-generic $L$-parameter. Gan and Ichino (Proposition B.1 in Appendix in \cite{iw}) proved that an $L$-parameter is generic if and only if its associated $L$-packet $\Pi_{\phi}$ contains a generic representation (i.e. one possessing a Whittaker model). Together with the Corollary 4.2.3 in \cite{Ge2}, which asserts that all elements in $\Pi_{\theta^n(\phi)}$ have no Whittaker models, we see that $\theta^n(\phi)$ is a non-generic $L$-parameter.
\end{rem}
\subsection{Case (ii)}
Now we shall consider the theta correspondence for $\U(V_2^{\epsilon'}) \times \U(W_{3}^\epsilon)$. The following summarises some results of \cite{gi}, \cite{iw}, which are specialised to this case.
\begin{thm} \label{Tae}
Let $\phi$ be an $L$-parameter for $\U(V_2^{\pm})$. Then we have:
\begin{enumerate}
\item Suppose that $\phi$ does not contain $\chi^3$.
\begin{enumerate}
\item For any $\pi \in \Pi_{\phi}^{\epsilon'}$ and any $\epsilon \in \{ \pm1\}$,
$\Theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ is nonzero and $\theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ has $L$-parameter
\[ \theta(\phi) = (\phi \otimes \chi^{-1}) \oplus \chi^2. \]
\item For each $\epsilon = \pm 1$, the theta correspondence $\pi \mapsto \theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ gives a bijection
\[ \Pi_{\phi} \longleftrightarrow \Pi_{\theta(\phi)}^{\epsilon}. \]
\end{enumerate}
\item Suppose that $\phi$ contains $\chi^3$.
\begin{enumerate}
\item For any $\pi \in \Pi_{\phi}^{\epsilon'}$, exactly one of $\Theta_{\psi, W_{3}^+, V_2^{\epsilon'}}(\pi)$ or $\Theta_{\psi, W_{3}^-, V_2^{\epsilon'}}(\pi)$ is nonzero.
\item If $\Theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ is nonzero, then $\theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ has $L$-parameter
\[ \theta(\phi) = (\phi \otimes \chi^{-1} ) \oplus \chi^2. \]
\item The theta correspondence $\pi \mapsto \theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ gives a bijection
\[ \Pi_{\phi} \longleftrightarrow \Pi_{\theta(\phi)}. \]
\end{enumerate}
\item We have fixed a bijection
\[ J_{\psi^E}(\phi): : \Pi_{\phi} \longleftrightarrow \Irr(S_{\phi}) ,\] where
\[ \psi^E(x) = \psi(\tfrac{1}{2} \Tr_{E/F}(\delta x)) \]
and there is the bijection
\[ \quad J^{\psi}(\theta(\phi)): \Pi_{\theta(\phi)} \longleftrightarrow \Irr(S_{\theta(\phi)}). \]
\begin{itemize}
\item If $\phi$ does not contain $\chi^3$, we have
\[ \ S_{\theta(\phi)} = S_{\phi} \times (\ZZ/2\ZZ)b_1, \]
where the extra copy of $\ZZ/2\ZZ$ of $S_{\theta(\phi)}$ arises from the summand $\chi^2$ in $\theta(\phi)$. \\Then for each $\epsilon$, using the above bijection $J$ and $J_{\psi^E}$, one has a canonical bijection
\begin{align*}
\Irr(S_{\phi}) & \longleftrightarrow \Irr^{\epsilon}(S_{\theta(\phi)}) \\
\eta & \longleftrightarrow \theta(\eta)
\end{align*}
induced by the theta correspondence, where $\Irr^{\epsilon}(S_{\theta(\phi)})$ is the set of irreducible characters $\eta'$ of $S_{\theta(\phi)}$ such that $\eta'(z_{\theta(\phi)}) =\epsilon$ and the bijection is determined by
\[ \theta(\eta)|_{S_{\phi}} =\eta. \]
\item If $\phi$ contains $\chi^3$, then $\phi \otimes \chi^{-1}$ contains $\chi^2$, and so
\[
S_{\theta(\phi)} = S_{\phi}. \]
Thus, one has a canonical bijection
\begin{align*}
\Irr(S_{\phi}) & \longleftrightarrow \Irr(S_{\theta(\phi)}) \\
\eta & \longleftrightarrow \theta(\eta)
\end{align*}
induced by the theta correspondence and it is given by \[ \theta(\eta) = \eta. \]
\end{itemize}
\item If $\pi$ is tempered and $\Theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ is nonzero, then $\Theta_{\psi, W_{3}^{\epsilon}, V_2^{\epsilon'}}(\pi)$ is irreducible and tempered.
\end{enumerate}
\end{thm}
\section{\textbf{Main Theorem}}
In this section, we prove our main theorem. To prove it, we first state the precise result of Beuzart-Plessis which we shall use in the proof of Theorem \ref{thm2}.\begin{footnote}{Recently, Gan and Ichino (\cite{iw}) extended Beuzart-Plessis's work to the generic case relating it to (FJ) case.}\end{footnote}
\begin{B}Let $\phi=\phi^{(n+1)} \times \phi^{(n)}$ be a tempered $L$-parameter of $U(V_{n+1}^{\pm}) \times U(V_n^{\pm})$ and write $S_{\phi^{(n+1)}}=\prod_{i} (\ZZ/2\ZZ)a_i$ and $S_{\phi^{(n)}}=\prod_{j} (\ZZ/2\ZZ)b_j$. Let $\Delta : U(V_n^{\pm}) \hookrightarrow U(V_{n+1}^{\pm}) \times U(V_n^{\pm})$ be the diagonal map. Then for $\pi(\eta) \in \Pi_{\phi}^{R,\ \pm}=\Pi_{\phi^{(n+1)}}^{\pm} \times \Pi_{\phi^{(n)}}^{\pm} $ where $\eta \in \Irr(S_{\phi})= \Irr (S_{\phi^{(n+1)}}) \times \Irr (S_{\phi^{(n)}})$,
$$\Hom_{\Delta (U(V_n^{\pm}))} ( \pi(\eta), \CC )=1 \Longleftrightarrow \eta=\eta^{\spadesuit} \text{ where }$$
\[
\begin{cases}
\eta^{\spadesuit}(a_i) = \epsilon(\frac{1}{2}, \phi^{(n+1)}_i \otimes \phi^{(n)}, \psi^E_{-2}); \\
\eta^{\spadesuit}(b_j) = \epsilon(\frac{1}{2}, \phi^{(n+1)} \otimes \phi^{(n)}_{j}, \psi^E_{-2}).
\end{cases}
\] where $\psi^E_{-2}(x) = \psi(-\Tr_{E/F}(\delta x))$.
\end{B}
\begin{thm}\label{thm2} Let $\phi^{(1)},\phi^{(2)}$ be tempered $L$-parameters of $U(V_1^{\pm})$ and $U(V_2^{\pm})$ respectively and suppose that $\phi^{(2)}$ does not contain $\chi^{-3}$. Let $$\theta^n(\phi^{(1)})=\chi |\cdot|_{E}^{\frac{1}{2}} \oplus \phi^{(1)} \cdot \chi^{-2} \oplus \chi |\cdot|_{E}^{-\frac{1}{2}},$$ $$\theta^s(\phi^{(1)})=\phi^{(1)} \cdot \chi^{-2} \oplus \chi \boxtimes \textbf{S}_2$$ be the two $L$-parameters of $U(W_3^{\pm})$ appearing in (\ref{non}) and let $$\theta(\phi^{(2)})=\phi^{(2)} \otimes \chi \oplus \chi^{-2} $$ be the $L$-parameters of $U(W_3^{\pm})$ appearing in Theorem \ref{Tae} (ii), in which $\chi$ is replaced by $\chi^{-1}$.
Write \begin{itemize}
\item $S_{\phi^{(1)}}=S_{\theta^n(\phi^{(1)})}=(\ZZ/2\ZZ)a_1;$ \item $S_{\theta^s(\phi^{(1)})}=S_{\phi^{(1)}} \times (\ZZ/2\ZZ)a_2;$
\item $S_{\phi^{(2)}}=\begin{cases}(\ZZ/2\ZZ)b_1 \quad \text{ if $\phi^{(2)}$ is irreducible}; \\ (\ZZ/2\ZZ)b_1 \times (\ZZ/2\ZZ)b_2 \quad \text{ if $\phi^{(2)}=\phi_1^{(2)} \oplus \phi_2^{(2)}$ is reducible} \end{cases}$
\item $S_{\theta(\phi^{(2)})}=S_{\phi^{(2)}} \times (\ZZ/2\ZZ)c_1$ \end{itemize} where $c_1$ comes from the component $\chi^{-2}$ of $\theta(\phi^{(2)})$.\\
We use the fixed character $\psi$ to fix the local Langlands correspondence for $\Pi_{ \phi^{(2)}} \leftrightarrow \Irr(S_{\phi^{(2)}}).$
\noindent For $x=n, s$, let $$\theta^{x}(\phi^{(1)},\phi^{(2)})=\theta^x(\phi^{(1)}) \times \theta(\phi^{(2)})$$ be a $L$-parameter of $G_3^{\pm}=\U(W_3^\pm) \times \U(W_3^\pm)$ and $\pi^{x}(\eta)\in \Pi^{R, \ \epsilon}_{\theta^{x}(\phi^{(1)},\phi^{(2)})}$ a representation of a relevant pure inner form of $G_3$.
Then,
\[ \Hom_{\Delta \U(W_3^\epsilon)}(\pi^{x}(\eta), \omega_{\psi,W_3^{\epsilon}}) \ne 0 \Longleftrightarrow \eta = \eta_{x}^{\blacklozenge}\]
where $\eta_{x}^{\blacklozenge}\in\Irr(S_{\theta^{x}(\phi^{(1)},\phi^{(2)})}) = \Irr(S_{\theta^x(\phi^{(1)})}) \times \Irr(S_{\theta(\phi^{(2)})}) $ is specified as follows;
\begin{enumerate}
\item When $\phi^{(2)}$ is irreducible,
$$\begin{cases}\eta_{n}^{\blacklozenge}(a_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E})\\ \eta_{n}^{\blacklozenge}(b_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)},\psi_2^{E})\\ \eta_{n}^{\blacklozenge}(c_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \end{cases} ,$$
$$\begin{cases}\eta_{s}^{\blacklozenge}(a_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E})\\ \eta_{s}^{\blacklozenge}(a_2)=-1\\
\eta_{s}^{\blacklozenge}(b_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)},\psi_2^{E})\\ \eta_{s}^{\blacklozenge}(c_1)=-\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \end{cases} .$$\\
\item When $\phi^{(2)}=\phi_1^{(2)}\oplus \phi_2^{(2)}$ is reducible,
$$\begin{cases}\eta_{n}^{\blacklozenge}(a_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E})\\ \eta_{n}^{\blacklozenge}(b_1)=\epsilon(\frac{1}{2}, \phi^{(1)} \otimes \phi_1^{(2)}, \psi_2^{E})\\ \eta_{n}^{\blacklozenge}(b_2)=\epsilon(\frac{1}{2}, \phi^{(1)} \otimes \phi_2^{(2)}, \psi_2^{E}) \\ \eta_{n}^{\blacklozenge}(c_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \end{cases},$$
$$\begin{cases}\eta_{s}^{\blacklozenge}(a_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E})\\ \eta_{s}^{\blacklozenge}(a_2)=-1\\
\eta_{s}^{\blacklozenge}(b_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi_1^{(2)}, \psi_2^{E})\\ \eta_{s}^{\blacklozenge}(b_2)=\epsilon(\frac{1}{2}, \phi^{(1)} \otimes \phi_2^{(2)}, \psi_2^{E}) \\ \eta_{s}^{\blacklozenge}(c_1)=-\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \end{cases}.$$
\end{enumerate}
Furthermore, \[ \dim_{\mathbb{C}}\Hom_{\Delta \U(W_3^\epsilon)}(\pi^{x}(\eta_{x}^{\blacklozenge}), \omega_{\psi,W_3^{\epsilon}}) =1 \]
\end{thm}
\begin{rem}When $x=n$ or $s$, we have $\eta_{x}^{\blacklozenge}(z_{\theta^x(\phi^{(1)})})=\eta_{x}^{\blacklozenge}(z_{\theta(\phi^{(2)})})$ so that $\eta_{x}^{\blacklozenge}$ always corresponds to a representation $\pi^x(\eta_{x}^{\blacklozenge})$ of a relevant pure inner form of $G_3$.
\end{rem}
\begin{proof} For each $x=n,s$, we first prove the existence of some $\epsilon_x \in \{\pm1\}$ and $\pi^{x}(\eta)\in \Pi^{R,\ \epsilon_x}_{\theta^{x}(\phi^{(1)},\phi^{(2)})}$ such that $$\Hom_{\Delta \U(W_3^{\epsilon_x})}(\pi^{x}(\eta), \omega_{\psi,W_3^{\epsilon_x}}) \ne 0.$$
For $a\in F^{\times}$, let $L_a$ be the 1-dimensional Hermitian space with form $a \cdot \N_{E/F}$. Then
\[ V_{2}^{+}/V_1^{+} \simeq V_{2}^{-}/V_1^{-} \simeq L_{-1}.
\]
We consider the following see-saw diagram : ($\epsilon, \epsilon'$ will be determined soon)
\begin{equation}\label{ss}
\xymatrix{
\U(W_3^{\epsilon}) \times \U(W_3^{\epsilon}) \ar@{-}[dr] \ar@{-}[d] & \U(V^{\epsilon'}_{2}) \ar@{-}[d] \\
\U(W_3^{\epsilon}) \ar@{-}[ur] & \U(V^{\epsilon'}_1) \times \U(L_{-1})}.
\end{equation}
In this diagram, we shall use three theta correspondences :
\begin{enumerate}
\item $U(V_2^{\epsilon'}) \times U(W_3^{\epsilon})$ relative to the pair of characters $(\chi^2, \chi^3)$;
\item $\U(V_1^{\epsilon'}) \times U(W_3^{\epsilon})$ relative to the pair of characters $(\chi, \chi^3)$;
\item $\U(L_{-1}) \times \U(W_3^{\epsilon})$ relative to the pair of characters $(\chi^{-1}, \chi^3)$.
\end{enumerate}
\noindent By $(B)_1$, there is a unique $\epsilon'\in \{\pm1\}$ and a unique pair of components characters $$ (\eta_2 ,\eta_1) \in \Irr^{\epsilon'}({S}_{(\phi^{(2)})^{\vee}}) \times \Irr^{\epsilon'}({S}_{(\phi^{(1)})^{\vee}})$$ such that $$\Hom_{\Delta U(V_1^{\epsilon'})}(\pi(\eta_2)\otimes \pi(\eta_1),\mathbb{C})\ne 0.$$
Moreover, $\epsilon'=\eta_1(a_1)=\epsilon(\frac{1}{2},(\phi^{(1)})^{\vee} \otimes (\phi^{(2)})^{\vee}, \psi^E_{-2})=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E})$.
By Theorem \ref{Tae} (i), (iv) and Theorem 4.1 in \cite{ag},$\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)$ is nonzero for any $\epsilon \in \{\pm1\}$. Since $\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) \boxtimes \Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)$ is the maximal $\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))$-isotypic quotient of $\omega_{\psi,V_2,W_3}$ and $\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) \boxtimes \pi(\eta_2)$ is a $\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))$-isotypic quotient of $\omega_{\psi,V_2,W_3}$, $\pi(\eta_2)$ should be a quotient of $\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)$. By Proposition 5.4 in \cite{ag}, $\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)$ is irreducible and thus we have $\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)=\pi(\eta_2)$.
Since $$\Hom_{\Delta U(V_1^{\epsilon'})}(\pi(\eta_2), \pi^{\vee}(\eta_1))\ne 0,$$ by the see-saw identity and Remark \ref{con}, we have $$\Hom_{\Delta U(W_3^{\epsilon})}\big(\Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \omega^{\vee}_{\psi,W_3^{\epsilon}}, \Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)\ne 0$$ and since
$\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))$ and $\omega^{\vee}_{\psi,W_3^{\epsilon}}$ are admissible, we have $$\Hom_{\Delta U(W_3^{\epsilon})}\big(\Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \Theta^{\vee}_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) , \omega_{\psi,W_3^{\epsilon}} \big)\ne 0.$$\\
By Theorem \ref{Te} (iii) and Theorem \ref{Tae} (i), the $L$-parameter of $ \Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \Theta^{\vee}_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) $ is \[ \begin{cases} \theta^{n}(\phi^{(1)},\phi^{(2)}) \quad \text{ if } \epsilon = \epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3}, \psi_2^{E})\cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E})\\ \theta^{s}(\phi^{(1)},\phi^{(2)}) \quad \text{ if } \epsilon = -\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3}, \psi_2^{E})\cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E}) \end{cases} \]
and by Theorem \ref{Te} (v), Theorem \ref{Tae} (iii), we see that their associated component characters are $\eta_{n}^{\blacklozenge}, \eta_{s}^{\blacklozenge}$ in each cases.
Next we shall prove that these are the unique representations which yield \emph{Fourier-Jacobi} model in each $L$-packets $\Pi_{\theta^{n}(\phi^{(1)},\phi^{(2)})}$ and $\Pi_{\theta^{s}(\phi^{(1)},\phi^{(2)})}$.\
\noindent Since $\theta^{s}(\phi^{(1)},\phi^{(2)})$ is tempered $L$-parameter, the uniqueness easily follows from (FJ)$_3$ in this case. Therefore, we shall only consider the non-tempered $L$-parameter $\theta^{n}(\phi^{(1)},\phi^{(2)})$.\\
Let $\pi_2 \otimes \pi_1 \in \Pi^{R, \ \epsilon}_{\theta^{n}(\phi^{(1)},\phi^{(2)})}=\Pi^{\epsilon}_{\theta^{n}(\phi^{(1)})}\times \Pi^{\epsilon}_{\theta(\phi^{(2)})}$ be a representation satisfying $$\Hom_{\Delta \U(W_3^{\epsilon})}(\pi_2 \otimes \pi_1, \omega_{\psi,W_3^{\epsilon}}) \ne 0$$ and in turn $$\Hom_{\Delta \U(W_3^{\epsilon})}(\pi_2 \otimes \omega^{\vee}_{\psi,W_3^{\epsilon}} , \pi^{\vee}_1)\ne0.$$
\noindent (The existence of such $\pi_2 \otimes \pi_1$ was ensured by the previous step.)\\
By Theorem \ref{Te} (iv), we can write $\pi_2=\Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{(1)})$ for some $\pi^{(1)}\in \Pi_{\theta^{(1)}}^{\epsilon'}$ where $$\epsilon'=\epsilon \cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3}, \psi_2^{E}).$$
Then by applying the see-saw duality in the see-saw diagram in (\ref{ss}), one has
$$\Hom_{\Delta \U(W_3^{\epsilon})}(\pi_2 \otimes \omega^{\vee}_{\psi,W_3^{\epsilon}} , \pi^{\vee}_1) \simeq \Hom_{U(V_1^{\epsilon'})}(\pi^{(2)},\pi^{(1)})\ne0$$ where $\pi^{(2)}=\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}(\pi^{\vee}_1)$. \\Note that $\pi^{(2)} \ne0$ and so it has tempered $L$-parameter $(\phi^{(2)})^{\vee}$. \\Then by the (B)$_1$, $(\pi^{(2)},\pi^{(1)})$ is the unique pair in the $L$-packet $\Pi_{(\phi^{(2)})^{\vee}} \times \Pi_{\phi^{(1)}}$ such that $$\Hom_{U(V_1^{\epsilon'})}(\pi^{(2)},\pi^{(1)})\ne0$$ and so $(\pi_2,\pi_1)$ should be $(\Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{(1)}),\Theta^{\vee}_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi^{(2)}))$.\\ This settles down the uniqueness issue.
\end{proof}
\begin{rem}When the $L$-parameter $\phi^{(2)}$ of $U(V_2^{\pm})$ contains $\chi^{-3}$, we can write $\phi^{(2)}=\phi_0 \oplus \chi^{-3}$ for a $L$-parameter $\phi_0$ of $U(V_1^{\pm})$. Then, we set $$\theta(\phi^{(2)})=\begin{cases}3\cdot \chi^{-2} \quad \quad \quad \quad \ \ \text{ if $\phi_0=\chi^{-3}$ }, \\ \phi_0 \cdot \chi\oplus 2 \cdot \chi^{-2} \quad \text{ if } \phi_0\ne \chi^{-3}\end{cases}$$ and $$S_{\theta(\phi^{(2)})}=\begin{cases}(\ZZ/2\ZZ)b_1 \ \quad \quad \quad \quad \quad \quad \text{ if $\phi_0=\chi^{-3}$ }, \\ (\ZZ/2\ZZ)b_1 \times (\ZZ/2\ZZ)c_1 \quad \text{ if } \phi_0\ne \chi^{-3}.\end{cases}$$
If one develops a similar argument in this case, one could also have a recipe as in Theorem \ref{mul} for non-tempered case. However, we need some assumption on the irreducibility of the theta lifts because we cannot apply Proposition 5.4 in \cite{ag}.
\end{rem}
\begin{thm}\label{mul} Let the notations be as in Thorem \ref{thm2} and assume this time that $\phi^{(2)}$ contains $\chi^{-3}$.\\
Assume that for $\pi \in \prod^{\epsilon}_{\theta((\phi^{(2)})^{\vee})}$, if $\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}(\pi)$ is nonzero, it is irreducible.
Then,
\[ \Hom_{\Delta \U(W_3^\epsilon)}(\pi^{n}(\eta), \omega_{\psi,W_3^{\epsilon}}) \ne 0 \Longleftrightarrow \eta = \eta_{n}^{\blacklozenge}\]
where $\eta_{n}^{\blacklozenge}\in\Irr(S_{\theta^{n}(\phi^{(1)},\phi^{(2)})}) = \Irr(S_{\theta^n(\phi^{(1)})}) \times \Irr(S_{\theta(\phi^{(2)})}) $ is specified as follows;
\begin{itemize}
\item When $\phi_0=\chi^{-3}$,
$$\begin{cases}\eta_{n}^{\blacklozenge}(a_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}) \\ \eta_{n}^{\blacklozenge}(b_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}). \end{cases}$$
\item When $\phi_0 \ne \chi^{-3}$,
$$\begin{cases}\eta_{n}^{\blacklozenge}(a_1)= \epsilon(\frac{1}{2}, \phi^{(1)} \otimes \phi_0,\psi_2^{E}) \\ \eta_{n}^{\blacklozenge}(b_1)=\epsilon(\frac{1}{2}, \phi^{(1)} \otimes \phi_0 ,\psi_2^{E}) \\ \eta_{n}^{\blacklozenge}(c_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3},\psi_2^{E}).\end{cases}$$
\end{itemize}
\end{thm}
\begin{proof}We first prove the existence of some $\epsilon \in \{\pm1\}$ and $\pi^{n}(\eta)\in \Pi^{R,\ \epsilon}_{\theta^{n}(\phi^{(1)},\phi^{(2)})}$ such that $$\Hom_{\Delta \U(W_3^{\epsilon})}(\pi^{n}(\eta), \omega_{\psi,W_3^{\epsilon}}) \ne 0.$$
By $(B)_1$, there is a unique $\epsilon'\in \{\pm1\}$ and a unique pair of components characters $$ (\eta_2 ,\eta_1) \in \Irr^{\epsilon'}({S}_{(\phi^{(2)})^{\vee}}) \times \Irr^{\epsilon'}({S}_{(\phi^{(1)})^{\vee}})$$ such that $$\Hom_{\Delta U(V_1^{\epsilon'})}(\pi(\eta_2)\otimes \pi(\eta_1),\mathbb{C})\ne 0.$$
Moreover, $\eta_2(b_1)=\epsilon((\phi^{(1)})^{\vee} \otimes (\phi_0)^{\vee} , \psi_{-2}^{E})=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi_0 , \psi_{2}^{E})$ and $$\epsilon'=\eta_1(a_1)=\epsilon(\frac{1}{2},(\phi^{(1)})^{\vee} \otimes (\phi^{(2)})^{\vee}, \psi^E_{-2})=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)}, \psi_2^{E}).$$
By Theorem \ref{Tae} (ii) \text{and} (iv), $\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))$ is a nonzero irreducible representation of $U(W_3^{\epsilon})$ for some $\epsilon \in \{\pm1\}$ and by Theorem 4.1 in \cite{ag}, $\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)$ is nonzero. So by our assumption, it is irreducible and we have $\Theta_{\psi,V_2^{\epsilon'},W_3^{\epsilon}}\big(\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)=\pi(\eta_2)$.
Since $$\Hom_{\Delta U(V_1^{\epsilon'})}(\pi(\eta_2), \pi^{\vee}(\eta_1))\ne 0,$$ by the see-saw identity and Remark \ref{con}, we have $$\Hom_{\Delta U(W_3^{\epsilon})}\big(\Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \omega^{\vee}_{\psi,W_3^{\epsilon}}, \Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))\big)\ne 0$$ and since
$\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))$ and $\omega^{\vee}_{\psi,W_3^{\epsilon}}$ are admissible, we have $$\Hom_{\Delta U(W_3^{\epsilon})}\big(\Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \Theta^{\vee}_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) , \omega_{\psi,W_3^{\epsilon}} \big)\ne 0.$$
Write $\Theta_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2))=\pi(\eta_3)$ for some $\eta_3 \in \prod_{\theta((\phi^{(2)})^{\vee})}$.\\ If $\phi_0=\chi^{-3}$, then $\eta_3(z_{\theta((\phi^{(2)})^{\vee})})=\eta_3(3\cdot b_1)=\eta_3(b_1)$ and if $\phi_0 \ne \chi^{-3}$, then $\eta_3(z_{\theta((\phi^{(2)})^{\vee})})=\eta_3(b_1)$.\\ Thus in both cases, we have $$\epsilon=\eta_3(z_{\theta((\phi^{(2)})^{\vee})})=\eta_3(b_1)=\eta_2(b_1)=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi_0, \psi_{2}^{E}).$$
Since $$\epsilon'=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi^{(2)} , \psi_{2}^{E})=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \phi_0 , \psi_{2}^{E}) \cdot \epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3} , \psi_{2}^{E}),$$ we have $\epsilon \cdot \epsilon'=\epsilon(\frac{1}{2},\phi^{(1)} \otimes \chi^{-3} , \psi_{2}^{E})$ and so by Theorem \ref{Te} (iii) and Theorem \ref{Tae} (i), the $L$-parameter of $ \Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \Theta^{\vee}_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) $ is $\theta^{n}(\phi^{(1)},\phi^{(2)})$.
Furthermore, by applying Theorem \ref{Te} (v), Theorem \ref{Tae} (iii), we see that the associated component character of $ \Theta_{\psi,W_3^{\epsilon},V_1^{\epsilon'}}(\pi^{\vee}(\eta_1)) \otimes \Theta^{\vee}_{\psi,W_3^{\epsilon},V_2^{\epsilon'}}(\pi(\eta_2)) $ is exactly $\eta_{n}^{\blacklozenge}$ in each cases and it proves the existence part. The proof of the uniqueness part is essentially same as the one in Theorem \ref{thm2}.
\end{proof}
\begin{rem}It is remarkable that for the supercuspidal $L$-parameter $\theta^{s}(\phi^{(1)},\phi^{(2)}) $ with $\theta(\phi^{(2)})$ as above, the recipe, which is sugegsted in (FJ)$_3$, does not occur from the theta lift from $U(V_1^{\pm})$ and $U(V_2^{\pm})$. This is quite similar with the Proposition 4.6 in \cite{Haan}, which concerns the non-generic aspect of \emph{Bessel} case of the GGP conjecture.
\end{rem}
|
\section{Introduction}
\label{Intro}
\setcounter{equation}{0}
The simplest form of the Cahn--Hilliard equation (see \cite{CahH, EllSt, EllSh}) reads as follows
\Begin{equation}
\partial_t y - \Delta w = 0
\quad\hbox{and}\quad
w = -\Delta y + f'(y)
\quad \hbox{in $\Omega\times (0,T) $},
\label{Icahil}
\End{equation}
where $\Omega$ is the domain where the evolution takes place,
and $y$~and $w$ denote the order parameter and the chemical potential, respectively.
Moreover, $f'$ represents the derivative of a double well potential~$f$,
and typical and important examples are the following
\Begin{eqnarray}
& f_{reg}(r) = \frac14(r^2-1)^2 \,,
\quad r \in {\mathbb{R}}
\label{regpot}
\\
& f_{log}(r) = ((1+r)\ln (1+r)+(1-r)\ln (1-r)) - c r^2 \,,
\quad r \in (-1,1),
\label{logpot}
\End{eqnarray}
where $c>0$ in the latter is large enough in order \pier{that $f_{log}$
be} nonconvex.
The potentials \eqref{regpot} and \eqref{logpot}
are usually called the classical regular potential
and the logarithmic double-well potential, respectively.
The present paper is devoted to the study of the control problem described below
for the initial--boundary value problem
obtained by complementing \eqref{Icahil} with an initial condition like $y(0)=y_0$
and the following boundary conditions
\Begin{equation}
\partial_n w = 0
\quad\hbox{and}\quad
\partial_n y + \partial_ty_\Gamma - \DeltaGy_\Gamma + f_\Gamma'(y_\Gamma) = u_\Gamma
\quad \hbox{on $\Gamma\times (0,T)$}
\label{Ibc}
\End{equation}
where $\Gamma$ is the boundary of~$\Omega$.
The former is very common in the literature and preserves mass conservation,
i.e., it implies that the space integral of $y$ is constant in time.
The latter is an evolution equation
for the trace $y_\Gamma$ of the order parameter on the boundary,
and the normal derivative $\partial_n y$ and $u_\Gamma$ act as forcing terms.
\pier{This condition enters the class of the so-called \gianni{dynamic boundary conditions
that have been widely used} in the literature in the last
twenty years, say: in particular, the study of dynamic boundary conditions
with Cahn--Hilliard type equations has been taken up by some
authors (let us quote \cite{CMZ, CGS, GiMiSchi, PRZ, RZ, WZ} and also refer to
the recent contribution \cite{CF} in which also a forced mass constraint
on the boundary is considered).}
\pier{The dynamic boundary condition in \eqref{Ibc} contains
the Laplace-Beltrami operator $\Delta_\Gamma$ and a
nonlinearity $f_\Gamma'$ which is analogous to $f'$
but is} now acting on the boundary values~$u_\Gamma$.
Even though some of our \juerg{results} hold under weaker hypotheses,
we assume from the very beginning that $f'$ and $f_\Gamma'$ have the same domain~$\calD$.
The main assumption we make is a compatibility condition between
\juerg{these} nonlinearities.
Namely, we suppose that $f_\Gamma'$ dominates $f'$ in the following sense:
\Begin{equation}
|f'(r)| \leq \eta \, |f_\Gamma'(r)| + C
\label{Icompatib}
\End{equation}
for some positive constants $\eta$ and~$C$
and for every $r\in\calD$.
\juerg{This} condition, earlier introduced in~\cite{CaCo}
in relation with the Allen--Cahn equation with dynamic boundary conditions \pier{(see also \cite{CS}),} is then used in~\cite{CGS} (as~well as in \cite{CFGS1} and~\cite{CGSvisc})
to deal with the Cahn--Hilliard system.
This complements~\cite{GiMiSchi},
where \juerg{some} kind of an opposite inequality is assumed.
As just said, this paper deals with a control problem for the state system
described above, the control being the source term $u_\Gamma$ that appears
in the dynamic boundary condition~\eqref{Ibc}.
Namely, the problem we want to address consists in minimi\sz ing
a proper cost functional depending on both the control $u_\Gamma$
and the associate state $(y,y_\Gamma)$.
Among several possibilities, we choose the cost functional
\Begin{equation}
\calJ(y,y_\Gamma,u_\Gamma)
:= \fracb_Q 2 \, \norma{y-z_Q}_{L^2(Q)}^2
+ \fracb_\Sigma 2 \, \norma{y_\Gamma-z_\Sigma}_{L^2(\Sigma)}^2
+ \fracb_0 2 \, \normau_\Gamma_{L^2(\Sigma)}^2\,,
\label{Icost}
\End{equation}
where the functions $z_Q, z_\Sigma$ and the nonnegative constants $b_Q,b_\Sigma,b_0$ are given.
The control problem then consists in minimi\sz ing \eqref{Icost}
subject to the state system and to the constraint $u_\Gamma\in\calU_{\rm ad}$,
where the control box $\calU_{\rm ad}$ is given~by
\Begin{eqnarray}
& \calU_{\rm ad} :=
& \bigl\{ u_\Gamma\in\H1H_\Gamma\cap\LS\infty:
\non
\\
&& \ u_{\Gamma,{\rm min}}\lequ_\Gamma\lequ_{\Gamma,{\rm max}}\ \checkmmode{a.e.\ on~$\Sigma$},\ \norma{\partial_tu_\Gamma}_{\LS2}\leqM_0
\bigr\}
\label{Iuad}
\End{eqnarray}
for some given functions $u_{\Gamma,{\rm min}},u_{\Gamma,{\rm max}}\in\LS\infty$
and some prescribed positive constant~$M_0$. \pier{Of course, the control box $\calU_{\rm ad}$ must be nonempty and this is guaranteed if, for instance, at least one of $u_{\Gamma,{\rm min}}$ or $u_{\Gamma,{\rm max}} $ is in $\H1H_\Gamma$} \juerg{and its time derivative satisfies the above $L^2(\Sigma)$ bound.}
This paper \pier{is a follow-up of} the recent
contributions \cite{CGS} and \cite{CGSvisc} already mentioned.
They deal with a similar system and a similar control problem.
\pier{The paper} \cite{CGS}~contains a number of results on the state system
obtained by considering
\Begin{equation}
w = \tau \, \partial_t y - \Delta y + f'(y)
\label{Ivisc}
\End{equation}
in place of the second \pier{condition in}~\eqref{Icahil}.
In~\eqref{Ivisc}, $\tau$~is a nonnegative parameter
and the case $\tau>0$ coupled with the first \pier{equation in}
\eqref{Icahil} yields the well-known viscous Cahn--Hilliard equation
(in~contrast, we term \eqref{Icahil} the pure Cahn--Hilliard system).
More precisely, existence, uniqueness and regularity results are proved in \cite{CGS}
for general potentials that include \accorpa{regpot}{logpot},
and are valid for both the viscous and pure cases, i.e., by assuming just $\tau\geq0$.
Moreover, if $\tau>0$, further regularity and properties of the solution are ensured.
\juerg{These} results are then used in \cite{CGSvisc}, where the boundary control problem
associated to a cost functional that generali\sz es \eqref{Icost}
is~addressed
and both the existence of an optimal control and \pier{first-order} necessary \pier{conditions} for optimality \pier{are proved and expressed}
in terms of the solution of a proper adjoint problem.
\pier{In fact, recently Cahn--Hilliard systems have been rather investigated from the viewpoint of optimal control. In this connection, we refer to \cite{HW1, WaNa, ZW} and to \cite{ZL1, ZL2} which deal with the convective Cahn--Hilliard equation;
the case with a nonlocal potential \pier{is} studied in \cite{RoSp}. There also exist contributions addressing some discretized versions of general Cahn--Hilliard systems, cf. \cite{HW2, Wang}. However, about the optimal control of viscous or non-viscous Cahn--Hilliard systems with dynamic boundary conditions of the form (\ref{Ibc}), we only know of the papers
\cite{CGSvisc} and \cite{CFGS1} dealing with the viscous case; to the best of our knowledge, the present contribution is the first paper treating
the optimal control of the pure Cahn--Hilliard system with dynamic boundary conditions.}
The technique used in \pier{our approach} essentially consists in
starting from the known results for $\tau>0$
and then letting the parameter $\tau$ tend to zero.
In doing that, we use some of the ideas of \cite{CFS} and \cite{CFGS1},
which deal with the Allen--Cahn and the viscous Cahn--Hilliard equations, respectively,
and address similar control problems related to the nondifferentiable
double obstacle potential by seeing it as a limit of logarithmic double-well potentials.
The paper is organized as follows.
In the next section, we list our assumptions, state the problem in a precise form
and present our results.
The corresponding proofs are given in the last section.
\section{Statement of the problem and results}
\label{STATEMENT}
\setcounter{equation}{0}
In this section, we describe the problem under study and give an outline of our results.
As in the Introduction,
$\Omega$~is the body where the evolution takes place.
We assume $\Omega\subset{\mathbb{R}}^3$
to~be open, bounded, connected, and smooth,
and we write $|\Omega|$ for its Lebesgue measure.
Moreover, $\Gamma$, $\partial_n$, $\nabla_{\!\Gamma}$ and $\Delta_\Gamma$ stand for
the boundary of~$\Omega$, the outward normal derivative, the surface gradient
and the Laplace--Beltrami operator, respectively.
Given a finite final time~$T>0$,
we set for convenience
\Begin{eqnarray}
&& Q_t := \Omega \times (0,t)
\quad\hbox{and}\quad
\Sigma_t := \Gamma \times (0,t)
\quad \hbox{for every $t\in(0,T]$}
\label{defQtSt}
\\
&& Q := Q_T \,,
\quad\hbox{and}\quad
\Sigma := \Sigma_T \,.
\label{defQS}
\End{eqnarray}
\Accorpa\defQeS defQtSt defQS
Now, we specify the assumptions on the structure of our system.
Even though some of the results we quote hold under rather mild hypotheses,
we give a list of assumptions that implies the whole set of conditions required in~\cite{CGS}.
We assume~that
\Begin{eqnarray}
& -\infty \leq r_- < 0 < r_+ \leq +\infty
\label{hpDf}\\[1mm]
& \hbox{$f,\,f_\Gamma:(r_-,r_+)\to[0,+\infty)$
are $C^3$ functions}
\label{hppot}
\\[1mm]
& f(0) = f_\Gamma(0) = 0
\quad\hbox{and}\quad
\hbox{$f''$ and $f_\Gamma''$ are bounded from below}
\label{hpfseconda}
\\[1mm]
\noalign{\allowbreak}
& |f'(r)| \leq \eta \,|f_\Gamma'(r)| + C
\quad \hbox{for some $\eta,\, C>0$ and every $r\in(r_-,r_+)$}
\label{hpcompatib}
\\[1mm]
\noalign{\allowbreak}
& \lim\limits_{r\mathrel{{\scriptscriptstyle\searrow}}r_-} f'(r)
= \lim\limits_{r\mathrel{{\scriptscriptstyle\searrow}}r_-} f_\Gamma'(r)
= -\infty
\quad\hbox{and}\quad
\lim\limits_{r\mathrel{{\scriptscriptstyle\nearrow}}r_+} f'(r)
= \lim\limits_{r\mathrel{{\scriptscriptstyle\nearrow}}r_+} f_\Gamma'(r)
= +\infty \, .
\label{fmaxmon}
\End{eqnarray}
\Accorpa\HPstruttura hpDf fmaxmon
We note that \HPstruttura\ imply the possibility
of splitting $f'$ as $f'=\beta+\pi$,
where $\beta$ is a monotone function that diverges at~$r_\pm$
and $\pi$ is a perturbation with a bounded derivative.
Moreover, the same is true for~$f_\Gamma$,
so that the assumptions of \cite{CGS} are satisfied.
Furthermore, the choices $f=f_{reg}$ and $f=f_{log}$
corresponding to \eqref{regpot} and \eqref{logpot} are allowed.
Next, in order to simplify notations, we~set
\Begin{eqnarray}
&& V := \Hx 1, \quad
H := \Lx 2, \quad
H_\Gamma := \LxG 2
\quad\hbox{and}\quad
V_\Gamma := \HxG 1
\qquad
\label{defVH}
\\
&& \calV := \graffe{(v,v_\Gamma)\in V\timesV_\Gamma:\ v_\Gamma=v{\vrule height 5pt depth 4pt\,}_\Gamma}
\quad\hbox{and}\quad
\calH := H \times H_\Gamma
\label{defcalVH}
\End{eqnarray}
\Accorpa\Defspazi defVH defcalVH
and endow these spaces with their natural norms.
If $X$ is any Banach space, \juerg{then} $\norma\,\cdot\,_X$~and $X^*$ denote
its norm and its dual space, respectively.
Furthermore, the symbol $\<\,\cdot\,,\,\cdot\,>$ usually stands
for the duality pairing between~$V^*$ and $V$ itself
and the similar notation $\<\,\cdot\,,\,\cdot\,>_\Gamma$ refers to the spaces $\VG^*$ and~$V_\Gamma$.
In the following, it is understood that $H$ is identified with $H^*$ and thus embedded in $V^*$
in the usual way, i.e., such that we have
$\<u,v>=(u,v)$ with the inner product $(\,\cdot\,,\,\cdot)$ of~$H$,
for every $u\in H$ and $v\in V$.
Thus, we introduce the Hilbert triplet $(V,H,V^*)$
and analogously behave with the boundary spaces $V_\Gamma$, $H_\Gamma$ and~$\VG^*$.
Finally, if $u\inV^*$ and $\underline u\in\L1V^*$,
we define their generali\sz ed mean values
$u^\Omega\in{\mathbb{R}}$ and $\underline u^\Omega\in L^1(0,T)$ by setting
\Begin{equation}
u^\Omega := \frac 1 {|\Omega|} \, \< u , 1 >
\quad\hbox{and}\quad
\underline u^\Omega(t) := \bigl( \underline u(t) \bigr)^\Omega
\quad \checkmmode{for a.a.~$t\in(0,T)$} .
\label{media}
\End{equation}
Clearly, \juerg{the relations in} \eqref{media} give the usual mean values when applied to elements
of~$H$ or $\L1H$.
At this point, we can describe the state problem.
For the data, we assume~that
\Begin{eqnarray}
&& y_0 \in \Hx 2
\quad\hbox{and}\quad
y_0{\vrule height 5pt depth 4pt\,}_\Gamma \in \HxG 2
\label{hpyz}
\\
&& r_- < y_0(x) < r_+
\quad \hbox{for every $x\in\overline\Omega$}
\label{hpyzbis}
\\
&& u_\Gamma \in \H1H_\Gamma \,.
\label{hpuG}
\End{eqnarray}
\Accorpa\HPdati hpyz hpuG
We look for a triplet $(y,y_\Gamma,w)$ satisfying
\Begin{eqnarray}
& y \in \H1V^* \cap \spazio L\infty V \cap \L2\Hx 2
\label{regy}
\\
& y_\Gamma \in \H1H_\Gamma \cap \spazio L\inftyV_\Gamma \cap \L2\HxG 2
\label{regyG}
\\
& y_\Gamma(t) = y(t){\vrule height 5pt depth 4pt\,}_\Gamma
\quad \checkmmode{for a.a.~$t\in(0,T)$}
\label{tracciay}
\\
& w \in \L2V\,,
\label{regw}
\End{eqnarray}
\Accorpa\Regsoluz regy regw
as well as, for almost every $t\in (0,T)$, the variational equations
\Begin{eqnarray}
&& \< \partial_t y(t) \, v >
+ \int_\Omega \nabla w(t) \cdot \nabla v = 0
\qquad \hbox{for every $v\in V$}
\label{prima}
\\
\noalign{\smallskip}
&& \int_\Omega w(t) \, v
= \int_\Omega \nabla y(t) \cdot \nabla v
+ \int_\Gamma \partial_ty_\Gamma(t) \, v_\Gamma
+ \int_\Gamma \nablaGy_\Gamma(t) \cdot \nablaGv_\Gamma
\qquad
\non
\\
&& \quad {}
+ \int_\Omega f'(y(t)) \, v
+ \int_\Gamma \bigl( f_\Gamma'(y_\Gamma(t)) - u_\Gamma(t) \bigr) \, v_\Gamma
\qquad \hbox{for every $(v,v_\Gamma)\in\calV$}
\qquad
\label{seconda}
\\
&& y(0) = y_0 \,. \vphantom\sum
\label{cauchy}
\End{eqnarray}
\Accorpa\State prima cauchy
\Accorpa\Pbl regy cauchy
Thus, we require that the state variables satisfy the variational counterpart
of the problem described in the Introduction in a strong form.
We note that an equivalent formulation of \accorpa{prima}{seconda}
is given~by
\Begin{eqnarray}
&& \int_0^t \< \partial_t y(t) \, v(t) > \, dt
+ \int_Q \nabla w \cdot \nabla v = 0
\label{intprima}
\\
\noalign{\smallskip}
&& \int_Q wv
= \int_Q \nabla y \cdot \nabla v
+ \int_\Sigma \partial_ty_\Gamma \, v_\Gamma
+ \int_\Sigma \nablaGy_\Gamma \cdot \nablaGv_\Gamma
\qquad
\non
\\
&& \quad {}
+ \int_Q f'(y) \, v
+ \int_\Sigma \bigl( f_\Gamma'(y_\Gamma) - u_\Gamma \bigr) \, v_\Gamma
\qquad
\label{intseconda}
\End{eqnarray}
\Accorpa\IntPbl intprima intseconda
for every $v\in\L2 V$ and every $(v,v_\Gamma)\in\L2\calV$, respectively.
Besides, we consider the analogous state system with viscosity.
Namely, for $\tau>0$ we replace \eqref{seconda}~by
\Begin{eqnarray}
&& \int_\Omega w(t) \, v
= \tau \int_\Omega \partial_t y(t) \, v
+ \int_\Omega \nabla y(t) \cdot \nabla v
+ \int_\Gamma \partial_ty_\Gamma(t) \, v_\Gamma
+ \int_\Gamma \nablaGy_\Gamma(t) \cdot \nablaGv_\Gamma
\qquad
\non
\\
&& \quad {}
+ \int_\Omega f'(y(t)) \, v
+ \int_\Gamma \bigl( f_\Gamma'(y_\Gamma(t)) - u_\Gamma(t) \bigr) \, v_\Gamma
\qquad \hbox{for every $(v,v_\Gamma)\in\calV$}
\qquad
\label{secondavisc}
\End{eqnarray}
\def\Pbltau{\eqref{prima}, \eqref{cauchy} and~\eqref{secondavisc}}%
in the above system.
We notice that a variational equation equivalent to \eqref{secondavisc}
is given by the analogue of \eqref{intseconda}, i.e.,
\Begin{eqnarray}
&& \int_Q wv
= \tau \int_Q \partial_t y \, v
+ \int_Q \nabla y \cdot \nabla v
+ \int_\Sigma \partial_ty_\Gamma \, v_\Gamma
+ \int_\Sigma \nablaGy_\Gamma \cdot \nablaGv_\Gamma
\qquad
\non
\\
&& \quad {}
+ \int_Q f'(y) \, v
+ \int_\Sigma \bigl( f_\Gamma'(y_\Gamma) - u_\Gamma \bigr) \, v_\Gamma
\quad
\hbox{for every $(v,v_\Gamma)\in\L2\calV$}.
\qquad
\label{intsecondavisc}
\End{eqnarray}
As far as existence, uniqueness, regularity and continuous dependence are concerned,
we directly refer to~\cite{CGS}.
From \cite[Thms.~2.2 and~2.3]{CGS}
(where $\calV$ has a slightly different meaning with respect to the present paper),
we~have the following results:
\Begin{theorem}
\label{daCGSpure}
Assume \HPstruttura\ and \HPdati.
Then, there exists a unique triplet $(y,y_\Gamma,w)$ satisfying
\Regsoluz\ and solving \State.
\End{theorem}
\Begin{theorem}
\label{daCGS}
Assume \HPstruttura\ and \HPdati.
Then, for every $\tau>0$,
there exists a unique triplet $(y^\tau,y^\tau_\Gamma,w^\tau)$ satisfying
\Regsoluz\ and solving \Pbltau.
Moreover, \juerg{this} solution satisfies
$\partial_ty^\tau\in\L2H$
and the estimate
\Begin{eqnarray}
&& \normay^\tau_{\H1V^* \cap \spazio L\infty V \cap \L2\Hx 2}
\non
\\[1mm]
&& \quad {}
+ \normay^\tau_\Gamma_{\H1H_\Gamma \cap \spazio L\inftyV_\Gamma \cap \L2\HxG 2}
\non
\\[1mm]
&& \quad {}
+ \normaw^\tau_{\pier{\L2V}}
+ \norma{f'(y^\tau)}_{\L2H}
+ \norma{f_\Gamma'(y^\tau_\Gamma)}_{\L2H_\Gamma}
\non
\\[1mm]
&& \quad {}
+ \tau^{1/2} \norma{\partial_ty^\tau}_{\L2H}
\leq \, C_0
\label{stab}
\End{eqnarray}
holds true for some constant $C_0>0$ that depends only on
$\Omega$, $T$, the shape of the nonlinearities $f$ and~$f_\Gamma$,
\pier{and the norms~$\bigl\Vert (y_0,{y_0}_{|_\Gamma})\bigr\Vert_{\calV}$,
$\Vert f' (y_0)\Vert_{L^1(\Omega)}$,
$\bigl\Vert f_\Gamma' \bigl({y_0}_{|_\Gamma}\bigr) \bigr\Vert_{L^1(\Gamma)}$,
and $\norma{u_\Gamma}_{\L2H_\Gamma}.$}
\End{theorem}
In fact, \pier{if the data are more regular, in particular\juerg{, if }
$u_\Gamma \in \H1H_\Gamma \cap\LS\infty $,}
\juerg{then} the solution $(y^\tau,y^\tau_\Gamma,w^\tau)$ is even smoother
(see \cite[Thms.~2.4 and~2.6]{CGS}) \pier{and, specifically,}
it satisfies
\Begin{equation}
r_-^\tau \leq y^\tau \leq r_+^\tau
\quad \checkmmode{a.e.\ in~$Q$}
\label{faraway}
\End{equation}
for some constants $r_-^\tau,r_+^\tau\in(r_-,r_+)$
that depend on~$\tau$, in addition.
It \juerg{follows} that the functions
$f''(y^\tau)$ and $f_\Gamma''(y^\tau_\Gamma)$
(which will appear as coefficients in a linear system later~on)
are~bounded.
We also notice that the stability estimate \eqref{stab} is not explicitely written in~\cite{CGS}.
However, as the proof of the regularity \Regsoluz\ of the solution
performed there relies on a~priori estimates and compactness arguments
and the dependence on $\tau$ in the whole calculation of~\cite{CGS} is always made explicit,
\eqref{stab}~holds as well,
and we stress that the corresponding constant $C_0$ does not depend on~$\tau$.
Once well-posedness for problem \State\ is established,
we can address the corresponding control problem.
As in the Introduction, given two functions
\Begin{equation}
z_Q \in \LQ2
\quad\hbox{and}\quad
z_\Sigma \in \LS2
\label{hpzz}
\End{equation}
and \pier{three} nonnegative constants $b_Q,\, b_\Sigma,\, b_0$,
we~set
\Begin{equation}
\calJ(y,y_\Gamma,u_\Gamma)
:= \fracb_Q 2 \, \norma{y-z_Q}_{L^2(Q)}^2
+ \fracb_\Sigma 2 \, \norma{y_\Gamma-z_\Sigma}_{L^2(\Sigma)}^2
+ \fracb_0 2 \, \normau_\Gamma_{L^2(\Sigma)}^2
\label{defcost}
\End{equation}
for, say, $y\in\L2H$, $y_\Gamma\in\L2H_\Gamma$ and $u_\Gamma\in\LS2$,
and consider the problem of minimi\sz ing the cost functional \eqref{defcost}
subject to the constraint $u_\Gamma\in\calU_{\rm ad}$,
where the control box $\calU_{\rm ad}$ is given~by
\Begin{eqnarray}
& \calU_{\rm ad} :=
& \bigl\{ u_\Gamma\in\H1H_\Gamma\cap\LS\infty:
\non
\\
&& \ u_{\Gamma,{\rm min}}\lequ_\Gamma\lequ_{\Gamma,{\rm max}}\ \checkmmode{a.e.\ on~$\Sigma$},\ \norma{\partial_tu_\Gamma}_{\LS2}\leqM_0
\bigr\}
\label{defUad}
\End{eqnarray}
and to the state system \State.
We simply assume~that
\Begin{equation}
M_0 > 0 , \quad
u_{\Gamma,{\rm min}} ,\, u_{\Gamma,{\rm max}} \in \LS\infty
\quad\hbox{and}\quad
\hbox{$\calU_{\rm ad}$ is nonempty}.
\label{hpUad}
\End{equation}
Besides, we consider the analogous control problem
of minimi\sz ing the cost functional \eqref{defcost}
subject to the constraint $u_\Gamma\in\calU_{\rm ad}$
and to the state system \Pbltau.
From \cite[Thm.~\pier{2.3}]{CGSvisc}, we~have
\juerg{the following result.}
\Begin{theorem}
\label{Optimumvisc}
Assume \HPstruttura\ and \HPdati,
and let $\calJ$ and $\calU_{\rm ad}$ be defined by \eqref{defcost} and \eqref{defUad}
under the assumptions \eqref{hpzz} and~\eqref{hpUad}.
Then, for every $\tau>0$, there exists $\overline u^{\,\pier{\tau}}_\Gamma\in\calU_{\rm ad}$ such~that
\Begin{equation}
\calJ(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline u^{\,\pier{\tau}}_\Gamma)
\leq \calJ(y^\tau,y^\tau_\Gamma,u_\Gamma)
\quad \hbox{for every $u_\Gamma\in\calU_{\rm ad}$}\,,
\label{optimumvisc}
\End{equation}
where $\overline y^{\,\pier{\tau}}$, $\overline y^{\,\pier{\tau}}_\Gamma$, $y^\tau$ and $y^\tau_\Gamma$
are the components of the solutions $(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline w^{\,\pier{\tau}})$ and $(y^\tau,y^\tau_\Gamma,w^\tau)$
to~the state system \Pbltau\
corresponding to the controls $\overline u^{\,\pier{\tau}}_\Gamma$ and~$u_\Gamma$, respectively.
\End{theorem}
In~\cite{CGSvisc} \pier{first-order} necessary conditions are obtained
in terms of the solution to a proper adjoint system.
More precisely, just the case $\tau=1$ is considered there.
However, by going through the paper with some care,
one easily reconstructs the version of the adjoint system corresponding
to an arbitrary~$\tau>0$.
Even though the adjoint problem considered in~\cite{CGSvisc}
involves a triplet $(p^\tau,q^\tau,q^\tau_\Gamma)$ as an adjoint state,
only the third component $q^\tau_\Gamma$ enters the necessary condition for optimality.
On the other hand, $q^\tau$~and $q^\tau_\Gamma$ are strictly related to each other.
Hence, we mention the result that deals with the pair~$(q^\tau,q^\tau_\Gamma)$.
To this end, we recall a tool, the generali\sz ed Neumann problem solver~$\calN$,
that is often used in connection with the Cahn--Hilliard equations.
With the notation for the mean value introduced in~\eqref{media}, we define
\Begin{equation}
\dom\calN := \graffe{v_*\inV^*: \ v_*^\Omega = 0}
\quad\hbox{and}\quad
\calN : \dom\calN \to \graffe{v \in V : \ v^\Omega = 0}
\label{predefN}
\End{equation}
by setting, for $v_*\in\dom\calN$,
\Begin{equation}
\calNv_* \in V, \quad
(\calNv_*)^\Omega = 0 ,
\quad\hbox{and}\quad
\int_\Omega \nabla\calNv_* \cdot \nabla z = \< v_* , z >
\quad \hbox{for every $z\in V$} .
\label{defN}
\End{equation}
Thus, $\calNv_*$ is the solution $v$ to the generali\sz ed Neumann problem for $-\Delta$
with datum~$v_*$ that satisfies~$v^\Omega=0$.
Indeed, if $v_*\in H$, the above variational equation means \juerg{that}
$-\Delta\calNv_* = v_*$ and $\partial_n\calNv_* = 0$.
As $\Omega$ is bounded, smooth, and connected,
it turns out that \eqref{defN} yields a well-defined isomorphism.
Moreover, we have
\Begin{equation}
\< u_* , \calN v_* >
= \< v_* , \calN u_* >
= \int_\Omega (\nabla\calNu_*) \cdot (\nabla\calNv_*)
\quad \hbox{for $u_*,v_*\in\dom\calN$},
\label{simmN}
\End{equation}
whence also
\Begin{equation}
2 \< \partial_tv_*(t) , \calNv_*(t) >
= \frac d{dt} \int_\Omega |\nabla\calNv_*(t)|^2
= \frac d{dt} \, \normaVp{v_*(t)}^2
\quad \checkmmode{for a.a.~$t\in(0,T)$}
\label{dtcalN}
\End{equation}
for every $v_*\in\H1V^*$ satisfying $(v_*)^\Omega=0$ \checkmmode{a.e.\ in~$(0,T)$},
where we have set
\Begin{equation}
\normaVpv_* := \normaH{\nabla\calNv_*}
\quad \hbox{for $v_*\inV^*$} .
\label{defnormaVp}
\End{equation}
One easily sees that $\normaVp\,\cdot\,$ is a norm in $V^*$
\juerg{which is}
equivalent to the usual dual norm.
Furthermore, we introduce the spaces $\calH_\Omega$ and $\calV_\Omega$ \juerg{by setting}
\Begin{equation}
\calH_\Omega := \graffe{(v,v_\Gamma) \in \calH :\ v^\Omega=0}
\quad\hbox{and}\quad
\calV_\Omega := \calH_\Omega \cap \calV\,,
\label{defHOVO}
\End{equation}
and endow them with their natural topologies as subspaces of $\calH$ and~$\calV$, respectively.
As in \cite[Thms.~\pier{2.5} and~5.4]{CGSvisc}, we~have \juerg{the following result.}
\Begin{theorem}
\label{Existenceadj}
Assume
\Begin{equation}
\lam \in \LQ\infty , \quad
\lam_\Gamma \in \LS\infty , \quad
\phi_Q \in \LQ2
\quad\hbox{and}\quad
\phi_\Sigma \in \LS2 .
\label{genlam}
\End{equation}
Then, for every $\tau>0$, there exists a unique pair $(q^\tau,q^\tau_\Gamma)$
satisfying the regularity conditions
\Begin{equation}
q^\tau \in \H1H \cap \L2{\Hx2}
\quad\hbox{and}\quad
q^\tau_\Gamma \in \H1H_\Gamma \cap \L2{\HxG2}
\label{regqqG}
\End{equation}
and solving the following adjoint problem:
\Begin{eqnarray}
&& (q^\tau,q^\tau_\Gamma)(t) \in \calV_\Omega
\quad \hbox{for every $t\in[0,T]$} \vphantom{\frac 1{|\Omega|}}
\label{primaN}
\\
&& - \int_\Omega \partial_t \bigl( \calN(q^\tau(t)) + \tau q^\tau(t) \bigr) \, v
+ \int_\Omega \nablaq^\tau(t) \cdot \nabla v
+ \int_\Omega \lam(t) \, q^\tau(t) \, v
\qquad
\non
\\
&& \qquad {}
- \int_\Gamma \partial_tq^\tau_\Gamma(t) \, v_\Gamma
+ \int_\Gamma \nablaGq^\tau_\Gamma(t) \cdot \nablaGv_\Gamma
+ \int_\Gamma \lam_\Gamma(t) \, q^\tau_\Gamma(t) \, v_\Gamma
\non
\\
&& = \int_\Omega \phi_Q(t) v
+ \int_\Gamma \phi_\Sigma(t) v_\Gamma
\quad \hbox{\checkmmode{for a.a.~$t\in(0,T)$}\ and every $(v,v_\Gamma)\in\calV_\Omega$}
\qquad
\label{secondaN}
\\
&& \int_\Omega \bigl( \calNq^\tau + \tau q^\tau \bigr)(T) \, v
+ \int_\Gamma q_\Gamma(T) \, v_\Gamma
= 0
\quad \hbox{for every $(v,v_\Gamma)\in\calV_\Omega$} \,.
\label{cauchyN}
\End{eqnarray}
\Accorpa\AggiuntoN primaN cauchyN
\End{theorem}
More precisely, in \cite{CGSvisc}
the above theorem is proved with the particular choice
\Begin{equation}
\lam = f''(\overline y^{\,\pier{\tau}}) , \quad
\lam_\Gamma = f_\Gamma''(\overline y^{\,\pier{\tau}}_\Gamma) , \quad
\phi_Q = b_Q (\overline y^{\,\pier{\tau}}-z_Q)
\quad\hbox{and}\quad
\phi_\Sigma = b_\Sigma (\overline y^{\,\pier{\tau}}_\Gamma-z_\Sigma)
\label{scelteCGS}
\End{equation}
where $\overline y^{\,\pier{\tau}}$ and $\overline y^{\,\pier{\tau}}_\Gamma$ are the components of the state
associated to an optimal control~$\overline u^{\,\pier{\tau}}_\Gamma$.
However, the same proof is valid under assumption~\eqref{genlam}.
Finally, \cite[Thm.~\pier{2.6}]{CGSvisc} gives a necessary condition
for $\overline u^{\,\pier{\tau}}_\Gamma$ to be an optimal control
in terms of the solution to the above adjoint system
corresponding to~\eqref{scelteCGS}.
\juerg{This} condition reads
\Begin{equation}
\int_\Sigma (q^\tau_\Gamma + b_0 \overline u^{\,\pier{\tau}}_\Gamma) (v_\Gamma - \overline u^{\,\pier{\tau}}_\Gamma) \geq 0
\quad \hbox{for every $v_\Gamma\in\calU_{\rm ad}$}.
\label{cnoptAV}
\End{equation}
In this paper, we first show the existence of an optimal control~$\overline u_\Gamma$.
Namely, we prove the following result.
\Begin{theorem}
\label{Optimum}
Assume \HPstruttura\ and \HPdati,
and let $\calJ$ and $\calU_{\rm ad}$ be defined by \eqref{defcost} and \eqref{defUad}
under the assumptions \eqref{hpzz} and~\eqref{hpUad}.
Then there exists \juerg{some} $\overline u_\Gamma\in\calU_{\rm ad}$ such~that
\Begin{equation}
\calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma)
\leq \calJ(y,y_\Gamma,u_\Gamma)
\quad \hbox{for every $u_\Gamma\in\calU_{\rm ad}$}\,,
\label{optimum}
\End{equation}
where $\overline y$, $\overline y_\Gamma$, $y$ and $y_\Gamma$
are the components of the solutions $(\overline y,\overline y_\Gamma,\overline w)$ and $(y,y_\Gamma,w)$
to~the state system \State\
corresponding to the controls $\overline u_\Gamma$ and~$u_\Gamma$, respectively.
\End{theorem}
Next, for every optimal control~$\overline u_\Gamma$,
we derive a necessary optimality condition like \eqref{cnoptAV}
in terms of the solution of a generali\sz ed adjoint system.
In order to make the last sentence precise,
we introduce the spaces
\Begin{eqnarray}
&& \calW := \L2\calV_\Omega \cap \bigl( \H1V^* \times \H1\VG^* \bigr)
\label{defZ}
\\
&& \calW_0 := \graffe{ (v,v_\Gamma) \in \calW :\ (v,v_\Gamma)(0) = (0,0) }
\qquad
\label{defZz}
\End{eqnarray}
and endow them with their natural topologies.
Moreover, we denote by $\[\,\cdot\,,\,\cdot\,]$
the duality product between $\calW_0^*$ and~$\calW_0$.
We will prove the following representation result for the elements
of the dual space~$\calW_0^*$.
\Begin{proposition}
\label{ReprWp}
A functional $F:\calW_0\to{\mathbb{R}}$ belongs to $\calW_0^*$ if and only if
there exist $\Lam$ and $\Lam_\Gamma$ satisfying
\Begin{eqnarray}
&& \Lam \in \bigl( \H1V^* \cap \L2V \bigr)^*
\quad\hbox{and}\quad
\Lam_\Gamma \in \bigl( \H1\VG^* \cap \L2V_\Gamma \bigr)^*
\qquad\quad
\label{goodLam}
\\
&& \[ F , (v,v_\Gamma) ]
= \< \Lam , v >_Q
+ \< \Lam_\Gamma , v_\Gamma >_\Sigma
\quad \hbox{for every $(v,v_\Gamma)\in\calW_0$}\,,
\label{reprWp}
\End{eqnarray}
where the duality products $\<\,\cdot\,,\,\cdot\,>_Q$ and $\<\,\cdot\,,\,\cdot\,>_\Sigma$
are related to the spaces $X^*$ and $X$ with
$X=\H1V^* \cap \L2V$ and $X=\H1\VG^* \cap \L2V_\Gamma$, respectively.
\End{proposition}
However, \juerg{this} representation is not unique,
since different pairs $(\Lam,\Lam_\Gamma)$ satisfying \eqref{goodLam}
could \juerg{generate} the same functional $F$ through formula~\eqref{reprWp}.
At this point, we are ready to present our result
on the necessary optimality conditions
for the control problem related to the pure Cahn--Hilliard equations, i.e.,
the analogue of \eqref{cnoptAV}
in terms of a solution to a generali\sz ed adjoint system.
\Begin{theorem}
\label{CNopt}
Assume \HPstruttura\ and \HPdati,
and let $\calJ$ and $\calU_{\rm ad}$ be defined by \eqref{defcost} and \eqref{defUad}
under the assumptions \eqref{hpzz} and~\eqref{hpUad}.
Moreover, let $\overline u_\Gamma$ be any optimal control
as in the statement of Theorem~\ref{Optimum}.
Then, there exist $\Lam$ and $\Lam_\Gamma$ satisfying \eqref{goodLam},
and a pair $(q,q_\Gamma)$ satisfying
\Begin{eqnarray}
&& q \in \spazio L\inftyV^* \cap \L2V
\label{regq}
\\
&& q_\Gamma \in \spazio L\inftyH_\Gamma \cap \L2V_\Gamma
\label{regqG}
\\
&& (q,q_\Gamma)(t) \in \calV_\Omega
\quad \hbox{for every $t\in[0,T]$}\,,
\label{primazero}
\End{eqnarray}
as well as
\Begin{eqnarray}
&& \int_0^T \< \partial_t v(t) , \calN q(t) > \, dt
+ \int_0^T \< \partial_tv_\Gamma(t) , q_\Gamma(t) >_\Gamma \, dt
\non
\\
&& \quad {}
+ \int_Q \nabla q \cdot \nabla v
+ \int_\Sigma \nablaGq_\Gamma \cdot \nablaGv_\Gamma
+ \< \Lam , v >_Q
+ \< \Lam_\Gamma , v_\Gamma >_\Sigma
\qquad
\non
\\
&& = \int_Q \phi_Q \, v
+ \int_\Sigma \phi_\Sigma \, v_\Gamma
\qquad \hbox{for every $(v,v_\Gamma)\in\calW_0$}\,,
\label{secondazero}
\End{eqnarray}
such that
\Begin{equation}
\int_\Sigma (q_\Gamma + b_0 \overline u_\Gamma) (v_\Gamma - \overline u_\Gamma) \geq 0
\quad \hbox{for every $v_\Gamma\in\calU_{\rm ad}$}.
\label{cnopt}
\End{equation}
In particular, if $b_0>0$, \juerg{then} the optimal control $\overline u_\Gamma$
is the $\LS2$-projection of $-q_\Gamma/b_0$ \juerg{onto} $\calU_{\rm ad}$.
\End{theorem}
One recogni\sz es in \eqref{secondazero}
a problem that is analogous to \accorpa{secondaN}{cauchyN}.
Indeed, if $\Lam$, $\Lam_\Gamma$ and the solution $(q,q_\Gamma)$ were regular functions,
\juerg{then its strong form} should contain
both a generali\sz ed \pier{backward} parabolic equation like \eqref{secondaN}
and a final condition for $(\calN q,q_\Gamma)$ of type \eqref{cauchyN},
since the definition of $\calW_0$ allows its elements to be free at $t=T$.
However, the terms $\lamtq^\tau$ and $\lamtGq^\tau_\Gamma$ are just replaced
by the functionals $\Lam$ and~$\Lam_\Gamma$ and
\juerg{cannot be} identified as products, unfortunately.
\section{Proofs}
\label{PROOFS}
\setcounter{equation}{0}
In the whole section, we assume that all \juerg{of} the conditions
\HPstruttura\ and \accorpa{hpyz}{hpyzbis}
on the structure and the initial datum of the state system,
as well as assumptions \eqref{hpzz} and~\eqref{hpUad}
that regard the cost functional \eqref{defcost} and the control box \eqref{defUad},
are satisfied.
We start with an expected result.
\Begin{proposition}
\label{Convergence}
Assume $u^\tau_\Gamma\in\H1H_\Gamma$ and let $(y^\tau,y^\tau_\Gamma,w^\tau)$
be the solution to \pier{the} problem \Pbltau\ associated to~$u^\tau_\Gamma$.
If $u^\tau_\Gamma$ converges to $u_\Gamma$ \pier{weakly} in $\H1H_\Gamma$ as $\tau\seto0$, then
\Begin{eqnarray}
& y^\tau \to y
& \quad \hbox{weakly star in $\H1V^*\cap\spazio L\infty V\cap\L2\Hx 2$}
\non
\\
&& \qquad \hbox{and strongly in \pier{$\C0H \cap \L2V$}}
\label{convyt}
\\
& y^\tau_\Gamma \to y_\Gamma
& \quad \hbox{weakly star in $\H1H_\Gamma\cap\spazio L\inftyV_\Gamma\cap\L2\HxG 2$}
\non
\\
&& \qquad \hbox{and strongly in \pier{$\C0H_\Gamma \cap \L2V_\Gamma$}}
\label{convytG}
\\
& w^\tau \to w
& \quad \hbox{weakly star in $\pier{\L2V}$}\,,
\label{convwt}
\End{eqnarray}
where $(y,y_\Gamma,w)$ is the solution to problem \State\
associated \juerg{with} $u_\Gamma$.
\End{proposition}
\Begin{proof}
The family $\graffe{u^\tau_\Gamma}$ is bounded in $\H1H_\Gamma$.
Thus, the solution $(y^\tau,y^\tau_\Gamma,w^\tau)$ satisfies~\eqref{stab}
for some constant~$C_0$, so that the weak or weak star convergence
specified in \accorpa{convyt}{convwt} holds for a subsequence.
In particular, the Cauchy condition \eqref{cauchy} for $y$ is satisfied.
Moreover, we also have
$\tau\,\partial_ty^\tau\to0$ strongly in $\L2H$ as well as
$f'(y^\tau)\to\xi$ and $f_\Gamma'(y^\tau_\Gamma)\to\xi_\Gamma$
weakly in $\L2H$ and in $\L2H_\Gamma$, respectively,
for some $\xi$ and $\xi_\Gamma$.
Furthermore, $y^\tau$~and $y^\tau_\Gamma$ converge to their limits strongly
in $\L2H$ and $\L2H_\Gamma$, respectively, thanks to the Aubin-Lions lemma
(see, e.g., \cite[Thm.~5.1, p.~58]{Lions}, which also implies a much better strong convergence \cite[Sect.~8, Cor.~4]{Simon}).
Now, as said in Section~\ref{STATEMENT},
we can split $f'$ as $f'=\beta+\pi$,
where $\beta$ is monotone \juerg{and} $\pi$ is Lip\-schitz\ continuous.
It follows that $\pi(y^\tau)$ converges to $\pi(y)$ strongly in $\L2H$,
whence \juerg{we obtain that also} $\beta(y^\tau)$ converges to $\xi-\pi(y)$ weakly in~$\L2H$.
Then, we infer that $\xi-\pi(y)=\beta(y)$ \checkmmode{a.e.\ in~$Q$}, i.e., $\xi=f'(y)$ \checkmmode{a.e.\ in~$Q$},
with the help of standard monotonicity arguments
(see, e.g., \cite[Lemma~1.3, p.~42]{Barbu}).
Similarly, we have $\xi_\Gamma=f_\Gamma'(y_\Gamma)$.
Therefore, by starting from \eqref{intprima} and \eqref{intsecondavisc}
written with $u^\tau_\Gamma$ in place of~$u_\Gamma$,
we can pass to the limit and obtain \IntPbl\
associated to the limit control~$u_\Gamma$.
As the solution to the limit problem is unique,
the whole family $(y^\tau,y^\tau_\Gamma,w^\tau)$ converges to $(y,y_\Gamma,w)$ in the sense of the statement
and the proof is complete.
\End{proof}
\Begin{corollary}
\label{CorConvergence}
Estimate \eqref{stab}, \juerg{written formally} with $\tau=0$,
holds for the solution to the pure Cahn--Hilliard system \State.
\End{corollary}
\Begin{proof}
By applying the above proposition with $u^\tau_\Gamma=u_\Gamma$
and using \eqref{stab} for the solution to the viscous problem,
we immediately conclude \juerg{the claim}.
\End{proof}
\step
Proof of Theorem~\ref{Optimum}
We use the direct method and start from a minimi\sz ing sequence $\graffe{u_{\Gamma,n}}$.
Then, $u_{\Gamma,n}$~remains bounded in $\H1H$, whence we have $u_{\Gamma,n}\to\overline u_\Gamma$ weakly in $\H1H_\Gamma$ for a subsequence.
By Corollary~\ref{CorConvergence},
the sequence of the corresponding states $(y_n,y_{\Gamma,n},w_n)$ satisfies the analogue of~\eqref{stab}.
Hence, by arguing as in the proof of Proposition~\ref{Convergence},
we infer that the solutions $(y_n,y_{\Gamma,n},w_n)$ converge in the proper topology
to the solution $(\overline y,\overline y_\Gamma,\overline w)$ associated to~$\overline u_\Gamma$.
In particular, there holds the strong convergence
specified by the analogues of \eqref{convyt} and~\eqref{convytG}.
Thus, by also owing to \juerg{the semicontinuity of $\calJ$ and the optimality of}
$u_{\Gamma,n}$, we have
\Begin{equation}
\calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma) \leq \liminf_{n\to\infty} \calJ(y_n,y_{\Gamma,n},u_{\Gamma,n})
\leq \calJ(y,y_\Gamma,u_\Gamma)
\non
\End{equation}
for every $u_\Gamma\in\calU_{\rm ad}$, where $y$ and $y_\Gamma$ are the components
of the solution to the Cahn--Hilliard system associated \juerg{with} $u_\Gamma$.
This means that $\overline u_\Gamma$ is an optimal control.\qed
\step
Proof of Proposition~\ref{ReprWp}
Assume \pier{that $\Lam$ and $\Lam_\Gamma$} satisfy~\eqref{goodLam}.
Then, formula \eqref{reprWp} actually defines a functional $F$ on~$\calW_0$.
Clearly, $F$~is linear.
Moreover, we have, for every $(v,v_\Gamma)\in\calW_0$,
\Begin{eqnarray}
&& |\<\Lam,v>_Q + \<\Lam_\Gamma,v_\Gamma>_\Sigma|
\non
\\
&& \leq \norma\Lam_{(\H1V^*\cap\L2V)^*} \, \norma v_{\H1V^*\cap\L2V}
\non
\\
&& \quad {}
+ \norma\Lam_\Gamma_{(\H1\VG^*\cap\L2V_\Gamma)^*} \, \norma v_{\H1\VG^*\cap\L2V_\Gamma}
\non
\\
&& \leq \bigl( \norma\Lam_{(\H1V^*\cap\L2V)^*} + \norma\Lam_\Gamma_{(\H1\VG^*\cap\L2V_\Gamma)^*} \bigr) \norma{(v,v_\Gamma)}_{\calW}\,,
\non
\End{eqnarray}
so that $F$~is continuous.
Conversely, assume that $F\in\calW_0^*$.
As $\calW_0$ is~a (closed) subspace of $\tilde\calW:=\bigl(\H1V^*\cap\L2V)\times(\H1\VG^*\cap\L2V_\Gamma)$,
we can extend $F$ to a linear continuous functional $\tilde F$ on~$\tilde\calW$.
Then, there exist $\Lam$ and~$\Lam_\Gamma$
(take $\Lam(v):=\tilde F(v,0)$ and $\Lam_\Gamma(v_\Gamma):=\tilde F(0,v_\Gamma)$)
satisfying \eqref{goodLam} such that
\Begin{equation}
\< \tilde F , (v,v_\Gamma) >
= \< \Lam , v >_Q
+ \< \Lam_\Gamma , v_\Gamma >_\Sigma
\quad \hbox{for every $(v,v_\Gamma)\in\tilde\calW$}\,,
\non
\End{equation}
where the duality product on the left-hand side\ refers to the spaces $(\tilde\calW)^*$ and~$\tilde\calW$.
Since $\[F,(v,v_\Gamma)]=\<\tilde F,(v,v_\Gamma)>$ for every $(v,v_\Gamma)\in\calW_0$,
\eqref{reprWp} immediately follows.\qed
\bigskip
The rest of this section is devoted to the proof of Theorem~\ref{CNopt}
on the necessary optimality conditions.
Therefore, \pier{besides} the general assumptions, we also suppose~that
\Begin{equation}
\hbox{\sl $\overline u_\Gamma$ is any optimal control as in Theorem~\ref{Optimum}},
\label{uoptfixed}
\End{equation}
that is, an arbitrary optimal control $\overline u_\Gamma$ is fixed once and for all.
In order to arrive at the desired necessary optimality condition for~$\overline u_\Gamma$,
we follow~\cite{Barbutrick} and introduce the \pier{modified} functional $\tilde\calJ$ defined \pier{by}
\Begin{equation}
\tilde\calJ(y,y_\Gamma,u_\Gamma)
:= \calJ(y,y_\Gamma,u_\Gamma) + \frac 12 \, \norma{u_\Gamma - \overline u_\Gamma}_{\LS2}^2 \,.
\label{modifiedcost}
\End{equation}
Then the analogue of Theorem~\ref{Optimumvisc} holds,
and we have:
\Begin{theorem}
\label{Optimummod}
For every $\tau>0$, there exists \juerg{some} $\tilde u^\tau_\Gamma\in\calU_{\rm ad}$ such~that
\Begin{equation}
\tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
\leq \tilde\calJ(y^\tau,y^\tau_\Gamma,u_\Gamma)
\quad \hbox{for every $u_\Gamma\in\calU_{\rm ad}$}\,,
\label{optimummod}
\End{equation}
where $\tilde y^\tau$, $\tilde y^\tau_\Gamma$, $y^\tau$ and $y^\tau_\Gamma$
are the components of the solutions $(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde w^\tau)$ and $(y^\tau,y^\tau_\Gamma,w^\tau)$
to~the state system \Pbltau\
corresponding to the controls $\tilde u^\tau_\Gamma$ and~$u_\Gamma$, respectively.
\End{theorem}
For the reader's convenience, we fix the notation just used
and introduce a new one
(\juerg{which was} already used with a different meaning earlier in this paper):
\Begin{eqnarray}
&& \hbox{\sl $\tilde u^\tau_\Gamma$ is an optimal control as in Theorem~\ref{Optimummod}}
\label{utmodfixed}
\\
&& \hbox{\sl $(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde w^\tau)$ is the solution to \Pbltau\ corresponding to $\tilde u^\tau_\Gamma$}
\qquad
\label{taustate}
\\
&& \hbox{\sl $(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline w^{\,\pier{\tau}})$ is the solution to \Pbltau\ corresponding to $\overline u_\Gamma$}.
\qquad
\label{newstate}
\End{eqnarray}
The next step consists in writing the proper adjoint system
and the corresponding necessary optimality condition,
\juerg{which} can be done by repeating the argument of~\cite{CGSvisc}.
However, instead of just stating the corresponding result,
we spend some words that can help the reader.
The optimality variational inequality is derived as a condition
on the Fr\'echet\ derivative of the map
(defined in a proper functional framework)
$u_\Gamma\mapsto\tilde\calJ(y,y_\Gamma,u_\Gamma)$,
where the pair $(y,y_\Gamma)$ is subjected to the state system.
Thus, \juerg{this} derivative depends
on the Fr\'echet\ derivative of the functional
$(y,y_\Gamma,u_\Gamma)\mapsto\tilde\calJ(y,u_\Gamma,u_\Gamma)$,
which is given~by
\Begin{equation}
[D\tilde\calJ(y,\pier{y_\Gamma},u_\Gamma)](k,k_\Gamma,h_\Gamma)]
= b_Q \int_Q (y-z_Q) k
+ b_\Sigma \int_\Sigma (y_\Gamma-z_\Sigma) k_\Gamma
+ \int_\Sigma \bigl( \bzu_\Gamma + (u_\Gamma-\overline u_\Gamma) \bigr) h_\Gamma \,.
\non
\End{equation}
Hence, the argument for $\tilde\calJ$ differs from the one for $\calJ$
only in relation to the last integral.
In other words, we just have to replace
$\bzu_\Gamma$ by $\bzu_\Gamma+(u_\Gamma-\overline u_\Gamma)$ in the whole argument of~\cite{CGSvisc}.
In particular, the adjoint system remains unchanged.
Here is the conclusion.
\Begin{proposition}
\label{CNoptmod}
With the notations \accorpa{utmodfixed}{taustate},
we have
\Begin{equation}
\int_\Sigma \bigl( q^\tau_\Gamma + b_0\tilde u^\tau_\Gamma + (\tilde u^\tau_\Gamma - \overline u_\Gamma) \bigr) (v_\Gamma - \tilde u^\tau_\Gamma) \geq 0
\quad \hbox{for every $v_\Gamma\in\calU_{\rm ad}$}\,,
\label{cnoptmod}
\End{equation}
where $q^\tau_\Gamma$ is the component of the solution $(q^\tau,q^\tau_\Gamma)$ to
\AggiuntoN\ corresponding \pier{to} $u_\Gamma=\tilde u^\tau_\Gamma$ with the
choices $\lam=\lambda^\tau$, $\lam_\Gamma=\lambda^\tau_\Gamma$, $\phi_Q=\phi_Q^\tau$ and $\phi_\Sigma=\phi_\Sigma^\tau$ specified~by
\Begin{equation}
\lambda^\tau=f''(\tilde y^\tau), \enskip
\lambda^\tau_\Gamma=f_\Gamma''(\tilde y^\tau_\Gamma), \enskip
\phi_Q^\tau=b_Q(\tilde y^\tau-z_Q) \enskip and \enskip
\phi_\Sigma^\tau=b_\Sigma(\tilde y^\tau_\Gamma-z_\Sigma) .
\label{sceltelam}
\End{equation}
\End{proposition}
Thus, our project for the proof of Theorem~\ref{CNopt} is the following:
we take the limit in \eqref{cnoptmod}
and in the adjoint system mentioned in the previous statement as $\tau$ tends to zero.
This will lead to the desired necessary optimality condition \eqref{cnopt}
provided that we prove that the optimal controls $\tilde u^\tau_\Gamma$ converge to~$\overline u_\Gamma$.
The \juerg{details of this project are} the following.
\noindent
$i)$~There hold
\Begin{eqnarray}
& \tilde u^\tau_\Gamma \to \overline u_\Gamma
& \quad \hbox{weakly \pier{star in $\H1H_\Gamma\cap L^\infty (\Sigma)$} and strongly in $\LS2$} \qquad
\label{convu}
\\
\noalign{\allowbreak}
& \tilde y^\tau \to \overline y
& \quad \hbox{weakly star in $\H1V^*\cap\spazio L\infty V\cap\L2\Hx 2$}
\non
\\
&& \qquad \hbox{and strongly in \pier{$\C0H \cap \L2V$}}
\label{convy}
\\
\noalign{\allowbreak}
& \tilde y^\tau_\Gamma \to \overline y_\Gamma
& \quad \hbox{weakly star in $\H1H_\Gamma\cap\spazio L\inftyV_\Gamma\cap\L2\HxG 2$}
\qquad
\non
\\
&& \qquad \hbox{and strongly in \pier{$\C0H_\Gamma \cap \L2V_\Gamma$}}
\label{convyG}
\\
& \tilde w^\tau \to \overline w
& \quad \hbox{weakly star in $\pier{\L2V}$}
\label{convw}
\\
& q^\tau \to q
& \quad \hbox{weakly star in $\spazio L\inftyV^*\cap\L2V$}
\label{convq}
\\
& q^\tau_\Gamma \to q_\Gamma
& \quad \hbox{weakly star in $\spazio L\inftyH_\Gamma\cap\L2V_\Gamma$}\,,
\qquad\qquad
\label{convqG}
\End{eqnarray}
as well as
\Begin{equation}
\tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
\to \calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma)\,,
\label{convJ}
\End{equation}
at least for a subsequence,
and $(\overline y,\overline y_\Gamma,\overline w)$ solves problem \State\ with $u_\Gamma=\overline u_\Gamma$.
\noindent
$ii)$~The functionals associated \juerg{with} the pair $(\lamtq^\tau,\lambda^\tau_\Gamma,q^\tau_\Gamma)$
through Proposition~\ref{ReprWp}
converge to some functional weakly in~$\calW_0^*$, at least for a subsequence,
and we then represent the limit by some pair~$(\Lam,\Lam_\Gamma)$,
so that we have
\Begin{equation}
\<\lamtq^\tau,v>_Q + \<\lamtGq^\tau_\Gamma,v_\Gamma>_\Sigma
\to \<\Lam,v>_Q + \<\Lam_\Gamma,v_\Gamma>_\Sigma
\quad \hbox{for every $(v,v_\Gamma)\in\calW_0$}.
\label{convlambda}
\End{equation}
\noindent
$iii)$~With such a choice of $(\Lam,\Lam_\Gamma)$,
the pair $(q,q_\Gamma)$ solves \accorpa{primazero}{secondazero}.
\noindent
$iv)$~Condition \eqref{cnopt} holds.
The main tool is proving a priori estimates.
To this concern, we use the following rule to denote constants
in order to avoid a boring notation.
The small-case symbol $c$ stands for different constants
\juerg{that neither} depend on $\tau$ nor on the functions whose norm we want to estimate.
Hence, the meaning of $c$ might
change from line to line and even in the same chain of equalities or inequalities.
Similarly, a~symbol like $c_\delta$ denotes different constants
that depend on the parameter~$\delta$, in addition.
\step
First a priori estimate
As $\overline u^{\,\pier{\tau}}_\Gamma\in\calU_{\rm ad}$ and Theorem~\ref{daCGS} holds, we have
\Begin{eqnarray}
\hskip-1cm && \norma\tilde u^\tau_\Gamma_{\H1H_\Gamma}
+ \norma\tilde y^\tau_{\H1V^* \cap \spazio L\infty V \cap \L2\Hx 2}
\non
\\[1mm]
\hskip-1cm && \quad {}
+ \norma\tilde y^\tau_\Gamma_{\H1H_\Gamma \cap \spazio L\inftyV_\Gamma \cap \L2\HxG 2}
+ \norma\tilde w^\tau_{\pier{\L2V}}
\non
\\[1mm]
\hskip-1cm && \quad {}
+ \norma{f'(\tilde y^\tau)}_{\L2H}
+ \norma{f_\Gamma'(\tilde y^\tau_\Gamma)}_{\L2H_\Gamma}
+ \tau^{1/2} \norma{\partial_t\tilde y^\tau}_{\L2H}
\leq c \,.
\label{primastima}
\End{eqnarray}
\step
Second a priori estimate
For the reader's convenience, we
\juerg{explicitly} write the adjoint system mentioned in Proposition~\ref{CNoptmod},
as well as the regularity of its solution,
\Begin{eqnarray}
&& q^\tau \in \H1H \cap \L2{\Hx2} , \quad
q^\tau_\Gamma \in \H1H_\Gamma \cap \L2{\HxG2}
\qquad
\label{regqGbis}
\\
&& (q^\tau,q^\tau_\Gamma)(s) \in \calV_\Omega
\quad \hbox{for every $s\in[0,T]$} \vphantom{\frac 1{|\Omega|}}
\label{primaNbis}
\\
\noalign{\allowbreak}
&& - \int_\Omega \partial_t \bigl( \calN(q^\tau(s)) + \tau q^\tau(s) \bigr) \, v
+ \int_\Omega \nablaq^\tau(s) \cdot \nabla v
+ \int_\Omega \lambda^\tau(s) \, q^\tau(s) \, v
\qquad
\non
\\
&& \qquad {}
- \int_\Gamma \partial_tq^\tau_\Gamma(s) \, v_\Gamma
+ \int_\Gamma \nablaGq^\tau_\Gamma(s) \cdot \nablaGv_\Gamma
+ \int_\Gamma \lambda^\tau_\Gamma(s) \, q^\tau_\Gamma(s) \, v_\Gamma
\non
\\
&& = \int_\Omega \phi_Q^\tau(s) v
+ \int_\Gamma \phi_\Sigma^\tau(s) v_\Gamma
\quad \hbox{for a.e.\ $s\in(0,T)$ and every $(v,v_\Gamma)\in\calV_\Omega$}
\qquad
\label{secondaNbis}
\\
\noalign{\allowbreak}
&& \quad \hbox{where
$\lambda^\tau=f''(\tilde y^\tau)$, \
$\lambda^\tau_\Gamma=f_\Gamma''(\tilde y^\tau_\Gamma)$, \
$\phi_Q^\tau=b_Q(\tilde y^\tau-z_Q)$ \ and \
$\phi_\Sigma^\tau=b_\Sigma(\tilde y^\tau_\Gamma-z_\Sigma)$
}\vphantom\int
\non
\\
&& \int_\Omega \bigl( \calNq^\tau + \tau q^\tau \bigr)(T) \, v
+ \int_\Gamma q_\Gamma(T) \, v_\Gamma
= 0
\quad \hbox{for every $(v,v_\Gamma)\in\calV_\Omega$} .
\label{cauchyNbis}
\End{eqnarray}
Now, we choose $v=q^\tau(s)$ and $v_\Gamma=q^\tau_\Gamma(s)$,
and integrate over $(t,T)$ with respect to~$s$.
\pier{Recalling} \eqref{dtcalN} and now reading
$Q_t:=\Omega\times(t,T)$ and $\Sigma_t:=\Gamma\times(t,T)$,
we~have
\Begin{eqnarray}
&& \frac 12 \, \normaVp{q^\tau(t)}^2
+ \frac \tau 2 \int_\Omega |q^\tau(t)|^2
+ \int_{Q_t} |\nablaq^\tau|^2
+ \int_{Q_t} \lambda^\tau |q^\tau|^2
\non
\\
&& \quad {}
+ \frac 12 \int_\Gamma |q^\tau_\Gamma(t)|^2
+ \int_{\Sigma_t} |\nablaGq^\tau_\Gamma|^2
+ \int_{\Sigma_t} \lambda^\tau_\Gamma |q^\tau_\Gamma|^2
\non
\\
&& = \int_{Q_t} \phi_Q^\tau\, q^\tau
+ \int_{\Sigma_t} \phi_\Sigma^\tau \, q^\tau_\Gamma
\leq \int_Q |\phi_Q^\tau|^2 + \int_{Q_t} |q^\tau|^2
+ \int_\Sigma |\phi_\Sigma^\tau|^2 + \int_{\Sigma_t} |q^\tau_\Gamma|^2
\non
\\
&& \leq \int_{Q_t} |q^\tau|^2
+ \int_{\Sigma_t} |q^\tau_\Gamma|^2
+ c
\label{persecondastima}
\End{eqnarray}
\juerg{where the last inequality follows from} \eqref{primastima}.
By accounting for~\eqref{hpfseconda}, we also have
\Begin{equation}
\int_{Q_t} \lambda^\tau |q^\tau|^2
\geq - c \int_{Q_t} |q^\tau|^2
\quad\hbox{and}\quad
\int_{Q_t} \lambda^\tau_\Gamma |q^\tau_\Gamma|^2
\geq - c \int_{\Sigma_t} |q^\tau_\Gamma|^2 .
\non
\End{equation}
We treat the volume integral (and the same on the right-hand side\ of~\eqref{persecondastima})
invoking the compact embedding $V\subset H$.
We have
\Begin{equation}
\int_\Omega |v|^2
\leq \delta \int_\Omega |\nabla v|^2 + c_\delta \normaVp v^2
\quad \hbox{for every $v\in V$ and $\delta>0$}.
\non
\End{equation}
Hence, we deduce \juerg{that}
\Begin{equation}
\int_{Q_t} |q^\tau|^2
\leq \delta \int_{Q_t} |\nablaq^\tau|^2 + c_\delta \int_t^T \normaVp{q^\tau(s)}^2 \, ds \,.
\non
\End{equation}
Therefore, by combining, choosing $\delta$ small enough and applying the backward Gronwall lemma,
we conclude~that
\Begin{equation}
\normaq^\tau_{\spazio L\inftyV^*\cap\L2V}
+ \normaq^\tau_\Gamma_{\spazio L\inftyH_\Gamma\cap\L2V_\Gamma}
+ \tau^{1/2} \normaq^\tau_{\spazio L\infty H}
\leq c \,.
\label{secondastima}
\End{equation}
\step
Third a priori estimate
\pier{Take an arbitrary pair} $(v,v_\Gamma)\in\H1\calH_\Omega\cap\L2\calV_\Omega$,
\pier{and} test \eqref{secondaNbis} by $v(s)$ and $v_\Gamma(s)$.
Then, we sum over $s\in(0,T)$ and integrate by parts
with the help of~\eqref{cauchyNbis},
so that no integral related to the time $T$ appears.
In particular, if $(v,v_\Gamma)\in\calW_0$, even the terms
evaluated at $t=0$ vanish
and we obtain \juerg{that}
\Begin{eqnarray}
&& \int_Q (\calNq^\tau+\tauq^\tau) \partial_t v
+ \int_Q \nablaq^\tau \cdot \nabla v
+ \int_Q \lambda^\tau q^\tau v
+ \int_\Sigma q^\tau_\Gamma \partial_tv_\Gamma
+ \int_\Sigma \nablaq^\tau_\Gamma \cdot \nablav_\Gamma
+ \int_\Sigma \lambda^\tau q^\tau_\Gamma v_\Gamma
\non
\\
&& = \int_Q \phi_Q^\tau \, v
+ \int_\Sigma \phi_\Sigma^\tau \, v_\Gamma \,.
\label{adjperparti}
\End{eqnarray}
Therefore, we have, for every $(v,v_\Gamma)\in\calW_0$,
\Begin{eqnarray}
&& \left| \int_Q \lambda^\tau q^\tau v + \int_\Sigma \lambda^\tau q^\tau_\Gamma v_\Gamma \right|
\non
\\
&& \leq \norma{\calNq^\tau+\tauq^\tau}_{\L2V} \, \norma{\partial_t v}_{\L2V^*}
+ \normaq^\tau_{\L2V} \, \norma v_{\L2V}
\non
\\
&& \quad {}
+ \normaq^\tau_\Gamma_{\L2V_\Gamma} \, \norma{\partial_t \pier{v_\Gamma}}_{\pier{\L2\VG^*}}
+ \norma{\pier{q^\tau_\Gamma}}_{\pier{\L2V_\Gamma}} \, \normav_\Gamma_{\L2V_\Gamma}
\non
\\
&& \quad {}
+ \norma\phi_Q^\tau_{\L2H} \, \norma v_{\L2H}
+ \norma\phi_\Sigma^\tau_{\L2H_\Gamma} \, \normav_\Gamma_{\L2H_\Gamma} \,.
\non
\End{eqnarray}
Now, by assuming $\tau\leq1$, we have
$\normaV{\calN v+\tau v}\leq c\normaVp v+\tau\normaV v\leq c\normaV v$
for every $v\in V$ with zero mean value
(see~\eqref{defnormaVp}).
Therefore, by accounting for \eqref{primastima} and \eqref{secondastima},
we conclude~that
\Begin{equation}
\left| \int_Q \lambda^\tau q^\tau v + \int_\Sigma \lambda^\tau q^\tau_\Gamma v_\Gamma \right|
\leq c \, \norma{(v,v_\Gamma)}_\calW
\quad \hbox{for every $(v,v_\Gamma)\in\calW_0$}.
\label{terzastima}
\End{equation}
\step
Conclusion of the proof of Theorem~\ref{CNopt}
From the above estimates, we infer that
\Begin{eqnarray}
& \tilde u^\tau_\Gamma \to u_\Gamma
& \quad \hbox{weakly \pier{star in $\H1H_\Gamma\cap L^\infty (\Sigma)$}}
\label{perconvu}
\\
\noalign{\allowbreak}
& \tilde y^\tau \to y
& \quad \hbox{weakly star in $\H1V^*\cap\spazio L\infty V\cap\L2\Hx 2$}\qquad
\non
\\
&& \qquad \hbox{and strongly in \pier{$\C0H \cap \L2V$}}
\label{perconvy}
\\
\noalign{\allowbreak}
& \tilde y^\tau_\Gamma \to y_\Gamma
& \quad \hbox{weakly star in $\H1H_\Gamma\cap\spazio L\inftyV_\Gamma\cap\L2\HxG 2$}
\qquad
\non
\\
&& \qquad \hbox{and strongly in \pier{$\C0H_\Gamma \cap \L2V_\Gamma$}}
\label{perconvyG}
\\
& \tilde w^\tau \to w
& \quad \hbox{weakly star in $\pier{\L2V}$}
\label{perconvw}
\\
& q^\tau \to q
& \quad \hbox{weakly star in $\spazio L\inftyV^*\cap\L2V$}
\label{perconvq}
\\
& q^\tau_\Gamma \to q_\Gamma
& \quad \hbox{weakly star in $\spazio L\inftyH_\Gamma\cap\L2V_\Gamma$}
\qquad\qquad
\label{perconvqG}
\\
& \tau q^\tau \to 0
& \quad \hbox{\pier{strongly} in $\spazio L\infty H$}
\label{pertauqt}
\End{eqnarray}
at least for a subsequence,
and $(y,y_\Gamma,w)$ is the solution to \pier{the} problem \Pbl\
corresponding to $u_\Gamma$,
thanks to Proposition~\ref{Convergence}.
Notice that \accorpa{perconvq}{perconvqG} coincide with \accorpa{convq}{convqG}
and that \accorpa{convu}{convw} hold once we prove that $u_\Gamma=\overline u_\Gamma$
and that $\tilde u^\tau_\Gamma$ \pier{converges} \juerg{strongly in~$\LS2$}.
\juerg{To this end, we} recall the notations \accorpa{utmodfixed}{newstate},
and it is understood that all the limits we write are referred to the selected subsequence.
By optimality, we have
\Begin{equation}
\calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma) \leq \calJ(y,y_\Gamma,u_\Gamma)
\quad\hbox{and}\quad
\tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma) \leq \tilde\calJ(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline u_\Gamma) .
\non
\End{equation}
On the other hand, \accorpa{perconvu}{perconvyG} and Proposition~\ref{Convergence}
applied with $u^\tau_\Gamma=\overline u_\Gamma$ yield
\Begin{equation}
\tilde\calJ(y,y_\Gamma,u_\Gamma) \leq \liminf \tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
\quad\hbox{and}\quad
\lim \calJ(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline u_\Gamma) = \calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma) .
\non
\End{equation}
By combining, we deduce that
\Begin{eqnarray}
&& \calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma) + \frac 12 \, \norma{u_\Gamma-\overline u_\Gamma}_{\LS2}^2
\leq \calJ(y,y_\Gamma,u_\Gamma) + \frac 12 \, \norma{u_\Gamma-\overline u_\Gamma}_{\LS2}^2
\non
\\
&& = \tilde\calJ(y,y_\Gamma,u_\Gamma)
\leq \liminf \tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
\leq \limsup \tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
\vphantom{\frac 12}
\non
\\
&& \leq \limsup \tilde\calJ(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline u_\Gamma)
= \limsup \calJ(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline u_\Gamma)
= \calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma) .
\non
\End{eqnarray}
By comparing the first and last terms of this chain,
we infer that the $\LS2$-norm of $u_\Gamma-\overline u_\Gamma$ vanishes,
whence $u_\Gamma=\overline u_\Gamma$, as desired.
In order to prove the strong convergence mentioned in~\eqref{convu},
we observe that the above argument also shows~that
\Begin{equation}
\liminf \tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
= \limsup \tilde\calJ(\tilde y^\tau,\tilde y^\tau_\Gamma,\tilde u^\tau_\Gamma)
= \calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma).
\non
\End{equation}
Notice that this coincides with \eqref{convJ}.
From the strong convergence given by \eqref{convy} and \eqref{convyG},
and by comparison, we deduce~that
\Begin{equation}
\lim \left(
\frac b_0 2 \int_\Sigma |\tilde u^\tau_\Gamma|^2 + \frac 12 \int_\Sigma |\tilde u^\tau_\Gamma-\overline u_\Gamma|^2
\right)
= \frac b_0 2 \int_\Sigma |\pier{\overline u_\Gamma}|^2\,,
\non
\End{equation}
whence also
\Begin{eqnarray}
&& \limsup \frac b_0 2 \int_\Sigma |\tilde u^\tau_\Gamma|^2
\leq \limsup \left(
\frac b_0 2 \int_\Sigma |\tilde u^\tau_\Gamma|^2 + \frac 12 \int_\Sigma |\tilde u^\tau_\Gamma-\overline u_\Gamma|^2
\right)
\non
\\
&& = \frac b_0 2 \int_\Sigma |\overline u_\Gamma|^2
\leq \liminf \frac b_0 2 \int_\Sigma |\tilde u^\tau_\Gamma|^2 .
\non
\End{eqnarray}
Therefore\pier{, we have}
\Begin{equation}
\lim \frac b_0 2 \int_\Sigma |\tilde u^\tau_\Gamma|^2
= \frac b_0 2 \int_\Sigma |\overline u_\Gamma|^2\,,
\quad \hbox{whence} \quad
\lim \frac 12 \int_\Sigma |\tilde u^\tau_\Gamma-\overline u_\Gamma|^2 = 0 \,,
\non
\End{equation}
and \accorpa{convu}{convw} are completely proved.
Now, we deal with the limit $(q,q_\Gamma)$ given
by~\accorpa{perconvq}{perconvqG}, i.e., \accorpa{convq}{convqG}.
Clearly, \eqref{primazero} holds as well.
Furthermore, as $\normaV{\calNv_*}\leq c\normaVpv_*$
for every $v_*\inV^*$ with zero mean value
(see~\eqref{defN}),
\juerg{and since the} convergence \eqref{pertauqt} holds,
we~also have
\Begin{equation}
\calNq^\tau + \tau q^\tau \to \calN q
\quad \hbox{weakly star in $\spazio L\infty H$}.
\non
\End{equation}
On the other hand, \eqref{terzastima} implies that
the functionals $F^\tau\in\calW_0^*$ defined~by
\Begin{equation}
\[ F^\tau , (v,v_\Gamma) ]
:= \< \lambda^\tau q^\tau , v >_Q
+ \< \lambda^\tau_\Gamma q^\tau_\Gamma , v_\Gamma >_\Sigma\,,
\non
\End{equation}
i.e., the functionals associated \juerg{with} $(\lamtq^\tau,\lambda^\tau_\Gamma,q^\tau_\Gamma)$
as in Proposition~\ref{ReprWp}, are bounded in~$\calW_0^*$.
Therefore, for a subsequence, we have $F^\tau\to F$ weakly star in~$\calW_0^*$,
where $F$ is some element of~$\calW_0^*$.
Hence, if we represent $F$ as stated in Proposition~\ref{ReprWp},
we find $\Lam$ and $\Lam_\Gamma$ satisfying \eqref{goodLam} and~\eqref{convlambda}.
At this point, it is straightforward\ to pass to the limit in~\eqref{adjperparti} and in \eqref{cnoptmod}
\juerg{to} obtain both \eqref{secondazero} and~\eqref{cnopt}.
This completes the proof of Theorem~\ref{CNopt}.\qed
\Begin{remark}\rm
\label{Fullconv}
The above proof can be repeated without any change
starting from any sequence $\tau_n\seto0$.
By doing that, we obtain:
there exists a subsequence $\graffe{\tau_{n_k}}$ such that
\accorpa{convu}{convJ} hold along the selected subsequence.
As the limits $\overline u_\Gamma$, $\overline y$, $\overline y_\Gamma$, $\overline w$ and $\calJ(\overline y,\overline y_\Gamma,\overline u_\Gamma)$
are always the same,
this proves that in fact \accorpa{convu}{convw} as well as \eqref{convJ}
hold for the whole family.
On the contrary, the limits $q$ and $q_\Gamma$ might depend on the selected subsequence
since no uniqueness result for the adjoint problem is known.
Nevertheless, the necessary optimality condition \eqref{cnopt}
holds for every solution $(q,q_\Gamma)$ to the adjoint problem
that can be found as a limit of pairs $(q^\tau,q^\tau_\Gamma)$ as specified in the above proof.
\End{remark}
\vspace{3truemm}
\Begin{thebibliography}{10}
\bibitem{Barbu}
V. Barbu,
``Nonlinear semigroups and differential equations in Banach spaces'',
Noord-hoff,
Leyden,
1976.
\bibitem{Barbutrick}
V. Barbu,
Necessary conditions for nonconvex distributed control problems
governed by elliptic variational inequalities,
{\it J. Math. Anal. Appl.} {\bf 80} (1981) \pier{566-597}.
\bibitem{CahH}
J. W. Cahn and J. E. Hilliard,
Free energy of a nonuniform system I. Interfacial free energy,
{\it J. Chem. Phys.\/}
{\bf 2} (1958) 258-267.
\bibitem{CaCo}
L. Calatroni and P. Colli,
Global solution to the Allen--Cahn equation with singular potentials and dynamic boundary conditions,
{\it Nonlinear Anal.\/} {\bf 79} (2013) 12-27.
\pier{
\bibitem{CMZ}
L.\ {C}herfils, A.\ {M}iranville and S.\ {Z}elik,
The {C}ahn--{H}illiard equation with logarithmic potentials,
{\it Milan J. Math.\/} {\bf 79} (2011) 561-596.%
}
\bibitem{CFGS1}
P. Colli, M.H. Farshbaf-Shaker, G. Gilardi and J. Sprekels,
Optimal boundary control of a viscous Cahn--Hilliard system with dynamic boundary condition and double obstacle potentials,
WIAS Preprint No.~2006 (2014) 1-29.
\bibitem{CFS}
P. Colli, M.H. Farshbaf-Shaker and J. Sprekels,
A deep quench approach to the optimal control of an Allen--Cahn equation
with dynamic boundary conditions and double obstacles,
{\it Appl. Math. Optim.\/} \pier{{\bf 71} (2015) 1-24.}
\pier{
\bibitem{CF}
P.\ {C}olli and T.\ {F}ukao,
{C}ahn--{H}illiard equation with dynamic boundary conditions
and mass constraint on the boundary,
preprint arXiv:1412.1932 [math.AP] (2014) 1-26.%
}
\bibitem{CGS}
P. Colli, G. Gilardi and J. Sprekels,
On the Cahn--Hilliard equation with dynamic
boundary conditions and a dominating boundary potential,
{\it J. Math. Anal. Appl.\/} {\bf 419} (2014) 972-994.
\bibitem{CGSvisc}
P. Colli, G. Gilardi and J. Sprekels,
A boundary control problem
for the viscous Cahn-Hilliard equation
with dynamic boundary conditions,
\pier{preprint arXiv:1407.3916 [math.AP] (2014) 1-27.}
\pier{
\bibitem{CS}
P. Colli and J. Sprekels,
Optimal control of an Allen--Cahn equation
with singular potentials and dynamic boundary condition,
{\it SIAM \juerg{J.} Control Optim.\/} {\bf 53} (2015) 213-234.%
}
\bibitem{EllSt}
C. M. Elliott and A. M. Stuart,
Viscous Cahn--Hilliard equation. II. Analysis,
{\it J. Differential Equations\/}
{\bf 128} (1996) 387-414.
\bibitem{EllSh}
C. M. Elliott and S. Zheng,
On the Cahn--Hilliard equation,
{\it Arch. Rational Mech. Anal.\/}
{\bf 96} (1986) 339-357.
\bibitem{GiMiSchi}
G. Gilardi, A. Miranville and G. Schimperna,
On the Cahn--Hilliard equation with irregular potentials and dynamic boundary conditions,
{\it Commun. Pure Appl. Anal.\/}
{\bf 8} (2009) 881-912.
\pier{
\bibitem{HW1}
M. Hinterm\"uller and D. Wegner, Distributed optimal control of the
Cahn--Hilliard system including the case of a double-obstacle
homogeneous free energy density, {\it SIAM J.
Control Optim.} {\bf 50} (2012) 388-418.
\bibitem{HW2}
M. Hinterm\"uller and D. Wegner, Optimal control of a semi-discrete
Cahn--Hilliard--Navier--Stokes system, {\it SIAM J. Control Optim.}
{\bf 52} (2014) 747-772.
\bibitem{Lions}
J.-L.~Lions,
``Quelques m\'ethodes de r\'esolution des probl\`emes
aux limites non lin\'eaires'',
Dunod; Gauthier-Villars, Paris, 1969.
\bibitem{PRZ}
J. Pr\"uss, R. Racke and S. Zheng,
Maximal regularity and asymptotic behavior of solutions for the Cahn--Hilliard equation with dynamic boundary conditions,
{\it Ann. Mat. Pura Appl.~(4)\/}
{\bf 185} (2006) 627-648.
\bibitem{RZ}
R. Racke and S. Zheng,
The Cahn--Hilliard equation with dynamic boundary conditions,
{\it Adv. Differential Equations\/}
{\bf 8} (2003) 83-110.
\bibitem{RoSp}
E. Rocca and J. Sprekels,
Optimal distributed control of a nonlocal convective Cahn--Hilliard equation
by the velocity in 3D, WIAS Preprint No. 1942 (2014), pp. 1-24.
\bibitem{Simon}
J. Simon,
{Compact sets in the space $L^p(0,T; B)$},
{\it Ann. Mat. Pura Appl.~(4)\/}
{\bf 146} (1987) 65-96.
\bibitem{Wang}
Q.-F. Wang, Optimal distributed control of nonlinear {C}ahn-{H}illiard
systems with computational realization,
{\it J. Math. Sci. (N. Y.)\/} {\bf 177} (2011) 440-458.
\bibitem{WaNa}
Q.-F. Wang and S.-I. Nakagiri, Optimal control of distributed parameter
system given by Cahn--Hilliard equation, {\it Nonlinear Funct. Anal. Appl.} {\bf 19} (2014) 19-33.
\bibitem{WZ} H. Wu and S. Zheng,
Convergence to equilibrium for the Cahn--Hilliard equation with dynamic boundary conditions, {\it J. Differential Equations\/}
{\bf 204} (2004) 511-531.
\bibitem{ZL1}
X. P. Zhao and C. C. Liu, Optimal control of the convective Cahn--Hilliard equation,
{\it Appl. Anal.\/} {\bf 92} (2013) 1028-1045.
\bibitem{ZL2}
X. P. Zhao and C. C. Liu, Optimal control for the convective Cahn--Hilliard equation
in 2D case, {\it Appl. Math. Optim.} {\bf 70} (2014) 61-82.
\bibitem{ZW}
J. Zheng and Y. Wang, Optimal control problem for {C}ahn--{H}illiard equations
with state constraint, {\it J. Dyn. Control Syst.\/}, {\bf 21} (2015) 257-272.%
}
\End{thebibliography}
\End{document}
\noindent
{\bf In the source file after {\tt\expandafter\string\ }\unskip end$\{$document$\}$,
one can find more references taken from~\cite{CGSvisc}.}
\bibitem{Brezis}
H. Brezis,
``Op\'erateurs maximaux monotones et semi-groupes de contractions
dans les espaces de Hilbert'',
North-Holland Math. Stud.
{\bf 5},
North-Holland,
Amsterdam,
1973.
\bibitem{BreGil}
F. Brezzi and G. Gilardi,
Part~1:
Chapt.~2, Functional spaces,
Chapt.~3, Partial differential equations,
in ``Finite element handbook'',
H. Kardestuncer and D. H.\ Norrie eds.,
McGraw-Hill Book Company,
NewYork,
1987.
\bibitem{CFP}
R. Chill, E. Fa\v sangov\'a and J. Pr\"uss,
Convergence to steady states of solutions of the Cahn--Hilliard equation with dynamic boundary conditions,
{\it Math. Nachr.\/}
{\bf 279} (2006) 1448-1462.
\bibitem{CF}
P. Colli and T. Fukao,
The Allen-Cahn equation with dynamic boundary conditions and mass constraints,
preprint arXiv:1405.0116~[math.AP] (2014), pp.~1-23.
\bibitem{CGPS}
P. Colli, G. Gilardi, P. Podio-Guidugli and J. Sprekels,
Distributed optimal control of a nonstandard system of phase field equations, {\it Cont. Mech. Thermodyn.\/}
{\bf 24} (2012) 437-459.
\bibitem{CGScon}
P. Colli, G. Gilardi and J. Sprekels,
Analysis and boundary control of a nonstandard system of phase field equations,
{\it Milan J. Math.\/} {\bf 80} (2012) 119-149.
\bibitem{Is}
H.\ Israel,
Long time behavior of an {A}llen-{C}ahn type equation with a
singular potential and dynamic boundary conditions,
{\it J. Appl. Anal. Comput.} {\bf 2} (2012) 29-56.
\bibitem{LioMag}
J.-L. Lions and E. Magenes,
``Non-homogeneous boundary value problems and applications'',
Vol.~I,
Springer, Berlin, 1972.
\bibitem{MirZelik}
A. Miranville and S. Zelik,
Robust exponential attractors for Cahn--Hilliard type equations with singular potentials,
{\it Math. Methods Appl. Sci.\/}
{\bf 27} (2004) 545--582.
\bibitem{Podio}
P. Podio-Guidugli,
Models of phase segregation and diffusion of atomic species on a lattice,
{\it Ric. Mat.} {\bf 55} (2006) 105-118.
\bye
is the following
\Begin{eqnarray}
&& \int_\Omega q(t) \, v
= \int_\Omega \nabla p(t) \cdot \nabla v
\qquad \hbox{for every $v\in V$ and $t\in [0,T]$}
\label{primaAV}
\\
&& - \int_\Omega \partial_t \bigl( p(t) + \tau q(t) \bigr) \, v
+ \int_\Omega \nabla q(t) \cdot \nabla v
+ \int_\Omega f''(\overline y^{\,\pier{\tau}}(t)) \, q(t) \, v
\qquad
\non
\\
&& \quad {}
- \int_\Gamma \partial_tq_\Gamma(t) \, v_\Gamma
+ \int_\Gamma \nablaGq_\Gamma(t) \cdot \nablaGv_\Gamma
+ \int_\Gamma f_\Gamma''(\overline y^{\,\pier{\tau}}_\Gamma(t)) \, q_\Gamma(t) \, v_\Gamma
\non
\\
&& = \int_\Omega b_Q \bigl( \overline y^{\,\pier{\tau}}(t) - z_Q(t) \bigr) v
+ \int_\Gamma b_\Sigma \bigl( \overline y^{\,\pier{\tau}}_\Gamma(t) - z_\Sigma(t) \bigr) v_\Gamma
\non
\\
&& \quad \hbox{for every $(v,v_\Gamma)\in\calV$ and \checkmmode{for a.a.~$t\in(0,T)$}}
\vphantom\int
\label{secondaAV}
\\
&& q^\tau_\Gamma(t) = q^\tau(t){\vrule height 5pt depth 4pt\,}_\Gamma
\quad \checkmmode{for a.a.~$t\in(0,T)$}
\vphantom\int
\label{qqG}
\\
&& \int_\Omega (p+\tau q)(T) \, v
+ \int_\Gamma q^\tau_\Gamma(T) \, v_\Gamma
= 0
\qquad \hbox{for every $(v,v_\Gamma)\in\calV$}
\label{cauchyAV}
\End{eqnarray}
\Accorpa\PblAV primaAV cauchyAV
where $\overline y^{\,\pier{\tau}}$ and $\overline y^{\,\pier{\tau}}_\Gamma$ are the components of the solution $(\overline y^{\,\pier{\tau}},\overline y^{\,\pier{\tau}}_\Gamma,\overline w^{\,\pier{\tau}})$
to \Pbltau\ corresponding to the choice $u_\Gamma=\overline u^{\,\pier{\tau}}_\Gamma$.
|
\section{Introduction and notation.}
\par Let $\mathbb D=\{z\in\mathbb C\,|\,|z|<1\}$, $\mathbb T=\{z\in\mathbb C\,|\,|z|=1\}=\mathbb R/2\pi$. We denote by $C(\mathbb T)$ the set of complex
continuous functions with the supremum norm $\|\cdot\|$ and, for $f\in C(\mathbb T)$, by $S_n(f,t)$ the $n$-th partial sum of the Fourier series of $f$ at the
point $t\in\mathbb T$,
$$ S_n(f,t)=\sum_{k=-n}^n\widehat f(k)e^{ikt},$$
where
$$\widehat f(k)=\frac 1{2\pi}\int_{-\pi}^{\pi}f(t)e^{-ikt}\,dt,\qquad k\in\mathbb Z,$$
is the $k$-th Fourier coefficient of $f$.
\par Also, let $A(\mathbb D)=\{f\in C(\mathbb T)\,|\,\widehat f(k)=0\,\,\text{for}\,\,k<0\}$ be the disc algebra with the supremum norm.
\par If $(X,d)$ is a complete metric space, a property is said to be satisfied at quasi all points of $X$ if it is satisfied at a $G_{\delta}$ and dense set,
i.e. at a topologically large set.
\par First we recall a few facts regarding the divergence of the partial sums $S_n(f,t)$ that the reader must have in mind.
\par The first, from \cite{K}, is that quasi all $f\in C(\mathbb T)$ have the property that their Fourier series diverge at quasi all points of $\mathbb T$.
\par The second (see \cite{Ka}) is the classical result that $E\subseteq\mathbb T$ is a set of divergence for $C(\mathbb T)$, i.e. there is a continuous
function whose Fourier series diverges at all points of $E$, if and only if $E$ is a set of infinite divergence for $C(\mathbb T)$, i.e. there is a continuous
function $f$ such that $\varlimsup|S_n(f,t)|=+\infty$ for all $t\in E$. Moreover, every set of Lebesgue measure zero is a set of divergence for $C(\mathbb T)$.
\par It is not hard to see that a set $E\subseteq\mathbb T$ may be $G_{\delta}$ and dense and simultaneously of Lebesgue measure zero (see \cite{Ox}). Of
course, by Carleson's theorem a set of divergence for $C(\mathbb T)$ has necessarily measure zero.
\par The third fact, from \cite{BH}, is that the set $\{t\,|\,\varlimsup|S_n(f,t)|=+\infty\}$ has Hausdorff dimension equal to $1$ for quasi all $f\in
C(\mathbb T)$.
\par Recently, in \cite{M} and \cite{HK}, a different notion of divergence has been studied. We present the definitions and the basic results of these papers.
\begin{definition}
Let $E\subseteq\mathbb T$. We say that $f\in C(\mathbb T)$ is pointwise universal on $E$ if for every $g:E\to\mathbb C$ belonging to the Baire-1 class there
exists a strictly increasing sequence of positive integers $(k_n)$ such that $S_{k_n}(f,t)\to g(t)$ for all $t\in E$.\newline
We denote the class of these functions by $U_p(E)$.
\end{definition}
\begin{definition}
Let $K\subseteq\mathbb T$ be a compact set. We say that $f\in C(\mathbb T)$ is uniformly universal on $K$ if for every continuous $g:K\to\mathbb C$ there
exists a strictly increasing sequence of positive integers $(k_n)$ such that $\|S_{k_n}(f,\cdot)-g\|_K\to 0$, where $\|\cdot\|_K$ is the supremum norm on
$K$.\newline
We denote the class of these functions by $U(K)$.
\end{definition}
\par Of course the notions of pointwise universality and uniform universality coincide when the set $E=K$ is finite. In this case we speak about universality
on $E=K$.
\par If we denote by $\mathcal K(\mathbb T)$ the complete metric space of all nonempty compact subsets of $\mathbb T$ with the Hausdorff metric then we know
that quasi all $f\in C(\mathbb T)$ and quasi all $f\in A(\mathbb D)$ are uniformly universal on quasi all sets $K\in\mathcal K(\mathbb T)$. (For $C(\mathbb
T)$ see \cite{M} and for $A(\mathbb D)$ see \cite{HK}.) And since quasi all sets in $\mathcal K(\mathbb T)$ are perfect sets (see \cite{Ko}) we know that there
are perfect $K\subseteq\mathbb T$ such that $A(\mathbb D)\cap U(K)\neq\emptyset$.
\par Finally, we know that for each countable $E\subseteq\mathbb T$ quasi all $f\in C(\mathbb T)$ and quasi all $f\in A(\mathbb D)$ are pointwise universal on
$E$. (For $C(\mathbb T)$ see \cite{M} and for $A(\mathbb D)$ see \cite{HK}.)
\par The proofs of the above results are not constructive. They use Baire's category theorem. Hence a first question which arises is to construct a uniformly or
pointwise universal function. A second question is what can we say about the convergence of the Fourier series outside $E$ of a function in $U_p(E)$. If for
example $f\in A(\mathbb D)\cap U_p(E)$, where $E$ is a countable dense set in $\mathbb T$, then the Fourier series of $f$ cannot converge at all points outside
$E$, since $E\subseteq G\subseteq D$, where $G=\bigcap_{N=1}^{+\infty}\bigcup_{n=1}^{+\infty}\{t\,|\,S_n(f,t)>N\}$ and
$D=\{t\,|\,S_n(f,t)\,\,\text{diverges}\}$, and $G$ is $G_{\delta}$ and dense and hence uncountable. Also, taking into account that the pointwise universal
functions are highly divergent, we may ask whether it is possible that the above set $D$ has Hausdorff dimension less than $1$. We deal with these questions in
sections 2 and 3. More precisely, we give a method to construct pointwise universal functions in $A(\mathbb D)$ on finite and countably infinite sets. We also
give a criterion for convergence of the Fourier series outside the set of pointwise universality and we see that the above set $D$ can even be of Hausdorff
dimension equal to $0$. Of course, the above method can also be applied for functions in $C(\mathbb T)$.
\par In section 4 we turn to the study of uniform universality (see definition 1.2). By Carleson's theorem the perfect sets which accept uniform universality
must have Lebesgue measure zero. Of course the first class of such perfect sets that comes to mind consists of the familiar Cantor type sets. We prove that
these sets do not accept uniform universality. Moreover, we prove that the same is true for a class of compact countable sets. This is in contrast with the
well known fact that all sets of measure zero are sets of divergence for $C(\mathbb T)$.
\par Finally, we close this paper with some open problems.
\par In the following the symbol $C$ will denote an absolute constant which may change from one relation to the next.
\section{Universality on finite $K\subseteq\mathbb T$ and convergence on $\mathbb T\setminus K$.}
\par Let $N>n$. We consider the Fejer polynomials
$$Q_{N,n}(t)=2\sin Nt\sum_{k=1}^n\frac{\sin kt}k=\sum_{m=-(N+n)}^{N+n}\widehat{Q_{N,n}}(m)e^{imt},$$
where
\begin{equation}
\widehat{Q_{N,n}}(\pm m)=\begin{cases}\frac 1{2k}, &\text{$m=N-k$, $k=1,\ldots,n$}\notag\\
-\frac 1{2k}, &\text{$m=N+k$, $k=1,\ldots,n$}\\
0, &\text{otherwise}\end{cases}
\end{equation}
\par Obviously,
$$Q_{N,n}(0)=0.$$
\par It is well known that the Fejer polynomials are uniformly bounded, i.e.
$$\|Q_{N,n}\|\leq C.$$
\par Also,
$$S_N(Q_{N,n},0)=\frac 1n+\cdots+1\sim\log n$$
and the sequences of the Fourier coefficients of $Q_{N,n}$ are of uniform bounded variation, i.e.
$$\sum_{m=-(N+n)}^{N+n+1}|\widehat{Q_{N,n}}(m-1)-\widehat{Q_{N,n}}(m)|\leq C.$$
\par Finally, we have the estimate
$$|S_k(Q_{N,n},t)|\leq\frac C{|t|}\qquad\text{for all}\,\,k\,\,\text{and}\,\,t\neq 0.$$
\par For all these properties see \cite{Ka}, Ch II, exercise 2.3.
\par Now, considering arbitrary $c\in\mathbb C$ and $\epsilon>0$ and setting
$$P_{N,n}=\frac c{S_N(Q_{N,n},0)}\,Q_{N,n},$$
it follows easily that if $N>n$ are large enough then
$$P_{N,n}(0)=0,$$
$$S_{N}(P_{N,n},0)=c,$$
$$\|P_{N,n}\|<\epsilon,$$
$$\sum_{m=-(N+n)}^{N+n+1}|\widehat{P_{N,n}}(m-1)-\widehat{P_{N,n}}(m)|\leq\epsilon,$$
$$|S_k(P_{N,n},t)|\leq\frac{\epsilon}{|t|}\qquad\text{for all}\,\,k\,\,\text{and}\,\,t\neq 0.$$
\par Let $\{c_j\,|\,j\in\mathbb N\}$ be a countable dense set in $\mathbb C$ and let $\epsilon, \epsilon_j>0$ with
$$\epsilon=\sum_{j=1}^{+\infty}\epsilon_j.$$
\par Now for each $j$ we can choose arbitrarily large $N_j>n_j$ and polynomials $P_{N_j,n_j}$ such that
\begin{equation}
P_{N_j,n_j}(0)=0,
\end{equation}
\begin{equation}
S_N(P_{N_j,n_j},0)=c_j,
\end{equation}
\begin{equation}
\|P_{N_j,n_j}\|<\epsilon_j,
\end{equation}
\begin{equation}
\sum_{m=-(N_j+n_j)}^{N_j+n_j+1}|\widehat{P_{N_j,n_j}}(m-1)-\widehat{P_{N_j,n_j}}(m)|\leq\epsilon_j.
\end{equation}
\begin{equation}
|S_k(P_{N_j,n_j},t)|\leq\frac{\epsilon_j}{|t|}\qquad\text{for all}\,\,k\,\,\text{and}\,\,t\neq 0.
\end{equation}
\par We also choose the $N_j>n_j$ to satisfy the inequalities
\begin{equation}
2N_j+(N_j+n_j)<2N_{j+1}-(N_{j+1}+n_{j+1})
\end{equation}
for all $j$ and we set
\begin{equation}
\begin{split}
P_j(t)&=e^{2iN_jt}P_{N_j,n_j}(t)\\
\mathcal B_j&=\{N_j-n_j,\ldots,3N_j+n_j\}\supseteq\spec(P_j).
\end{split}
\end{equation}
\par From (6) we have
\begin{equation}
\mathcal B_j\prec \mathcal B_{j+1},
\end{equation}
where $\prec$ means that the block $\mathcal B_j$ lies to the left of $\mathcal B_{j+1}$, that is $\max\mathcal B_j<\min\mathcal B_{j+1}$.
\par Hence (3) implies that the series
$$\sum_{j=1}^{+\infty}P_j=g$$
converges uniformly to a function $g\in C(\mathbb T)$ such that
$$\|g\|<\epsilon.$$
Also, from (7) and (8) we get
\begin{equation}
\begin{split}
\widehat{g}(n)=\begin{cases}\widehat{P_j}(n)=\widehat{P_{N_j,n_j}}(n-2N_j), &\text{if $n\in\mathcal B_j$, $j\in\mathbb N$}\\ 0, &\text{otherwise}\end{cases}
\end{split}
\end{equation}
In particular $\widehat{g}(n)=0$ for $n<0$ and thus $g$ belongs to $A(\mathbb D)$.
\par Also, (1) and (2) imply
\begin{equation}
S_{3N_j}(g,0)=c_j.
\end{equation}
Consequently, $g$ is universal on $K=\{0\}$.
\par Moreover, (4), (8) and (9) imply
$$\sum_{n=0}^{+\infty}|\widehat{g}(n-1)-\widehat{g}(n)|<\epsilon$$
i.e. the sequence $(\widehat{g}(n))$ is of bounded variation. Hence the Fourier series $\sum_{n=0}^{+\infty}\widehat{g}(n)e^{int}$ converges for $t\neq 0$
and uniformly in each closed interval of $\mathbb T$ which does not contain $0$. See \cite{Z}, Ch I, Theorem (2.6). Moreover, (5) implies
\begin{equation}
|S_k(P_j,t)|\leq\frac{\epsilon_j}{|t|}\qquad\text{for all}\,\,j,k\,\,\text{and}\,\,t\neq 0.
\end{equation}
\par From (7) and (8) and the uniform convergence of the series $\sum_{j=1}^{+\infty}P_j$ we get
$$S_n(g,t)\to g(t)\qquad\text{as}\,\,n\to+\infty\,\,\text{and}\,\,n\notin\bigcup_{j=1}^{+\infty}\mathcal B_j.$$
\par The previous constructions can be extended for any finite number of points $t_1,\ldots,t_m\in\mathbb T$. For simplicity we present the construction for two
points.
\par Let $t_1\neq t_2$ and $\{(a_j,b_j)\,|\,j\in\mathbb N\}$ be a countable dense set in $\mathbb C^2$. We consider the functions
$$f_1(t)=g_1(t-t_1)=\sum_{j=1}^{+\infty}P_j^{(1)}(t),\quad f_2(t)=g_2(t-t_2)=\sum_{j=1}^{+\infty}P_j^{(2)}(t),$$
where $g_1$ is the function $g$ constructed above with $(a_j)$ in place of $(c_j)$ and $g_2$ is the function $g$ with $(b_j)$ in place of $(c_j)$. Note that
the polynomials $P_j^{(1)},P_j^{(2)}$ have their spectrum in $\mathcal B_j$ and satisfy (11) with $t-t_1,t-t_2$ repsectively at the denominator of the right
side.
\par Relation (10) becomes
$$S_{3N_j}(f_1,t_1)=a_j,\qquad S_{3N_j}(f_2,t_2)=b_j.$$
\par Note that, since $N_j, n_j$ in the previous constructions can be chosen arbitrarily large, we may consider them to be the same for the functions $g_1,g_2$.
\par Then
$$f_1\in A(\mathbb D)\cap U(\{t_1\}),\qquad f_2\in A(\mathbb D)\cap U(\{t_2\}).$$
\par Also $S_n(f_1,t)\to f_1(t)$ for $t\neq t_1$ and $S_n(f_2,t)\to f_2(t)$ for $t\neq t_2$ and uniformly on every closed interval in $\mathbb T$ which
does not contain $t_1,t_2$ and $S_n(f_1,t)\to f_1(t)$ and $S_n(f_2,t)\to f_2(t)$ for every $t$ as $n\to+\infty$ and $n\notin\bigcup_{j=1}^{+\infty}\mathcal
B_j$.
\par Taking into account that the set $\{(a_j+f_2(t_1),b_j+f_1(t_2))\}$ is dense in $\mathbb C^2$ we get the following for the function $f=f_1+f_2$ or, more
generally, for the function $f=f_1+\cdots+f_m$ when $K=\{t_1,\ldots,t_m\}$.
\begin{theorem}
Let $K=\{t_1,\ldots,t_m\}$ be a finite set in $\mathbb T$. For every $\epsilon_j>0$ with $\sum_{j=1}^{+\infty}\epsilon_j<+\infty$ there are blocks $\mathcal
B_j$ in $\mathbb N$ such that $\mathcal B_1\prec\mathcal B_2\prec\ldots\,$ and corresponding polynomials $P_j$ with $\spec(P_j)\subseteq\mathcal B_j$ and
$\|P_j\|<\epsilon_j$ so that the function $f=\sum_{j=1}^{+\infty}P_j$ has $\spec(f)\subseteq\bigcup_{j=1}^{+\infty}\mathcal B_j$ and the following
properties:\newline
$(i)$ $f\in A(\mathbb D)\cap U(K)$,\newline
$(ii)$ $\|f\|\leq\epsilon=\sum_{j=1}^{+\infty}\epsilon_j$,\newline
$(iii)$ $S_n(f,t)\to f(t)$ for every $t\notin K$ and uniformly on every closed interval in $\mathbb T$ which does not intersect $K$.\newline
$(iv)$ $|S_k(P_j,t)|\leq\epsilon_j\sum_{l=1}^m\frac 1{|t-t_l|}$ for all $j,k$ and $t\neq t_1,\ldots,t_m$.\newline
$(v)$ $S_n(f,t)\to f(t)$ for every $t$ as $n\to+\infty$ and $n\notin\bigcup_{j=1}^{+\infty}\mathcal B_j$.
\end{theorem}
\par We note that the $\min\mathcal B_j$ can be taken arbitrarily large.
\par
\section{Pointwise universality on countably infinite $E\subseteq\mathbb T$ and convergence on $\mathbb T\setminus E$.}
\par Let $E=\{t_l\,|\,l\in\mathbb N\}\subseteq\mathbb T$ be a countably infinite set. We begin with the construction of a function $f\in A(\mathbb D)\cap
U_p(E)$.
\par Let $E_m=\{t_1,\ldots,t_m\}$. By Theorem 2.1 we have that for each $\epsilon_{m,j}>0$ with $\sum_{j=1}^{+\infty}\epsilon_{m,j}<+\infty$ there are blocks
$\mathcal B_{m,j}$ in $\mathbb N$ such that $\mathcal B_{m,1}\prec\mathcal B_{m,2}\prec\ldots$ and corresponding polynomials $P_{m,j}$ with
$\spec(P_{m,j})\subseteq\mathcal B_{m,j}$ and $\|P_{m,j}\|<\epsilon_{m,j}$ so that the
function
$$f_m=\sum_{j=1}^{+\infty}P_{m,j}$$
belongs to $f\in A(\mathbb D)\cap U(E_m)$ and satisfies
\begin{equation}
\|f_m\|\leq\epsilon_m=\sum_{j=1}^{+\infty}\epsilon_{m,j}
\end{equation}
and
\begin{equation}
|S_k(P_{m,j},t)|\leq\epsilon_{m,j}\sum_{l=1}^m\frac 1{|t-t_l|}\qquad\text{for all}\,\,k,j\,\,\text{and}\,\,t\neq t_1,\ldots,t_m.
\end{equation}
\par Since $\min\mathcal B_{m,j}$ can be taken arbitrarily large, we may take $\mathcal B_{m,j}$ in the following diagonal order:
$$\mathcal B_{1,1}\prec\mathcal B_{2,1}\prec\mathcal B_{1,2}\prec\mathcal B_{3,1}\prec\mathcal B_{2,2}\prec\mathcal B_{1,3}\prec\ldots\,.$$
We may also assume that
\begin{equation}
\sum_{m=1}^{+\infty}\epsilon_m<+\infty.
\end{equation}
\par We set
$$f=\sum_{m=1}^{+\infty}f_m=\sum_{m=1}^{+\infty}\sum_{j=1}^{+\infty}P_{m,j}.$$
\par By (12), (14) it follows that $f\in A(\mathbb D)$. We now prove that $f$ is pointwise universal on $E$.
\begin{theorem}
Let $f$ be the function constructed above.\newline
(a) $f\in A(\mathbb D)\cap U_p(E)$.\newline
(b) For each $t\in\mathbb T\setminus E$ satisfying the condition
\begin{equation}
\max_j\epsilon_{m,j}\sum_{l=1}^m\frac 1{|t-t_l|}\to 0
\end{equation}
we have
$$ S_n(f,t)\to f(t).$$
In particular, if $m\max_j\epsilon_{m,j}\to 0$ and the distance of $t$ from $E$ is positive, then $S_n(f,t)\to f(t)$.
\end{theorem}
\begin{proof}
(a) Let $h:E\to\mathbb C$ be an arbitrary function.
\par We first choose $m_1$ such that
$$\sum_{m=m_1+1}^{+\infty}\epsilon_m<\frac{\delta_1}3,$$
where $\delta_1=1$.
\par From the fact that the blocks $\mathcal B_{m,j}$ are mutually disjoint and from Theorem 2.1, we get
\begin{equation}
\begin{split}
S_n(f_m,t_l)\to f_m(t_l)\qquad\text{as}\,\,n&\to+\infty,\,\,n\in\bigcup_{j=1}^{+\infty}\mathcal B_{m_1,j},\\
&1\leq l\leq m_1,\,\,1\leq m\leq m_1-1.
\end{split}
\end{equation}
\par From (12) it follows that
\begin{equation}
\begin{split}
\sum_{m=m_1+1}^{+\infty}|S_n(f_m,t_l)|\leq\sum_{m=m_1+1}^{+\infty}&\sum_{j=1}^{+\infty}\epsilon_{m,j}<\frac{\delta_1}3,\\
&1\leq l\leq m_1,\,\,n\in\bigcup_{j=1}^{+\infty}\mathcal B_{m_1,j}.
\end{split}
\end{equation}
\par Also, from the universality of $f_{m_1}$ on $E_{m_1}=\{t_1,\ldots,t_{m_1}\}$ and from (15) we get that there exists $n_1\in\bigcup_{j=1}^{+\infty}\mathcal
B_{m_1,j}$ so that
\begin{equation}
\Big|S_{n_1}(f_{m_1},t_l)-\Big(h(t_l)-\sum_{m=1}^{m_1-1}f_m(t_l)\Big)\Big|<\frac{\delta_1}3,\qquad 1\leq l\leq m_1
\end{equation}
\begin{equation}
\begin{split}
\big|S_{n_1}(f_m,t_l)-f_m(t_l)\big|<&\frac{\delta_1}{3(m_1-1)},\\
& 1\leq l\leq m_1,\,\,1\leq m\leq m_1-1.
\end{split}
\end{equation}
\par Now we observe that the $n_1$-th Fourier sum of $f$ is a finite sum of $n_1$-th Fourier sums of the functions $f_1,f_2,\ldots,f_{m_1'}$ for some $m_1'\geq
m_1$. Hence
\begin{equation}
\begin{split}
S_{n_1}(f,t_l)&-h(t_l)\notag\\
&=\big(S_{n_1}(f_1,t_l)-f_1(t_l)\big)+\cdots+\big(S_{n_1}(f_{m_1-1},t_l)-f_{m_1-1}(t_l)\big)\notag\\
&\relphantom{=}{}+\Big(S_{n_1}(f_{m_1},t_l)-\Big(h(t_l)-\sum_{m=1}^{m_1-1}f_m(t_l)\Big)\Big)\notag\\
&\relphantom{=}{}+S_{n_1}(f_{m_1+1},t_l)+\cdots+S_{n_1}(f_{m_1'},t_l).
\end{split}
\end{equation}
\par Finally, from (17), (18), (19) we get
$$|S_{n_1}(f,t_l)-h(t_l)|<\delta_1,\qquad 1\leq l\leq m_1.$$
\par Similarly, for $\delta_2=\frac 12$ there exists $m_2>m_1$ such that $\sum_{m=m_2+1}^{+\infty}\epsilon_m<\frac{\delta_2}3$ and there exists $n_2>n_1$,
$n_2\in\bigcup_{j=1}^{+\infty}\mathcal B_{m_2,j}$ such that
$$|S_{n_2}(f,t_l)-h(t_l)|<\delta_2,\qquad 1\leq l\leq m_2.$$
\par Continuing in this manner, we construct strictly increasing sequences of positive integers $(m_N)$, $(n_N)$ such that
$$|S_{n_N}(f,t_l)-h(t_l)|<\delta_N=\frac 1N,\qquad 1\leq l\leq m_N.$$
\par This implies that
$$S_{n_N}(f,t)\to h(t),\qquad t\in E $$
and the proof of pointwise universality is complete.\newline
(b) By our diagonal ordering of the blocks $\mathcal B_{m,j}$ every $n\in\mathbb N$ lies in some diagonal $\mathcal B_{k,1},\mathcal
B_{k-1,2},\ldots,\mathcal B_{1,k}$. I.e. $\max\mathcal B_{1,k-1}<n\leq\max\mathcal B_{k,1}$ or $\max\mathcal B_{k-j+1,j}<n\leq\max\mathcal B_{k-j,j+1}$
for some $j=1,2,\ldots,k-1$.
\par Hence by the definition of $f$ we have
\begin{equation}
\begin{split}
S_n(f,t)&=P_{1,1}(t)+P_{2,1}(t)+P_{1,2}(t)+\cdots\notag\\
&\relphantom{=}{}+P_{k,1}(t)+P_{k-1,2}(t)+\cdots+P_{k-j+1,j}(t)+S_n(P_{k-j,j+1},t)
\end{split}
\end{equation}
if $n\in\mathcal B_{k-j,j+1}$. Note that the final term in the last sum is missing in case $\max\mathcal B_{k-j+1,j}\leq n<\min\mathcal B_{k-j,j+1}$.
\par Similarly, if $m>n$ then $m$ lies in some diagonal $\mathcal B_{p,1},\mathcal B_{p-1,2},\ldots,\mathcal B_{1,p}$ with $p\geq k$.
\par Therefore,
\begin{equation}
\begin{split}
|S_n(f,t)-S_m(f,t)|&\leq|P_{k-j,j+1}(t)-S_n(P_{k-j,j+1},t)|\notag\\
&\relphantom{\leq}{}+\big(|P_{k-j-1,j+2}(t)|+\cdots+|P_{p-q+1,q}(t)|\big)\notag\\
&\relphantom{\leq}{}+|S_m(P_{p-q,q+1},t)|
\end{split}
\end{equation}
\par Now (13) implies
$$|S_m(P_{p-q,q+1},t)|\leq\epsilon_{p-q,q+1}\sum_{l=1}^{p-q}\frac 1{|t-t_l|}$$
and, together with (12),
$$|P_{k-j,j+1}(t)-S_n(P_{k-j,j+1},t)|\leq\epsilon_{k-j,j+1}+\epsilon_{k-j,j+1}\sum_{l=1}^{k-j}\frac 1{|t-t_l|}.$$
\par Hence
\begin{equation}
\begin{split}
|S_n(f,t)-S_m(f,t)|&\leq\epsilon_{k-j,j+1}+\epsilon_{k-j,j+1}\sum_{l=1}^{k-j}\frac 1{|t-t_l|}\notag\\
&\relphantom{\leq}{}+\epsilon_{k-j-1,j+2}+\cdots+\epsilon_{p-q+1,q}\notag\\
&\relphantom{\leq}{}+\epsilon_{p-q,q+1}\sum_{l=1}^{p-q}\frac 1{|t-t_l|}
\end{split}
\end{equation}
\par If $n,m\to+\infty$ then $k,p\to+\infty$ and since the double series $\sum_{r,s}\epsilon_{r,s}$ converges, we get
$\epsilon_{k-j,j+1}+\epsilon_{k-j-1,j+2}+\cdots+\epsilon_{p-q+1,q}\to 0$.
\par Regarding the term $r=\epsilon_{k-j,j+1}\sum_{l=1}^{k-j}\frac 1{|t-t_l|}$ we have two cases. If $k-j$ is bounded then $j+1\to+\infty$ and thus $r\to 0$.
If $k-j\to+\infty$ then (15) implies $r\to 0$. Similarly, $\epsilon_{p-q,q+1}\sum_{l=1}^{p-q}\frac 1{|t-t_l|}\to 0$.
\par We conclude that $(S_n(f,t))$ converges. The proof will be complete if we show that $S_n(f,t)\to f(t)$ when $n\to+\infty$ and does not belong to any of
the blocks $\mathcal B_{m,j}$. Indeed, in this case $(v)$ of Theorem 2.1 implies
$$S_n(f,t)=\sum_{m=1}^{+\infty}S_n(f_m,t)\to\sum_{m=1}^{+\infty}f_m(t)=f(t)$$
due to the uniform bound $\sum_{m=1}^{+\infty}|S_n(f_m,t)|\leq\sum_{m=1}^{+\infty}\epsilon_m$.
\end{proof}
\par As an application we get the following for the function $f$ of Theorem 3.1.
\begin{theorem}
Let $(\delta_m)$ be a decreasing sequence of positive numbers such that $\sum_{m=1}^{+\infty}\delta_m^a<+\infty$ for all $a>0$. Now, if
$\max_j\epsilon_{m,j}\sum_{l=1}^m\frac 1{\delta_l}\to 0$, then we have $S_n(f,t)\to f(t)$ outside a set of Hausdorff dimension zero.
\end{theorem}
\begin{proof}
We consider
$$I_m=(t_m-\delta_m,t_m+\delta_m),\qquad D=\bigcap_{m=1}^{+\infty}\bigcup_{l=m}^{+\infty}I_l.$$
\par Since $\sum_{m=1}^{+\infty}\delta_m^a<+\infty$ for all $a>0$, the set $D$ is of Hausdorff dimension zero.
\par Now, if $t$ is not in the countable set $E$ neither in $D$, then it satisfies (15). Indeed, let $t\notin E$ and $t\notin\bigcup_{l=m_0}^{+\infty}I_l$ for
some $m_0$. Then if $m\geq m_0$ we have
$$\max_j\epsilon_{m,j}\sum_{l=1}^m\frac 1{|t-t_l|}\leq\max_j\epsilon_{m,j}\sum_{l=1}^{m_0-1}\frac 1{|t-t_l|}+\max_j\epsilon_{m,j}\sum_{l=m_0}^m\frac
1{|t-t_l|}.$$
\par The first term of the right side obviously tends to $0$ as $m\to+\infty$. As for the second term we have
$$\max_j\epsilon_{m,j}\sum_{l=m_0}^m\frac 1{|t-t_l|}\leq\max_j\epsilon_{m,j}\sum_{l=m_0}^m\frac 1{\delta_l}\leq\max_j\epsilon_{m,j}\sum_{l=1}^m\frac
1{\delta_l}$$
which by our hypothesis tends to $0$ as $m\to+\infty$.
\end{proof}
\par Note that, if $E$ is dense in $\mathbb T$, the set of divergence of $S_n(f,t)$ is necessarily $G_{\delta}$ and dense and hence uncountable, although by
the proper choice of $f$ it can be of Hausdorff dimension zero.
\section{Subsets of $\mathbb T$ not accepting uniform universality.}
\par Let $K$ be a compact subset of $\mathbb T$, $K\neq\mathbb T$. In Theorem 2.2 of \cite{KNP} and in \cite{N} it was shown that $H(\mathbb D)\cap
U(K)$ is a dense-$G_{\delta}$ subset of $H(\mathbb D)$ when the latter has the topology of uniform convergence on compacta.
\par We now prove a special case of Theorem 4.3 of \cite{KNP} replacing the condition appearing there with a much simpler one (number (20) in what follows).
\begin{proposition}
(a) Let $K$ be a compact proper subset of $\mathbb T$ and $t_0\in K$ with the following property:
\begin{equation}
\begin{split}
&\text{for each infinite}\,\,\mathbb M\subseteq\mathbb N\,\,\text{there are}\,\,a,b\,\,\text{with}\,\,0<a<b<2\pi\\
&\text{and}\,\,[a/m,b/m]\cap(K-t_0)\neq\emptyset\,\,\text{for infinitely many}\,\,m\in\mathbb M.
\end{split}
\end{equation}
If $f\in H(\mathbb D)\cap U(K)$, then the Taylor series of $f$ is not (C,1) summable at $t_0$.\newline
(b) Now let $K$ and $t_0$ be as in (a) and moreover let $K$ be symmetric with respect to $t_0$, i.e. $2t_0-t\in K$ for every $t\in K$.\newline
If $f\in L^1(\mathbb T)\cap U(K)$, then the Fourier series of $f$ is not (C,1) summable at $t_0$.
\end{proposition}
\begin{proof}
(a) We denote by $\sum_{k=0}^{+\infty}c_kz^k$ the Taylor series of $f$ and by $S_n(f,t)=\sum_{k=0}^nc_ke^{ikt}$ the partial sums on $\mathbb T$ and let us
assume that the Taylor series is (C,1) summable at $t_0$.
\par Since $f\in U(K)$, for every $h\in C(K)$ there is a subsequence $S_{n_j}(f,t)$ which converges to $h$ uniformly on $K$. Property (20) implies that there
is a subsequence of $(n_j)$, which without loss of generality we may assume equal to $(n_j)$, and corresponding $\theta_j\in K-t_0$ such that $a\leq
n_j\theta_j\leq b$ for all $j$.
\par Then Theorem (12.16) of \cite{Z}, Vol I, Ch III implies
$$S_{n_j}(f,t_0+\theta_j)-(S_{n_j}(f,t_0)-s)e^{in_j\theta_j}\to s,$$
where $s$ is the (C,1) sum of the Taylor series of $f$ at $t_0$.
\par Taking a further subsequence of $n_j$, we may assume that $n_j\theta_j\to\phi$ for some $\phi\in[a,b]$. Now the uniform convergence $S_{n_j}(f,t)\to
h(t)$ on $K$ implies
$$h(t_0)-(h(t_0)-s)e^{i\phi}=s.$$
Choosing $h(t_0)\neq s$ we arrive at a contradiction.\newline
(b) The proof is the same as before. If $S_n(f,t)=\sum_{k=-n}^n\widehat f(k)e^{ikt}$ are the partial sums of the Fourier series of $f$, we assume that their
(C,1) means at $t_0$ converge to some $s$. We consider $h\in C(K)$, the sequence $S_{n_j}(f,t)$ converging to $h$ uniformly on $K$ and the corresponding
$\theta_j$ so that $\pm\theta_j\in K-t_0$ and $a\leq n_j\theta_j\leq b$ for all $j$.
\par We now apply Theorem (12.9) of \cite{Z}, Vol I, Ch III and get
$$\frac 12\big(S_{n_j}(f,t_0+\theta_j)+S_{n_j}(f,t_0-\theta_j)\big)-(S_{n_j}(f,t_0)-s)\cos n_j\theta_j\to s.$$
Exactly as before we get $h(t_0)-(h(t_0)-s)\cos\phi=s$ and we arrive at a contradiction choosing $h(t_0)\neq s$.
\end{proof}
\par We note that property (20) can be stated with ``left hand'' intervals $[t_0-(b/m),t_0-(a/m)]$ and the result of Proposition 4.1(a) remains unchanged.
\begin{corollary}
If $K$ is a compact proper subset of $\mathbb T$ which satifies property (20) for some $t_0\in K$ then $A(\mathbb D)\cap U(K)=\emptyset$. If moreover
$K$ is symmetric with respect to $t_0$ then $C(\mathbb T)\cap U(K)=\emptyset$.\newline
In particular the one third Cantor set $C\subseteq[0,\pi]$ satisfies property (20) for $t_0=0\in C$ and hence $A(\mathbb D)\cap U(C)=\emptyset$. For the
symmetric Cantor set $C^*=C\cup(-C)\subseteq[-\pi,\pi]$ we have that $C(\mathbb T)\cap U(C^*)=\emptyset$.
\end{corollary}
\begin{proof}
Since the Fourier series of any $f\in A(\mathbb D)$ is (C,1) summable at every point of $\mathbb T$, the first statement is obvious.
\par Regarding the Cantor set $C$ and $t_0=0\in C$ we shall prove a stronger version of property (20): if $0<a<b<2\pi$ and $\frac ba>2$, then
$[\frac an,\frac bn]\cap K\neq\emptyset$ for every $n\in\mathbb N$, $n\geq 3b$.
\par We take $N\in\mathbb N$ so that $3^N\leq\frac nb<3^{N+1}$ and then $(\frac bn)/(\frac an)>2$ implies that the interval $(\frac an,\frac bn)$ contains at
least one of the points $\frac 1{3^{N+1}}$ and $\frac 2{3^{N+1}}$ of $C$.
\end{proof}
\par In fact it is not very difficult to see that the one third Cantor set satisfies property (20) for every $t_0\in C$.
\par The question whether $C(\mathbb T)\cap U(C)=\emptyset$ is true is only slightly more complicated. The one third Cantor set $C\subseteq[0,\pi]$ is not
symmetric with respect to $0$ and in fact it is not symmetric with respect to any of its points. Nevertheless $C(\mathbb T)\cap U(C)=\emptyset$ is true.
\begin{proposition}
Let $K$ be a compact proper subset of $\mathbb T$ and $t_0\in K$ with the following property:
\begin{equation}
\begin{split}
&\text{for each infinite}\,\,\mathbb M\subseteq\mathbb N\,\,\text{there are}\,\,t_m\in K\,\,\text{for all}\,\,m\in\mathbb M\\
&\text{so that}\,\,t_m\to t_0\,\,\text{and there are}\,\,a,b\,\,\text{with}\,\,0<a<b<2\pi\\
&\text{and}\,\,[a/m,b/m]\cap(K-t_m)\cap(t_m-K)\neq\emptyset\\
&\text{for infinitely many}\,\,m\in\mathbb M.
\end{split}
\end{equation}
If $f\in L^1(\mathbb T)\cap U(K)$, then the (C,1) means of the Fourier series of $f$ do not converge uniformly on $K$.
\end{proposition}
\begin{proof}
The proof is a variation of the proof of Proposition 4.1.
\par Let $S_n(f,t)=\sum_{k=-n}^n\widehat f(k)e^{ikt}$ be the partial sums of the Fourier series of $f$, and assume that their (C,1) means converge to some
{\it function} $s$ uniformly on $K$. We consider an arbitrary $h\in C(K)$ and a sequence $S_{n_j}(f,t)$ converging to $h$ uniformly on $K$.
\par Now there are $t_j\in K$ so that $t_j\to t_0$ and (considering a subsequence) corresponding $\theta_j$ so that $t_j\pm\theta_j\in K$ and $a\leq
n_j\theta_j\leq b$ for all $j$.
\par We apply again Theorem (12.9) of \cite{Z}, Vol I, Ch III and get
$$\frac 12\big(S_{n_j}(f,t+\theta_j)+S_{n_j}(f,t-\theta_j)\big)-(S_{n_j}(f,t)-s(t))\cos n_j\theta_j\to s(t)$$
uniformly for $t\in K$. In fact here we apply the proof rather than the statement of the above Theorem (12.9) of \cite{Z}. One should read the small paragraph a
few lines before the statement of Theorem (12.6): ``we also observe that if the terms of $u_0+u_1+\ldots$\, depend on a parameter, and if the hypotheses
concerning this series are satisfied uniformly, the conclusions also hold uniformly''. The relevant hypothesis in our case is the uniform (C,1) summability of
the Fourier series of $f$.
\par Since $t_j\to t_0$ and $t_j\pm\theta_j\to t_0$ and since we may assume that $n_j\theta_j\to\phi$ for some $\phi\in[a,b]$, we finally get
$$h(t_0)-(h(t_0)-s)e^{i\phi}=s$$
and, choosing $h(t_0)\neq s$, we arrive at a contradiction.
\end{proof}
\begin{corollary}
If $K$ is a compact proper subset of $\mathbb T$ which satifies property (21) for some $t_0\in K$ then $C(\mathbb T)\cap U(K)=\emptyset$.\newline
In particular the one third Cantor set $C\subseteq[0,\pi]$ satisfies property (21) for every $t_0\in C$ and hence $C(\mathbb T)\cap U(C)=\emptyset$.
\end{corollary}
\begin{proof}
For the first statement we recall that if $f\in C(\mathbb T)$ then the (C,1) means of the Fourier series of $f$ converge to $f$ uniformly on $\mathbb T$.
\par The one third Cantor set satisfies property (21) in a slightly stronger form: there are $a,b$ with $0<a<b<2\pi$ so that for every $t_0\in C$ there are
$t_n\in C$ for all $n\in\mathbb N$ with $t_n\to t_0$ and $[\frac an,\frac bn]\cap(K-t_n)\cap(t_n-K)\neq\emptyset$ for every $n\in\mathbb N$.
\par In fact, take arbitrary $a,b$ with $0<a<b<2\pi$ and $\frac ba>3$ and any $t_0\in C$. For each $N\in\mathbb N$ we consider the interval $I_N$ of length
$\frac 1{3^N}$ which appears at the $N$-th step of the construction of $C$ and which contains $t_0$. If $I_{N,-}$ and $I_{N,+}$ are the left and right
subintervals of $I_N$ of length $\frac 1{3^{N+1}}$, then $t_0$ belongs to one of them (and this is $I_{N+1}$). If $t_0\in I_{N,-}$ then we define $t_n$ to be
the right endpoint of $I_{N,-}$ for every $n$ with $3^Nb\leq n<3^{N+1}b$. If $t_0\in I_{N,+}$ then we define $t_n$ to be the left endpoint of $I_{N,+}$ for
every $n$ with $3^Nb\leq n<3^{N+1}b$. In the case $t_0\in I_{N,-}$ the three points $t_n$ and $t_n\pm\frac 1{3^{N+1}}$ belong to $C$ and $\frac
1{3^{N+1}}\in(\frac an,\frac bn)$ since $3^{N+1}a<3^Nb\leq n<3^{N+1}b$. The case $t_0\in I_{N,+}$ is similar.
\end{proof}
\par It is not hard to construct countably infinite sets satistfying (20). Take for example $K=\{0\}\cup\{\frac 1n\,|\,n\in\mathbb N\}$. Hence,
\begin{corollary}
There are countably infinite compact $K\subseteq\mathbb T$ such that $A(\mathbb D)\cap U(K)=\emptyset$.
\end{corollary}
\par We close with the following open problems.
\par I. {\it Construct a perfect set $K$ and a function $f\in A(\mathbb D)\cap U(K)$}.
\par II. {\it Study pointwise universality on uncountable sets $E$ with Lebesgue measure zero for functions in $C(\mathbb T)$ or $A(\mathbb D)$}.\newline
Of course the functions $g$ in Definition 1.1 of pointwise universality must belong to the first class of Baire.
\par III. {\it Is it true that, if $E$ has positive Hausdorff dimension, then $E$ does not accept universality?}
\par {\bf Acknowledgement}. We would like to thank the referee for valuable comments which cleared certain subtle points of the proofs and helped us improve the
presentation of this paper. Proposition 4.1 in our original paper contained nothing about sets $K$ symmetric with respect to $t_0$ and hence our original
Corollary 4.1 considered only the result $A(\mathbb D)\cap U(C)=\emptyset$. The referee asked whether we could prove that $C(\mathbb T)\cap U(C)=\emptyset$ and
this prompted us to work first with symmetric sets $K$ and get that $C(\mathbb T)\cap U(C^*)=\emptyset$. After a while we proved the stronger $C(\mathbb T)\cap
U(C)=\emptyset$. We include the weaker result since it shows that we do not need extra hypotheses (i.e. the uniform (C,1) summability) in the presence of the
symmetry of the set $K$.
|
\section{Introduction}
Let $G = (V, E)$ denote a simple undirected graph and let $k$ be a positive integer. A \emph{$k$-colouring} of~$G$ is a mapping $\phi: V \rightarrow \{1, \dots, k\}$ such that $\phi(u) \not= \phi(v)$ if $uv \in E$. The \emph{chromatic number} of $G$, denoted by $\chi(G)$, is the smallest $k$ such that $G$ has a $k$-colouring.
If $a$ and $b$ are distinct colours, then $G(a, b)$ denotes the subgraph of $G$ induced by vertices with colour $a$ or $b$. An \emph{$(a, b)$-component} of $G$ is a connected component of $G(a, b)$ and is known as a \emph{Kempe chain}. A \emph{Kempe change} is the operation of interchanging the colours of some $(a, b)$-component of $G$. Let $C_k(G)$ be the set of all $k$-colourings of $G$. Two colourings $\alpha, \beta \in C_k(G)$ are \emph{Kempe equivalent}, denoted by $\alpha \sim_k \beta$, if each can be obtained from the other by a series of Kempe changes. The equivalence classes $C_k(G)/\sim_k$ are called \emph{Kempe classes}.
Kempe changes were first introduced by Kempe in his well-known failed attempt at proving the Four-Colour Theorem. The Kempe change method has proved to be a powerful tool with applications to several areas such as timetables~\cite{rolf}, theoretical physics~\cite{wang1, wang2}, and Markov chains~\cite{vigoda}. The reader is referred to~\cite{mohar1, sokal} for further details. From a theoretical viewpoint, Kempe equivalence was first addressed by Fisk~\cite{fisk} who proved that all $4$-colourings of an Eulerian triangulation of the plane are Kempe equivalent. This result was later extended by Meyniel~\cite{meyniel1} who showed that all $5$-colourings of a planar graph are Kempe equivalent, and by Mohar~\cite{mohar1} who proved that all $k$-colourings, $k > \chi(G)$, of a planar graph $G$ are Kempe equivalent. Las Vergnas and Meyniel~\cite{meyniel3} extended Meyniel's result by proving that all $5$-colourings of a $K_5$-minor free graph are Kempe equivalent. Bertschi~\cite{marc} also showed that all $k$-colourings of a perfectly contractile graph are Kempe equivalent, thus answering a conjecture of Meyniel~\cite{meyniel2}. We note that Kempe equivalence with respect to edge-colourings has also been investigated~\cite{mohar1, mohar2, ruth}.
Here we are concerned with a conjecture of Mohar~\cite{mohar1} on {\it $k$-regular} graphs, that is, graphs in which every vertex has degree $k$ for some $k\geq 0$. Note that, for every 2-regular graph $G$ that is not an odd cycle, it holds that $C_2(G)$ is a Kempe class. Mohar conjectured the following (where $K_{k+1}$ is the complete graph on $k+1$ vertices).
\begin{conjecture1}[\cite{mohar1}]\label{c-1}
Let $k\geq 3$. If $G$ is a $k$-regular graph that is not $K_{k+1}$ then $C_k(G)$ is a Kempe class.
\end{conjecture1}
\tikzstyle{vertex}=[circle,draw=black, fill=black, minimum size=5pt, inner sep=1pt]
\tikzstyle{edge} =[draw,-,black,>=triangle 90]
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.75]
\foreach \pos/\name in {{(0,2)/p1}, {(4.55,2)/p2}, {(4.55,0)/p3}, {(0,0)/p4}, {(3.14,1)/p5}, {(1.41,1)/p6}}
\node[vertex] (\name) at \pos {};
\foreach \source/ \dest in {p1/p2, p2/p3,p3/p4, p2/p5, p3/p5, p5/p6, p1/p6, p1/p4, p4/p6}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\end{tikzpicture}
\end{center}
\caption{The 3-prism.}\label{fig:0}
\end{figure}
\noindent We address Conjecture~\ref{c-1} for the case $k = 3$. For this case the conjecture is known to be false. A counter-example is the 3-prism displayed in Figure~\ref{fig:0}. The fact that some $3$-colourings of the $3$-prism are not Kempe equivalent was already observed by van den Heuvel~\cite{heuvel}. Our contribution is that the 3-prism is the {\it only} counter-example for the case $k=3$, that is, we completely settle the case $k=3$ by proving the following result for 3-regular graphs also known as {\it cubic} graphs.
\begin{theorem}\label{mainthm}
If $G$ is a cubic graph that is neither $K_4$ nor the $3$-prism then $C_3(G)$ is a Kempe class.
\end{theorem}
\noindent
We give the proof of our result in the next section. Besides exploiting the 3-regularity, our proof also takes into account the fact that one additional graph, namely the 3-prism, is forbidden. We did not find any counter-examples for $k\geq 4$ and believe Conjecture~\ref{c-1} may well hold for $k\geq 4$. As such, new techniques are necessary to tackle the remaining cases.
Our result is an example of a type of result that has received much recent attention: that of determining the structure of a \emph{reconfiguration graph}. A reconfiguration graph has as vertex set all solutions to a search problem and an edge relation that describes a transformation of one solution into another. Thus Theorem~\ref{mainthm} is concerned with the reconfiguration graph of 3-colourings of cubic graph with edge relation $\sim_k$ and shows that it is connected except in two cases. To date the stucture of reconfiguration graphs of colourings has focussed~\cite{BB13, BJLPP14, BP14, CHJ06, CHJ06b, FJP14} on the case where vertices are joined by an edge only when they differ on just one colour (that is, when one colouring can be transformed into another by a Kempe change of a Kempe chain that contains only one vertex). For a survey of recent results on reconfiguration graphs see~\cite{heuvel}.
\section{The Proof of Theorem~\ref{mainthm}}\label{cubic}
We first give some further definitions and terminology. Let $G=(V,E)$ be a graph. Then~$G$ is {\it $H$-free} for some graph~$H$ if $G$ does not contain an induced subgraph isomorphic to $H$. A \emph{separator} of $G$ is a set $S \subset V$ such that $G - S$ has more components than~$G$. We say that $G$ is \emph{$p$-connected} for some integer~$p$ if $|V|\geq p+1$ and every separator of $G$ has size at least~$p$. Some small graphs that we will refer to are defined by their illustrations in Figure~\ref{fig:1}.
\tikzstyle{vertex}=[circle,draw=black, fill=black, minimum size=5pt, inner sep=1pt]
\tikzstyle{edge} =[draw,-,black,>=triangle 90]
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=1]
\foreach \pos/\name / \label / \posn / \dist in {{(1,0)/c1/u/above/2}, {(1,-1.2)/c2/w/{right}/1}, {(1.7,-2)/c3/v/{right}/1}, {(0.3,-2)/c4/s/{left}/1}}
{ \node[vertex] (\name) at \pos {};
\node [\posn=\dist] at (\name) {$\label$};
}
\foreach \source/ \dest in {c2/c1, c2/c4, c2/c3}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\foreach \pos/\name / \label / \posn / \dist in {{(4,0)/h1/x/{above}/2}, {(3.3,-0.8)/h2/y/{left}/1},
{(4.7,-0.8)/h3/z/{right}/1}, {(3.3,-2)/h4/w/{left}/1},{(4.7,-2)/h5/s/{right}/1}}
{ \node[vertex] (\name) at \pos {};
\node [\posn=\dist] at (\name) {$\label$};
}
\foreach \source/ \dest in {h1/h2, h1/h3, h2/h3, h2/h4, h4/h5, h5/h3}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\foreach \pos/\name / \label in {{(7.2,0)/n1/x}, {(7.2,-2)/n2/y}, {(6.3,-1)/n3/w}, {(8.1,-1)/n4/z}}
{ \node[vertex] (\name) at \pos {};}
\foreach \source/ \dest in {n1/n3, n1/n4, n2/n3, n2/n4, n3/n4}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\node at (1,-2.6) { claw};
\node at (4,-2.6) { house};
\node at (7.2,-2.6) {diamond};
\begin{scope}[xshift=3cm]
\foreach \pos/\name / \label / \posn / \dist in {{(6.3,-2)/n1/y'/{left}/1}, {(6.8,-1.5)/n2/y/{left}/1}, {(7.2,-0.8)/n3/x/{right}/1}, {(7.2,0)/n4/x'/{right}/1}, {(7.6,-1.5)/n5/z/{right}/1}, {(8.1,-2)/n6/z'/{right}/1}}
{ \node[vertex] (\name) at \pos {};
\node [\posn=\dist] at (\name) {$\label$};
}
\foreach \source/ \dest in {n1/n2, n2/n3, n3/n4, n2/n5, n3/n5, n5/n6}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\node at (7.2,-2.6) { net};
\end{scope}
\end{tikzpicture}
\end{center}
\caption{A number of special graphs used in our paper.}\label{fig:1}
\end{figure}
Besides three new lemmas, we will need the aforementioned result of van den Heuvel, which follows from the fact that for the 3-prism $T$, the subgraphs$T(1, 2)$, $T(2, 3)$ and $T(1, 3)$ are connected so that the number of Kempe classes is equal to the number of different 3-colourings of $T$ up to colour permutation, which is two.
\begin{lemma}[\cite{heuvel}]\label{pprism}
If $G$ is the 3-prism then $C_3(G)$ consists of two Kempe classes.
\end{lemma}
\begin{lemma}\label{l1}
If $G$ is a cubic graph that is not $3$-connected then $C_3(G)$ is a Kempe~class.
\end{lemma}
\begin{lemma}\label{l2}
If $G$ is a $3$-connected cubic graph that is claw-free but that is neither $K_4$ nor the $3$-prism then $C_3(G)$ is a Kempe class.
\end{lemma}
\begin{lemma}\label{l3}
If $G$ is a $3$-connected cubic graph that is not claw-free then $C_3(G)$ is a Kempe class.
\end{lemma}
Observe that Theorem~\ref{mainthm} follows from the above lemmas, which form a case distinction. Hence it suffices to prove Lemmas~\ref{l1}--\ref{l3}. These proofs form the remainder of the paper.
\subsection{Proof of Lemma~\ref{l1}}\label{s-lemma1}
In order to prove Lemma~\ref{l1} we need three auxiliary results and one more definition: a graph $G$ is {\it $d$-degenerate} if every induced subgraph of $G$ has a vertex with degree at most $d$.
\begin{lemma}[\cite{meyniel3, mohar1}]\label{degprop1}
Let $d$ and $k$ be any two integers with $d\geq 0$ and $k\geq d+1$. If $G$ is a $d$-degenerate graph then $C_k(G)$ is a Kempe class.
\end{lemma}
\begin{lemma}[\cite{meyniel3}]\label{ldeg}
Let $k\geq 1$ be an integer. Let $G_1, G_2$ be two graphs such that $G_1 \cap G_2$ is complete. If both $C_k(G_1)$ and $C_k(G_2)$ are Kempe classes then $C_k(G_1 \cup G_2)$ is a Kempe class.
\end{lemma}
\begin{lemma}[\cite{mohar1}]\label{ledge}
Let $k\geq 1$ be an integer and let $G$ be a subgraph of a graph $G'$. Let~$c_1$ and $c_2$ be the restrictions, to $G$, of two $k$-colourings $c_1'$ and $c_2'$ of $G'$. If $c_1'$ and $c_2'$ are Kempe equivalent then $c_1$ and $c_2$ are Kempe equivalent.
\end{lemma}
For convenience we restate Lemma~\ref{l1} before we present its proof.
\medskip
\noindent
{\bf Lemma~\ref{l1} (restated).}
{\it If $G$ is a cubic graph that is not $3$-connected then $C_3(G)$ is a Kempe~class.}
\medskip
\noindent {\it Proof.}
As disconnected graphs can be considered component-wise, we assume that $G$ is connected. As $G$ is cubic, $G$ has at least four vertices. Because $G$ is not 3-connected, $G$ has a separator $S$ of size at most~2. Let $S$ be a minimum separator of $G$ such that $G =G_1 \cup G_2$ and $G_1 \cap G_2=S$. As every vertex in $S$ has degree at most~2 in each $G_i$ and $G$ is cubic, $G_1$ and $G_2$ are $2$-degenerate. Hence, by Lemma~\ref{degprop1}, $C_3(G_1)$ and $C_3(G_2)$ are Kempe classes. If $S$ is a clique, we apply Lemma~\ref{ldeg}. Thus we assume that $S$, and any other minimum separator of $G$, is not a clique. Then $S=\{x, y\}$ for two distinct vertices $x$ and $y$ with $xy \not\in E(G)$.
Because $S$ is a minimum separator, $x$ and $y$ are non-adjacent and $G$ is cubic, $x$ has either one neighbour in $G_1$ and two in $G_2$, or the other way around; the same holds for vertex~$y$. For $i=1,2$, let $N_i(x)$ and $N_i(y)$ be the set of neighbours of $x$ and~$y$, respectively, in $G_i$. Then we have that either $|N_1(x)|=1$ and $|N_2(x)|=2$, or $|N_1(x)|=2$ and $|N_2(x)|=1$, and similarly, that either $|N_1(y)|=1$ and $|N_2(y)|=2$, or $|N_1(y)|=2$ and $|N_2(y)|=1$. Let $x_1\in N_1(x)$ for some $x_1\in V(G_1)$.
We may assume that $|N_1(x)|\neq |N_1(y)|$; if not we can do as follows. Assume without loss of generality that $N_1(x)=\{x_1\}$ and that $|N_1(y)|=1$. Then $\{x_1,y\}$ is a separator. By our assumption that $G$ has no minimum separator that is a clique, we find that $\{x_1,y\}$ is a minimum separator with $x_1y\notin E(G)$. As $G$ is cubic, $x_1$ has two neighbours in $V(G_1)\setminus \{x,x_1\}$. As $|N_1(y)|=1$ and $x_1$ and $y$ are not adjacent,~$y$ has exactly one neighbour in $V(G_1)\setminus \{x,x_1\}$. Hence we could take $\{x_1,y\}$ as our minimum separator instead of~$S$ in order to get the desired property. We may thus assume that $|N_1(x)|\not=|N_1(y)|$. As this means that $|N_2(x)|\not=|N_2(y)|$, we can let $N_1(x) = \{x_1\}$ and $N_2(y)=\{y_1\}$ for some $y_1\in V(G_2)$.
It now suffices to prove the following two claims.
\medskip
\noindent
{\it Claim 1. All colourings $\alpha$ such that $\alpha(x) \not=\alpha(y)$ are Kempe equivalent in $C_3(G)$. }
\medskip
\noindent
We prove Claim~1 as follows. We add an edge $e$ between $x$ and $y$. This results in graphs $G_1+e$, $G_2+e$ and $G+e$. We first prove that $C_3(G+e)$ is a Kempe class. Because $x$ and $y$ have degree~1 in $G_1$ and $G_2$, respectively, and $G$ is cubic, we find that the graphs $G_1 +e$ and $G_2 +e$ are $2$-degenerate. Hence, by Lemma~\ref{degprop1}, $C_3(G_1+e)$ and $C_3(G_2+e)$ are Kempe classes. By Lemma~\ref{ldeg}, it holds that $C_3(G+e)$ is a Kempe class. Applying Lemma~\ref{ledge} completes the proof of Claim~1.
\medskip
\noindent
{\it Claim 2. For every colouring $\alpha$ such that $\alpha(x) = \alpha(y)$, there exists a colouring $\beta$ with $\beta(x) \not= \beta(y)$ such that $\alpha$ and $\beta$ are Kempe equivalent in $C_3(G)$.}
\medskip
\noindent
We assume without loss of generality that $\alpha(x) = \alpha(y) = 1$ and $\alpha(y_1) = 2$. If $\alpha(x_1)=2$ then we apply a Kempe change on the $(1,3)$-component of $G$ that contains $x$. Note that $y$ does not belong to this component. Hence afterwards we obtain the desired colouring $\gamma$. If $\alpha(x_1)=3$ then we first apply a Kempe change on the $(2,3)$-component of $G$ that contains $x_1$. Note that this does not affect the colours of $x$, $y$ and~$y_1$ as they do not belong to this component. Afterwards we proceed as before. This completes the proof of Claim~2 (and the lemma).\qed
\subsection{Proof of Lemma~\ref{l2}}\label{s-l2}
We require some further terminology and three lemmas. We {\it identify} two vertices $x$ and~$y$ in a graph~$G$ if we replace them by a new vertex adjacent to all neighbours of $x$ and $y$ in~$G$. Two colourings $\alpha$ and $\beta$ of a graph $G$ \emph{match} if there exists two vertices $x, y$ with a common neighbour in $G$ such that $\alpha(x) = \alpha(y)$ and $\beta(x) = \beta(y)$.
\begin{lemma}\label{lplanar}
Let $k\geq 1$ and $G'$ be the graph obtained from a graph $G$ by identifying two non-adjacent vertices $x$ and $y$. If $C_k(G')$ is a Kempe class then all $k$-colourings $c$ of $G$ with $c(x) = c(y)$ are Kempe equivalent.
\end{lemma}
\begin{proof}
Let $\alpha$ and $\beta$ be two $k$-colourings of $G$ with $\alpha(x)=\alpha(y)$ and $\beta(x)=\beta(y)$. Let~$z$ be the vertex of $G'$ that is obtained after identifying $x$ and $y$. Let $\alpha'$ and $\beta'$ be the $k$-colourings of $G'$ that agree with $\alpha$ and $\beta$, respectively, on $V(G)\setminus \{x,y\}$ and for which $\alpha'(z)=\alpha(x)(=\alpha(y))$ and $\beta'(z)=\beta(x)(=\beta(y))$. By our assumption, there exists a Kempe chain from $\alpha'$ to $\beta'$ in $G'$. We mimic this Kempe chain in $G$. Note that any $(a,b)$-component in $G'$ that contains $z$ corresponds to at most two $(a,b)$-components in~$G$, as $x$ and $y$ may get separated. Hence, every Kempe change on an $(a,b)$-component corresponds to either one or two Kempe changes in $G$ (if $x$ and $y$ are in different $(a,b)$-components then we apply the corresponding Kempe change in $G'$ on each of these two components). In this way we obtain a Kempe chain from $\alpha$ to $\beta$ as required. \qed
\end{proof}
\begin{lemma}\label{lidentify1}
Let $k\geq 3$. If $\alpha$ and $\beta$ are matching $k$-colourings of a 3-connected graph $G$ of maximum degree~$k$ then $\alpha \sim_k \beta$.
\end{lemma}
\begin{proof}
If $G$ is $(k-1)$-degenerate then $\alpha \sim_k \beta$ by Lemma~\ref{degprop1}. Assume that $G$ is not $(k-1)$-degenerate. Then $G$ is $k$-regular. Since $\alpha$ and $\beta$ match, there exist two vertices~$u$ and $v$ of $G$ that have a common neighbour $w$ such that $\alpha(u) = \alpha(v)$ and $\beta(u) = \beta(v)$. Let $x$ denote the vertex of $G'$ obtained by identifying $u$ and $v$.
Let $S$ be a separator of $G'$. If $S$ does not contain $x$ then $S$ is a separator of $G$. Then $|S|\geq 3$ as $G$ is 3-connected. If $S$ contains $x$ then $S$ must contain another vertex as well; otherwise $\{u,v\}$ is a separator of size~2 of~$G$, which is not possible. Hence, $|S|\geq 2$ in this case. We conclude that $G'$ is 2-connected.
We now prove that $G'$ is $(k-1)$-degenerate. Note that, in $G'$, $w$ has degree $k-1$, $x$ has degree at least~$k$ and all other vertices have degree $k$. Let $u_1, \dots, u_r$ for some $r\geq k-1$ be the neighbours of $x$ not equal to $w$. Since $G'$ is $2$-connected, the graph $G'' = G' \backslash x$ is connected. This means that every $u_i$ is connected to $w$ via a path in $G''$, which corresponds to a path in $G'$ that does not contain~$x$. Since $w$ has degree $k-1$ and every vertex not equal to $x$ has degree~$k$, we successively delete vertices of these paths starting from $w$ towards $u_i$ so that each time we delete a vertex of degree at most $k-1$. Afterwards we can delete $x$ as $x$ has degree~0. The remaining vertices form an induced subgraph of $G'$ whose components each have maximum degree at least~$k$ and at least one vertex of degree at most~$k-1$. Hence, we can continue deleting vertices of degree at most~$k-1$ and thus find that $G'$ is $(k-1)$-degenerate. Then, by Lemma~\ref{degprop1}, $C_k(G')$ is a Kempe class. Hence, by Lemma~\ref{lplanar}, we find that $\alpha \sim_k \beta$ as required. This completes the proof. \qed
\end{proof}
\begin{lemma}\label{l-net}
Every 3-connected cubic claw-free graph $G$ that is neither $K_4$ nor the 3-prism is house-free, diamond-free and contains an induced net (see also Figure~\ref{fig:1}).
\end{lemma}
\begin{proof}
First suppose that $G$ contains an induced diamond $D$. Then, since $G$ is cubic, the two non-adjacent vertices in $D$ form a separator and $G$ is not 3-connected, a contradiction. Consequently, $G$ is diamond-free.
Now suppose that $G$ contains an induced house $H$. We use the vertex labels of Figure~\ref{fig:1}. So, $s,w,x$ are the vertices that have degree~2 in $H$, and $s$ and $w$ are adjacent. As $G$ is cubic, $w$ has a neighbour $t\in V(G)\setminus V(H)$. Since $G$ is cubic and claw-free,~$t$ must be adjacent to $s$. If $tx \in E$ then $G$ is the $3$-prism. If $tx \notin E$ then $t$ and $x$ form a separator of size~2. In either case we have a contradiction. Consequently, $G$ is house-free.
We now prove that $G$ has an induced net. As $G$ is cubic and claw-free, it has a triangle and each vertex of the triangle has one neighbour in $G$ outside the triangle. Because $G$ is not $K_4$ and diamond-free, these neighbours are distinct. Then, because~$G$ is house-free, no two of them are adjacent. Hence, together with the vertices of the triangle, they induce a net.\qed
\end{proof}
We restate Lemma~\ref{l2} before we present its proof.
\medskip
\noindent
{\bf Lemma~\ref{l2} (restated).}
{\it If $G$ is a $3$-connected cubic graph that is claw-free but that is neither $K_4$ nor the $3$-prism then $C_3(G)$ is a Kempe class.}
\medskip
\noindent
{\it Proof.} By Lemma~\ref{l-net}, $G$ contains an induced net~$N$. For the vertices of $N$ we use the labels of Figure~\ref{fig:1}. In particular, we refer to $x$, $y$ and $z$ as the \emph{t-vertices} of $N$, and $x'$, $y'$ and $z'$ as the \emph{p-vertices}. Let $\alpha$ and~$\beta$ be two $3$-colourings of~$G$. In order to show that $\alpha \sim_3 \beta$ we distinguish two cases.
\medskip
\noindent
{\bf Case 1.} There are two p-vertices with identical colours under $\alpha$ or $\beta$.\\
Assume that $\alpha(x') = \alpha(y') = 1$. Then $\alpha(z) = 1$ as the t-vertices form a triangle, so colour 1 must be used on one of them. Assume without loss of generality that $\alpha(z') = \alpha(x)=2$ and so $\alpha(y)=3$. If $\beta(z') = \beta(x)$ then $\alpha$ and $\beta$ match (as $x$ and $z'$ have $z$ as a common neighbour). Then, by Lemma~\ref{lidentify1}, we find that $\alpha \sim_3 \beta$. Otherwise $\beta(z') = \beta(y)$, since the colour of $z'$ must appear on one of $x$ and $y$. Note that the $(2, 3)$-component containing $x$ under $\alpha$ consists only of $x$ and $y$. Then a Kempe exchange applied to this component yields a colouring $\alpha'$ such that $\alpha'(z') = \alpha'(y)$. As $y$ and $z'$ have $z$ as a common neighbour as well, this means that $\alpha'$ and $\beta$ match. Hence, it holds that $\alpha\sim_3 \alpha' \sim_3 \beta$, where the second equivalence follows from Lemma~\ref{lidentify1}.
\medskip
\noindent
{\bf Case 2.} All three p-vertices have distinct colours under both $\alpha$ and $\beta$.\\
Assume without loss of generality that $\alpha(x) = \alpha(z') = 1$, $\alpha(y) = \alpha(x') = 2$, and $\alpha(z) = \alpha(y') = 3$. Note that Kempe chains of $G$ are paths or cycles, as no vertex in a chain can have degree~3 since all its neighbours in a chain are coloured alike and $G$ is claw-free. So, we will refer to \emph{$(a,b)$-paths} rather than $(a,b)$-components.
We will now prove that there exists a colouring $\alpha'$ with $\alpha \sim_3 \alpha'$ that assigns the same colour to two p-vertices of~$N$. This suffices to complete the proof of the lemma, as afterwards we can apply Case~1.
Consider the $(1,2)$-path $P$ that contains $x'$. If $P$ does not contain $z'$, then a Kempe exchange on $P$ gives us a desired colouring $\alpha'$ (with $x'$ and $z'$ coloured alike). So we can assume that $x'$ and $z'$ are joined by a $(1,2)$-path $P_{12}$, and, similarly, $x'$ and $y'$ by a $(2,3)$-path $P_{23}$, and $y'$ and $z'$ by a $(1,3)$-path $P_{13}$.
Let $G'$ be the subgraph of $G$ induced by the three paths. Note that $P_{12}$ has end-vertices $y$ and $z'$, $P_{23}$ has end-vertices $z$ and $x'$ and $P_{13}$ has end-vertices $x$ and $y'$. Hence, $G'$ contains the vertices of $N$ and every vertex in $G'-N$ is an internal vertex of one of the three paths. As $G$ is cubic, this means that each vertex in $G'-N$ belongs to exactly one path. Moreover, as $G$ is claw-free and cubic, two vertices in $G'-N$ that are on different paths are adjacent if and only if they have a $p$-vertex as a common neighbour.
In Figure~\ref{fig:4} are illustrations of $G'$ and the colourings of this proof that we are about to discuss. Let $x'' \neq x$ be the vertex in $P_{12}$ adjacent to $x'$. From the above it follows that~$x''$ is adjacent to the neighbour of $x'$ on $P_{23}$ and that no other vertex of $P_{12}$ (apart from $x'$) is adjacent to a vertex of $P_{23}$. As $G$ is cubic, this also means that $x''$ has no neighbour outside $G'$. Apply a Kempe exchange on $P_{12}$ and call the resulting colouring~$\gamma$. By the arguments above, the new $(2,3)$-path $Q_{23}$ (under $\gamma$) that contains $y'$ has vertex set $(V(P_{23}) \cup \{x''\}) \backslash \{x', y, z\}$. Apply a Kempe exchange on $Q_{23}$. This results in a colouring $\alpha'$ with $\alpha'(y')=\alpha'(z')=2$, hence $\alpha'$ is a desired colouring. This completes the proof of Case~2 and thus of the lemma. \qed
\tikzstyle{vertex}=[circle,draw=black, fill=black, minimum size=5pt, inner sep=1pt]
\tikzstyle{edge} =[draw,-,black,>=triangle 90]
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.4]
\foreach \pos/\name / \label / \posn / \dist /\labelb /\posnb /\distb in {{(0,0.85)/c1/x/{left}/1/1/right/1}, {(0,3.85)/c2/x'/{above}/1/2/{below left}/-1}, {(-2,3.85)/c3/{}/{left}/1/3/{below right}/-1}, {(2,3.85)/c4/x''/{above right}/-1/1/{below left}/-1}, {(-1,-0.85)/d1/y/{above left}/-1/2/below/1}, {(-3.68,-2.35)/d2/y'/{below left}/-1/3/{above}/1}, {(-4.68,-0.63)/d3/{}/{left}/1/2/left/1}, {(-2.68,-4.08)/d4/{}/{left}/1/1/{below left}/-1}, {(1,-0.85)/e1/z/{above right}/-1/3/below/1}, {(3.68,-2.35)/e2/z'/{below right}/-2/1/above/1}, {(4.68,-0.63)/e3/{}/{left}/1/2/right/1}, {(2.68,-4.08)/e4/{}/{left}/1/3/{below right}/-1}}
{ \node[vertex] (\name) at \pos {};
\node [\posn=\dist] at (\name) {$\label$};
\node [\posnb=\distb] at (\name) {$\labelb$};
}
\foreach \source/ \dest in {c1/c2, d1/d2, e1/e2, e1/c1, c1/d1, d1/e1, d2/d4, e2/e4, c2/c3, c2/c4, d2/d3, e2/e3}
\path[edge, very thick] (\source) -- (\dest);
\draw[edge, very thick] (c3) .. controls (0,5.2) .. (c4);
\draw[edge, very thick] (d3) .. controls (-5.24,-3.35) .. (d4);
\draw[edge, very thick] (e3) .. controls (5.24,-3.35) .. (e4);
\draw[edge, thick, dashdotted] (c2) .. controls (-5, 4) and (-7.04, 3.2) .. (d2) node [midway, above, sloped] (TextNode) {$(2,3)$-path};
\draw[edge, thick, dashdotted] (c2) .. controls (5, 4) and (7.04, 3.2) .. (e2) node [midway, above, sloped] () {$(1,2)$-path};
\draw[edge, thick, dashdotted] (d2) .. controls (-0.32, -7.9) and (0.32, -7.9) .. (e2) node [midway, below, sloped] () {$(1,3)$-path};
\node at (0,-8.7) {colouring $\alpha$};
\node at (16,-8.7) {colouring $\gamma$};
\node at (8,-24.5) {colouring $\alpha'$};
\begin{scope}[xshift=16cm, yshift=0cm]
\foreach \pos/\name / \label / \posn / \dist /\labelb /\posnb /\distb in {{(0,0.85)/c1/x/{left}/1/2/right/1}, {(0,3.85)/c2/x'/{above}/1/1/{below left}/-1}, {(-2,3.85)/c3/{}/{left}/1/3/{below right}/-1}, {(2,3.85)/c4/x''/{above right}/-1/2/{below left}/-1}, {(-1,-0.85)/d1/y/{above left}/-1/1/below/1}, {(-3.68,-2.35)/d2/y'/{below left}/-1/3/{above}/1}, {(-4.68,-0.63)/d3/{}/{left}/1/2/left/1}, {(-2.68,-4.08)/d4/{}/{left}/1/1/{below left}/-1}, {(1,-0.85)/e1/z/{above right}/-1/3/below/1}, {(3.68,-2.35)/e2/z'/{below right}/-2/2/above/1}, {(4.68,-0.63)/e3/{}/{left}/1/1/right/1}, {(2.68,-4.08)/e4/{}/{left}/1/3/{below right}/-1}}
{ \node[vertex] (\name) at \pos {};
\node [\posn=\dist] at (\name) {$\label$};
\node [\posnb=\distb] at (\name) {$\labelb$};
}
\foreach \source/ \dest in {c1/c2, d1/d2, e1/e2, e1/c1, c1/d1, d1/e1, d2/d4, e2/e4, c2/c3, c2/c4, d2/d3, e2/e3}
\path[edge, very thick] (\source) -- (\dest);
\draw[edge, very thick] (c3) .. controls (0,5.2) .. (c4);
\draw[edge, very thick] (d3) .. controls (-5.24,-3.35) .. (d4);
\draw[edge, very thick] (e3) .. controls (5.24,-3.35) .. (e4);
\draw[edge, thick, dashdotted] (c2) .. controls (-5, 4) and (-7.04, 3.2) .. (d2) node [midway, above, sloped] (TextNode) {$(2,3)$-path};
\draw[edge, thick, dashdotted] (c2) .. controls (5, 4) and (7.04, 3.2) .. (e2) node [midway, above, sloped] () {$(1,2)$-path};
\draw[edge, thick, dashdotted] (d2) .. controls (-0.32, -7.9) and (0.32, -7.9) .. (e2) node [midway, below, sloped] () {$(1,3)$-path};
\end{scope}
\begin{scope}[xshift=8cm, yshift=-16cm]
\foreach \pos/\name / \label / \posn / \dist /\labelb /\posnb /\distb in {{(0,0.85)/c1/x/{left}/1/2/right/1}, {(0,3.85)/c2/x'/{above}/1/1/{below left}/-1}, {(-2,3.85)/c3/{}/{left}/1/2/{below right}/-1}, {(2,3.85)/c4/x''/{above right}/-1/3/{below left}/-1}, {(-1,-0.85)/d1/y/{above left}/-1/1/below/1}, {(-3.68,-2.35)/d2/y'/{below left}/-1/2/{above}/1}, {(-4.68,-0.63)/d3/{}/{left}/1/3/left/1}, {(-2.68,-4.08)/d4/{}/{left}/1/1/{below left}/-1}, {(1,-0.85)/e1/z/{above right}/-1/3/below/1}, {(3.68,-2.35)/e2/z'/{below right}/-2/2/above/1}, {(4.68,-0.63)/e3/{}/{left}/1/1/right/1}, {(2.68,-4.08)/e4/{}/{left}/1/3/{below right}/-1}}
{ \node[vertex] (\name) at \pos {};
\node [\posn=\dist] at (\name) {$\label$};
\node [\posnb=\distb] at (\name) {$\labelb$};
}
\foreach \source/ \dest in {c1/c2, d1/d2, e1/e2, e1/c1, c1/d1, d1/e1, d2/d4, e2/e4, c2/c3, c2/c4, d2/d3, e2/e3}
\path[edge, very thick] (\source) -- (\dest);
\draw[edge, very thick] (c3) .. controls (0,5.2) .. (c4);
\draw[edge, very thick] (d3) .. controls (-5.24,-3.35) .. (d4);
\draw[edge, very thick] (e3) .. controls (5.24,-3.35) .. (e4);
\draw[edge, thick, dashdotted] (c2) .. controls (-5, 4) and (-7.04, 3.2) .. (d2) node [midway, above, sloped] (TextNode) {$(2,3)$-path};
\draw[edge, thick, dashdotted] (c2) .. controls (5, 4) and (7.04, 3.2) .. (e2) node [midway, above, sloped] () {$(1,2)$-path};
\draw[edge, thick, dashdotted] (d2) .. controls (-0.32, -7.9) and (0.32, -7.9) .. (e2) node [midway, below, sloped] () {$(1,3)$-path};
\end{scope}
\end{tikzpicture}
\end{center}
\caption{Colourings of $G'$ in the proof of Lemma~\ref{l2}. The dotted lines indicate paths of arbitrary length.}\label{fig:4}
\end{figure}
\subsection{Proof of Lemma~\ref{l3}}\label{a-c}
We first need another lemma.
\begin{lemma} \label{lem-W}
Let $W$ be a set of three vertices in a 3-connected cubic graph $G$ such that every 3-colouring of $G$ colours at least two vertices of $W$ alike. Let every 3-colouring~$c$ of $G$ be Kempe equivalent to another 3-colouring $c'$ such that $c$ and $c'$ colour alike distinct pairs of $W$. Then $C_3(G)$ is a Kempe class.
\end{lemma}
\begin{proof}
Let $\alpha$ and $\beta$ be two 3-colourings of $G$. To prove the lemma we show that \mbox{$\alpha \sim_3 \beta$}. By Lemma~\ref{lidentify1}, it is sufficient to find a matching pair of colourings that are Kempe equivalent to $\alpha$ and $\beta$ respectively.
Let $W=\{x,y,z\}$. We can assume that $\alpha(x)=\alpha(y)$. If $\beta(x)=\beta(y)$, then $\alpha$ and $\beta$ match and we are done. So we can instead assume that $\beta(y)=\beta(z)$. Let $\alpha'$ be a 3-colouring Kempe equivalent to $\alpha$ that colours alike a different pair of $W$. If $\alpha'(y) = \alpha'(z)$, then $\alpha'$ and $\beta$ match. Otherwise we must have that $\alpha'(x)=\alpha'(z)$. Let~$\beta'$ be a 3-colouring Kempe equivalent to $\beta$ that colours alike a different pair of $W$. So $\beta'(x) \in \{\beta'(y), \beta'(z) \}$ and $\beta'$ matches either $\alpha$ or $\alpha'$. \qed
\end{proof}
We restate Lemma~\ref{l3} before we present its proof.
\medskip
\noindent
{\bf Lemma~\ref{l3} (restated).}
{\it If $G$ is a $3$-connected cubic graph that is not claw-free then $C_3(G)$ is a Kempe class.}
\medskip
\noindent
{\it Proof.}
Note that if a vertex has three neighbours coloured alike it is a single-vertex Kempe chain. We will write that such a vertex can be recoloured to refer to the exchange of such a chain.
We make repeated use of Lemma~\ref{lidentify1}: two colourings are Kempe equivalent if they match.
Let $C$ be a claw in $G$ with vertex labels as in Figure~\ref{fig:1}. Note that in every 3-colouring of $G$, two of $s$, $u$ and $v$ are coloured alike since they have a common neighbour. If some fixed pair of $u$, $v$ and $s$ is coloured alike by every 3-colouring of $G$, then every pair of colourings matches and we are done. So let $\alpha$ be a 3-colouring of $G$ and assume that $\alpha(u)=\alpha(v)=1$ and that there are colourings for which $u$ and $v$ have distinct colourings, or, equivalently, colourings for which $s$ has the same colour as either $u$ or~$v$. By Lemma~\ref{lem-W}, it is sufficient to find such a 3-colouring that is Kempe equivalent to~$\alpha$. Our approach is to divide the proof into a number of cases, and, in each case, start from~$\alpha$ and make a number of Kempe changes until a colouring in which $s$ agrees with either $u$ or $v$ is obtained. We will denote such a colouring $\omega$ to indicate a case is complete.
First some simple observations. If $\alpha(s)=1$, then let $\omega=\alpha$ and we are done. So we can assume instead that $\alpha(s)=2$ (and so, of course, $\alpha(w)=3$). If it is possible to recolour one of $u$, $v$ or $s$, then we can let $\omega$ be the colouring obtained. Thus we can assume now that each vertex of $u$, $v$ and $s$ has two neighbours that are not coloured alike.
For a colouring $c$, vertex $x$, and colours $a$ and $b$ let $F_{c,x}^{ab}$ denote the $(a, b)$-component at $s$ under $c$. We can assume that $F_{\alpha,s}^{12}$ contains both $u$ and $v$ as otherwise exchanging $F_{\alpha,s}^{12}$ results in a colouring in which~$s$ agrees with either $u$ or $v$.
Let $N(u) = \{w, u_1, u_2\}$, $N(v) = \{w, v_1, v_2\}$, and $N(s) = \{w, s_1, s_2\}$. Note that the vertices $u_1, u_2, v_1, v_2, s_1, s_2$ are not necessarily distinct.
\medskip
\noindent
{\bf Case 1.} $\alpha(u_1)\not=\alpha(u_2)$, $\alpha(v_1)\not=\alpha(v_2)$ and $\alpha(s_1)\not=\alpha(s_2)$. \\
\noindent So each of $u$, $v$ and $s$ has degree 1 in $F_{\alpha,s}^{12}$ and therefore $F_{\alpha,s}^{12}$ has at least one vertex of degree 3. Let $x$ be the vertex of degree 3 in $F_{\alpha,s}^{12}$ that is closest to $u$ and let $\alpha'$ be the colouring obtaining by recolouring $x$. Then $u$ is not in $F_{\alpha',s}^{12}$ which can be exchanged to obtain $\omega$.
\medskip
\noindent
{\bf Case 2.} $\alpha(s_1) = \alpha(s_2)$. \\
\noindent Then $\alpha(s_1) = \alpha(s_2) = 1$ else $\omega$ can be obtained by recolouring $s$.
\smallskip
\noindent
\textbf{Subcase 2.1:} $\alpha(u_1) = \alpha(u_2)$ or $\alpha(v_1) = \alpha(v_2)$. \\
\noindent The two cases are equivalent so we consider only the first. We have $\alpha(u_1) = \alpha(u_2) = 2$ else $u$ is not in $F_{\alpha,s}^{12}$. Note that $F_{\alpha,s}^{23}$ consists only of $s$ and $w$. If $F_{\alpha,s}^{23}$ is exchanged, $u$ has three neighbours coloured $2$, and can be recoloured to obtain $\omega$ (as $u$ and $s$ are both now coloured 3).
\smallskip
\noindent
\textbf{Subcase 2.2:} $\alpha(u_1) \not= \alpha(u_2)$ and $\alpha(v_1)\not=\alpha(v_2)$. \\
\noindent We can assume that $\alpha(u_1) = \alpha(v_1)= 2$, and $\alpha(u_2) = \alpha(v_2) = 3$.
In this case, we take a slightly different approach. Let $\omega$ now be some fixed 3-colouring with $\omega(s) \in \{\omega(u),\omega(v)$\}.We show that $\alpha \sim_3 \omega$ by making Kempe changes from $\alpha$ until a colouring that matches $\omega$ (or a colouring obtained from $\omega$ by a Kempe change) is reached.
Let $\{a,b,c\}=\{1,2,3\}$. If $\omega(s_1)=\omega(s_2)$, then $\omega$ matches $\alpha$ (recall $\alpha(s_1)=\alpha(s_2)$ in this case). So assume that $\omega(s_1)=a$ and $\omega(s_2)=b$. Then $\omega(s)=c$, and we can assume, without loss of generality, that $\omega(w)=a$. Note that we can assume that $\mbox{$\omega(u) \neq \omega(v)$}$ else $\alpha$ and $\omega$ match and we are done. So, as $u$ and $v$ are symmetric under~$\alpha$, we can assume that $\omega(u)=b$ and $\omega(v)=c$. If $\omega(u_2)=a$ or $\mbox{$\omega(v_2)=a$}$, then, again, $\alpha$ and $\omega$ match (recall that $\alpha(w) = \alpha(u_2) = \alpha(v_2)$) so we assume otherwise (noting that this implies $\omega(u_2)=c$ and $\omega(v_2)=b$) and consider two cases. For convenience, we first illustrate our current knowledge of $\alpha$ and $\omega$ in Figure~\ref{fig:2}. (Though it is not pertinent in this case, we again observe that the six vertices of degree 1 in the illustraton might not, in fact, be distinct.)
\tikzstyle{vertex}=[circle,draw=black, minimum size=13pt, inner sep=0pt]
\tikzstyle{edge} =[draw,-,black,>=triangle 90]
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.65]
\node at (0,-4.15) {colouring $\alpha$};
\node at (10,-4.15) {colouring $\omega$};
\begin{scope}[xshift=0cm]
\foreach \pos/\name / \label / \posn / \dist / \vertexlabel in {{(0,3)/c1/1/above/5/u}, {(0,0)/c2/3/{above right}/3/w}, {(2.6,-1.5)/c3/1/{right}/5/v}, {(-2.6,-1.5)/c4/2/{left}/5/s}, {(-1.74,3)/c5/2/{above left}/3/{u_1}}, {(1.74,3)/c6/3/{above right}/3/{u_2}}, {(-1.74,-3)/c7/1/{above right}/3/{s_2}}, {(-3.46,0)/c8/1/{above right}/3/{s_1}}, {(1.74,-3)/c9/2/{above left}/3/{v_1}}, {(3.46,0)/c10/3/{above left}/3/{v_2}}}
{ \node[vertex] (\name) at \pos {$\vertexlabel$};
\node [\posn=\dist] at (\name) {$\label$};
}
\foreach \source/ \dest in {c2/c1, c2/c4, c2/c3, c1/c5, c1/c6, c4/c7, c4/c8, c3/c9, c3/c10}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\end{scope}
\begin{scope}[xshift=10cm]
\foreach \pos/\name / \label / \posn / \dist / \vertexlabel in {{(0,3)/c1/b/above/5/u}, {(0,0)/c2/a/{above right}/3/w}, {(2.6,-1.5)/c3/c/{right}/5/v}, {(-2.6,-1.5)/c4/c/{left}/5/s}, {(-1.74,3)/c5/{}/{above left}/3/{u_1}}, {(1.74,3)/c6/c/{above right}/3/{u_2}}, {(-1.74,-3)/c7/b/{above right}/3/{s_2}}, {(-3.46,0)/c8/a/{above right}/3/{s_1}}, {(1.74,-3)/c9/{}/{above left}/3/{v_1}}, {(3.46,0)/c10/b/{above left}/3/{v_2}}}
{ \node[vertex] (\name) at \pos {$\vertexlabel$};
\node [\posn=\dist] at (\name) {$\label$};
}
\foreach \source/ \dest in {c2/c1, c2/c4, c2/c3, c1/c5, c1/c6, c4/c7, c4/c8, c3/c9, c3/c10}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\end{scope}
\end{tikzpicture}
\end{center}
\caption{The colourings of Subcase 2.2 of Lemma~\ref{l3}.}\label{fig:2}
\end{figure}
\smallskip
\noindent
\textbf{Subcase 2.2.1:} $\omega(w) =a \in \{\omega(u_1), \omega(v_1) \}$. \\
\noindent Notice that $F_{\alpha,s}^{23}$ contains only $s$ and $w$. If it is exchanged then a colouring is obtained where $w$, $u_1$ and $v_1$ are coloured alike and this colouring matches $\omega$.
\smallskip
\noindent
\textbf{Subcase 2.2.2:} $\omega(w) =a \not\in \{\omega(u_1), \omega(v_1) \}$. \\
\noindent So $\omega(u_1)=c$ and $\omega(v_1)=b$. Thus $F_{\omega,w}^{ab}$ contains only $u$ and $w$, and the colouring obtained by its exchange matches $\alpha$ as $w$ and $v_1$ are both coloured $b$.
\medskip
\noindent
{\bf Case 3.} $\alpha(u_1) = \alpha(u_2)$, $\alpha(v_1) \not= \alpha(v_2)$, and $\alpha(s_1)\not= \alpha(s_2)$. \\
\noindent If $\alpha(u_1) = \alpha(w)$ then the three neighbours of $u$ are coloured alike and it can be recoloured to obtain $\omega$. So suppose $\alpha(u_1) = \alpha(u_2)= 2$. We may assume that $\alpha(s_1) = 1$, $\alpha(s_2) = \alpha(v_2) = 3$, and $\alpha(v_1) = 2$; see the illustration of Figure~\ref{fig:3}.
\tikzstyle{vertex}=[circle,draw=black, minimum size=13pt, inner sep=0pt]
\tikzstyle{edge} =[draw,-,black,>=triangle 90]
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.8]
\foreach \pos/\name / \label / \posn / \dist / \vertexlabel in {{(0,3)/c1/1/above/5/u}, {(0,0)/c2/3/{above right}/3/w}, {(2.6,-1.5)/c3/1/{right}/4/v}, {(-2.6,-1.5)/c4/2/{left}/4/s}, {(-1.74,3)/c5/2/{above left}/3/{u_1}}, {(1.74,3)/c6/2/{above right}/3/{u_2}}, {(-1.74,-3)/c7/3/{above right}/3/{s_2}}, {(-3.46,0)/c8/1/{above right}/3/{s_1}}, {(1.74,-3)/c9/2/{above left}/3/{v_1}}, {(3.46,0)/c10/3/{above left}/3/{v_2}}}
{ \node[vertex] (\name) at \pos {$\vertexlabel$};
\node [\posn=\dist] at (\name) {$\label$};
}
\foreach \source/ \dest in {c2/c1, c2/c4, c2/c3, c1/c5, c1/c6, c4/c7, c4/c8, c3/c9, c3/c10}
\path[edge, black!50!white, thick] (\source) -- (\dest);
\end{tikzpicture}
\end{center}
\caption{The colouring $\alpha$ of Case 3 of Lemma~\ref{l3}.}\label{fig:3}
\end{figure}
\smallskip
\noindent We continue to assume that $F_{\alpha,s}^{12}$ contains $u$ and $v$ and note that $s$ and $v$ have degree 1 therein.
\smallskip
\noindent
\textbf{Subcase 3.1:} $F_{\alpha,s}^{12}$ is not a path. \\
\noindent Let $t$ be vertex of degree $3$ closest to $s$ in $F_{\alpha,s}^{12}$. Then $t$ can be recoloured to obtain a colouring $\alpha'$ such that $F_{\alpha',s}^{12}$ does not contain $v$. Exchanging $F_{\alpha',s}^{12}$, we obtain $\omega$.
\smallskip
\noindent
\textbf{Subcase 3.2:} $F_{\alpha,s}^{12}$ is a path. \\
\noindent Note that $F_{\alpha,s}^{12}$ is a path from $s$ to $v$ through $s_1$ and $u$.
\smallskip
\noindent
\textbf{Subcase 3.2.1:} $F_{\alpha,s_2}^{13}$ is a path from $s_1$ to $s_2$. \\
\noindent Note that $F_{\alpha,u}^{13} \neq F_{\alpha,s_2}^{13}$ since if $F_{\alpha,u}^{13}$ is a path then $u$ would be an endvertex coloured~1 implying $u=s_1$ contradicting that $C$ is a claw. As $G$ is cubic a vertex can belong to both $F_{\alpha,s}^{12}$ and $F_{\alpha,s_2}^{13}$ if it is an endvertex of one of them, and we note that $s_1$ is the only such vertex.
Let $\alpha'$ be the colouring obtained from $\alpha$ by the exchange of $F_{\alpha,s_2}^{13}$. If $s \not\in F_{\alpha',v}^{12}$, then let $\omega$ be the colouring obtained by the further exchange of $F_{\alpha',v}^{12}$.
Otherwise, $F_{\alpha',v}^{12}=F_{\alpha',s}^{12}$, $s$ and $v$ each have degree 1 therein, and we can assume it is a path (else, as in Subcase 3.1, there is a vertex of degree 3 that can be recoloured to obtain $\alpha''$ and $F_{\alpha'',s}^{12}$ does not contain $v$ and can be exchanged to obtain $\omega$). We can also assume that $F_{\alpha',s}^{12}$ contains $F_{\alpha,s}^{12} \setminus \{s_1\}$: if not, then $F_{\alpha, s_2}^{13} \backslash \{s_1, s_2\} \cap F_{\alpha, v}^{12} \not= \emptyset$ (recall that $F_{\alpha,s}^{12}$ is a path from $s$ to $v$ through $s_1$ and $u$) but their common vertices would have degree $4$. Thus, in particular, $F_{\alpha',s}^{12}$ contains $u$ and the vertex~$t$ at distance 2 from $s$ in~$F_{\alpha,s}^{12}$.
As $t$ is not an endvertex in $F_{\alpha',s}^{12}$, $s_1$ is its only neighbour coloured 3 under $\alpha'$. So $F_{\alpha',w}^{23}$ contains four vertices: $w$, $s$, $s_1$ and $t$. Let $\alpha''$ be the colouring obtained from $\alpha'$ by the exchange of $F_{\alpha',w}^{23}$. If $t \not\in \{u_1, u_2\}$, then $u$ has three neighbours with colour $2$ with $\alpha''$ and so can be recoloured to obtain $\omega$. Otherwise the conditions of Case 1 are now met.
\smallskip
\noindent
\textbf{Subcase 3.2.2:} $F_{\alpha,s_2}^{13}$ is not a path from $s_1$ to $s_2$. \\
\noindent If $s_1 \notin F_{\alpha,s_2}^{13}$, then the exchange of $F_{\alpha,s_2}^{13}$ gives a colouring in which $s_1$ and $s_2$ are coloured alike (the colour of $s$ is not affected by the exchange and either both or neither of $u$ and $v$ change colour). Thus Case 2 can now be used.
So we can assume that $s_1 \in F_{\alpha,s_2}^{13}$ has degree 1 in $F_{\alpha,s_2}^{13}$ (recall that $s_1$ has degree~2 in $F_{\alpha,s}^{12}$). If $s_2$ has degree 2 in $F_{\alpha,s_2}^{13}$, then $F_{\alpha, s}^{23}$ contains only $w$, $s$ and $s_2$. If it is exchanged, $u$ has three neighbours with colour $2$ and can be recoloured to $\omega$.
Thus $s_1$ and $s_2$ both have degree 1 in $F_{\alpha,s_2}^{13}$. Let $x$ be the vertex of $F_{\alpha,s_2}^{13}$ closest to~$s_2$. Then $x$ can be recoloured to obtain a colouring $\alpha'$ such that $F_{\alpha',s_2}^{13}$ does not contain $s_1$. Exchanging $F_{\alpha',s_2}^{13}$ again takes us to Case 2. This completes Case 3.
By symmetry, we are left to consider the following case to complete the proof of the lemma.
\medskip
\noindent
{\bf Case 4.} $\alpha(u_1) = \alpha(u_2)$, $\alpha(v_1) = \alpha(v_2)$, and $\alpha(s_1)\not= \alpha(s_2)$. \\
\noindent If $\alpha(v_1) = \alpha(v_2) = 3$, then $v$ can be recoloured to obtain $\omega$. So we can assume that $\alpha(v_1) = \alpha(v_2) = 2$, and, similarly, that $\alpha(u_1) = \alpha(u_2) = 2$. We can also assume that $F_{\alpha,s}^{23}$ is a path since otherwise the vertex of degree 3 closest to $s$ can be recoloured. Define $S = \{u_1, u_2, v_1, v_2\}$. We distinguish two cases.
\smallskip
\noindent
\textbf{Subcase 4.1:} $|S \cap F_{\alpha,s}^{23}| \geq 2$. \\
\noindent As $F_{\alpha,s}^{23}$ is a path and $w$ is an endvertex, one vertex of $S$, say $v_1$, has degree $2$ in $F_{\alpha,s}^{23}$. Consider $F_{\alpha,w}^{13}$: it consists only of vertices $w$, $u$, and $v$. After it is exchanged, $v_1$ has three neighbours with colour $3$ and recolouring $v_1$ allows us to apply Case 3.
\smallskip
\noindent
\textbf{Subcase 4.2:} $|S \cap F_{\alpha,s}^{23}| \leq 1$. \\
\noindent It follows, without loss of generality, that $\{u_1, u_2\} \cap F_{\alpha,s}^{23} = \emptyset$. Exchange $F_{\alpha, u_1}^{23}$and $F_{\alpha, u_2}^{23}$ (which might be two distinct components or just one) to obtain a colouring $\alpha'$. As $w \in F_{\alpha',s}^{23}$ (and hence $w \not\in F_{\alpha, u_1}^{23} \cup F_{\alpha, u_2}^{23}$), every neighbour of $u$ is coloured 3 and it can be recoloured to obtain $\omega$. This completes Case 4 and the proof of Lemma~\ref{l3}.
\qed
|
\section{Introduction}
The relations between the black hole mass $(M_{BH})$ and its host galaxy bulge properties such as luminosity $L$, stellar velocity dispertion $\sigma*$ and mass $M$, (Gebhardt et al. 2000, Graham 2007, Kotilainen et al. 2007), suggest a coeval black hole-host galaxy evolution. It is widely believed that the relativistic jet launched by the black hole might play an important role in this co-evolution, by heating and forcing the cold gas out of the bulge, thus affecting star formation. So the question arises: Is there any relation between the mass of a black hole and the jet it produces? To effectively address this issue, we need reliable $M_{BH}$ estimations. Black hole masses can be estimated by assuming virialization of the clouds in the BLR. However, studies by Arshakian et al. 2010 and Le\'on-Tavares (2010, 2013) have shown that the gas in the BLR can be accelerated and ionized by non-thermal emission from the jet. Moreover, absorption features are swamped by non-thermal emission from the jet, thus making impossible to measure $\sigma*$. Finally, the empirical relation between the $M_{BH}$ and the bulge luminosity ($L_{bulge}$) can be used to obtain homogeneous and bias-free estimations of $M_{BH}$ in the overall population of AGN with prominent relativistic jets.
\section{$M_{BH}-jet$ relations}
By using the $M_{BH}-L_{bulge}$ relations (Graham 2007), we estimate the $M_{BH}$ of 29 radio loud AGN. Bulge luminosities were estimated by modeling the surface brightness of the host galaxy (figure \ref{profiles}) with GALFIT (Peng et al. 2011). Thus, allowing us to investigate the connection between $M_{BH}$ and jet properties such as luminosity (Torrealba et al. 2012) and Lorentz factor (Hovatta et al. 2009). We find statistically significant correlations between $M_{BH}$, jet luminosity and Lorentz factor.
\begin{figure}[H]
\centering
\includegraphics[scale=.53]{olguin_1.eps}
\caption{2D surface brightness decomposition of 1641+399 (3C345) at J band taken with with the NIR camera NOTCam on the Nordic Optical Telescope (NOT). Top left subpanel: the observed image. Middle top subpanel: Shows the model used to describe the surface brightness distribution, which is a nuclear unresolved component (PSF) convolved with a S\'ersic model. Right top subpanel: The residual image. Bottom: Radial profile of the surface brightness distribution.}
\label{profiles}
\end{figure}
\section{Implications}
Statistically significant correlations found between $M_{BH}$ and its jet in this work, strongly suggest that the more massive the black hole is, the faster and brighter the jet it produces. These result could be interpreted as an evolutionary connection between the black hole and its relativistic jet. Moreover, these relations could be useful to constraint high energy production models in radio-loud AGN. The manuscript reporting these findings in detail is under preparation.
|
\section{Introduction} It is well known (see, e.g., \cite{EM} or \cite{Lax}) that the canonical commutation relations in the Weyl form \cite{W}
\begin{equation}\label{niniintro}
U_tV_s=e^{ist}V_{s}U_t, \quad s, t\in \mathbb R,
\end{equation}
between two strongly continuous unitary groups $U_t=e^{itB}$ and $V_s=e^{isA}$ in a (separable) Hilbert space are satisfied if and only if
\begin{equation}\label{1/2W}
U_tAU_t^*=A+t I \quad \text{on } \quad \mathrm{Dom}(A), \quad t\in \mathbb R.
\end{equation}
It is not well known, but trivial (being an immediate corollary of the Stone-von Neumann uniqueness result \cite{JvN}), that if a self-adjoint operator $A$ satisfies \eqref{1/2W}, then $A$ always admits a symmetric restriction $\dot A\subset A$ with deficiency indices (1,1)
such that
\begin{equation}\label{1/2Wdot}
U_t\dot AU_t^*=\dot A+t I \quad \text{on } \quad \mathrm{Dom}(\dot A), \quad t\in \mathbb R.
\end{equation}
In this setting the following natural question arises. Suppose that \eqref{1/2Wdot} holds for some symmetric operator $\dot A$. More generally, assume that the following commutation relations
\begin{equation}\label{1/2Wdotm}
U_t\dot AU_t^*=g_t(\dot A)\quad \text{on } \quad \mathrm{Dom}(\dot A), \quad t\in \mathbb R,
\end{equation}
are satisfied, where $g_t$ is a one-parameter group of affine transformations of the real line and $g_t\mapsto U_t$ is its strongly continuous representation by unitary operators.
{\bf The Extension Problem}: Suppose that \eqref{1/2Wdotm} holds for some symmetric operator $\dot A$.
Classify all maximal dissipative, in particular, self-adjoint (if any) solutions $A $
of the semi-Weyl
relations
\begin{equation}\label{mmm}
U_t AU_t^*=g_t(A)\quad \text{on } \quad \mathrm{Dom}( A)
\end{equation}
that extend $\dot A$ such that $\dot A\subset A\subset (\dot A)^*$.
Given the fact that any one-parameter subgroup of the affine group is either of the form \begin{equation}\label{g2intro}
g_t(x)=x+vt, \quad t\in \mathbb R, \quad \text{for some }v\in \mathbb R,\,\,\,(v\ne 0),
\end{equation}
or
\begin{equation}\label{g1intro}
g_t(x)=a^{ t}(x-\gamma )+\gamma , \quad t\in \mathbb R, \quad \text{for some} \,\,\,\,0<a\ne 1 \quad \text{and } \gamma \in \mathbb R,
\end{equation}
a formal differentiation of \eqref{mmm}
yields commutation relations for the generators $B$ and $A$ of the unitary group $U_t=e^{iBt}$ and the semi-group of contractions $V_s=e^{isA}$, $s\ge 0$, respectively. These relations are of the form
\begin{equation}\label{ins}
[A,B]=ivI,
\end{equation}
in the case of the subgroup $g_t$ of translations \eqref{g2intro},
and
\begin{equation}\label{zwai}
[A, B]=i\lambda A-i\mu I,\quad \lambda=\log a,\quad \mu =\gamma \log a,
\end{equation} for the subgroup $g_t$ given by \eqref{g1intro}.
Here $[A, B]$ stands for the commutator $[A,B]=AB-BA$ on $\mathrm{Dom} (A)\cap \mathrm{Dom}(B)$.
If we make the change of variables:
take $X=A$, $Y=B$ and $Z=ivI$, in the first case, and set $X=A-\gamma I$ and $Y=\frac{1}{i\lambda}B$, in the second, we arrive at the commutation relations
\begin{equation}\label{li1}
[X,Y]=Z, \quad [X,Z]=0, \quad [Y, Z]=0,
\end{equation}
and
\begin{equation}\label{li2}
[X, Y]=X,
\end{equation}
respectively. Therefore, solving the semi-Weyl relations \eqref{mmm} may be considered a variant of non-self-adjoint quantization of the three- and two-dimensional Lie algebras \eqref{li1} and \eqref{li2}, respectively. For an alternative approach towards quantization of low-dimensional Lie algebras and their functional models we refer to \cite{Zol}.
In this paper we solve a slightly more general extension problem. Namely, we assume that
$G\ni g\mapsto U_g$ is a unitary representation of an arbitrary subgroup of the affine group ${\mathcal G}$, the group of affine transformation of the real axis preserving the orientation.
We also suppose that $\dot A$ is a densely defined symmetric operator satisfying the semi-Weyl relations
\begin{equation}\label{mm}
U_g\dot AU_g^*=g(\dot A)\quad \text{on } \quad \mathrm{Dom}( \dot A), \quad g\in G.
\end{equation}
Those $\dot A$'s will be called $G$-invariant operators (with respect to the unitary representation $G\ni g\mapsto U_g$ of the group $G$).
We remark that the $G$-invariant bounded operator colligations (with respect to the group of linear fractional transformations and the Lorentz group) were studied in \cite{Dub1}, \cite[Ch. 10]{LJ}, \cite{PT}.
Our main results are as follows.
We show that in the semi-bounded case, that is, if $\dot A$ from \eqref{1/2Wdotm} is semi-bounded, the extension problem
\eqref{mmm}
is always solvable and that the solution can be given by
$G$-invariant self-adjoint operators.
In particular, if a nonnegative symmetric operator
is $G$-invariant with respect to the group of all scaling transformations of the real axis into itself $(g(x)=ax$, $ a>0$, $ g\in G )$, then
its Friedrichs and Krein-von Neumann extensions are
also
$G$-invariant. In fact, if the
symmetric operator has deficiency indices $(1, 1)$, then those extensions are the only ones that are $G$-invariant.
To treat the general case, we study a flow of
transformations on the set ${\mathcal V}$ of contractive mappings
from the deficiency subspace $\Ker( (\dot A)^*-iI)$ to the
deficiency subspace $\Ker( (\dot A)^*+iI)$ induces by the unitary representation $G\ni g\mapsto U_g$. Based on this study, we show
that
$\dot A$ admits a $G$-invariant
maximal dissipative extension if and only if the flow has a fixed
point.
If the deficiency indices are finite, applying the Schauder fixed point theorem, we show that the extension problem
\eqref{mmm}
is always solvable
in the space of maximal dissipative operators.
If, in addition, the indices are equal, the flow restricted to the set ${\mathcal U}\subset {\mathcal V}$ of all isometries from
$\Ker( (\dot A)^*-iI)$ onto $\Ker( (\dot A)^*+iI)$ leaves this set invariant. In this case, one can
reduce the search for $G$-invariant self-adjoint solutions to the extension problem to the one of
fixed points of the restricted flow. We remark that if $\dot A$ is not semi-bounded, then
self-adjoint $G$-invariant extensions of $\dot A$ may not exist in general,
even if the deficiency indices of $\dot A$ are equal and finite.
Special attention is paid to the case of deficiency
indices $(1,1)$. In particular,
we show that if $G$ is a one-parametric continuous
subgroup of the affine group, the extension problem always has a self-adjoint affine invariant solution if not with respect to the whole group $G$, but at least
with respect to a discrete subgroup of $G$.
This phenomenon, see Remark \ref{fall}, can be considered an abstract operator-theoretic counterpart of the
fall to the center ``catastrophe" in Quantum Mechanics \cite{LL}.
For discussion of the topological origin for this effect see Remark \ref{top}.
In this connection,
we also refer to \cite{Ef} and \cite{FM} for a related discussion of the Efimov Effect in three-body systems and to \cite{FMin}
where the collapse in a three-body system with point interactions has been discovered (also see \cite{MM} and references therein).
\section{G-invariant operators}
Throughout this paper ${\mathcal G}$ denotes the ``$ax+b$''-group which is the non-commutative group of non-degenerate
affine transformations (with respect to composition) of the real axis preserving the
orientation.
Recall that the group ${\mathcal G}$ consists of linear transformations of the real axis
followed by translations
\begin{equation}\label{affine}
x\mapsto g(x)=ax+b, \quad x\in \mathbb R,
\end{equation}
with $a>0$ and $b\in \mathbb R$.
Introduce the concept of a unitarily affine invariant operator.
Suppose that $G$ is a
subgroup of ${\mathcal G}$.
Assume, in addition, that $G\ni g\mapsto U_g$ is a
unitary representation of $G$ on a separable Hilbert space
${\mathcal H}$,
$$
U_fU_g=U_{fg}, \quad f, g \in G.
$$
If $G$ is a continuous group, assume that the representation
$G\ni g\mapsto U_g$ is strongly continuous.
\begin{definition}\label{self}
A densely defined closed operator $ A$ is said to be unitarily affine invariant, or more specifically, $G$-invariant with respect to a
unitary representation $G\ni g\mapsto U_g$, if $ \text{ for all } g\in G$
$$
U_g(\mathrm{Dom} (A))= \mathrm{Dom} ( A)
$$
and
\begin{equation}\label{g(T)}
U_gAU_g^*=g(A) \quad \text{ on }\quad \mathrm{Dom}(A).
\end{equation}
\end{definition}
We notice that the operator equality \eqref{g(T)} should be
understood in the sense that
\begin{equation}\label{afa}
U_gAU_g^*f=g(A)f=aAf+bf, \quad \text{ for all } f\in \mathrm{Dom} (A),
\end{equation}
whenever the transformation $g\in G$ is of the form $
x\mapsto g(x)=ax+b $.
The case of one-parameter continuous subgroups of the affine group deserves a special discussion.
Recall that if $g_t$ is a one-parameter subgroup of the affine group, then either
the group $G=\{g_t\}_{t\in \mathbb R}$ consists of
affine transformations of $\mathbb R$ of the form
\begin{equation}\label{g2}
g_t(x)=x+vt, \quad t\in \mathbb R,
\end{equation}
for some $v\in \mathbb R$, $v\ne 0,$
or
\begin{equation}\label{g1}
g_t(x)=a^{ t}(x-\gamma )+\gamma , \quad t\in \mathbb R,
\end{equation}
for some $a>0$, $a\ne 1$, and $ \gamma \in \mathbb R$.
\begin{remark}\label{suda}
Note that in case \eqref{g1},
in contrast to \eqref{g2},
all transformations $g_t$, $t\in \mathbb R$,
have a finite fixed point:
$$
g_t(\gamma)=\gamma \quad \text{ for all } t\in \mathbb R.
$$
\end{remark}
The concept of $G$-invariant self-adjoint operators in the case
of the group of translations,
$$
g_t(x)=x+t,
$$ is naturally arises in connection with the canonical commutation relations in the Weyl form.
\begin{theorem}\label{una}
Suppose that $G=\{g_t\}_{t\in \mathbb R}$ is a one-parameter continuous subgroup of the
affine transformations
\begin{equation}\label{shift}
g_t(x)=x+t.
\end{equation}
Assume that $A$ is a self-adjoint
$G$-invariant operator with respect to a strongly continuous unitary representation $ t\mapsto U_t$
in a separable Hilbert space. That is,
$$
U_tAU^*_t=A+tI \quad \text{on } \quad \mathrm{Dom} (A).
$$
Then the unitary group $U_t$ and
the unitary group $V_s$ generated by $A$,
$$
V_s=e^{iAs}, \quad s\in \mathbb R,
$$
satisfy the Weyl commutation relations
\begin{equation}\label{nini}
U_tV_s=e^{ist}V_{s}U_t, \quad s, t\in \mathbb R.
\end{equation}
The converse is also true. That is, if two strongly continuous unitary groups $U_t$ and $V_s$ satisfy the Weyl relations \eqref{nini}, then the generator $A$ of the group
$V_s$ is $G$-invariant with respect to the shift group \eqref{shift} and its unitary representation $g_t\mapsto U_t$.
\end{theorem}
As for the proof of this result we refer to \cite[Chapter II, Sec. 7]{Lax}.
\begin{remark} We remark that the Stone-von Neumannn uniqueness result \cite{JvN} states that the shift invariant self-adjoint operator $A$ and
the generator $B$ of the unitary group $U_t$ are mutually unitarily equivalent to a finite or infinite direct sum of the momentum and position operators from the Schr\"odinger representation:
$$
(A, B)\approx \bigoplus_{n=1}^\ell (P, Q), \quad \ell=1,2,\dots, \infty.
$$
Here
$$
(Pf)(x)=i\frac{d}{dx}f(x), \quad \mathrm{Dom} (P)=W_2^1(\mathbb R),
$$
is the momentum operator, and
$$
(Qf)(x)=xf(x), \quad \mathrm{Dom}(Q)=\{f\in L^2(\mathbb R)\,|\, xf(\cdot )\in L^2(\mathbb R) \},
$$
is the position operator, respectively.
\end{remark}
Theorem \ref{una} admits a generalization to the case of arbitrary continuous one-parameter affine subgroups.
\begin{theorem}\label{dos}
Suppose that $G$ is a one-parameter continuous subgroup of the affine
group ${\mathcal G}$. Assume that $A$ is a self-adjoint
$G$-invariant operator with respect to a strongly continuous unitary representation $G\ni g_t\mapsto U_t$
in a separable Hilbert space.
Then the unitary group $U_t$ and
the unitary group $V_s$ generated by $A$,
$$
V_s=e^{iAs}, \quad s\in \mathbb R,
$$
satisfy the generalized Weyl commutation relations
\begin{equation}\label{ninii}
U_tV_s=e^{isg_t(0)}V_{g'_t(0)s}U_t, \quad s,\,t\in \mathbb R,
\end{equation}
where
$$
g'_t(0)=\frac{d}{dx}g_t(x)|_{x=0}.
$$
The converse is also true. That is, if two strongly continuous unitary groups $U_t$ and $V_s$ satisfy the generalized Weyl relations \eqref{ninii}, then the generator $A$ of the group
$V_s$ is $G$-invariant with respect to the one-parameter continuous group $G=\{g_t\}_{t\in\mathbb R}$ and its unitary representation $G\ni g_t\mapsto U_t$.
\end{theorem}
It is easy to see that if $G=\{g)t\}_{t\in \mathbb R}$ is the group of translations,
$$
g_t(x)=x+t,
$$
the commutation relations \eqref{ninii} turn into the standard Weyl commutation relations \eqref{nini}.
Indeed, in this case, $g_t(0)=t$ and $ g'_t(0)=1$, so, \eqref{ninii} simplifies to \eqref{nini}.
For the further generalizations of the Stone-von Neumann result we refer to \cite{Mac}.
Along with the Weyl commutation relations \eqref{ninii}, one can also introduce the concept of restricted generalized Weyl commutation relations
$$
U_tV_s=e^{isg_t(0)}V_{g'_t(0)s}U_t, \quad t\in \mathbb R, \,\,\, s\ge0,
$$
involving a strongly continuous unitary group $U_t$ and a strongly continuous semi-group of contractions $V_s$, $s
\ge 0$. See \cite{BL,Jorg1,J79,J80,J80a,J81,JM, Sch1, Sch2} where the concept of restricted Weyl commutation relations in the case of the group of affine translations has been discussed.
The following result, which is an immediate generalization of Theorem \ref{dos} to the case of maximal dissipative $G$-invariant generators,
characterizes non-unitary solutions to the
restricted generalized Weyl commutation relations
(see Section 7, Subsections \ref{can} and \ref{gencan}, for a number of examples of such solutions).
\begin{theorem}\label{bccr2}
Suppose that $G=\{g_t\}_{t\in \mathbb R}$ is a one-parameter continuous group of the affine
group ${\mathcal G}$. Assume that $A$ is a maximal dissipative
$G$-invariant operator with respect to unitary representation $G\ni g_t\mapsto U_t$
in a Hilbert space.
Then the strongly continuous unitary group $U_t$ and
the strongly continuous semi-group $V_s$ of contractions generated by $A$,
$$
V_s=e^{iAs}, \quad s\ge0,
$$
satisfy the restricted
generalized Weyl commutation relations
\begin{equation}\label{nini2}
U_tV_s=e^{isg_t(0)}V_{g'_t(0)s}U_t, \quad t\in \mathbb R, \,\,\, s\ge0,
\end{equation}
where
$$
g'_t(0)=\frac{d}{dx}g_t(x)|_{x=0}.
$$
The converse is also true. That is, if a strongly continuous unitary group $U_t$ and a strongly continuous semi-group of contractions $V_s$, $s\ge 0$, satisfy the generalized restricted
Weyl commutation relations \eqref{nini2}, then the generator $A$ of the semi-group
$V_s$ is $G$-invariant with respect to the one-parameter continuous group $G=\{g_t\}_{t\in \mathbb R}$ and its unitary representation $g_t\mapsto U_t$.
\end{theorem}
\begin{proof}
To prove the assertion,
assume that $f\in \mathrm{Dom} (A)$. Then (see, e.g., \cite[Theorem 1.3]{KrSel})
$$
V_sf=-\frac{1}{2\pi i}\int_\Gamma e^{i\lambda s} (A-\lambda I)^{-1}f d\lambda, \quad s\ge 0,
$$
where $\Gamma$ is a(ny) contour in the lower half-plane parallel to the real axis and the integral is understood in the principal value sense.
Since $A$ is a $G$-invariant operator, it is easy to see that
$$
U_t (A-\lambda I)^{-1}f=g'_{-t}(\lambda)(A-g_{-t}(\lambda) I)^{-1}U_t f
$$
and therefore (after a simple change of variable)
\begin{align}
U_tV_sf
&=-\frac{1}{2\pi i}\int_\Gamma e^{i\lambda s} U_t(A-\lambda I)^{-1}f d\lambda \nonumber
\\ &=-\frac{1}{2\pi i}\int_\Gamma e^{i\lambda s}
g'_{-t}(\lambda)(A-g_{-t}(\lambda) I)^{-1}U_t fd\lambda
\nonumber
\\
&=-\frac{1}{2\pi i}\int_{\Gamma'} e^{ig_t(\lambda) s} (A-\lambda I)^{-1}f d\lambda
\nonumber
\\
&=-\frac{1}{2\pi i}\int_{\Gamma'} e^{i(g'_t(0)\lambda+g_t(0)) s} (A-\lambda I)^{-1}f d\lambda
\nonumber
\\
&=e^{isg_t(0)}V_{g'_t(0)s}U_tf, \label{qqqq}
\end{align}
with $\Gamma'=g_{-t}(\Gamma)$, a contour in the lower half-plane.
Thus, \eqref{qqqq} shows that the representation \eqref{nini2} holds in the strong sense on the dense set $\mathrm{Dom} (A)$. Taking into account that the operators $U_t$,
$t\in \mathbb R$, and $V_s$, $s\ge 0$, are bounded,
one extends \eqref{qqqq} from the dense set to the whole Hilbert space which
proves the claim.
The converse follows by differentiation of the commutation relations.
\end{proof}
\section{$G$-invariant symmetric operators and the extension problem}
The search for self-adjoint or, more generally, maximal dissipative $G$-invariant operators can be
accomplished solving the following extension problem:
Given a symmetric $G$-invariant operator,
find its all maximal dissipative $G$-invariant extensions.
We remark that the search for $G$-invariant self-adjoint realizations of the symmetric operator can be undertaken only if the deficiency indices are equal.
We start with the following elementary observation.
\begin{lemma}\label{defpot}
Assume that $\dot A$ is a $G$-invariant symmetric operator.
Suppose that $A$ is a maximal dissipative extension of $\dot A$.
Then the restriction $A_g$ of the adjoint operator $(\dot A)^*$ onto $D_g=U_g(\mathrm{Dom}(A))$,
\begin{equation}\label{ag}
A_g=(\dot A)^*\vert_{ D_g},
\end{equation}
is a maximal dissipative extension of $\dot A$.
\end{lemma}
\begin{proof}
Since $\dot A$ is $G$-invariant,
the operator
$U_gAU_g^*, $ $ g\in G,
$ is a maximal dissipative extension of $g(\dot A)$, and hence
$g^{-1}(U_gAU_g^*)$ is a maximal dissipative extension of $\dot A$.
Since
$$
\mathrm{Dom}(g^{-1}(U_gAU_g^*))=\mathrm{Dom}(U_gAU_g^*)=U_g(\mathrm{Dom}(A)),
$$
one concludes that the restriction $A_g$ given by \eqref{ag} is a maximal dissipative operator.
\end{proof}
Recall that the set of all maximal dissipative extensions of $\dot A$ is in a one-to-one correspondence with the set ${\mathcal V}$ of contractive mappings from the deficiency subspace ${\mathcal N}_+=\Ker((\dot A)^*-iI)$ into the deficiency subspace ${\mathcal N}_-=\Ker((\dot A)^*+iI)$.
Note that the set ${\mathcal V}$ is an operator unit ball
in the Banach space ${\mathcal L}( {\mathcal N}_+, {\mathcal N}_-)$.
\begin{proposition}\label{vNET}
Let $\dot A$ be a closed symmetric operator and $V\in {\mathcal V}$ a contractive mapping from
${\mathcal N}_+=\Ker((\dot A)^*-iI)$ into ${\mathcal N}_-=\Ker((\dot A)^*+iI)$. Then the
restriction $ A$ of the adjoint operator $(\dot A)^*$ on
\begin{equation}\label{domdom}
\mathrm{Dom}(A)=\mathrm{Dom}(\dot A)\dot + (I-V) \Ker ((\dot A)^*-iI)
\end{equation}
is a maximal dissipative extension of $\dot A$.
Moreover, the domain of any
maximal dissipative extension $A$ of $\dot A$ such that
$\dot A \subset A\subset (\dot A)^*$ has a decomposition of the form \eqref{domdom},
where $V$ is the restriction of the Cayley transform $(A-iI)(A+iI)^{-1}$
onto the deficiency subspace ${\mathcal N}_+=\Ker ((\dot A)^*-iI)$.
\end{proposition}
\begin{remark}\label{simetr}
If $\dot A$ has equal deficiency indices and
$V$ is an isometric mapping from $\Ker ((\dot A)^*-iI)$ onto $\Ker ((\dot A)^*+iI)$, then the representation \eqref{domdom}, known as von Neumann's formula,
provides a parametric description of all self-adjoint extensions of a symmetric operator $\dot A$.
The extension of von Neumann's formulae to the dissipative case can be found in \cite{TS} (also see \cite[Theorem ~4.1.9]{ABT}).
\end{remark}
\section{The flow on the set ${\mathcal V}$}
Given a symmetric $G$-invariant operator $\dot A$,
the von Neumann's formula \eqref{domdom} from Proposition \ref{vNET} determines a flow of transformations on the set of contractive mappings $V\in {\mathcal V}$. More precisely,
the mapping
$$
G\ni g\mapsto A_g,
$$
with $A_g$ given by \eqref{ag},
can naturally be ``lifted'' to an action of the group $G$
on the operator unit ball ${\mathcal V}$.
In order to describe the action of the group $G$ on ${\mathcal V}$ in more detail, suppose that $V\in {\mathcal V}$ and let $A$ be a unique maximal dissipative extension of $\dot A$
such that
\begin{equation}\label{flow2}
\mathrm{Dom} (A)=\mathrm{Dom}(\dot A)\dot +(I-V)\Ker( (\dot A)^*-iI).
\end{equation}
Define the extension $A_g$ as in \eqref{ag} by
$$
A_g=(\dot A)^*|_{U_g (\mathrm{Dom} (A))},
$$
and let, in accordance with Lemma \ref{defpot} and Proposition \ref{vNET}, the contractive mapping $V_g\in {\mathcal V}$ be such that
\begin{equation}\label{flow3}
\mathrm{Dom} (A_g)=\mathrm{Dom}(\dot A)\dot +(I-V_g)\Ker( (\dot A)^*-iI),
\quad g\in G.
\end{equation}
Introduce the action of the group $G$ on the set of a contractive mappings ${\mathcal V}$ by
\begin{equation}\label{flow1}
\Gamma_g( V)=V_g, \quad g\in G.
\end{equation}
From the definition of $\Gamma_g$ it follows
that
$$
\Gamma_{f\cdot g}=\Gamma_f\circ \Gamma_g, \quad f,g\in G,
$$
so that
\begin{equation}\label{ccdd}
\varphi(f\cdot g, V)=\varphi(f, \Phi(g,V)), \quad f, g\in {\mathcal G}, \quad V\in {\mathcal V},
\end{equation}
with
\begin{equation}\label{babyflow}
\varphi(g, V)=\Gamma_g(V).
\end{equation}
Clearly, the maximal dissipative extension $A$ given by \eqref{flow2} is $G$-invariant if and
only if $V\in {\mathcal V}$ is a fixed point of the flow $G\ni g\mapsto
\Gamma_g$. That is,
\begin{equation}\label{fixp}
\Gamma_g(V)=V \quad \text{ for all}\quad g\in G.
\end{equation}
Therefore,
the set of maximal dissipative extensions of $\dot A$
is in a one-to-one correspondence with the set of all fixed points of the flow $\Gamma_g$, and hence,
the search for all $G$-invariant
maximal dissipative extensions of $\dot A$ can be reduced to the one for the fixed points of the flow \eqref{flow1}.
If, in addition, the symmetric operator $\dot A$ has equal deficiency indices, the restriction of the flow $\Gamma_g$ to the set ${\mathcal U}\subset {\mathcal V}$ of
all
isometric mappings from the deficiency subspace
$\Ker ((\dot A)^*-iI)$ onto the deficiency subspace
$\Ker ((\dot A)^*+iI)$ leaves the set ${\mathcal U}$ invariant. In turn, the set of all self-adjoint extensions of $\dot A$
is in a one-to-one correspondence with the set of fixed points of the restricted flow $\Gamma_g|_{{\mathcal U}}$.
The following technical result provides a formula representation for the
transformations
$\Gamma_g$, $g\in {\mathcal G}$, of the operator unit ball ${\mathcal V}$ into itself.
\begin{lemma}\label{gammy}
Assume that $\dot A$ is a $G$-invariant closed symmetric densely defined operator and $\Gamma_g$, $g\in G$, is the flow defined by \eqref{flow2} and \eqref{flow1}.
Then,
\begin{equation}\label{final}
\Gamma_g(V)=-[P_-U_g(\gamma I-\delta V)]
[P_+U_g(\alpha I-\beta V)]^{-1},
\end{equation}
where $P_\pm$ denotes the orthogonal projections
onto $\Ker ((\dot A)^*\mp iI)$ and $
\alpha= g^{-1}(i)+i$,
$
\beta=g^{-1}(-i)+i
$,
$
\gamma=g^{-1}(i)-i
$,
$
\delta=g^{-1}(-i)-i
$.
\end{lemma}
\begin{proof}
Given $V\in{\mathcal V}$, denote by $A$ the maximal dissipative operator obtained
by the restriction of the adjoint operator $(\dot A)^*$ onto the
domain
$$
\mathrm{Dom} (A)=\mathrm{Dom}(\dot A)\dot +(I-V)\Ker( (\dot A)^*-iI).
$$
Thus, any element $f\in D(A)$ admits the representation
\begin{equation}\label{dliaf}
f=f_0\dot +h -Vh,
\end{equation}
where $f_0\in D(\dot A)$, $h\in \Ker ((\dot A)^*-iI)$ and $Vh\in
\Ker ((\dot A)^*+iI)$.
By the definition of the mapping $\Gamma_g$, $\Gamma_g(V)\in {\mathcal V}$, the domain of the operator
$A_g$ given by (cf. \eqref{ag})
$$
A_g=(\dot A)^*\vert_{ U_g(\mathrm{Dom}(A))}.
$$
is of the form
\begin{equation}\label{dag}
\mathrm{Dom} (A_g)=\mathrm{Dom} (\dot A)\dot +(I-\Gamma_g(V))\Ker ((\dot A)^*-iI).
\end{equation}
Suppose that $f\in D(A)$ and that \eqref{dliaf} holds.
Taking into account that $$U_gf\in
\mathrm{Dom} (U_gAU_g^*)=\mathrm{Dom}(A_g),$$
from \eqref{dag} it follows that
\begin{equation}\label{nono}
U_gf=U_gf_0 +U_gh -U_gVh=k_0+m-\Gamma_g(V)m
\end{equation}
for some $k_0\in D(\dot A)$ and some $m\in \Ker ((\dot A)^*-iI)$.
Then
\begin{equation}\label{vagno}
m-\Gamma_g(V)m=U_gh -U_gVh +\ell_0,
\end{equation}
with $\ell_0=U_gf_0-k_0\in D(\dot A)$.
Applying
$
(\dot A)^*+iI $
to both sides of \eqref{vagno} one gets
\begin{equation}\label{2im}
2i m=((\dot A)^*+iI)U_gh-((\dot A)^*+iI)U_gVh+(\dot A+iI)\ell_0.
\end{equation}
Here we used that $m\in \Ker ((\dot A)^*-iI)$ and $\Gamma_g(V)m\in
\Ker ((\dot A)^*+iI)$. Since by hypothesis $\dot A$ is
$G$-invariant, the adjoint
operator $(\dot A )^*$ is $G$-invariant as well. One obtains
\begin{equation}\label{ugh}
((\dot A)^*+iI)U_gh=U_g(g^{-1}((\dot A)^* +iI))h=(g^{-1}(i)+i)U_gh.
\end{equation}
Similarly, one gets that
\begin{equation}\label{gvh}
((\dot A)^*+iI)U_gVh=(g^{-1}(-i)+i)U_gVh.
\end{equation}
Combining \eqref{2im}, \eqref{ugh} and \eqref{gvh}, one derives the representation
\begin{equation}\label{2iml}
2i m=(g^{-1}(i)+i)U_gh-(g^{-1}(-i)+i)U_gVh+(\dot A+iI)\ell_0.
\end{equation}
Recall that $P_\pm$ is the orthogonal projection onto $\Ker
((\dot A)^*\mp iI)$. From \eqref{2iml} it
follows that
\begin{equation}\label{m1}
2i m=P_+U_g\left ( (g^{-1}(i)+i)h- (g^{-1}(-i)+i)Vh\right ).
\end{equation}
Applying $(\dot A)^*-iI$ to both sides of \eqref{vagno}, in a
completely analogous way one arrives at the identity
\begin{equation}\label{m2}
-2i\Gamma_g(V) m=P_-U_g((g^{-1}(i)-i)h-(g^{-1}(-i)-i)Vh).
\end{equation}
Combining \eqref{m1} and \eqref{m2} proves the equality
\begin{equation}\label{pochti}
-\Gamma_g(V)P_+U_g(\alpha h- \beta Vh )=P_-U_g(\gamma h-\delta Vh),
\end{equation}
where
$
\alpha= g^{-1}(i)+i$,
$
\beta=g^{-1}(-i)+i
$,
$
\gamma=g^{-1}(i)-i
$, and
$
\delta=g^{-1}(-i)-i
$.
Since
$$
U_g(\mathrm{Dom}(A))=\mathrm{Dom}(U_gAU_g^*),
$$
from \eqref{nono} and \eqref{m1} it follows that the operator
$P_+U_g(\alpha I-\beta V)$ maps the deficiency subspace $\Ker((\dot A)^*-iI)$ onto itself.
Moreover, the null space of $P_+U_g(\alpha I-\beta V)$ is trivial.
Indeed, let
$$
P_+U_g(\alpha I-\beta V)h=0
$$
for some $h\in \Ker((\dot A)^*-iI)$ and $m$ be as above
(cf. \eqref{nono}). From \eqref{m1} it follows that $m=0$.
Therefore, coming back to \eqref{nono}, one concludes that
$$
U_g(\alpha I-\beta V)h\in \mathrm{Dom}(\dot A).
$$
Since $\dot A$ is $G$-invariant, this means that $(\alpha I-\beta V)h\in
\mathrm{Dom}(\dot A)$ which is only possible if $h=0$ ($\alpha$ is never zero). By the closed graph
theorem the bounded mapping $P_+U_g(\alpha I-\beta V)$ from the deficiency subspace $\Ker((\dot A)^*-iI)$ onto
itself has a bounded inverse and hence from
\eqref{pochti} one concludes that
\begin{equation*}
\Gamma_g(V)=-[P_-U_g(\gamma I-\delta V)][
P_+U_g(\alpha I-\beta V)]^{-1},
\end{equation*}
completing the proof.
\end{proof}
\begin{corollary}\label{homoho} Assume that $G$ is a continuous subgroup of the affine group ${\mathcal G}$. Assume, in addition, that
$\dot A$ is a $G$-invariant closed symmetric operator with finite deficiency indices. Then the mapping
$$(g, V)\mapsto \Gamma_g(V)
$$
from $G\times {\mathcal V}$ to ${\mathcal V}$ is continuous. In particular, if $ G=\{g_t\}_{t\in \mathbb R}$
is continuous one-parameter subgroup of the affine group ${\mathcal G}$,
then for any $V\in {\mathcal V}$ the trajectory $\mathbb R\ni t\mapsto
\Gamma_{g_t}(V)$ is continuous.
\end{corollary}
\begin{remark}To validate the usage of the term ``flow" in our general considerations, recall that a flow on a closed, oriented manifold ${\mathcal M}$ is a one-parameter group of homomorphisms of ${\mathcal M}$. That is, a flow is a function
$$
\varphi:\mathbb R\times {\mathcal M}\mapsto {\mathcal M}
$$
which is continuous and satisfies
\begin{itemize}
\item[(i)] $ \varphi(t_1+t_2,x)=\varphi(t_1, \varphi(t_2, x))$,
\item[(ii)] $\varphi(0,x)=x$
\end{itemize}
for all $t_1, t_2\in \mathbb R$ and $x\in {\mathcal M}$.
If $G=\{g_t\}_{t\in \mathbb R}$ is a continuous one-parameter group of affine transformations and the deficiency indices of the symmetric operator $\dot A$ are finite, then
by Corollary \ref{homoho}, relation
\eqref{babyflow} defines a flow, in the standard sense, on the $n$-manifold ${\mathcal V}$, where
$$n=\dim(\Ker((\dot A)^*-iI))\times\dim(\Ker((\dot A)^*+iI)).$$
\end{remark}
\section{Fixed point theorems}
The main goal of this section is to present several fixed point theorems for the flow $\Gamma_g$ given by \eqref{flow1}
on the operator unit ball ${\mathcal V}$ introduced in the previous section.
We start with recalling the following general fixed point theorem obtained in \cite{MT}.
This theorem solves the problem of the existence of fixed points for the flow \eqref{flow1} in the case where the symmetric operator $\dot A$ is semi-bounded.
To formulate the result we need some preliminaries.
Recall that if $\dot A$ is a densely defined (closed) positive operator, then the
set of all positive self-adjoint extensions of $\dot A$ has the minimal $A_{\mathrm{KvN}}$
(Krein--von Neumann extension) and
the maximal $A_F$ (Friedrichs extension) elements. That is, for any positive self-adjoint extension
$\widetilde A$ the following two-sided operator inequality holds \cite{Kr1}
$$
(A_F+\lambda I)^{-1}\le (\widetilde A+\lambda I)^{-1}\le (A_{\mathrm{KvN}}+\lambda I)^{-1}, \quad
\text{ for all } \lambda >0.
$$
For the convenience of the reader, recall the following fundamental result characterizing the Friedrichs and
Krein--von Neumann extensions \cite{AN}.
\begin{proposition} \label{ando}(\cite{AN})
Let $\dot A$ be a closed densely defined positive symmetric operator. Denote by $\bf a$ the
closure of the quadratic form
\begin{equation}\label{kvf}
{\dot {\bf a}}[f]=(\dot A
f, f), \quad \mathrm{Dom}[{\dot {\bf a}}]=\mathrm{Dom}(\dot A).
\end{equation}
Then the Friedrichs extension $A_F$ coincides with
the restriction of the adjoint operator
$(\dot A)^*$ on the domain
$$
\mathrm{Dom} (A_F)=\mathrm{Dom} ((\dot A)^*)\cap \mathrm{Dom} [{\bf a}].
$$
The Krein--von Neumann extension $A_{\mathrm{KvN}}$ coincides with
the restriction of the adjoint operator
$(\dot A)^*$ on the domain
$
\mathrm{Dom} (A_{\mathrm{KvN}})
$
which consists of the set of elements $f$ for which there exists a
sequence $\{f_n\}_{n\in \mathbb N}$, $f_n\in \mathrm{Dom}(\dot A)$, such that
$$
\lim_{n,m\to \infty}{\bf a}[f_n-f_m]=0
$$
and
$$
\lim_{n\to \infty}\dot Af_n=(\dot A)^*f.
$$
\end{proposition}
Now we are ready to present the aforementioned fixed point theorem in the case where the underlying $G$-invariant symmetric operator is semi-bounded.
\begin{theorem}\label{FK} Let $G$ be a subgroup of the affine group ${\mathcal G}$. Suppose that $\dot A$ is a
closed, densely defined, semi-bounded from below
symmetric $G$-invariant operator with the greatest lower bound $\gamma$. Then both the Friedrichs extension $A_F$ of the operator $\dot
A$ and the extension $A_K=(\dot A-\gamma I)_{\mathrm{KvN}}+\gamma I$, where
$(\dot A-\gamma I)_{\mathrm{KvN}}$ denotes
the Krein--von Neumann extension of the positive symmetric operator
$\dot A-\gamma I$, are $G$-invariant.
\end{theorem}
We provide a short proof the idea of which is due to Gerald Teschl \cite{GeT} (see \cite{MT} for the original reasoning).
\begin{proof}
Suppose that $\dot A$ is $G$-invariant with respect to a unitary representation $G\ni g\mapsto U_g$.
Denote by ${\mathcal I}$ the operator interval of self-adjoint extensions of $\dot A$ with the greatest lower bound greater or equal to $\gamma$.
Recall that for any $A\in {\mathcal I}$ one has the operator inequality $A_K\le A\le A_F $.
Since the symmetric operator $\dot A$ is $G$-invariant, the correspondence
$$
A \to g^{-1} (U_g A U_g^*), \quad \quad g(x)=ax+b,
$$
with $A$ a self-adjoint extension of $\dot A$,
maps the operator interval ${\mathcal I}=[A_K, A_F] $ onto itself.
Moreover, this mapping is operator monotone, that is, it preserves the order.
Therefore, the end-points $A_K$ and $A_F$ of the operator interval ${\mathcal I}$ has to be fixed points of this map which completes the proof.
\end{proof}
\begin{remark}\label{rem53} We remark that in the situation of Theorem \ref{FK}, the greatest lower bound $\gamma$ has to be a fixed point for
all transformations from the group $G$. Therefore, the hypothesis that $\dot A$ is semi-bounded is rather restrictive. In particular, this hypothesis implies that the group $G$ is necessarily a proper subgroup of the whole affine group ${\mathcal G}$. For instance, the group $G$ does not contain the transformations $g(x)=x+b$, $b\ne 0,$ without (finite) fixed points.
\end{remark}
Next, we claim (under mild assumptions) that for finite deficiency indices of $\dot A$
the flow \eqref{flow1} always has a fixed point. Therefore,
any $G$-invariant symmetric operator with finite deficiency indices possesses
either a self-adjoint or a maximal dissipative $G$-invariant extension of $\dot A$ (cf. \cite[Theorem 15]{J80}).
\begin{theorem}\label{Shauder}
Let $G$ be either a cyclic or a one-parameter continuous subgroup of affine
transformations of the real axis into itself preserving the orientation.
Suppose that $G\ni g\mapsto U_g$ is a unitary representation in a Hilbert space
${\mathcal H}$.
Assume that $\dot A$ is a
closed symmetric $G$-invariant operator with finite deficiency indices.
Then the operator $\dot A$
admits a $G$-invariant maximal dissipative extension.
\end{theorem}
\begin{proof}
For any $g\in G$, the mapping $\Gamma_g: {\mathcal V}\to {\mathcal V}$ is a continuous mapping
from the unit ball ${\mathcal V}$ into itself. Since the deficiency indices of $\dot A$ are finite,
${\mathcal V}$ is a compact convex set and therefore, by the Schauder fixed point theorem (see, e.g., \cite{LuS}, page 291), the mapping $\Gamma_g$ has a fixed point
$V\in {\mathcal V}$. That is,
$$
\Gamma_g(V)=V.
$$
In particular,
$$
\Gamma_{g^n}(V)=V \quad \text{ for all } n \in \mathbb Z.
$$
Therefore, the restriction $A$ of the adjoint operator $(\dot A)^*$ onto the domain
$$
\mathrm{Dom} (A)=\mathrm{Dom} (\dot A)+(I-V)\Ker((\dot A)^*-iI)
$$
is a $G[g]$-invariant operator, where $G[g]$ is the cyclic group generated by the element $g\in G$, proving the claim in the case of cyclic groups $G$ .
In order to treat the case of one-parameter continuous subgroups $G=\{g_t\}_{t\in \mathbb R}$, we remark that
by the first part of the proof for any $ n\in \mathbb N$ there exists a $V_n\in {\mathcal V}$
such that
\begin{equation}\label{sec0}
\Gamma_{g_q}(V_n)=V_n \text{ for all } q\in \mathbb Z/n.
\end{equation}
Since by hypothesis $\dot A$ has finite deficiency indices and hence the unit ball ${\mathcal V}$ is compact,
one can find a $V\in {\mathcal V}$ and a subsequence $\{n_k\}_{k\in \mathbb N}$ such that
\begin{equation}\label{sec1}
\lim_{k\to \infty}V_{n_k}=V.
\end{equation}
Given $t\in \mathbb R$, choose a sequence of rationals $\{q_k\}_{k\in \mathbb N}$
such that
\begin{equation}\label{sec2}
\lim_{k\to \infty}q_k=t
, \quad g_k\in \mathbb Z/{n_k} .
\end{equation}
Combining \eqref{sec0} and \eqref{sec1},
one obtains that
$$
\Gamma_{g_{q_k}}(V_{n_k})=V_{n_k}.
$$
Going to the limit $k \to \infty$ in this equality,
from \eqref{sec1} and \eqref{sec2} one concludes that
$$
\Gamma_{g_t}(V)=V \quad \text{ for all } t\in \mathbb R
$$
upon applying Corollary \ref{homoho}. The proof is complete.
\end{proof}
\begin{remark}\label{pervaia} If $\dot A$ has equal deficiency indices and if a fixed point $V$ is
an isometry from $\Ker((\dot A)^*-iI)$ onto
$\Ker((\dot A)^*+iI)$, then the corresponding maximal dissipative extension is self-adjoint.
We also remark that Theorem \ref{Shauder} does not guarantee the existence
of a $G$-invariant self-adjoint extensions of $\dot A$ in general. In other words, the restricted flow $\Gamma_g|_{{\mathcal U}}$,
with ${\mathcal U}$ the set of all isometries from $\Ker((\dot A)^*-iI)$ onto
$\Ker((\dot A)^*+iI)$, may not have fixed points at all
(see Section 7 for the corresponding examples and Remark \ref{fall} for a detailed discussion of the group-theoretic descent of those examples).
\end{remark}
\begin{remark} It is worth mentioning
that Theorem \ref{Shauder} is a simple consequence of the following Lefschetz fixed point theorem for flows on manifolds.
\begin{proposition}(see, eg., \cite[Theorem 6.28]{Vick})\label{Lef}
If ${\mathcal M}$ is a closed oriented manifold such that the Euler characteristic $\chi ({\mathcal M})$ of ${\mathcal M}$ is not zero, then any flow on ${\mathcal M}$ has a fixed point.
\end{proposition}
Indeed, take ${\mathcal M}={\mathcal V}$ and notice that ${\mathcal V}$ is a closed convex set with $\chi({\mathcal V})\ne 0$.
\end{remark}
%
\section{The case of deficiency indices $(1,1)$}
In this subsection we provide several results towards the solution of the extension problem \eqref{mmm}
in the dissipative as well as in the self-adjoint settings that are specific to the case
of deficiency indices $(1,1)$ only.
\begin{hypothesis}\label{kiki}
Assume that $G$ is a subgroup of the affine group ${\mathcal G}$. Suppose, in addition, that
$\dot A$ is
a $G$-invariant closed symmetric operator
with deficiency indices $(1,1)$.
\end{hypothesis}
Under Hypothesis \ref{kiki},
based on the results of Proposition \ref{vNET} and Lemma \ref{gammy} one can identify the flow $G\ni g\mapsto \Gamma_g$ with the representation of the group $G$ into the group $\text{Aut}(\mathbb D)$
of automorphisms of the
unit disk $\mathbb D$ (consisting of linear-fraction transformations
$z\mapsto e^{i\theta}\frac{z-a}{1-\bar a z}$ for some $|a|<1 $ and $\theta \in \mathbb R$,
mapping the unit circle $\mathbb T$ into itself).
Recall the following elementary result.
\begin{proposition}\label{nado} Any automorphism $\Gamma$ of the (closed) unit disk $\mathbb D$ different from the identical map
has at most two different fixed points in $\mathbb D$.
If $\Gamma$ has exactly two fixed points, then both of them lie on the boundary $\mathbb T$ of $\mathbb D$.
\end{proposition}
\begin{theorem}\label{tri}
Assume Hypothesis \ref{kiki}.
Suppose that $\dot A$ has either
\begin{itemize}
\item[(i)] at least three different self-adjoint $G$-invariant
extensions
\item[]
or
\item[(ii)] at least two $G$-invariant maximal dissipative extensions one of which is not self-adjoint.
\end{itemize}
Then any maximal dissipative extension of $\dot A$ is $G$-invariant. Moreover, in this case,
if the group $G$ is nontrivial, then $\dot A$ is not semi-bounded.
\end{theorem}
\begin{proof}
In case (i), for any $g\in G$ the automorphism $\Gamma_g$ of the unit disk $\mathbb D$
has at least three different fixed points. Hence, by Proposition \ref{nado}, $ \Gamma_g $ is
the identical map on $\mathbb D$ and thus any maximal dissipative extension $A$
of
$\dot A$ such that $\dot A \subset A\subset (\dot A)^*$ is $G$-invariant.
In case (ii), for any $g\in G$ the automorphism $ \Gamma_g $ has at least two fixed points, one of which is not on the boundary of the unit disk and hence, again, by Proposition \ref{nado},
the automorphism $ \Gamma_g $, $g\in G$, is the identical map, proving that
any maximal dissipative extension of $\dot A$ is $G$-invariant.
To prove the last statement of the theorem, assume, to the
contrary, that $\dot A$ is a semi-bounded symmetric operator.
Suppose, for definiteness, that $\dot A$ is semi-bounded from
below.
Assume that one can find a point $\lambda_0\in \mathbb R$ to the left
from the greatest lower bound of $\dot A$ such that the set $\Sigma=\bigcup_{g\in
G}g(\lambda_0)$ is unbounded from below, that is,
\begin{equation}\label{Si}
\inf \Sigma =-\infty.
\end{equation}
By Krein's theorem (see \cite {Kr1}, Theorem 23) there exists a
self-adjoint extension $A$ of the symmetric operators $\dot A$
such that $\lambda_0$ is a simple eigenvalue of $A$. Since any maximal dissipative extension of $\dot A$ is $G$-invariant, the self-adjoint operator $A$ is
also $G$-invariant. Since the spectrum of a $G$-invariant
operator is a $G$-invariant set, one concludes that the set $\Sigma$
belongs to the pure point spectrum
of $A$ and hence $A$ is not semi-bounded from below for
\eqref{Si} holds. The latter is impossible: any self-adjoint
extension of a semi-bounded symmetric operator with finite
deficiency indices is semi-bounded from below.
To justify the choice of $\lambda_0$ with the property \eqref{Si} one
can argue as follows.
By hypothesis, the group $G$ is nontrivial and, therefore, $G$
contains a transformations $g$ either of the form $g(x)=x+b$ for
some $b\ne 0$ or $ g(x)=a(x-\gamma)+\gamma $ for some $a>0$, $a\ne 1$, and
$ \gamma \in \mathbb R$.
In the first case, the set $\Sigma $ contains the orbit $
\bigcup_{n\in \mathbb Z} \{ \lambda_0+nb\}$ and hence one can take any
$\lambda_0$ to the left from the greatest lower bound of $\dot A$
to meet the requirement \eqref{Si}. In the second case,
$$
\bigcup_{n\in \mathbb Z}\{a^n(\lambda_0-\gamma)+\gamma\}\subset \bigcup_{g\in
G}g(\lambda_0)= \Sigma.
$$
Hence, one can choose any $\lambda_0<\gamma$ to the left from the
greatest lower bound of $\dot A$
for \eqref{Si} to hold.
\end{proof}
Our next result shows that a semi-bounded $G$-invariant symmetric operator with deficiency indices $(1,1)$ possesses self-adjoint $G$-invariant extensions only.
\begin{theorem}\label{FKnet} Assume Hypothesis \ref{kiki}.
Suppose, in addition,
that $\dot A$ is semi-bounded from below.
If $A$ is a maximal dissipative $G$-invariant extension of $\dot A$, then $A$ is necessarily self-adjoint.
Moreover, either
$A=A_F$ or $A=A_K$ where $A_F$ and $A_K$ are the operators referred to in Theorem \ref{FK}.
\end{theorem}
\begin{proof} Assume to the contrary that $ A$, with $\dot A \subset A\subset (\dot A)^*$,
is a $G$-invariant maximal dissipative extension of $\dot A$
different from $A_F$ and $A_K$.
If $A_F\ne A_K$, then $\dot A$
has at least three different $G$-invariant extensions.
Therefore, by Theorem
\ref{tri}, the operator $\dot A$ is not semi-bounded from below which contradicts the hypothesis.
Thus, either
$A=A_F$ or $A=A_K$.
If the extensions $A_F$ and $A_K$ coincide, that is,
\begin{equation}\label{coin}
A_F=A_K,
\end{equation}
one proceeds as follows.
By hypothesis, $A\ne A_F=A_K$,
the $G$-invariant maximal dissipative extension $A$ is necessarily self-adjoint.
Indeed, otherwise, $\dot A$ has
at least two $G$-invariant maximal dissipative extensions, $A_F$ and $A$, one of which is not self-adjoint. Applying
Theorem \ref{tri} shows that the symmetric operator $\dot A$ is not semi-bounded which contradicts the hypothesis. Thus,
$A=A_F=A_K$.
Denote by $\widetilde \gamma$
the greatest lower bound of $ A$. Since $A_K=A_F$, the greatest lower bound $\widetilde \gamma$ of $ A$
is strictly less than the greatest lower bound $\gamma$
of the symmetric operator $\dot A$. Since $\dot A$ has deficiency indices $(1,1)$, $\widetilde \gamma$ is
an eigenvalue of $ A$. Following the strategy of proof of the second statement
of Theorem \ref{tri}, one concludes that the orbit $\Sigma=\bigcup_{g\in G}g(\widetilde \gamma)
$
belongs to the spectrum of the $G$-invariant self-adjoint operator $ A$.
By Remark \ref{rem53}, the greatest lower bound $\gamma$ is a fixed point for any transformation $g$ from the group $G$,
\begin{equation}\label{hold}
g(\gamma)=\gamma, \quad g\in G,
\end{equation}
that is, $g(x)=a(x-\gamma)+\gamma$ for some $a>0$ (cf. \eqref{g1}).
Since $\widetilde \gamma<\gamma$ and \eqref{hold} holds, it is also clear that $\inf \Sigma=-\infty$.
Therefore, $ A$ is not semi-bounded from
below, so is $\dot A$. A contradiction.
\end{proof}
We recall, see Remark \ref{pervaia}, that the fixed point result, Theorem \ref{Shauder}, does not guarantee
the existence of a self-adjoint unitarily affine invariant extension of a symmetric invariant operator in general.
However, a $G$-invariant symmetric operator with deficiency indices $(1, 1) $ always has an affine invariant self-adjoint extension if not with respect to the whole group $G$ but
at least
with respect to a cyclic subgroup of $G $, provided that $G$ is a continuous one-parameter group.
\begin{theorem}\label{discrete-invariant}
Assume Hypothesis \ref{kiki}.
Suppose that
$G=\{g_t\}_{t\in \mathbb R}$ is a continuous one-parameter group and that $\dot A$ has no self-adjoint $G$-invariant extension.
Then
there exists a discrete cyclic subgroup $G[g]$ generated by some element $g\in G$
such that any maximal dissipative, in particular, any self-adjoint extension of $\dot A$ is $G[g]$-invariant.
\end{theorem}
\begin{proof}
Let $V\in \mathbb T$ be the von Neumann parameter on the unit circle and $A_V$ the
associated self-adjoint extension of $\dot A$.
Introduce the continuous flow on the circumference $\mathbb T$ by
\begin{equation}\label{defff}
\varphi(t,V)=\Gamma_{g_t}(V), \quad t\in \mathbb R, \quad g_t\in G.
\end{equation}
To prove the assertion, it is sufficient to show that
the map $t\mapsto \varphi(t,V)$ is not injective for some $V\in \mathbb T$.
Indeed, if this map is not injective, then
there exist two different values $t_1 $ and $t_2$ of the parameter $t$ such that
\begin{equation}\label{vse}
\varphi(t_1,V)=\varphi(t_2,V).
\end{equation}
Therefore, $V$ is a fixed point of the transformation
$\varphi(t_2-t_1, \cdot )$. Hence, for any $n\in \mathbb Z$
$$
V=\varphi(n(t_2-t_1),V),
$$
and, therefore, the extension $A[g]=A_V$
is $G[g]$-invariant with respect to the
discrete subgroup $G[g]$ generated by the element
$$g=g_{t_1-t_2}.$$
By hypothesis the self-adjoint operator $A[g]$ is not $G$-invariant (with respect to the whole subgroup $G$)
and therefore $V$ is not a fixed point for the whole family of the transformations $\gamma_t$. In particular, the orbit
$S_V=\cup_{t\in \mathbb R}\varphi(t,V)$ is an infinite set. Next, if $W=\varphi(t,V)\in S_V$ for some $t\in \mathbb R$, then
$$
\varphi(T,W)=\varphi(T+t,V)=\varphi(t,V)=W,
$$
where
\begin{equation}\label{period}T=t_2-t_1.
\end{equation}
Therefore
$\gamma_{nT}(W)=W$ for all $W\in S_V$ and all $n\in \mathbb Z$ and hence the corresponding self-adjoint operators $A_W$, $W\in S_V$, are $G[g]$-invariant. Thus, since the set $S_V$ is infinite, the operator
$\dot A$ has infinitely many, and, therefore, at least three different $G[g]$-invariant self-adjoint extensions. Therefore, by Theorem \ref{tri}, every maximal dissipative extension of $\dot A$ (obtained as a restriction of $(\dot A)^*$) is $G[g]$-invariant.
To complete the proof it remains to justify the claim that the map $t\mapsto \varphi(t,V)$ is not injective for some $V\in \mathbb T$.
To do that assume the opposite. That is, suppose that the map
$t\mapsto \varphi(t,V)$ is injective for all $V\in \mathbb T$. Then, by Brouwer's invariance of domain theorem \cite{Br}, the set
$\displaystyle{S_{V}=\bigcup_{t\in \mathbb R}\varphi(t,V)}$
is an open arc.\footnote{We are indebted to N. Yu. Netsvetaev who attracted
our attention to this fact.} Thus,
$$
\mathbb T=\bigcup_{V\in \mathbb T}S_{V}
$$
is an open cover. Since $\mathbb T$ is a compact set, choosing a finite sub cover,
$$
\mathbb T=\bigcup_{k=1}^NS_{V_k},\quad \text{for some }\quad N\in \mathbb N,
$$
we get that $\mathbb T$ is covered by finitely many open arcs and therefore some of them must intersect. If so, the circumference $\mathbb T$ would be
the trajectory of some point $V_*\in \mathbb T$. That is, $\mathbb T=S_{V_*}$ which means that the map $t\mapsto \varphi(t,V_*)$ is not injective.
A contradiction.
The proof is complete.
\end{proof}
\begin{remark}\label{unnonself} It follows from Theorems \ref{Shauder} and \ref{tri}
that under hypothesis of Theorem \ref{discrete-invariant}, that is, if $\dot A$ has no $G$-invariant self-adjoint extensions, the symmetric operator $\dot A$ admits a unique non-self-adjoint maximal dissipative $G$-invariant extension $A$
with the property $\dot A \subset A\subset (\dot A)^*$.
\end{remark}
\begin{remark}\label{top} Concrete examples show (see Examples \ref{ex1} and \ref{ex2} and Subsections \ref{7.1} and \ref{7.5}, respectively) that our hypothesis that $\dot A$ has no self-adjoint $G$-invariant extension is not empty.
Accordingly, this hypothesis implies that the flow
$\varphi(t, V)=\gamma_t(V)$ on the circumference ${\mathcal M}=\mathbb T$ given by \eqref{defff} has no fixed point. It is no surprise: the Euler characteristics of $\mathbb T$ is zero, $\chi (\mathbb T)=0$,
and hence Proposition \ref{Lef} cannot be applied. Instead, the flow is periodic with period $T$ given by \eqref{period} and thus
$$
\varphi(T, \cdot )=\text{Id}_{\mathbb T},
$$
So, any point $V\in \mathbb T$ is a fixed point,
$$
\varphi(nT, V)=V, \quad n\in \mathbb Z.
$$
\end{remark}
\section{examples}
In this section we provide several examples that illustrate our main results.
In particular, we show that, even in the simplest cases, the restricted standard or generalized Weyl commutation relations possess a variety of non-unitarily equivalent
solutions.
\begin{example}\label{ex1}
Suppose that $G=\{g_t\}_{t\in \mathbb R}$ is the one-parameter group of translations
\begin{equation}\label{shi}
g_t(x)=x+t.
\end{equation}
Let $\dot A$ be the differentiation operator in $L^2(0, \ell)$,
$$
(\dot A) f(x)=i\frac{d}{dx} f(x)\quad \text{on} \quad \mathrm{Dom} (\dot A) =\{f\in W^1_2(0,\ell)\, |\, f(0)=f(\ell)=0\}.
$$
It is well know \cite{AkG} that $\dot A$ has deficiency indices $(1,1)$. Moreover, $\dot A$ is obviously $G$-invariant with respect to
the group of unitary transformations given by
$$
(U_tf)(x)=e^{ixt}f(x), \quad f\in L^2(0,\ell).
$$
\end{example}
\subsection{}\label{7.1} First, we notice that the symmetric $G$-invariant operator $\dot A$ from Example \ref{ex1} does not admit self-adjoint $G$-invariant extensions.
Indeed, if $A$ is a self-adjoint extension of $\dot A$, then there exists a $\Theta
\in [0, 2\pi)$ such that $A$ coincides with the differentiation operator on
$$
\mathrm{Dom}(A)=\{f\in W^{1}_2(0,\ell)\, | f(\ell)=e^{i\Theta} f(0)\}.
$$
Therefore, as a simple computation shows,
the spectrum of the operator $A$
is discrete and it consists of simple eigenvalues forming the arithmetic progression
\begin{equation}\label{latt}
{\ensuremath{\rm spec}} (A)=\bigcup_{n\in \mathbb Z}\left \{\frac{\Theta +2\pi n}{\ell}\right \}.
\end{equation}
However, the spectrum of a $G$-invariant operator is a $G$-invariant set:
if $\lambda $
is an eigenvalue of $A$, then for all $t\in \mathbb R$ the point $g_t(\lambda)=\lambda+t$ is also an eigenvalue of $A$ which is impossible, for the underlying Hilbert space is separable.
Thus, $ \dot A$ has no $G$-invariant self-adjoint extension.
This observation justifies the claim from Remark \ref{pervaia}
that the restricted flow $\Gamma_g|_{{\mathcal U}}$ may not have fixed points at all, even in the case of equal deficiency indices.
\subsection{}Next, since $\dot A$ from Example \ref{ex1} has no $G$-invariant self-adjoint extension, our general fixed point Theorem \ref{Shauder} states that there exists a maximal dissipative
$G$-invariant extension. In turn, Theorem \ref{tri} shows then that this extension is unique.
Clearly, that $G$-invariant extension is given by the differentiation operator with the Dirichlet boundary condition at the left end-point of the interval,
\begin{equation}\label{disdom}
(Af)(x)=i\frac{d}{dx}f(x) \quad \text{on} \quad \mathrm{Dom}(A)=\{f\in W_2^1(0, \ell)\, |\, f(0)=0\}.
\end{equation}
\subsection{}\label{can}
It is worth mentioning that the dissipative extension $A$ defined by \eqref{disdom} generates a nilpotent semi-group of shifts
$V_s$
with index $\ell$. Indeed,
$$(V_sf)(x)=\begin{cases}f(x-s),&x>s\\
0,&x\le s\end{cases},
\quad f\in
L^2(0,\ell),
$$
and therefore
$V_s=0$ for $ s\ge \ell.
$
An immediate computation shows that the restricted Weyl commutation relations
\begin{equation}\label{nunu}
U_tV_s=e^{ist}V_{s}U_t, \quad t\in \mathbb R, \quad s\ge0,
\end{equation}
for the unitary group $U_t$ and the semi-group of contractions $V_s$ are satisfied.
We remark that the generators $A$, the differentiation operators on a finite interval defined by \eqref{disdom},
are not unitarily equivalent for different values of the parameter $\ell$, the length of the interval
$[0, \ell]$ (see \cite{AkG}). Therefore, the restricted Weyl
commutation relations
\eqref{nunu}
admit a variety of non-unitarily equivalent solutions.
\subsection{}Finally, we illustrate Theorem \ref{discrete-invariant}.
One observes that any maximal dissipative, in particular, self-adjoint extension of $\dot A$
has the property that
\begin{equation}\label{inv}
U_t\left ( \mathrm{Dom} (A)\right )=\mathrm{Dom} (A)\quad \text{for all }\quad t=\frac{2\pi }{\ell}n, \quad n \in \mathbb Z.
\end{equation}
Therefore, $A$ is $G[g]$-invariant, where $G[g]$ is the discrete subgroup
of the one-parameter group $G$ \eqref{shi} generated by the translations
$$
g(x)=x+\frac{2\pi}{\ell}.
$$
Indeed, as it is shown in \cite{AkG}, the domain of any maximal dissipative extensions admits the representation
$$\mathrm{Dom} (A)= \{f\in W^{1}_2(0,\ell)\, | f(0)=\rho f(\ell)\}$$
for some $|\rho|\le 1$. Therefore, the domain invariance property \eqref{inv} automatically holds. In particular, the spectrum of $A$
is a lattice in the upper half-plane with period
$ \frac{2\pi}{\ell}$, whenever $A$ is maximal dissipative non-self-adjoint.
If $|\rho|=1$, and hence the operator $A$ is self-adjoint, the spectrum coincides with the set of points forming
the arithmetic progression \eqref{latt} (for some $\Theta$).
Our second example relates to the homogeneous Schr\"odinger operator with a singular potential. In this example,
the underlying one-parameter group coincides with the group of scaling transformations of the real axis.
\begin{example}\label{ex2}
Suppose that $G=\{g_t\}_{t\in \mathbb R}$ is the group of scaling transformations
$$
g_t(x)=e^{t}x.
$$
Let $\dot A$ be
the closure in $L^2(0,\infty)$ of the
symmetric operator given by the differential expression
\begin{equation}\label{formself}
\tau_\nu=-\frac{d^2}{dx^2}+\frac{\gamma}{x^2}
\end{equation}
initially defined on $C_0^\infty(0, \infty)$.
It is known, see, e.g., \cite[Ch. X]{Simon}, that
\begin{itemize}
\item[(i)] if $\gamma <\frac34$, the operator $\dot A$ has deficiency indices $(1,1)$,
\item[] and
\item[(ii)]
it
is non-negative for $-\frac14\le \gamma $,
\item[]
while
\item[(iii)] $\dot A $ is not semi-bounded from below for $\gamma<-\frac14$.
\end{itemize}
Regardless of the magnitude of the coupling constant $\gamma$, the operator $\dot A$ is $G$-invariant (homogeneous) with respect to the unitary group of scaling transformations
\begin{equation}\label{give9}
(U_tf)(x)=e^{\frac{t}{4}}f(e^{\frac{t}{2}}x), \quad f\in L^2(\mathbb R_+).
\end{equation}
\end{example}
\subsection{}\label{7.5} First, we notice that as in Example \ref{ex1}, the symmetric $G$-invariant operator $\dot A$ from Example \ref{ex2} with
$$
\gamma<-\frac14
$$
does not admit self-adjoint $G$-invariant extensions at all.
Indeed,
it is well known (see, e.g., \cite{PZ}) that if $A$ is a self-adjoint extension of $\dot A$, then the domain of $\dot A$ can be characterized by the following asymptotic
boundary condition: there exists a $\Theta \in [0, 2\pi )$ such that for any
$$
f\in \mathrm{Dom} (A)\subset W^{2}_{2, \text{loc}}(0,\infty) $$
the asymptotic representation
\begin{equation}\label{asbound}
f(x)\sim C\sqrt{x}\sin\left ( \nu \log x +\Theta\right ) \quad \text{as}\,\, x\downarrow 0,
\end{equation}
holds for some $C\in \mathbb C$, where
$$
\nu= \sqrt{\left |\gamma+\frac14\right |}.
$$
Clearly,
$$
U_t \left (\mathrm{Dom} (A)\right )=\mathrm{Dom} (A) \quad \text{for all }\quad t=\frac{4\pi n}{\nu}, \quad n \in \mathbb Z,
$$
where the scaling group $U_t$ is given by \eqref{give9}. Therefore, the extension $ A$ is $G[g]$-invariant where
$G[g]$ denotes the discrete cyclic subgroup generated by the linear transformation $g\in {\mathcal G}$,
\begin{equation}\label{general}
g(x)=\kappa x \quad \text{for } \quad \kappa = e^{4\pi/\nu}.
\end{equation}
In particular, if $\lambda_0$ is a negative eigenvalue of the self-adjoint extension $A$, then the two-sided geometric progression
\begin{equation}\label{2side}
\lambda_n=\kappa^n \lambda_0, \quad n \in \mathbb Z,
\end{equation}
fills in the (negative) discrete spectrum of A.
It remains to argue exactly as in the case of the differentiation operator on a finite interval from Example \ref{ex1}
to show that $\dot A$ has no $G$-invariant self-adjoint extension.
Also, applying Theorem \ref{discrete-invariant} shows that any maximal extension of $\dot A$ is invariant with respect to the cyclic group $G[g]$, where $g$ is given by \eqref{general}.
\begin{remark}\label{fall} We remark that the self-adjoint realizations of the Schr\"odinger operator
$$
H=-\frac{d^2}{dx^2}+\frac{\gamma}{x^2}, \quad \gamma<-\frac14,
$$
defined by the asymptotic boundary conditions \eqref{asbound}, are well-suited for modeling the fall to the center phenomenon in quantum mechanics \cite[Sec. 35]{LL}, see also \cite{PP} and \cite{PZ}.
From the group-theoretic standpoint this phenomenon can be expressed as follows. The initial symmetric homogeneous operator (with respect to the scaling transformations) is neither semi-bounded from below nor admits self-adjoint homogeneous realizations, while any self-adjoint realization of the corresponding Hamiltonian is homogeneous with respect to a cyclic subgroup of the transformations only.
As a consequence, its negative discrete spectrum as a set is invariant with respect to the natural action of this subgroup of scaling transformations of the real axis.
That is,
$$
{\ensuremath{\rm spec}}(A)\cap (-\infty, 0)=\bigcup_{n\in \mathbb Z}\{\kappa^n\lambda_0\}.
$$
\end{remark}
\subsection{}\label{8.6} Theorem \ref{Shauder} combined with Theorem \ref{tri} shows that
$\dot A$
has a unique maximal dissipative $G$-invariant extension which coincides with the following
homogeneous operator $A$, the restriction of the adjoint operator
$(\dot A)^*$ on
\begin{equation}\label{domai}
\mathrm{Dom}(A)=\left \{f\in \mathrm{Dom} ((\dot A)^* )\, \bigg |\,
\lim_{x\downarrow 0}\frac{f(x)}{x^{1/2+i\sqrt{\left |\gamma+\frac14\right |} }}\, \text{ exists}\right \}.
\end{equation}
It is clear that $A$
is $G$-invariant. We claim without proof that $A$ is a maximal dissipative operator.
\subsection{}\label{gencan} The dissipative Schr\"odingier operator $A$ with a singular potential from subsection \ref{8.6} generates
the contractive semi-group $ V_s=e^{iAs}$, $s\ge0$.
In accordance with Theorem \ref{bccr2}, the restricted generalized Weyl commutation relations \begin{equation}\label{ccrcr}
U_tV_s=\exp (ise^{t})V_{e^{t}s}U_t, \quad t\in \mathbb R, \,\,\, s\ge0,
\end{equation}
hold.
One can also show that the homogeneous extensions $A$ are not unitarily equivalent for different values of the coupling constant $\gamma $ satisfying the inequality $\gamma<-\frac14$.
Therefore, the restricted generalized Weyl commutation relations \eqref{ccrcr} admit
a continuous family of non-unitarily equivalent solutions (cf. Subsection \ref{can}).
\subsection{} Finally, we illustrate Theorem \ref{FKnet} which deals with the case of semi-bounded $G$-invariant operators.
Assuming that
\begin{equation}\label{buss}
-\frac14\le \gamma <\frac34,
\end{equation} one observes that
the symmetric operator $\dot A$ is non-negative. Moreover, it is clear that the following two homogenous extensions,
the restrictions of the adjoint operator
$(\dot A)^*$ on
$$
\mathrm{Dom}(A_F)=\left \{f\in \mathrm{Dom} ((\dot A)^* )\, \bigg |\,
\lim_{x\downarrow 0}\frac{f(x)}{x^{1/2+\sqrt{\gamma+\frac14} }}\, \text{ exists}\right \}
$$
and on
$$
\mathrm{Dom}(A_K)=\left \{f\in \mathrm{Dom} ((\dot A)^* )\, \bigg |\,
\lim_{x\downarrow 0}\frac{f(x)}{x^{1/2-\sqrt{\gamma+\frac14} }}\, \text{ exists}\right \},
$$
respectively,
are $G$-invariant with respect to the unitary representation \eqref{give9}.
We remark that the self-adjoint realization $A_F$ and $A_K$ (cf. \cite{EK,GP,Ka}) can be recognized as
the Friedrichs and Krein-von Neumann extensions of $\dot A$, respectively, which agrees with the statement of Theorem \ref{FKnet}.
We also remark that under hypothesis \eqref{buss}, the corresponding flow has two different fixed point on the unit circle.
In the critical case of
$$\gamma=-\frac14,$$ the extensions $A_K$ and $A_F$ coincide. Hence, the flow \eqref{flow1} has only one fixed point on the boundary $\mathbb T$ of the unit disk $\mathbb D$ (cf. Proposition \ref{nado}).
We conclude this section by an example of a $G$-invariant generator that has deficiency indices $(0,1)$.
We will see below that this example is universal in the following sense. The semi-group of unilateral shifts $V_s=e^{isA}$, $s\ge 0$, generated by $A$ solves the restricted standard as well as generalized Weyl commutation relations \eqref{ninii} and \eqref{nini2}.
\begin{example}\label{ex3}
Let $\dot A$ be the Dirichlet differentiation operator on the positive semi-axis (in $L^2(0, \infty)$),
$$
(\dot A) f(x)=i\frac{d}{dx} f(x) \quad \text{on} \quad \mathrm{Dom} (\dot A) =\{f\in W^1_2(0,\infty)\,|\, f(0)=0\}.
$$
It is well known, see, e.g., \cite{AkG}, that the symmetric operator $\dot A$ has deficiency indices $(0,1)$ and it is simultaneously a maximal dissipative operator.
Suppose that $G^{(1)}=\{g_t\}_{t\in \mathbb R}$ is the group of affine transformations
$$
g_t(x)=x+t.
$$
Clearly, $\dot A$ is $G^{(1)}$-invariant with respect to
the group of unitary transformations given by
$$
(U_t^{(1)}f)(x)=e^{ixt}f(x), \quad f\in L^2(0,\infty).
$$
Denoting by $G^{(2)}=\{g_t\}_{t\in \mathbb R}$ the group of scaling transformations
$$
g_t(x)=e^tx,
$$
one also concludes that $\dot A$ is $G^{(2)}$-invariant with respect to
the group of unitary transformations given by
\begin{equation}\label{give10}
(U_t^{(2)}f)(x)=e^{\frac{t}{2}}f(e^{t}x), \quad f\in L^2(0,\infty).
\end{equation}
In particular, for the semi-group of unilateral right shifts $V_s=e^{isA}$ generated by the differentiation operator $A=\dot A$, one gets
the restricted Weyl commutation relations
$$
U_t^{(1)}V_s=e^{ist}V_{s}U_t^{(1)}, \quad t\in \mathbb R, \quad s\ge0,
$$
as well as the restricted generatlized Weyl commutation relations
$$
U_t^{(2)}V_s=\exp (ise^{t})V_{e^{t}s}U_t^{(2)}, \quad t\in \mathbb R, \,\,\, s\ge0.
$$
We also remark that the operator $A=\dot A$ is ${\mathcal G}$-invariant with respect to the unitary representation
${\mathcal G}\ni g\mapsto U_g$ of the whole affine group ${\mathcal G}$ given by
$$
(U_g)f(x)=a^{1/2}e^{ibx}f(ax),\quad f\in L^2(0,\infty), \quad g\in {\mathcal G}, \quad g(x)=ax+b.
$$
\end{example}
|
\section{INTRODUCTION}
It is known that low-mass stars comprise a significant fraction of stars in the Solar neighborhood
and the Galaxy as a whole. The Galaxy may be teeming with even smaller mass brown dwarfs. Therefore,
studying the abundance and properties of low-mass stars and brown dwarfs is of fundamental importance.
There have been surveys searching for very low-mass (VLM) objects \citep{reid08,aberasturi14}, but
these surveys are limited to the immediate solar neighborhood. As a result, the sample of VLM objects
is small despite their intrinsic numerosity and thus our understanding about VLM objects is poor.
Microlensing surveys detect objects through their gravitational fields rather than their radiation
and thus microlensing can provide a powerful probe of VLM objects. However, the weakness of microlensing
is that it is difficult to determine the lens mass for general microlensing events. This difficulty
arises due to the fact that the time scale of an event, which is the only observable related to the
physical lens parameters, results from the combination of the lens mass, distance and the relative
lens-source transverse speed. As a result, it is difficult to identify and characterize VLM objects
although a significant fraction of lensing events are believed to be produced by these objects.
However, it is possible to uniquely determine the physical lens parameters and thus identify VLM
objects for a subset of lensing events produced by lenses composed of two masses. For unique
determinations of the physical lens parameters, it is required to simultaneously measure the
angular Einstein radius $\theta_{\rm E}$ and the microlens parallax $\pi_{\rm E}$ that are related
to the lens mass $M$ and distance to the lens $D_{\rm L}$ by
\begin{equation}
M_{\rm tot}={\theta_{\rm E} \over \kappa\pi_{\rm E}};\qquad
D_{\rm L}={{\rm AU}\over \pi_{\rm E}\theta_{\rm E}+\pi_{\rm S}},
\label{eq1}
\end{equation}
respectively \citep{gould00}. Here $\kappa=4G/(c^{2}{\rm AU})$, $\pi_{\rm S}={\rm AU}/D_{\rm S}$
is the parallax of the source star, and $D_{\rm S}$ is the distance to the source. The angular
Einstein radius is estimated by analyzing deviations in lensing light curves caused by the finite
size of the lensed source stars \citep{gould94,nemiroff94,witt94}. For a single-lens events,
finite-source effects can be detected only for a very small fraction of extremely high-magnification
events where the lens-source separation at the peak magnification is equivalent to the source size.
On the other hand, light curves of binary-lens events usually result from caustic crossings during
which finite-source effects become important and thus the chance to detect finite-source effects
and measuring the Einstein radius is high. Furthermore, binary-lens events tend to have longer
time scales than single-lens events and this also contributes to the higher chance to measure the
lens parallax. In fact, most known VLM lensing objects were identified through the channel of
binary-lens events \citep{hwang10,shin12,choi13,han13,park13,jung15}.
Despite the usefulness of binary-lens events, the chance to identify VLM objects has been low.
One important reason for the low chance is that caustic crossings last for very short periods of
time. The duration of a caustic crossing is
\begin{equation}
t_{\rm cc} = {2\rho \over \sin\phi}t_{\rm E},
\label{eq2}
\end{equation}
where $t_{\rm E}$ is the Einstein time scale of the event, $\phi$ is the entrance angle of the
source star with respect to the caustic line, and $\rho=\theta_{\ast}/\theta_{\rm E}$ is the
angular source radius $\theta_{\ast}$ normalized to the angular Einstein radius $\theta_{\rm E}$.
Considering that the Einstein time scale is $\sim({\cal O}) 10$ days and the Einstein radius is
$\sim({\cal O})$ milli-arcsec
for typical Galactic microlensing events, the duration of a caustic crossing is $\sim({\cal O})$ hours for
Galactic bulge source stars with angular stellar radii of $\sim({\cal O})$ 1--10 $\mu$-arcsec. Therefore,
it is difficult to densely resolve caustic crossings from surveys that are being carried out with
over hourly observational cadences.
Another reason for the low chance of resolving caustic crossings is the difficulty in predicting
their occurrence. Caustics produced by a binary lens form a single or multiple sets of closed curves
and thus caustic crossings always come in pairs. Although it is difficult to predict the first
crossing (caustic entrance) based on the fraction of the light curve before the caustic entrance,
the second crossing (caustic exit) is guaranteed after the caustic entrance. To resolve the short-lasting
caustic exit, it is required to precisely predict the time of the caustic crossing so that observation
can be focused to resolve caustic crossings. This requires vigilant modeling of a lensing event conducted
with the progress of the event followed by intensive follow-up observation.
In this paper, we report the discovery of a VLM binary that was detected from the caustic-crossing
binary-lens microlensing event OGLE-2013-BLG-0578. The caustic exit of the event was precisely predicted
by real-time modeling, enabling dense resolution and complete coverage of the caustic crossing. Combined
with the Einstein radius measured from the caustic-crossing part of the light curve and the lens
parallax measured from the long-term deviation induced by the Earth's parallactic motion, we uniquely
measure the lens mass and identify that the lens is a VLM binary composed of a low-mass star and
a brown dwarf.
\section{Observation}
The microlensing event OGLE-2012-BLG-0578 occurred on a star located toward the Galactic Bulge direction.
The equatorial coordinates of the lensed star are $(\alpha,\delta)_{\rm J2000}=(17^{\rm h}59^{\rm m}
59^{\rm s}\hskip-2pt.85, -29^{\circ}44'06''\hskip-2pt.9)$, that correspond to the Galactic coordinates
$(l,b)=(0^{\circ}\hskip-2pt.90, -3^{\circ}\hskip-2pt.10)$. The event was first noticed on April 22,
2013 from survey observations conducted by the Optical Gravitational Lensing Experiment
\citep[OGLE:][]{udalski03} using the 1.3m Warsaw telescope at the Las Campanas Observatory in Chile.
Images were taken using primarily in {\it I}-band filter and some {\it V}-band images were also taken
to constrain the lensed star (source).
In Figure~\ref{fig:one}, we present the light curve of the event. The light curve shows two distinctive
spikes that are characteristic features of a caustic-crossing binary-lens event. The caustic crossings
occurred at ${\rm HJD}'={\rm HJD}-2450000\sim6426.0$ and $6461.8$. Although the event was first noticed
before the caustic crossings, the caustic entrance was missed. With the progress of the event, it became
clear that the event was produced by a binary lens from the characteristic ``U''-shape trough between the
caustic crossings.
\begin{figure}[ht]
\epsscale{1.15}
\plotone{fig1.eps}
\caption{\label{fig:one}
Light curve of OGLE-2013-BLG-0578. Also plotted is the best-fit model
(``orbit+parallax'' model with $u_{0}>0$). The two bottom panels show
the residuals from the modeling with and without considering
lens-orbital and parallax effects. The upper panel shows the enlargement
of the caustic-exit region and the times of caustic-crossing alerts.
}\end{figure}
Although the first caustic crossing was missed, the second crossing was densely resolved. Resolving the
crossing became possible with the prediction of the caustic crossing from vigilant modeling of the light
curve followed by intensive follow-up observation. It is known that reliable prediction of the second
caustic crossing is difficult based on the light curve before the minimum between the two caustic crossings
\citep{jaro01}. With the emergence of the correct model after passing the caustic trough, we focused on
the prediction of the exact caustic-crossing time. The first alert of the caustic exit was issued on June 17, 2013,
1.3 days before the actual caustic exit. On the next day, the second alert was issued to predict more
refined time of the caustic exit. In response to the alert, the OGLE experiment, which is usually operated
in survey mode, entered into ``following-up'' mode observation by increasing the observation cadence.
Thanks to the intensive follow-up observation, the caustic exit was completely and densely covered. See the
upper panel of Figure~\ref{fig:one}.
In order to securely measure the baseline magnitude and detect possible higher-order effects, observation
was continued after the caustic exit to the end of the Bulge season. From these observations, we obtain
2284 and 27 images taken in {\it I} and {\it V} bands, respectively. Photometry of the event was done by
using the customized pipeline \citep{udalski03} that is based on the Difference Imaging Analysis method
\citep{alard98,wozniak00}. We note that the {\it I}-band data are used for light curve analysis, while
the {\it V}-band data are used for the investigation of the source type.
It is known that photometric errors estimated by an automatic pipeline are often underestimated and thus
errors should be readjusted. We readjust error bars by
\begin{equation}
e'=k(e^{2}+e_{\rm min}^{2})^{1/2}.
\label{eq3}
\end{equation}
Here $e_{\rm min}$ is a term used to make the cumulative distribution function of $\chi^{2}$ as a function
of lensing magnification becomes linear. This process is needed to ensure that the dispersion of data points
is consistent with error bars of the source brightness. The other term $k$ is a scaling factor used to make
$\chi^{2}$ per degree of freedom (dof) becomes unity.
\section{Analysis}
We analyze the event by searching for the set of lensing parameters (lensing solution) that best describe
the observed light curve. Basic description of a binary-lens event requires seven standard lensing parameters.
Three of these parameters describe the source-lens approach including the time of the closest approach of the
source to a reference position of the lens, $t_{0}$, the lens-source separation at $t_{0}$ in units of the
Einstein radius, $u_{0}$, and the time required for the source to cross the Einstein radius, $t_{\rm E}$ (Einstein
time scale). In our analysis, we set the center of the mass of the binary lens as a reference position in the
lens plane. Another two parameters describe the binary lens including the projected binary separation in units
of the Einstein radius, $s$ (normalized separation), and the mass ratio between the lens components, $q$. Due
to the asymmetry of the gravitational field around the binary lens, it is needed to define the angle between
the source trajectory and the binary axis, $\alpha$ (source trajectory angle). The last parameter is the
normalized source radius $\rho$, which is needed to describe the caustic-crossing parts of the light curve
that are affected by finite-source effects.
It is often needed to consider higher-order effects in order to precisely describe lensing light curves and
this requires to include additional lensing parameters. For long time-scale events, such effects are caused
by the positional change of an observer induced by the orbital motion of the Earth around the Sun
\citep[``parallax effect'':][]{gould92} and/or the change of the binary separation and orientation caused by the
orbital motion of the lens \citep[``lens orbital effect'':][]{dominik98,albrow00}. The analyzed event lasted
throughout the whole Bulge season and thus these effects can be important. The parallax effect is described
by two parameters, $\pi_{\rm E,\it N}$ and $\pi_{\rm E,\it E}$, that are the two components of the lens
parallax vector ${\mbox{\boldmath $\pi$}}_{\rm E}$ projected onto the sky along the north and east equatorial coordinates,
respectively. The direction of the lens parallax vector corresponds to the relative lens-source proper motion
and its magnitude corresponds to the relative lens-source parallax $\pi_{\rm rel}={\rm AU}(D_{\rm L}^{-1}
-D_{\rm S}^{-1})$ scaled to the Einstein radius of the lens, i.e.,
\begin{equation}
\pi_{\rm E} = { \pi_{\rm rel} \over \theta_{\rm E} }.
\label{eq4}
\end{equation}
To the first-order approximation, the lens orbital effect is described by two parameters, $ds/dt$ and $d\alpha/dt$,
which are the change rates of the normalized binary separation and the source trajectory angle, respectively.
\begin{figure}[ht]
\epsscale{1.15}
\plotone{fig2.eps}
\caption{\label{fig:two}
The source trajectory (the curve with an arrow) with respect to
the caustic (the closed curve with 6 cusps). The two small dots
marked by $M_1$ and $M_2$ represent the positions of the binary
lens components. All lengths are scaled by the angular Einstein
radius corresponding to the total mass of the binary lens. Due
to the change of the relative positions of the binary lens components
caused by the orbital motion, the caustic varies in time. We present
4 caustics and lens positions corresponding to the times marked
in right upper corner of the panel. The source trajectory is curved
due to the combination of the lens-orbital and parallax effects.
}\end{figure}
To model caustic-crossing parts of the light curve, it is needed to compute magnifications affected by
finite-source effects. To compute finite-source magnifications, we use the numerical method of the inverse
ray-shooting technique \citep{kayser86,schneider86} in the immediate neighboring region around caustics and
the semi-analytic hexadecapole approximation \citep{pejcha09,gould08} in the outer region surrounding caustics.
We consider the effects of the surface brightness variation of the source star. The surface brightness is
modeled as
\begin{equation}
S_{\lambda}\propto 1-\Gamma_{\lambda} \left(1-{3\over 2}\cos \psi \right),
\label{eq5}
\end{equation}
where $\Gamma_{\lambda}$ is the linear limb-darkening coefficient, $\lambda$ is the passband, and $\psi$ is
the angle between the line of sight toward the source star and the normal to the source surface. We adopt
the limb-darkening coefficients $(\Gamma_{V}, \Gamma_{I})=(0.62, 0.45)$ from \citet{claret00} based on the
source type. The source type is determined based on its de-reddened color and brightness. See Section 4
for details about how the source type is determined.
Searching for the best-fit solution of the lensing parameters is carried out based on the combination of
grid-search and downhill approaches. We set $(s,q,\alpha)$ as grid parameters because lensing magnifications
can vary dramatically with small changes of these parameters. On the other hand, magnifications vary smoothly
with changes of the remaining parameters, and thus we search for these parameters by using a downhill approach.
We use the Markov Chain Monte Carlo (MCMC) method for the downhill approach. Searching for solutions throughout
the grid-parameter spaces is important because it enables one to check the possible existence of degenerate
solutions where different combinations of the lensing parameters result in similar light curves.
In the initial search for solutions, we conduct modeling of the light curve based on the 7 basic binary-lensing
parameters (``standard model''). From this, it is found that the event was produced by a binary with a projected
separation very close to the Einstein radius, i.e. $s\sim1.0$. Caustics for such a resonant binary form a single
big closed curve with 6 cusps. The two spikes of the light curve were produced by the source trajectory passing
diagonally through the caustic. See Figure~\ref{fig:two} where we present the source trajectory with respect to
the caustic\footnotemark[7]. The estimated mass ratio between the binary components is $q\sim0.2-0.3$.
\footnotetext[7]{
We note that the source trajectory is curved due to the combination of
the lens-orbital and parallax effects. We also note that the caustic
varies in time due to the positional change of the binary-lens components
caused by the orbital motion.}
Although the standard model provides a fit that matches the overall pattern of the observed light curve, it is
found that there exists some residual in the region around ${\rm HJD}'\sim6410$. See the bottom panel of
Figure~\ref{fig:one}. We, therefore, check whether higher-order effects improve fit. We find that separate
consideration of the parallax effect (``parallax model'') and the lens-orbital effect (``orbital model'')
improves fit by $\Delta\chi^{2}=579$ and $344$, respectively. When both effects are simultaneously considered
(``orbit+parallax model''), the fit improves by $\Delta\chi^{2}=616$, implying that both effects are important.
In Figure~\ref{fig:three}, we present the contours of $\Delta\chi^{2}$ in the space of the higher-order lensing
parameters. Contours marked in different colors represent the regions with $\Delta\chi^{2}<1$ (red),
$4$ (yellow), $9$ (green), $16$ (sky blue), $25$ (blue), and $36$ (purple). It shows that the higher-order
effects are clearly detected. Considering the time gap between the caustic crossings that is approximately
a month and the long duration of the event that lasted throughout the whole Bulge season, the importance of
the higher-order effects is somewhat expected.
\begin{deluxetable}{lrr}
\tablecaption{Lensing parameters \label{table:one}}
\tablewidth{0pt}
\tablehead{
\multicolumn{1}{c}{Parameter} &
\multicolumn{1}{c}{$u_{0}>0$} &
\multicolumn{1}{c}{$u_{0}<0$}
}
\startdata
$\chi^2/{\rm dof}$ & 2303.1/2300 & 2328.1/2300 \\
$t_{0}$ (HJD) & 2456440.13 $\pm$ 0.11 & 2456439.99 $\pm$ 0.08 \\
$u_{0}$ & 0.109 $\pm$ 0.004 & -0.120 $\pm$ 0.003 \\
$t_{\rm E}$ (days) & 72.11 $\pm$ 0.84 & 70.79 $\pm$ 0.10 \\
$s$ & 1.027 $\pm$ 0.004 & 1.035 $\pm$ 0.002 \\
$q$ & 0.260 $\pm$ 0.004 & 0.241 $\pm$ 0.003 \\
$\alpha$ (rad) & 0.676 $\pm$ 0.007 & -0.647 $\pm$ 0.007 \\
$\rho$ ($10^{-3}$) & 0.96 $\pm$ 0.01 & 0.97 $\pm$ 0.01 \\
$\pi_{\rm E,\it N}$ & 0.54 $\pm$ 0.08 & -0.51 $\pm$ 0.05 \\
$\pi_{\rm E,\it E}$ & -0.53 $\pm$ 0.03 & -0.72 $\pm$ 0.02 \\
$ds/dt$ (${\rm yr}^{-1}$) & 0.48 $\pm$ 0.14 & 0.40 $\pm$ 0.03 \\
$d\alpha/dt$ (${\rm rad}$ ${\rm yr}^{-1}$) & 1.89 $\pm$ 0.24 & -1.75 $\pm$ 0.13
\enddata
\end{deluxetable}
It is known that lensing events with higher-order effects are subject to the degeneracy caused by the mirror
symmetry of the source trajectory with respect to the binary axis \citep{skowron11}. This so-called ``ecliptic
degeneracy'' is important for Galactic Bulge events that occur near the ecliptic plane. The pair of the solutions
resulting from this degeneracy have almost identical parameters except $(u_{0}, \alpha, \pi_{\rm E,\it N}, d\alpha/dt)
\rightarrow -(u_0,\alpha,\pi_{{\rm E}, N},d\alpha/dt)$. It is found that $u_{0}>0$ is marginally preferred
over the $u_{0}<0$ solution by $\Delta\chi^{2}=25.0$, which corresponds to formally $\sim5\sigma$ level difference.
However, this level of $\Delta\chi^{2}$ can often occur due to systematics in data and thus one cannot
completely rule out the $u_{0}<0$ solution. In Table~\ref{table:one}, we present the best-fit parameters of both
$u_{0}>0$ and $u_{0}<0$ solutions. We note that the uncertainties of the lensing parameters are estimated based
on the distributions of the parameters obtained from the MCMC chain of the solution. We also present the best-fit
model light curve ($u_{0}>0$ solution) and the corresponding source trajectory with respect to the lens and
caustic in Figures~\ref{fig:one} and~\ref{fig:two}, respectively.
\begin{figure}[ht]
\epsscale{1.15}
\plotone{fig3.eps}
\caption{\label{fig:three}
Contours of $\Delta\chi^{2}$ in the space of the parallax and
lens-orbital parameters for the best-fit model ($u_{0}>0$
model). Contours marked in different colors represent the
regions with $\Delta\chi^{2}<1$ (red), $4$ (yellow), $9$
(green), $16$ (sky blue), $25$ (blue), and $36$ (purple).
}\end{figure}
\begin{figure}[ht]
\epsscale{1.15}
\plotone{fig4.eps}
\caption{\label{fig:four}
Location of the lensed star with respect to the centroid of giant
clump in the color-magnitude diagram of neighboring stars.
}\end{figure}
\section{Physical Parameters}
By detecting both finite-source and parallax effects, one can measure the angular Einstein radius and the lens
parallax, which are the two quantities needed to determine the mass and distance to the lens. The lens parallax
is estimated by $\pi_{\rm E}=(\pi_{\rm E,\it N}^{2}+\pi_{\rm E,\it E}^{2})^{1/2}$ from the parallax parameters
determined from light-curve modeling.
In order to estimate the angular Einstein radius, it is needed to convert the measured normalized source radius
$\rho$ into $\theta_{\rm E}$ by using the angular radius of the source star, i.e. $\theta_{\rm E}=\theta_{\ast}/\rho$.
The angular source radius is estimated based on the de-reddened color $(V-I)_{0}$ and brightness $I_{0}$ of the
source star which are calibrated by using the centroid of the Bulge giant clump on the color-magnitude diagram
as a reference \citep{yoo04}. By adopting the color and brightness of the clump centroid $(V-I,I)_{0,{\rm c}}=
(1.06, 14.45)$ \citep{bensby11,nataf13}, we estimate that $(V-I,I)_{0}=(0.72, 17.68)$ for the source star,
implying that the source star is a G-type subgiant. Figure~\ref{fig:four} shows the locations of the source and
the centroid of giant clump in the color-magnitude diagram of stars around the source star. We then translate
$V-I$ into $V-K$ using the color-color relation of \citet{bessell88} and obtain the angular source radius by
using the relation between $V-K$ and $\theta_{\ast}$ of \citet{kervella04}. The determined angular source radius
is $\theta_{\ast}=0.93\pm0.06$ $\mu$as. Then, the Einstein radius is estimated as $\theta_{\rm E}=0.97\pm0.07$
mas for the best-fit solution ($u_{0}>0$ model). We note that $u_{0}<0$ model results in a similar Einstein
radius due to the similarity in the measured values of $\rho$. Combined with the measured Einstein time scale
$t_{\rm E}$, the relative lens-source proper motion is estimated as $\mu=\theta_{\rm E}/t_{\rm E}=4.90\pm0.35$
mas ${\rm yr}^{-1}$.
With the measured lens parallax and Einstein radius, we determine the mass and the distance to the lens using
Equation (\ref{eq1}). In Table~\ref{table:two}, we list the physical parameters of the lens system corresponding
to the $u_{0}>0$ and $u_{0}<0$ solutions. We note that the estimated parameters from the two solutions are
similar. According to the estimated mass, the lens system is composed of a substellar brown-dwarf companion
and a late-type M-dwarf primary. The distance to the lens is $D_{\rm L}<1.2$ kpc and thus the lens is located
in the Galactic disk. The projected separation between the lens components is $r_{\perp}=sD_{\rm L}\theta_{\rm E}$
is slightly greater than 1 AU. In order to check the validity of the obtained lensing solution, we compute
the projected kinetic to potential energy ratio $\rm (KE/PE)_{\perp}$ by
\begin{equation}
\left( {\rm KE \over \rm PE} \right)_{\perp} = {(r_{\perp}/{\rm AU})^{2} \over 8{\pi}^{2}(M/M_{\odot})}
\left[ \left( {1 \over s}{ds \over dt} \right)^{2} + \left( { d\alpha \over dt} \right)^{2} \right]
\label{eq6}
\end{equation}
\citep{dong09}. The estimated ratio is $\rm (KE/PE)_{\perp}<1$ for both $u_{0}>0$ and $u_{0}<0$ solutions and thus
meets the condition of a bound system.
\begin{deluxetable}{lrr}[h]
\tablecaption{Physical parameters\label{table:two}}
\tablewidth{0pt}
\tablehead{
\multicolumn{1}{c}{Parameter} &
\multicolumn{1}{c}{$u_0>0$} &
\multicolumn{1}{c}{$u_0<0$}
}
\startdata
Einstein radius (mas) & 0.97 $\pm$ 0.07 & 0.96 $\pm$ 0.07 \\
Proper motion (mas $\rm yr^{-1}$) & 4.90 $\pm$ 0.35 & 4.96 $\pm$ 0.35 \\
Total mass ($M_\odot$) & 0.156 $\pm$ 0.017 & 0.133 $\pm$ 0.011 \\
Mass of primary ($M_\odot$) & 0.124 $\pm$ 0.014 & 0.107 $\pm$ 0.009 \\
Mass of companion ($M_\odot$) & 0.032 $\pm$ 0.004 & 0.026 $\pm$ 0.002 \\
Distance (kpc) & 1.16 $\pm$ 0.11 & 1.02 $\pm$ 0.08 \\
Projected separation (AU) & 1.16 $\pm$ 0.11 & 1.02 $\pm$ 0.08 \\
$({\rm KE}/{\rm PE})_{\perp}$ & 0.48 $\pm$ 0.20 & 0.32 $\pm$ 0.07
\enddata
\end{deluxetable}
\section{Conclusion}
We presented the analysis of a caustic-crossing binary-lens microlensing event OGLE-2013-BLG-0578 that led to
the discovery of a binary system composed of a substellar brown-dwarf companion and a late-type M-dwarf primary.
Identification of the lens became possible due to the prediction of the caustic crossing from vigilant real-time
modeling and resolution of the caustic from prompt follow-up observation. The event demonstrates the capability
of current real-time modeling and the usefulness of microlensing in detecting and characterizing
faint or dark objects in the Galaxy.
\acknowledgments
Work by C.H. was supported by Creative Research Initiative
Program (2009-0081561) of National Research Foundation of
Korea. We acknowledge funding from the European Research
Council under the European Community's Seventh Framework
Programme (FP7/2007-2013)/ERC grant agreement No. 246678
to A.U. for the OGLE project.
|
\section{Introduction}
The study of nanoparticles (or colloids) dispersed in a sea of solvent particles is a popular topic of research in the statistical physics of liquids. One reason
for this popularity is the possibility of applying coarse-graining techniques integrating out certain degrees of freedom of the complex system, which may obey a
relatively simple interaction model that, on one hand, still possesses most of the crucial physics and, on the other, is computationally tractable via well developed
methods from the statistical theory of simple fluids. A well-known example presents the simple model introduced originally by Asakura and Oosawa for a mixture of
colloids and non-adsorbing polymers \cite{ao1,ao2, vrij}. This model represents an extreme case of a non-additive hard-sphere mixture where the colloids are treated
as hard spheres and the polymers are modelled as point particles that are excluded from colloids by a centre-of-mass distance reflecting the colloid diameter and the
radius of gyration of the polymer coils. Another important model is a binary mixture of highly size-asymmetric hard spheres, such that $q=\sigma_b/\sigma\gtrsim10$
where $\sigma_b$ and $\sigma$ are the diameters of the large and the small species, respectively. Here, the smaller species can again represent non-adsorbing polymers
or a second species of the colloidal mixture. In both cases, the depletion potential arises between a pair of large particles due to excluding volume effects that
give rise to the osmotic pressure between the bulk region of the small species and the depleted region between two adjacent large spheres \cite{lek1}. For the
Asakura-Oosawa model, the depletion potential is attractive within a short distance $r<\sigma_b+\sigma$ between the centres of the colloidal particles and vanishes
beyond this separation. For a binary mixture of hard spheres, the additional excluded volume effects between the small particles may significantly change this
scenario especially at a high concentration of the small spheres and as a result, the depletion potential has a more complicated oscillatory structure, especially at
high densities. In any case, the depletion potential embodies an entropically driven interaction between large particles complementing the bare hard-sphere potential.
This extra effective attraction plays an important role for phase separation \cite{lek, biben, dijk} and flocculation \cite{jenkins, somas} in colloidal and
colloid-polymer mixtures.
Another type of solvent mediated interactions between two large bodies (nanoparticles) can occur in solvents exhibiting fluid coexistence. Depending on the
thermodynamic conditions, wetting layers of a given thickness may adsorb at the nanoparticle surfaces, giving rise to a solvent-mediated potential, quite distinct
from the depletion potential. First of all, because the wetting layers can be fairly thick at or near the solvent two phase coexistence and for large radii of the
solute particles, the range of this interaction can be substantially longer than that of the depletion potential. Furthermore, in view of the much higher number of
solvent particles contributing to the interaction, the resulting net force between the nanoparticles is expected to be much stronger than that of the depletion force.
This, indeed, has been corroborated in relatively recent theoretical studies \cite{archer, hopkins} for soft-core models. Finally, and perhaps most importantly, upon
drawing the nanoparticles coated by their wetting layers closer to each other, the bridging transition occurs as the wetting layers merge into a single film
encapsulating both nanoparticles. The transition is generally deemed to be first order (at least, on a mean-field level) and leads to a strongly attractive bridged
configuration. Here, it should be noted that by the term ``wetting layer'' we do not necessarily mean a liquid layer; it can be either a liquid or a gas phase
intruding between a nanoparticle and a bulk solvent if the solvent exhibits a liquid-gas coexistence or it can be a preferentially adsorbed phase if the solvent
exhibits a fluid-fluid separation.
In this work we investigate the effective interactions between two spherical nanoparticles dispersed in a one-component fluid (solvent) exhibiting liquid-vapour
coexistence using a microscopic fundamental measure density functional theory (FM-DFT). The solvent particles are modelled by a square-well (SW) fluid that is treated
grand canonically. The pair of nanoparticles is at a fixed mutual separation and acts as an external field, which leads to a 2D axially-symmetric problem. Attempts
to describe the model as a 1D problem led to a formulation of the so-called insertion method \cite{roth}, in which case the use of the exact potential distribution
theorem allows to treat the system as a binary mixture in an external field exerted by only a single nanoparticle. Although this approach simplifies the numerics of
the problem significantly and provides a good approximation for the effective interaction for the state points far from the coexistence, the insertion method was
found inadequate to account for the effects of bridging \cite{archer}. For the solvent models interacting via dispersion forces, the use of the sharp-kink
approximation was made to determine the solvent mediated interaction between the nanoparticles \cite{yeomans, yeomans2, bauer}.
Within Model 1, the nanoparticles are represented by two identical hard-sphere (HS) particles of sizes much larger than that of the solvent SW fluid. Model 2 mimics
the situation where the two nanoparticles are competing in the sense that different phases are preferably adsorbed at their surfaces. In this case, one of the
nanoparticles is again a hard sphere but the other one is the square-well particle, such that their hard cores are of identical diameters (and again much larger than
that of the solvent particles). We fix a subcritical temperature of the bulk solvent and immerse the nanoparticles into either of the coexisting fluid phases of the
solvent. Within Model 1 and the liquid solvent, the HS particles, when sufficiently far apart, experience complete drying (wetting by gas). The finite curvature of
the nanoparticles prevents the drying films from divergening even at the two-phase coexistence, as it is well known that the wetting layer thickness grows
logarithmically with (large) $R$ for systems with short-range potentials \cite{holyst}. The situation here is similar to the one studied in Ref. \cite{hopkins} for a
binary mixture of soft-core solvent particles exhibiting fluid-fluid separation and treated by a simple DFT where a mean-field approximation was used for the entire
part of the excess free energy functional (which corresponds to the random phase approximation for the direct correlation function). Upon bringing the large particles
closer together, the authors found a first-order transition associated with the connection of the wetting into a single bridge. This abrupt transition exhibits itself
as a cusp in the effective potential $W$ as a function of the distance between the large particles. When the bridge is formed, $W$ decreases steeply with the
distance, indicating a very strong attraction between the large particles in this regime. If, on the other hand, nanoparticles of the same type (Model 1) are inserted
into the gas phase of the bulk solvent then, of course, no wetting layers form at their surfaces. Therefore, one does not expect any substantial effective interaction
between the nanoparticles except, perhaps, for very small separations of the nanoparticles in which case depletion effects presumably play some role.
For Model 2, the situation is clearly quite different. Here, in either case, a wetting film forms at one of the nanoparticle surfaces. That is, in the case of the
liquid ambient phase, a single drying film forms around the HS particle whilst if the bulk phase is gas, the situation is just reversed with a wetting layer formed at
the SW nanoparticle. In both cases, the asymmetry in adsorption preferences must be reflected as the two nanoparticles are brought close together.
The paper proceeds as follows. In section II, we set the molecular model and formulate the FM-DFT in terms of weighted densities expressed as two-dimensional
integrals in cylindrical coordinates. In section III, we present our numerical results for the effective potentials of both Model 1 and Model 2. In particular, we
determine a critical value for the hard-sphere diameter $R_c$ below which the first-order bridging transition disappears. We also show that for Model 2 the effective
interaction between the nanoparticles is harshly repulsive for small separations and interpret our results by using simple arguments based on the interfacial
Hamiltonian approach. In section IV, we summarise and discuss our main results and conclude by outlining a perspective for future work.
\section{Density functional theory}
Within classical density functional theory the equilibrium density profile is found by minimising the grand potential functional
\bb
\Omega[\rho]={\cal{F}}[\rho]+\int\dr\rhor\left[V(\rr)-\mu\right]\,.\label{om}
\ee
where $V(\rr)$ is the external field and $\mu$ is the chemical potential. The intrinsic free energy functional ${\cal{F}}$ characterises a given model fluid
regardless of $V(\rr)$, and it is convenient to split this term into the contribution due to the ideal gas and the remaining excess part
\bb
{\cal{F}}[\rho]={\cal{F}}_{\rm id}[\rho]+{\cal{F}}_{\rm ex}[\rho]\,,
\ee
where ${\cal{F}}_{\rm id}[\rho]= \beta^{-1}\int\dr\rhor\left[\ln(\Lambda^3\rhor)-1\right]$ with $\beta=1/k_BT$ and $\Lambda$ being the thermal de Broglie wavelength
that can be set to unity.
In the spirit of the van der Waals theory, the excess term is treated in a perturbative manner, such that we separate the repulsive and attractive contributions:
\bb
{\cal{F}}_{\rm ex}[\rho]={\cal{F}}_{\rm hs}[\rho]+\frac{1}{2}\int\dr\rhor\int\dr'\rho(\rr')u(|\rr-\rr')\,.\label{fex}
\ee
Here, the attractive portion of the excess free energy functional is treated in a mean-field fashion, while the repulsive bit is approximated by the excess HS free
energy using the fundamental measure theory \cite{ros}
\bb
\beta{\cal{F}}_{\rm hs}[\rho]=\int\dr\Phi(\{n_\alpha\})\,.\label{fhs}
\ee
We consider two models characterised by their external potentials. Within Model 1, the external potential $V_1(\rr)$ is exerted by two large hard spheres, each of a
diameter $\sigma_b$, placed a distance $\sigma_b+h$ apart. Owing to the axial symmetry of the model, the potential is expressed in the cylindrical coordinates $(z,
r)$:
\bb
V_1(z,r)=V_{\rm hs}(z-\sigma_b/2-h/2,r)+V_{\rm hs}(z+\sigma_b/2+h/2,r)\,,\label{v1}
\ee
where
\bb
V_{\rm hs}(z,r)=\left\{\begin{array}{cc}
\infty& z^2+r^2\leq R^2\,,\\
0\,;&z^2+r^2>R^2\,,
\end{array}\right.
\ee
and $R=\sigma_b/2$.
In Model 2, one of the HS particles has an additional SW attraction, such that the entire external potential is:
\bb
V_2(z,r)=V_{\rm hs}(z-\sigma_b/2-h/2,r)+V_{\rm sw}(z+\sigma_b/2+h/2,r)\,,\label{v2}
\ee
where
\bb
V_{\rm sw}(z,r)=\left\{\begin{array}{cc}
\infty& z^2+r^2\leq R^2\,,\\
-\varepsilon\,;&R^2<z^2+r^2<(R+\sigma)^2\,.\\
0\,;&z^2+r^2\geq(R+\sigma)^2\,.
\end{array}\right.
\ee
is expressed in terms of the hard core diameter $\sigma$ and the depth of the well $\varepsilon$ of the solvent SW fluid. The attractive portion of the fluid-fluid
interaction that enters Eq.~(\ref{fex}) is:
\bb
u(\tilde{r})=\left\{\begin{array}{cc}
-\varepsilon& \tilde{r}<\lambda\sigma\,,\\
0\,;&\tilde{r}\geq\lambda\sigma\,,
\end{array}\right.
\ee
where $\tilde{r}=\sqrt{(z_2-z_1)^2+(r_2-r_1)^2}$ is the distance between the centres of two SW particles and $\lambda=1.5$, as usual.
The repulsive part of the intrinsic free energy given by Eq.~(\ref{fhs}) corresponds to a one-component HS fluid of diameter $\sigma$. The free energy density
$\Phi$ is a function of a set of weighted densities $\{n_\alpha\}$, five of which are given by surface integrals that must be expressed as two-dimensional integrals.
The remaining weighted density $n_3$ is given by integrating the density profile $\rho(z, r)$ over the volume of a hard core of diameter $\sigma$. To minimise the
dimensionality of this volume integral by exploiting the axial symmetry of the system, we can express $n_3$ as a double integral as follows:
\bb n_3(z,r)=2\int_{-R}^R\dd z' \int_{r-\tilde{R}(z')}^{r+\tilde{R}(z')}\dd r' r'
\arccos\left[\frac{r'^2+r^2-\tilde{R}^2(z')}{2r'r}\right]\rho(z'+z,r');\;\;\;r>R \label{n3a}
\ee
\bb
n_3(z,r)=2\int_{-R}^{R}\dd z' \Psi(z';z,r)+2\pi\int_{-\sqrt{R^2-r^2}}^{\sqrt{R^2-r^2}}\dd z'\int_0^{\tilde{R}(z')-r}\dd r'r'\rho(z'+z,r')\nonumber;\;\;\;r<R\,, \label{n3b}
\ee
with
\bb
\Psi(z';r,z)\equiv\int_{|r-\tilde{R}(z')|}^{r+\tilde{R}(z')}\dd r' r' \arccos\left[\frac{r'^2+r^2-\tilde{R}^2(z')}{2r'r}\right]\rho(z'+z,r')\,.
\ee
and where $\tilde{R}(z)=\sqrt{R^2-z^2}$.
The minimisation of Eq.~(\ref{om}) leads to the Euler-Lagrange equation
\bb
\frac{1}{\beta}\ln\Lambda^3\rho(z,r)+\frac{\delta F_{\rm ex}}{\delta\rho(z,r)}+V(z,r)=\mu \label{el}
\ee
which is solved iteratively on a two-dimensional grid of a cylindrical system of volume $V=2L_z\pi L_r^2$, where $L_r=20\sigma$ and $L_z=40\sigma$, with a uniform
spacing of $0.02\sigma$ in both $z$ and $r$ directions. Although the grid does not follow the symmetry of the external field, the fine spacing avoids any noticeable
artifacts in the structure of the fluid near the nanoparticles. By determining the equilibrium density profile, the excess (over ideal) grand potential $\Omega_{\rm
ex}$ can be calculated. The effective potential between the pair of immersed particles a distance $h$ apart is given by
\bb
W(h)=\Omega_{\rm ex}(h)-2\Omega_{\rm ex}^{(\rm hs)}
\ee
for Model 1 and by
\bb
W(h)=\Omega_{\rm ex}(h)-\Omega_{\rm ex}^{(\rm hs)}-\Omega_{\rm ex}^{(\rm sw)}
\ee
for Model 2. Here, we defined $\Omega_{\rm ex}^{\rm hs}$ and $\Omega_{\rm ex}^{\rm sw}$ as the excess grand potentials corresponding to a system with only one
HS or one large SW particle, respectively.
\section{Results}
Throughout this work, we consider a subcritical temperature $T/T_c=0.89$ where $k_BT_c\approx0.9\varepsilon$ is the critical temperature of the bulk SW fluid, as
determined from Eq.~({\ref{el}) in the case of zero external potential. We begin with Model 1, such that two hard spheres of diameter $\sigma_b=10\,\sigma$ are
immersed in a saturated gas of the square-well fluid. In this case, no wetting/drying films develop around the HS bodies, so that the effective potential displayed in
Fig.~1 as a function of the separation between the hard-sphere surfaces is very short-range and attractive due to the depletion effect. In view of the low-density
state of the ambient fluid, the depletion potential is weak and monotonic.
\begin{figure}
\includegraphics{d10g}
\caption{Effective potential of a pair of hard spheres of diameters $\sigma_b=10\,\sigma$ as a function of the distance between their surfaces. The pair is immersed
in a saturated vapour of the SW fluid at a temperature $T/T_c=0.89$. } \label{fig1}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{d10l}\includegraphics[width=0.5\textwidth]{force}
\caption{ Left panel: Effective potential of a pair of hard spheres of diameters $\sigma_b=10\,\sigma$ as a function of the distance between their surfaces. The pair is immersed
in a saturated liquid of the SW fluid at a temperature $T/T_c=0.89$. The inset shows an oscillatory behaviour of the potential at larger distances. Right panel:
Effective force $-{dW}/{dh}$ between the nanoparticles. As the bridge between the nanoparticles is formed the force changes abruptly but still continuously.}
\label{fig2}
\end{figure}
\begin{figure}
\includegraphics[width=0.4\textwidth]{p_D10_6}\hspace*{0.2cm}\includegraphics[width=0.4\textwidth]{p_D10_64}
\includegraphics[width=0.4\textwidth]{p_D10_68}\hspace*{0.2cm}\includegraphics[width=0.4\textwidth]{p_D10_7}
\caption{Two-dimensional density profiles of a system consisting of two hard spheres of diameters $\sigma_b=10\,\sigma$ immersed in a saturated liquid of the SW fluid
at a temperature $T/T_c=0.89$. The separations between the hard spheres are (from top left to right bottom): $h=6\,\sigma$, $h=6.4\,\sigma$, $h=6.8\,\sigma$ and
$h=7\,\sigma$. A low density gas film develops around the spheres and the films get connected at lower separations to form a bridge. As the separation between the
hard spheres increases, the bridging bond disappears continuously.} \label{fig3}
\end{figure}
Considerably stronger and longer-ranged interactions are induced when the hard spheres are inserted into a disfavoured liquid phase. Here, we have considered four
different hard-sphere diameters $\sigma_b/\sigma=5,10,15$ and $20$. Qualitatively similar results have been obtained for the diameters $\sigma_b/\sigma\leq15$, as
illustrated in Fig.~2 for $\sigma_b=10\,\sigma$. Here, beyond a certain distance $h_b\approx6\sigma$, the effective potential is only very slightly negative but below
$h_b$ the potential rapidly decreases with the distance. Concomitant with the development of the strong interaction between the hard spheres is a formation of a
bridging bond between the particles at a distance of approximately $h_b$. Although the curvature of the function $W(h)$ changes abruptly near $h_b$, the function
itself is smooth and does not exhibit a hysteresis. This has been checked by performing two sets of calculations with bridged and unbridged initial configurations
that both resulted in identical minima of the grand potential functional. In Fig.~2 we also display the effective force $-dW/dh$ as a function of $h$ that changes
abruptly yet continuously near $h_b$. The reversible process of bridging/unbridging is illustrated in Fig.~3, where the development/termination of the bridge bond is
shown for the separations between the hard spheres near $h_b$.
\begin{figure}
\includegraphics{d20l}
\caption{Effective potential of a pair of hard spheres of diameters $\sigma_b=20\,\sigma$ as a function of the distance between their surfaces. The pair is immersed
in a saturated liquid of the SW fluid at a temperature $T/T_c=0.89$. The line with circles denotes the unbridged states, and the line with crosses denotes the bridged
states. The first-order bridging transition occurs at a distance $h\approx8\,\sigma$.} \label{fig4}
\end{figure}
\begin{figure}
\includegraphics[width=0.4\textwidth]{p_D20_b}
\includegraphics[width=0.4\textwidth]{p_D20_unb}
\caption{Two-dimensional density profiles of a system consisting of two hard spheres of diameter $\sigma_b=20\,\sigma$ immersed in a saturated liquid of the SW fluid
at a temperature $T/T_c=0.89$. In both cases, the separation between the hard spheres is $h_b=8\,\sigma$, which corresponds to the equilibrium between the bridged
(left panel) and unbridged (right panel) configurations. } \label{fig5}
\end{figure}
The situation, however, becomes different for larger solute/solvent size ratios. In Fig.~4, we display the separation dependence of the effective potential between
two hard spheres of diameters $\sigma_b=20\,\sigma$. Starting from two different (bridged and unbridged) configurations, the minimisation of the grand potential
functional converges to two different density profiles in the vicinity of the separation $h\approx8\,\sigma$. Associated with this hysteresis are the metastable
extensions of both configurations terminating at their spinodals, such that the equilibrium state corresponds to the global minimum of the grand potential. The
intersection of these bridged and unbridged branches produces a cusp at $h_b$ corresponding to a location of a first-order bridging transition. The coexisting bridged
and unbridged configurations are shown in Fig.~5. It should be noted that the bridge does {\emph not} form as the two wetting layers meet. Instead, in pure analogy
to a first-order capillary condensation with the walls covered by wetting films, the transition occurs at a distance $h>2\ell$ where $\ell$ is the thickness of the
wetting layer. This contrasts with the continuous process of bridging displayed in Fig.~3.
\begin{figure}
\includegraphics[width=0.22\textwidth]{cw_gas_10}\hspace*{0.1cm}\includegraphics[width=0.22\textwidth]{cw_gas_5}\hspace*{0.1cm}\includegraphics[width=0.22\textwidth]{cw_gas_2}\hspace*{0.1cm}\includegraphics[width=0.22\textwidth]{cw_gas_0}
\includegraphics[width=0.22\textwidth]{cw_liq_10}\hspace*{0.1cm}\includegraphics[width=0.22\textwidth]{cw_liq_5}\hspace*{0.1cm}\includegraphics[width=0.22\textwidth]{cw_liq_2}\hspace*{0.1cm}\includegraphics[width=0.22\textwidth]{cw_liq_0}
\caption{Two-dimensional density profiles of a system consisting of a HS (left) and a SW (right) particle of (hard core) diameter $\sigma_b=20\,\sigma$ immersed in a
saturated gas (top row) and a saturated liquid (bottom row) of the SW fluid at a temperature $T/T_c=0.89$. The distances between the hard-core surfaces of the
particles are (from left to right): $h/\sigma=10, 5, 2$ and $0$.} \label{fig6}
\end{figure}
\begin{figure}
\includegraphics{cw}
\caption{Effective potential between a HS and a SW particle of (hard core) diameter $\sigma_b=20\,\sigma$ as a function of the distance between the the hard-core
surfaces. The pair is immersed in a saturated gas (line with crosses) and saturated liquid (line with open circles) of the SW fluid at a temperature $T/T_c=0.89$.}
\label{fig7}
\end{figure}
We next consider the case of two ``competing'' particles that favour opposite phases. We describe this situation within Model 2, where a pair of nanoparticles is
represented by a HS and a SW particle. In Fig.~6 (upper row), we display several representative density profiles of different particles separations. The upper row
of Fig.~6 corresponds to a saturated vapour of the bulk SW fluid. In this case, the SW nanoparticle, if sufficiently far from the hard-sphere nanoparticle, is
surrounded by a wetting layer of thickness $\ell$. As the two particles are brought closer to each other, the wetting film is interrupted by the excluded volume of
the hard sphere. At a sufficiently small separation between the two nanoparticles, the liquid-vapour interface of the wetting film becomes the subject of two
competing forces: while the hard-sphere particle, which prefers to remain dry, pushes the interface toward the SW particle, the SW particle strives to restore the
location of the interface to a distance $\ell$ from its surface. This competition is less dramatic than in the case of a confined fluid between two parallel plates,
which would lead to the delocalisation of the liquid-vapour interface \cite{comp_walls, stewart}. As for our model with short-range interactions the thickness of the
wetting layer increases with the radius of the adsorbate rather slowly, $\ell\sim \ln R$ as $R\to \infty$, the equilibrium wetting thickness is of the order of
several solvent diameter units and so is the range of the effective force between the particles due to this mechanism. If the previous arguments are reversed and the
liquid-vapour interface is thought fixed, it becomes clear that both nanoparticles are repelled from the interface and hence from each other. As a result, the
effective potential is strongly repulsive over the range of $\ell$, as illustrated in Fig.~7.
If inserted into the saturated liquid phase of the solvent, the drying film adsorbed at the HS becomes disturbed by the presence of the SW particle that endeavours to
remain wet, as shown in Fig.~6 (lower row). This, again, induces a repulsive force between the nanoparticles over the range of the drying film thickness. The
situation is thus largely analogous to the previous case where the particles were dissolved in the gas but there are some subtle differences. Firstly, the magnitude
and the range of the effective potential for the compound dissolved in the liquid is smaller. This can be explained as follows:
Consider a system of one HS particle of radius $R$ dissolved in the saturated liquid of the SW fluid. Using a standard coarse-grained treatment, the interfacial
potential can be expressed as an excess grand potential per area of the hard sphere (see, e.g., Ref.\cite{hend}):
\bb
\Omega_{\rm ex}^{(\rm hs)}(\ell)/(4\pi R^2)=\gamma_{wg}(R)+\gamma_{lg}(R+\ell)\left(1+\frac{2\ell}{R}\right)+\omega(\ell)\,, \label{om_hs}
\ee
where $\gamma_{wg}(R)$ and $\gamma_{lg}(R+\ell)$ are the surface tensions between the hard sphere and the gas and between the gas and liquid, respectively, for a
given radius of curvature of the interface. The last term in Eq.~(\ref{om_hs}) represents interaction between the liquid-gas interface and the particle-gas interface;
for our model where all the molecular interactions are short-ranged, $\omega(\ell)=a(T)\exp\left(-\frac{\ell}{\xi_g}\right)$, where $a(T)>0$ and $\xi_g$ is the bulk
correlation length of the gas. For a large SW particle dissolved in the saturated vapour of the SW fluid, the interfacial potential reads:
\bb
\Omega_{\rm ex}^{(\rm sw)}(\ell)/(4\pi R^2)=\gamma_{wl}(R)+\gamma_{lg}(R+\ell)\left(1+\frac{2\ell}{R}\right)+\tilde{\omega}(\ell)\,, \label{om_sw}
\ee
where $\tilde{\omega}(\ell)=\tilde{a}(T)\exp\left(-\frac{\ell}{\xi_l}\right)$ with $\tilde{a}>0$ and where $\xi_l$ is the bulk correlation length of the liquid.
Going back to Model 2, we can now compare the influence of the second particle on the interfacial potentials given above. To this end, we estimate how the
disruption of the fluid structure due to the presence of the second particle affects the interfacial potential pertinent to a single particle in both cases. Assuming
that the thicknesses of the equilibrium wetting and drying layers are similar (as can be inferred from Fig.~6), the two interfacial potentials given by
Eq.~(\ref{om_hs}) and Eq.~(\ref{om_sw}) differ in only the first and the third terms. At the distance of $h\approx\ell$, the repulsion between the pair of particles
will be dominated by the final terms, the asymptotic decay of which are controlled by the correlation length of the adsorbed phase. Assuming $\xi_g\approx\sigma$, the
repulsion in the liquid becomes vanishingly small for $h\gtrsim 5\,\sigma$, in accordance with Fig.~7. The particles are correlated to a larger distance in the liquid
than in the gas, i.e., $\xi_l>\xi_g$, and thus $\tilde{\omega}(\ell)$ can be expected to be slightly longer-ranged than $\omega(\ell)$. This, in turn, imply that
$W(h)$ for Model 2 is longer-ranged when the nanoparticles are dissolved in the gas.
For small separations $h<\sigma$, the dominant contribution to the effective force is due to the surface interactions between the adsorbed molecules and the
adsorbent. The work required to remove the surface particles from the region between the nanoparticles is clearly considerably higher for the high-density molecules
attracted to the SW particle than for the gas-like fluid interacting with its adsorbent only repulsively. This explains why $W(h)$ is larger for small $h$ when the
bulk fluid is gas.
The second difference between the two cases can be inferred from Fig.~6 by comparison of the fluid structure in the vicinity of the non-adsorbing particle. In the
liquid state, the additional attraction of the SW particle gives rise to a rich, oscillating structure which in turn leads to mild oscillations of $W(h)$ for large
$h$ in contrast to a monotonic decay of $W(h)$ if the bulk is vapour. The asymptotic behaviour of $W(h)$ indicates that the effective interactions can also be weakly
attractive at sufficiently large distances. The damped oscillatory structure of $W(h)$ suggests that the liquid state lies above the Fisher-Widom line \cite{fw},
while the gas state, for which the structure exhibits monotonic decay, lies below.
\section{Concluding remarks}
In this paper, we have employed non-local density functional theory to study the effective interactions between a pair of nanoparticles in a solvent exhibiting
liquid-vapour coexistence. The pair of nanoparticles plays a role of an external, axially-symmetric field, and we determined equilibrium two-dimensional density
profiles of the solvent fluid modelled grand-canonically by a SW potential. Within Model 1, the nanoparticles are identical large hard spheres. If immersed in the
vapour phase, the effective interaction corresponds to a short-range depletion potential induced by the excluded volume effects of the hard spheres. However, when
inserted into the reservoir of the saturated liquid, drying films form around the particles and mediate the effective interaction between the particles over an
arbitrarily long distance determined by the HS curvature. A strong attraction arises when the two films connect to form a bridge between the nanoparticles. The main
results of this bridging process as obtained by our microscopic DFT can be summarised as follows:
\begin{itemize}
\item The magnitude of the bridging potential found in this work was on the order of $10^2\,k_bT$. This is one order of magnitude less than the value found
in Ref. \cite{archer} and two orders less than that found in Ref. \cite{hopkins}. This reveals that the strength of the bridging interactions is strongly
model-dependent. We attribute these differences to the disparate number of solvent particles that contribute to the bridging bond. Ref. \cite{archer} and Ref.
\cite{hopkins} address effective solvent interactions for polymers and charged particles, respectively. This means that the considered potentials are soft and the
particles may overlap. This is of course not the case for simple fluid models such as the square-well potential considered in this work. The hard core interaction
between the particles obviously reduces the number of solvent particles forming the bridge compared to soft core models. Moreover, the bridging formation considered
here was due to a connection of low density films, which strongly contrasts with the high density bridges of Refs. \cite{archer} and \cite{hopkins}. Note that the
simple mean field analysis used in Refs. \cite{archer} and \cite{hopkins} is justified for only these high density states.
\item Supported by theoretical studies for models with short-range \cite{archer, hopkins}
and long-range \cite{yeomans, yeomans2, bauer} interactions, the bridging is generally considered to be a first-order transition. From a macroscopic viewpoint, this
picture is very reasonable because the symmetries of the bridged and unbridged states are different. However, our microscopic study reveals that the bridged state may
also develop continuously. In this case, the bridging bond forms as the wetting/drying films meet. We have found that it is so for only relatively high radii of the
large spheres ($\sigma_b\approx20\,\sigma$) that the bridging transition becomes first-order; here, the bridging occurs before the films meet. A discussion on the
order of the transition may appear somewhat academic in view of the fact that an inclusion of fluctuations would necessarily round the transition anyway \cite{note}.
However, the study of bridging between a pair of particles is only a first step to understanding phenomena such as the flocculation in colloidal suspensions, where a
bridging connection between a pair of particles plays a role of only a single constituent in a percolating network. The entire collective (flocculation) transition
can then be genuinely first-order, provided that the bridging bond is sufficiently strong.
\item Based on the fact that the bridging transition is first order (on a mean-field level) for only large radii of the particles, one may identify the bridging
transition as just a reminiscence of capillary condensation. In fact, for macroscopically large particles one can, in the spirit of Derjaguin's approximation, think
of the interaction between the nanoparticles in terms of the interaction between a pair of two plates. This picture becomes less satisfactory as the curvature of the
particles increases. Ultimately, the capillary condensation disappears completely below some critical value $R=R_c$, which we have identified to lay in the interval
of $(15\,\sigma,20\,\sigma)$. For radii smaller than the critical radius (corresponding to the given temperature), the bridge can still be formed \cite{note2} as the
two wetting layers merge, but the process is no longer accompanied by any phase transition. The role of the critical radius $R_c$ in bridging is thus analogous to the
role of the capillary critical distance $L_c(T)$ for capillary condensation between plates.
\end{itemize}
We have further investigated the effective interaction of two different (competing) nanoparticles using Model 2 represented by a pair of a HS and a SW particle.
Dissolved in a saturated vapour or liquid, only one wetting layer forms around a respective particle. As the particles are brought together to the distance of the
thickness of the wetting layer, the interaction between the liquid-vapour interface of the layer and the other particle leads to a strongly repulsive effective force
between the nanoparticles. Owing to the dissimilarity of the liquid and gas bulk correlation lengths, the repulsive potential is slightly longer-ranged when the
reservoir is a vapour. In this case, the magnitude of the potential is also distinctly higher than for the liquid bulk phase near a zero separation of the
nanoparticles in view of the strong interaction between the solvent molecules and the SW particle. In the gas phase, the effective potential decays monotonically,
while in the liquid phase the potential exhibits damped oscillatory structure at large distances.
In summary, we have considered two possible mechanisms that induce an effective interaction between a pair of nanoparticles mediated by a fluid solvent, a depletion
potential due to the excluded volume of the region near the contact of the nanoparticles and the effective potential induced by the presence of wetting films. The
latter is both longer-ranged and stronger than the depletion potential but of course less universal, as it requires that the solvent be near its phase separation. The
potential is strongly attractive when the nanoparticles are identical and are covered by wetting layers at separations where the wetting layers merge and form a
bridge between the nanoparticles. The bridging gives rise to a first-order phase transition if the radii of the particles are larger than a certain critical value
$R_c$. If the adsorption preferences of the nanoparticles are different and the wetting layer forms around only one nanoparticle, the effective interaction is
strongly repulsive. The third possibility that could induce the effective interaction that we have not discussed here is the critical Casimir effect. The required
thermodynamic conditions are even more restrictive in this case (the critical point of the solvent) but a qualitative picture of this interaction would be most likely
similar to the previous case, such that the interaction is attractive for identical and repulsive for opposite adsorption preferences of the nanoparticles
\cite{casimir, okamoto}.
Although we have tried to provide a rather complete microscopic description of the effective interactions near a two phase coexistence by examining four different
model situations, our results instigate further interesting questions. Most pertinently, one can ask what is behind the rather large value of the critical radius
$R_c$ and how to predict the separation at which the bridging transition occurs. To answer the latter, one needs to find a generalisation of Kelvin's equation for
spherical bodies. One can also inquire how the results would change for other molecular models, particularly for those including long-range interactions. The
analysis of the bridging transitions can also be extended by considering two identical particles exerting attractive potentials, which would allow to address the role
of the temperature in the light of wetting transition. Although, e.g., the occurrence of the mean-field pre-wetting transition that does not occur in our Model 1
would make the problem somewhat more complex, in the main, we do not expect qualitatively different behaviour of such a model, since, as we have demonstrated, the
bridging transition is a capillary condensation pertinent to spherical walls. Finally, it would also be interesting to examine to what extent the shape of the bridge
given by the microscopic DFT matches with the macroscopic requirements of the minimal surface (catenoid) of the bridge. These and other related problems are subjects
of the current study and will be published elsewhere.
\begin{acknowledgments}
\hspace*{0.01cm}
\noindent Dedicated to the so-called Czech school of the statistical mechanics of liquids. In particular, I express my gratitude to my father, who was the first to
introduce me to the fascinating field of the liquid matter theory. I also acknowledge the financial support from the Czech Science Foundation, Grant No. 13-02938S.
\end{acknowledgments}
|
\section{Introduction}
Throughout this paper, $R$ denotes an associative ring with unity. Let $Id(R)$ be the set of all idempotent elements of $R$. A ring $R$ is called abelian if all idempotent of it are central (i.e., $re=er$ for all $r\in R$ and $e\in Id(R)$). Following Hashemi and Moussavi \cite{hashemi/quasi}, a ring $R$ is $\sigma$-{\it compatible} if for each $a,b\in R$, $ab=0$ if and only if $a\sigma(b)=0$. Agayev et al. \cite{agayev/2009}, introduced $\sigma$-abelian rings as a generalization of abelian rings. A ring $R$ is called $\sigma$-abelian if it is abelian and $\sigma$-compatible. A left ideal $I$ is said to be {\it reflexive} \cite{mason}, if $aRb\subseteq I$ implies $bRa\subseteq I$ for $a,b\in R$. A ring $R$ is called {\it reflexive} if $0$ is a reflexive ideal. The reflexive property for ideals was first studied by Mason \cite{mason}, this concept was generalized by Kim and Baik \cite{kim/2005,kim/2006}, and they introduced the concept of idempotent reflexive right ideals and rings. A left ideal $I$ is called {\it idempotent reflexive} \cite{kim/2005}, if $aRe\subseteq I$ implies $eRa\subseteq I$ for $a,e=e^2\in R$. A ring $R$ is called {\it idempotent reflexive} if $0$ is an idempotent reflexive ideal. Also, there is an example of an idempotent reflexive ring which is not reflexive \cite[Example 5]{kim/2005}. Kwak and Lee \cite{kwak/2012}, introduced the concept of left and right idempotent reflexive rings. A ring $R$ is called {\it left idempotent reflexive} if $eRa=0$ implies $aRe=0$ for $a,e=e^2\in R$. Right idempotent reflexive rings are defined similarly. If a ring $R$ is left and right idempotent reflexive then it is called an idempotent reflexive ring. It can be easily checked that every reflexive ring is an idempotent reflexive ring. Abelian rings (i.e., its idempotents are central) are idempotent reflexive and hence semicommutative rings are idempotent reflexive since semicommutative rings are abelian. For a subset $X$ of $R$, $r_R(X)=\{a\in R|Xa=0\}$ and $\ell_R(X)=\{a\in R|aX=0\}$ will stand for the right and the left annihilator of $X$ in $R$ respectively.
\par For an endomorphism $\sigma$ of a ring $R$, a {\it skew polynomial} ring (also called an {\it Ore extension of endomorphism type}) $R[x;\sigma]$ of $R$ is the ring obtained by giving the polynomial ring over $R$ with the new multiplication $xr=\sigma(r)x$ for all $r\in R$. Also, a {\it skew power series ring} $R[[x;\sigma]]$ is the ring consisting of all power series of the form $\sum_{i=0}^{\infty}a_ix^i\;(a_i\in R)$, which are multiplied using the distributive law and the Ore commutation rule $xa=\sigma(a)x$, for all $a\in R$.
\par In this paper, we study skew polynomial rings and skew power series rings over idempotent reflexive rings and abelian rings. Also, we used the idea of Agayev et al. \cite{agayev/2009} to introduce the concept of right (resp., left) $\sigma$-idempotent reflexive rings which is a generalization of right (resp., left) idempotent reflexive rings and $\sigma$-abelian rings. Agayev et al. \cite{agayev/2009}, studied abelian and $\sigma$-abelian properties on skew polynomial rings. Also, Kwak and Lee \cite{kwak/2012}, investigated reflexive and idempotent reflexive properties on ordinary polynomial extensions. In this note, we continue studding abelian and idempotent reflexive properties on skew polynomial and skew power series extensions. For a ring $R$ and $\sigma$ an endomorphism of $R$.
\par $1.$ We introduce the concept of right (resp., left) $\sigma$-idempotent reflexive rings which generalizes right (resp., left) idempotent reflexive rings and $\sigma$-abelian rings.
\par $2.$ We discuss when (left and right) idempotent reflexive property passes form $R$ to $R[x;\sigma]$ (or $R[[x;\sigma]]$) and conversely, under the conditions ``$\sigma$-compatible" and ``$\sigma(Re)\subseteq Re$ for all $e\in Id(R)$".
\par$3.$ We show that, if $R$ satisfy the condition $(\mathcal{C_{\sigma}})$ then, abelian property passes from $R$ to $R[x;\sigma]$ and $R[[x;\sigma]]$. Some results are obtained as corollaries of our results.
\section{Idempotent reflexive property}
In \cite[Theorem 3.1]{kwak/2013}, a ring $R$ is idempotent reflexive if and only if $R[x]$ is idempotent reflexive if and only if $R[[x]]$ is idempotent reflexive. One might expect if idempotent reflexive property can be preserved under skew polynomial and skew power series extensions.
\begin{proposition}\label{prop1}Let $R$ be a ring and $\sigma$ an endomorphism of $R$ such that $\sigma(Re)\subseteq Re$ for all $e\in Id(R)$.
\par$(1)$ If $R$ is left idempotent reflexive then $R[x,\sigma]$ and $R[[x,\sigma]]$ are left idempotent reflexive.
\par$(2)$ If $R[x,\sigma]$ or $R[[x,\sigma]]$ is right idempotent reflexive then $R$ is right idempotent reflexive.
\end{proposition}
\begin{proof}$(1)$ Let $e(x)=\sum_{j=0}^me_jx^j\in Id(R[x,\sigma])$ and $f(x)=\sum_{i=0}^nf_ix^i\in R[x,\sigma]$. Assume that $e(x)R[x,\sigma]f(x)=0$, then $e(x)Rf(x)=0$. Thus for any $b\in R$, we have the following system of equations.
$$e_0bf_0=0\eqno(0)$$
$$e_0bf_1+e_1\sigma(bf_0)=0\qquad\qquad\;\;\;\eqno(1)$$
$$e_0bf_2+e_1\sigma(bf_1)+e_2\sigma^2(bf_0)=0\qquad\qquad\qquad\qquad\qquad\eqno(2)$$
$$\cdots$$
From equation $(0)$, $f_0\in r_R(e_0R)$. Let $s\in R$ and take $b=e_0s$ in equation $(1)$. Then $e_0sf_1+e_1\sigma(e_0sf_0)=0$, but $f_0\in r_R(e_0R)$. So $e_0sf_1=0$, hence $f_1\in r_R(e_0R)$. Now, in equation $(2)$, take $b=e_0s$ for each $s\in R$. Then $e_0sf_2+e_1\sigma(e_0sf_1)+e_2\sigma^2(e_0sf_0)=0$, since $f_0,f_1\in r_R(e_0R)$ so $e_0sf_2=0$. Hence $f_2\in r_R(e_0R)$. Continuing this procedure yields $f_i\in r_R(e_0R)$ for all $i=0,1,2,\cdots,n$. Hence $e_0Rf_i=0$ for all $i=0,1,2,\cdots,n$. Thus $f_iRe_0=0$ for all $i=0,1,\cdots,n$ because $R$ is left idempotent reflexive. On the other hand, $$f(x)Re_0=f_0Re_0+f_1xRe_0+\cdots+f_nx^nRe_0$$
$$\qquad\qquad\qquad\quad\;=f_0Re_0+f_1\sigma(Re_0)x+\cdots+f_n\sigma^n(Re_0)x^n$$
Since $\sigma(Re_0)\subseteq Re_0$, it follows that $f(x)Re_0\subseteq \sum_{i=0}^nf_iRe_0=0$, then $f(x)Re_0=0$. With the same method as above, we obtain $$f(x)Rxe_0=f(x)Rx^2e_0=\cdots=f(x)Rx^ne_0=0.$$ So $f(x)R[x,\sigma]e_0=0$. However, since $e(x)^2=e(x)$, we have $e_0^2=e_0,\; e_0e_1+e_1\sigma(e_0)=e_1,\: \cdots$. Using the hypothesis $``\sigma(Re_0)\subseteq Re_0"$, we have $e_j\in Re_0R$ for any $1\leq j\leq m$. Thus from the fact $f(x)R[x;\sigma]e_0=0$, we have $f(x)R[x;\sigma]e_j=0$, and therefore $f(x)R[x;\sigma]e(x)=0$. We use similar argument for the case of skew power series rings.
\par$(2)$ Suppose that $R[x;\sigma]$ (resp., $R[[x;\sigma]]$) is right idempotent reflexive. If $aRe=0$ for some $e^2=e,a\in R$ then $aR[x;\sigma]e=0$, by the hypothesis ``$\sigma(Re)\subseteq Re$". Since $R[x;\sigma]$ (resp., $R[[x;\sigma]]$) is right idempotent reflexive, we get $eR[x;\sigma]a=0$, then $eRa=0$. Therefore $R$ is right idempotent reflexive.
\end{proof}
We can easily see that, if a ring $R$ is $\sigma$-compatible then $\sigma(e)=e$ for all $e\in Id(R)$, and so the condition ``$\sigma(Re)\subseteq Re$ for all $e\in Id(R)$", is satisfied.
\begin{proposition}\label{prop2}Let $R$ be a ring and $\sigma$ an endomorphism of $R$ such that $R$ is $\sigma$-compatible.
\par$(1)$ If $R$ is right idempotent reflexive then $R[x,\sigma]$ and $R[[x,\sigma]]$ are right idempotent reflexive.
\par$(2)$ If $R[x,\sigma]$ or $R[[x,\sigma]]$ is left idempotent reflexive then $R$ is left idempotent reflexive.
\end{proposition}
\begin{proof}$(1)$ Assume that $R$ is right idempotent reflexive and let $e(x)=\sum_{i=0}^ne_ix^i\in Id(R[x,\sigma])$ and $f(x)=\sum_{j=0}^mf_jx^j\in R[x,\sigma]$ such that $f(x)R[x,\sigma]e(x)=0$. Thus we have the following system of equations where $b$ is an arbitrary element of $R$.
$$f_0be_0=0\eqno(0)$$
$$f_0be_1+f_1\sigma(be_0)=0\qquad\qquad\;\;\;\eqno(1)$$
$$f_0be_2+f_1\sigma(be_1)+f_2\sigma^2(be_0)=0\qquad\qquad\qquad\qquad\qquad\eqno(2)$$
$$\cdots$$
Equation $(0)$ yields $f_0\in \ell_R(Re_0)$. In equation $(1)$ substitute $se_0$ for $b$ to obtain $f_1\sigma(se_0)=0$, but $R$ is $\sigma$-compatible then $f_1se_0=0$ and so $f_1\in \ell_R(Re_0)$. Continuing this procedure yields $f_i\in \ell_R(Re_0)$ for all $i=0,1,\cdots,n$.
Since $R$ is right idempotent reflexive, it follows that $e_0Rf_i=0$ for all $i=0,1,\cdots,n$. We claim that $e_0R[x,\sigma]f(x)=0$. Observe that $e_0Rf(x)=\sum_{j=0}^me_0Rf_jx^j=0$ and $e_0Rxf(x)=\sum_{j=0}^me_0R\sigma(f_j)x^{j+1}=0$
because $R$ is $\sigma$-compatible. Also, we have $e_0Rx^kf(x)=0$ for all nonnegative integers $k$. Consequently $e_0\varphi(x)f(x)=0$ for all $\varphi(x)\in R[x,\sigma]$. Thus $e_0R[x,\sigma]f(x)=0$. As, in the proof of Proposition \ref{prop1}, we have $e_i\in Re_0R$ for any $1\leq i\leq n$. It follows that $e(x)R[x,\sigma]f(x)=0$. Therefore, $R[x,\sigma]$ is right idempotent reflexive.
\par$(2)$ As in the proof of Proposition \ref{prop1}(2), it suffices to replace the hypothesis ``$\sigma(Re)\subseteq Re$" by ``$\sigma$-compatible".
\end{proof}
\begin{Theorem}\label{theorem}Let $R$ be a ring and $\sigma$ an endomorphism of $R$ such that $R$ is $\sigma$-compatible. Then
\par$(1)$ $R$ is right idempotent reflexive if and only if $R[x,\sigma]$ is right idempotent reflexive if and only if $R[[x,\sigma]]$ is right idempotent reflexive.
\par$(2)$ $R$ is left idempotent reflexive if and only if $R[x,\sigma]$ is left idempotent reflexive if and only if $R[[x,\sigma]]$ is left idempotent reflexive.
\par$(3)$ $R$ is idempotent reflexive if and only if $R[x,\sigma]$ is idempotent reflexive if and only if $R[[x,\sigma]]$ is idempotent reflexive.
\end{Theorem}
\begin{proof}Obvious from Propositions \ref{prop1} and \ref{prop2}.
\end{proof}
Following Agayev et al. \cite{agayev/2009}. A ring $R$ is called $\sigma$-abelian if it is abelian and $\sigma$-compatible. In the next, we use the same idea to introduce $\sigma$-idempotent reflexive rings as a generalization of idempotent reflexive rings and $\sigma$-abelian rings.
\begin{definition}Let $R$ be a ring and $\sigma$ an endomorphism of $R$.
\NL$(1)$ $R$ is called right $($resp., left$)$ $\sigma$-idempotent reflexive if it is right $($resp., left$)$ idempotent reflexive and $\sigma$-compatible.
\NL$(2)$ $R$ is called $\sigma$-idempotent reflexive if it is both right and left $\sigma$-idempotent reflexive.
\end{definition}
A ring $R$ is right (resp., left) idempotent reflexive if $R$ is right (resp., left) $id_R$- idempotent reflexive, where $id_R$ denotes the identity endomorphism of $R$. We can easily see that $\sigma$-abelian rings are $\sigma$-idempotent reflexive.
\begin{example}\label{ex1}Consider the ring $R=\set{\left(
\begin{array}{cc}
a & b \\
0 & a \\
\end{array}
\right)\mid a,b\in \Z_4}$, where $\Z_4$ is the ring of integers modulo $4$. Let $\sigma$ an endomorphism of $R$ defined by $$\sigma\parth{\left(
\begin{array}{cc}
a & b \\
0 & a \\
\end{array}
\right)}=\left(
\begin{array}{cc}
a & -b \\
0 & a \\
\end{array}
\right)\mathrm{for\;all}\;\left(
\begin{array}{cc}
a & b \\
0 & a \\
\end{array}
\right)\in R$$
By \cite[Example 2.2]{agayev/2009}. The ring $R$ is abelian so it is right and left idempotent reflexive. Also, $R$ is $\sigma$-compatible. Thus $R$ is right and left $\sigma$-idempotent reflexive.
\end{example}
\begin{corollary}\label{corollary1}Let $R$ be a ring and $\sigma$ an endomorphism of $R$. Then
\par$(1)$ If $R$ is left $\sigma$-idempotent reflexive then $R[x,\sigma]$ and $R[[x,\sigma]]$ are left idempotent reflexive.
\par$(2)$ If $R$ is right $\sigma$-idempotent reflexive then $R[x,\sigma]$ and $R[[x,\sigma]]$ are right idempotent reflexive.
\par$(3)$ If $R$ is $\sigma$-idempotent reflexive then $R[x,\sigma]$ and $R[[x,\sigma]]$ are idempotent reflexive.
\end{corollary}
\begin{proof}Immediate from Proposition \ref{prop1}.
\end{proof}
\section{Abelian property}
In this section, $\sigma$ is an endomorphism of $R$ such that $\sigma(1)=1$. We can easily see that $Id(R)\subseteq Id(R[x])$. However $Id(R)\neq Id(R[x])$. We can take the next example.
\begin{example}Let $F$ be a field and consider the ring
$R=\left(
\begin{array}{cc}
F & F \\
0 & F \\
\end{array}
\right)$, then
$e=\left(\begin{array}{cc}
1 & 0 \\
0 & 0 \\
\end{array}
\right)+\left( \begin{array}{cc}
0 & 1 \\
0 & 0 \\
\end{array}
\right)x$ is an element of $Id(R[x])$. But $e\not\in Id(R)$.
\end{example}
According to Kanwara et al. \cite[Corollary 6]{matzuk/2013}, a ring $R$ is abelian if and only if $R[x]$ is abelian. In the next, we give a generalization to skew polynomial rings and skew power series rings.
\begin{lemma}\label{lemma}Let $R$ be an abelian ring such that $\sigma(Re)\subseteq Re$ for all $e\in Id(R)$. For any $e=\sum_{i=0}^ne_ix^i\in Id(R[x,\sigma])$ $($resp., $e=\sum_{i=0}^{\infty}e_ix^i\in Id(R[[x,\sigma]])$$)$, we have $ee_0=e_0$ and $e_0e=e$.
\end{lemma}
\begin{proof}Let $e=\sum_{i=0}^ne_ix^i\in Id(R[x,\sigma])$, then we have the following system of equations
$$e_0^2=e_0\eqno(0)$$
$$e_0e_1+e_1\sigma(e_0)=e_1\qquad\qquad\;\quad\eqno(1)$$
$$e_0e_2+e_1\sigma(e_1)+e_2\sigma^2(e_0)=e_2\qquad\qquad\qquad\qquad\qquad\eqno(2)$$
$$\vdots$$
$$e_0e_n+e_1\sigma(e_{n-1})+\cdots+e_n\sigma^n(e_0)=e_n\qquad\qquad\qquad\qquad\qquad\qquad\;\eqno(n)$$
Since abelian rings are idempotent reflexive, then by the proof of Proposition \ref{prop1}, we have $e_i\in Re_0R=e_0R$ for all $0\leq i\leq n$, because $e_0$ is central. Hence $e_i=e_0e_i=e_ie_0$ for all $0\leq i\leq n$ and so $e_0e=e\;\;(1')$.
\NL On the other hand, from equations $(1)$ and $(1')$, we have $e_1\sigma(e_0)=0$ . Also, we have $e_1\sigma(e_1)+e_2\sigma^2(e_0)=0\;(2')$ from equations $(2)$ and $(1')$. But $\sigma^2(e_0)$ is an idempotent, then equation $(2')$ becomes $e_1\sigma(e_1)\sigma^2(e_0)+e_2\sigma^2(e_0)=e_1\sigma[e_1\sigma(e_0)]+e_2\sigma^2(e_0)=0$, since $e_1\sigma(e_0)=0$ we get $e_2\sigma^2(e_0)=0$. Continuing this procedure yields $e_i\sigma^i(e_0)=0$ for all $1\leq i\leq n$. Thus $ee_0=e_0$. With similar method we get the result for $R[[x,\sigma]]$.
\end{proof}
Recall that a ring $R$ is satisfying the condition $(\mathcal{C_{\sigma}})$ if whenever $a\sigma(b)=0$ with $a,b\in R$, then $ab=0$ \cite{louzari}. Clearly, $\sigma$-compatible rings satisfy the condition $(\mathcal{C_{\sigma}})$.
\begin{lemma}\label{lemma2}Let $R$ be a ring and $\sigma$ an endomorphism of $R$. If $R$ satisfies the condition $(\mathcal{C_{\sigma}})$ then for any $e^2=e\in R$, we have $\sigma(e)=e$.
\end{lemma}
\begin{proof}Let $e\in Id(R)$, by the condition $(\mathcal{C_{\sigma}})$ and from $\sigma(e)(1-\sigma(e))=(1-\sigma(e))\sigma(e)=0$, we get $\sigma(e)=\sigma(e)e$ and $e=\sigma(e)e$. Therefore $\sigma(e)=e$.
\end{proof}
\begin{proposition}\label{prop3}If $R$ satisfies the condition $(\mathcal{C_{\sigma}})$. Then $R$ is abelian if and only if $R[x,\sigma]$ is abelian if and only if $R[[x,\sigma]]$ is abelian.
\end{proposition}
\begin{proof}It suffices to show the necessary condition, let $e=\sum_{i=0}^ne_ix^i\in Id(R[x,\sigma])$. From Lemma \ref{lemma}, we get $e_i=e_ie_0$ and $e_i\sigma^i(e_0)=0$ for all $1\leq i\leq n$. By the condition $(\mathcal{C_{\sigma}})$, we have $e_i=0$ for all $1\leq i\leq n$, then $e=e_0\in Id(R)$. Therefore $R[x,\sigma]$ is abelian. The same method for $R[[x,\sigma]]$.
\end{proof}
\begin{example}\label{ex4}Let $F$ be a field. Consider the ring
$$R=\set{\left(
\begin{array}{cccc}
a & b & 0 & 0 \\
0 & a & 0 & 0 \\
0 & 0 & 0 & v \\
0 & 0 & u & u \\
\end{array}
\right)\mid a,b,u,v\in F
}$$ and $\sigma$ an endomorphism of $R$ defined by $$\sigma\parth{\left(
\begin{array}{cccc}
a & b & 0 & 0 \\
0 & a & 0 & 0 \\
0 & 0 & 0 & v \\
0 & 0 & u & u \\
\end{array}
\right)}=\left(
\begin{array}{cccc}
u & v & 0 & 0 \\
0 & u & 0 & 0 \\
0 & 0 & 0 & b \\
0 & 0 & a & a \\
\end{array}
\right)\mathrm{for\;all}\;\left(
\begin{array}{cccc}
a & b & 0 & 0 \\
0 & a & 0 & 0 \\
0 & 0 & 0 & v \\
0 & 0 & u & u \\
\end{array}
\right)\in R$$
\NL$(1)$ $R$ is abelian. But $R[x;\sigma]$ is not abelian by \cite[Example 2.17]{agayev/2009}.
\NL$(2)$ $R$ does not satisfy the condition $(\mathcal{C_{\sigma}})$. Let $e_{ij}$ denote the $4\times 4$ matrix units having alone $1$ as its $(i,j)$-entry and all other entries $0$. Consider $$e=\left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
\end{array}
\right)\in Id(R)$$
Then
$$\sigma(e)=\left(
\begin{array}{cccc}
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 \\
0 & 0 & 1 & 1 \\
\end{array}
\right)\neq e$$
By Lemma \ref{lemma2}, the condition $(\mathcal{C_{\sigma}})$ is not satisfied.
\end{example}
We can see from Example \ref{ex4}, that the condition $(\mathcal{C_{\sigma}})$ in Proposition \ref{prop3} is not superfluous.
\begin{corollary}\label{corollary}Let $R$ be a $\sigma$-compatible ring. Then $R$ is abelian if and only if $R[x,\sigma]$ is abelian if and only if $R[[x,\sigma]]$ is abelian
\end{corollary}
\begin{proof}Obvious from Proposition \ref{prop3}.
\end{proof}
\begin{corollary}[{\cite[Corollary 6]{matzuk/2013}}]Let $R$ be a ring. Then
$R$ is abelian if and only if $R[x]$ is abelian if and only if $R[[x]]$ is abelian.
\end{corollary}
\begin{proof}It suffices to take $\sigma=id_R$ in Proposition \ref{prop3}.
\end{proof}
\begin{corollary}[{\cite[Theorem 2.21]{agayev/2009}}]\label{corollary2}If $R$ is $\sigma$-abelian then $R[x,\sigma]$ and $R[[x,\sigma]]$ are abelian.
\end{corollary}
\begin{proof}Clearly from Proposition \ref{prop3}.
\end{proof}
In the next, we will give an example of a ring $R$ and an endomorphism $\sigma$ of $R$ such that $R[x,\sigma]$ is abelian, but $R$ is not $\sigma$-abelian. Thus the converse of Corollary \ref{corollary2} is not true.
\begin{example}\label{ex2}Let $\K$ be a field and $R=\K[t]$ a polynomial ring over $\K$
with the endomorphism $\sigma$ given by $\sigma(f(t))=f(0)$ for all
$f(t)\in R$. Since $R$ is commutative then it is abelian.
\NL$(1)$ $R$ is not $\sigma$-abelian. Take $f=a_0+a_1t+a_2t^2+\cdots+a_nt^n$ and
$g=b_1t+b_2t^2+\cdots+b_mt^m$, since $g(0)=0$ so, $f\sigma(g)=0$ but $fg\neq 0$. Therefore $R$ is not $\sigma$-compatible.
\NL$(2)$ Since $R$ has only two idempotents $0$ and $1$ then $Id(R[x,\sigma])=Id(R)$ and thus $R[x,\sigma]$ is abelian.
\end{example}
\begin{remark} If for a ring $R$ we have $Id(R)=\{0,1\}$, then $Id(R[x,\sigma])=\{0,1\}$. Indeed, let $e=\sum_{i=0}^ne_ix^i\in Id(R[x,\sigma])$ and consider the system cited in the proof of Lemma \ref{lemma}. If $e_0=0$ or $e_0=1$, we can easily see that $e_i=0$ for all $1\leq i\leq n$.
\end{remark}
\begin{example}\label{ex3}Let $\Z_2$ is the ring of integers modulo $2$, take $R=\Z_2\oplus\Z_2$ with the usual addition and multiplication. Consider $\sigma\colon R\rightarrow R$ defined by $\sigma((a,b))=(b,a)$. Let $e=(1,0)+(0,1)x,\; f=(0,1)+(0,1)x\in Id(R[x,\sigma])$. Then
\NL$(1)$ $R$ is abelian because it is commutative.
\NL$(2)$ Since $(0,1)e=(0,1)x\neq e(0,1)=0$, then $e$ is not central. Thus $R[x,\sigma]$ is not abelian.
\end{example}
\begin{example}Consider the ring $R=\set{\begin{pmatrix}
a & t \\
0 & a
\end{pmatrix}|\; a\in \Z\;,t\in \Q},$ where $\Z$ and $\Q$ are the
set of all integers and all rational numbers, respectively. The ring
$R$ is commutative then it's abelian, let $\sigma\colon R\rightarrow R$ be an
automorphism defined by $\sigma\left(\begin{pmatrix}
a & t \\
0 & a
\end{pmatrix}\right)=\begin{pmatrix}
a & t/2 \\
0 & a
\end{pmatrix}$.
\NL Since $R$ has only two idempotents $\begin{pmatrix}
0 & 0 \\
0 & 0
\end{pmatrix}$ and $\begin{pmatrix}
1 & 0 \\
0 & 1
\end{pmatrix}$, it follows that $Id(R[x,\sigma])=Id(R)$. Hence $R[x,\sigma]$ is abelian.
\end{example}
|
\subsection{Applications}
We demonstrate the robustness of our method in a variety of real-world applications.
\input{tshirt_results}
The strength of our approach is demonstrated on a sparsely-textured sail surface
with a few dot markers, shown in Fig.~\ref{fig:sail}. Thanks to the large basin
of convergence of our algorithm, we can simply initialize the registration from
a very rough initial estimate without having first to establish correspondences.
Our algorithm naturally exploits line features, which feature point-based
methods usually do not.
Fig.~\ref{fig:bird} depicts another application of our method for animation
capture from a monocular camera stream. In this setting, we capture the
animations of a bird whose animations can be transferred to another character.
The video of captured animations is provided in the supplementary material.
\begin{figure}
\centering
\includegraphics[height=3.4cm]{fig/sail/initial.pdf} \hspace{4mm}
\includegraphics[height=3.4cm]{fig/sail/reconstruct.pdf} \hspace{4mm}
\includegraphics[height=3.4cm]{fig/sail/view.jpg}
\caption{Image registration and surface reconstruction on the sparsely textured sail surface. From left to right: input image and initialization; final registration and reconstruction; the sail shape seen from a different viewpoint.}
\label{fig:sail}
\vspace{-2mm}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=1.9cm]{fig/bird/frame_088.pdf}
\includegraphics[height=1.9cm]{fig/bird/frame_096.pdf}
\includegraphics[height=1.9cm]{fig/bird/frame_104.pdf} \\ \vspace{0.1cm}
\includegraphics[height=3.2cm]{fig/bird/mesh_0088.jpg} \hfill
\includegraphics[height=3.2cm]{fig/bird/mesh_0096.jpg} \hfill
\includegraphics[height=3.2cm]{fig/bird/mesh_0104.jpg}
\caption{Surface reconstruction of an animation capture from a monocular camera stream.}
\label{fig:bird}
\vspace{-4mm}
\end{figure}
\subsection{Basin of Convergence}
\label{basin_of_convergence}
\begin{figure*}
\vspace{-3mm}
\centering
\newcommand{2.1cm}{2.1cm}
\newcommand{2.5cm}{1.5cm}
{\subfigure[] { \includegraphics[width=2.1cm,height=2.5cm]{fig/align/align_temp.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/sc2.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/mc2.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/mi2.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/ncc2.png} }} \\ \vspace{-4mm}
{\subfigure[] { \includegraphics[width=2.1cm,height=2.5cm]{fig/align/align_inp.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/sc.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/mc.png} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/mi_new.pdf} }} \hspace{2mm}
{\subfigure[] { \includegraphics[width=3cm,height=2.5cm]{fig/align/ncc.png} }}
\vspace{-1mm}
\caption{Robustness of alignment functions {\it w.r.t.} translations between (a) template, and (f) input image, showing the basin of convergence of the alignment costs around the correct position using {\bf Top row:} weak Gaussian smoothing, {\bf Bottom row:} strong Gaussian smoothing, over (b)(g) Intensity, (c)(h) GBDF, (d)(i) MI, (e)(j) NCC.}
\label{fig:exp_align}
\vspace{-1mm}
\end{figure*}
To understand the limitations of the various cost functions, we conducted a simple alignment experiment to test their respective sensitivities to initial position and image distractors; results are presented in Fig.~\ref{fig:exp_align}. The red image window in the input image is translated in x and y about the known best alignment to the green template window, and the cost to compare each window pair is plotted, to allow the basins of convergence to be inspected visually. \ac{MI} and \ac{NCC} both reach a maximum value close to 1 at the point of best alignment, but we invert these functions so that a minimum cost of all functions is expected at the point of best alignment.
The GBDF descriptors from Eq.~\ref{MC_eqn} are seen to have a strong minimum at the point of best alignment, with a reasonably wide, smooth basin of convergence, the desired property of a good cost function. However, Intensity, MI, and NCC all have several nearby local minima. Mutual Information is further seen to have a very narrow basin of convergence around the correct point of best alignment, meaning that it is very likely to converge to an incorrect alignment given an imperfect initial position.
This experiment only tests translation sensitivity because this is the variation best understood visually, but the similar results are likely from other types of misalignment.
\subsection{Well-Textured Surfaces}
We performed experiments using the well-textured paper dataset presented in \cite{Varol12a} consisting of $193$ consecutive images, for example see Fig.~\ref{fig:capture_paper_failure}. Quantitative results are presented in Fig.~\ref{fig:capture_paper_errors}. For this well-textured dataset, all the feature point-based methods work well and dense matching methods are only slightly better. The biggest errors are due to lighting changes, where intensity features using SSD occasionally fail to track part of the surface, and hence have a higher error.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth,height=3cm]{fig/capture_paper/error_graph}
\caption{Reconstruction results on the well-textured paper dataset, no occlusions. All feature-points based methods work reasonably well.}
\label{fig:capture_paper_errors}
\vspace{-0.3cm}
\end{figure}
To evaluate the robustness of each method to occlusion, we add artificial hand image occlusions to the image sequence. The reconstruction results are presented in Fig.~\ref{fig:capture_paper_oc_errors}.
Feature-based methods still produce reasonably good 2D reprojection results in this dataset, but the recovered depths under the occlusion are not very accurate.
Fig.~\ref{fig:capture_paper_failure} provides the output for a single frame where it can be seen that the reconstruction fails when using the strong GBDF without occlusion handling and also when using M-estimators to attempt to handle occlusion, while the proposed framework is still able to able track the surface accurately. In this situation, Mutual Information and both the Tukey and Huber M-estimators are confused by the edges created by the finger and converge to incorrect locations.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth,height=3cm]{fig/capture_paper_occlusion/error_graph}
\caption{Reconstruction results on the well-textured paper dataset, with occlusions. Feature-based methods are largely robust to occlusion, however the overall depths recovered are not as accurate as the proposed framework that includes occlusion handling.}
\vspace{-0.4cm}
\label{fig:capture_paper_oc_errors}
\end{figure}
\begin{figure}
\centering
\newcommand{2.1cm}{2.4cm}
{\subfigure[Ostlund12]
{\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/cvpr14_f65}
\hspace{.001cm}}}
{\subfigure[GBDF alone]
{\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/mc_f65}
\hspace{.001cm}}}
{\subfigure[MI]
{\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/mutualinfo_f65}
\hspace{.001cm}}} \\ \vspace{-2mm}
{\subfigure[Tukey]
{\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/tukey_f65}
\hspace{.001cm}}}
{\subfigure[Huber]
{\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/huber_f65}
\hspace{.001cm}}}
{\subfigure[GBDF+Oc]
{\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/our_method_f65} }} \\
\vspace{-1mm}
\caption{Output for a single frame showing relative reconstruction accuracies. Mutual Information and M-estimators fail to correctly handle the occlusion, while the proposed framework is successful.}
\label{fig:capture_paper_failure}
\vspace{-0.4cm}
\end{figure}
We also demonstrate that the proposed rotation handling technique described in Section \ref{features} that overcomes the rotation sensitivity of the GBDF descriptors can successfully track a rotating deformable object.
\comment{We ran our tracking algorithm with and without rotation handling on the well-textured paper dataset with images rotated 5 degree every frame. This rotation in addition to original surface deformations introduce large frame-to-frame motions.}
Fig.~\ref{fig:rotation_handling} shows that without rotation handling, the original GBDF descriptors can only track up to 50 degrees of rotation, while the proposed rotation handling technique can track the whole sequence including a full 360 degrees of rotation.
\begin{figure*}
\vspace{-0.1cm}
\centering
\newcommand{2.1cm}{2.0cm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_149} \hspace{2mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_151} \hspace{2mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_153} \hspace{2mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_156} \hspace{2mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_159} \hspace{2mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_161} \hspace{2mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_167}
\\
\hspace{5mm}
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_149} \hfill
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_156} \hfill
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_163} \hfill
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_175} \hfill
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_115} \hfill
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_125} \hfill
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_141} \hspace{5mm}
\caption{Tracking a rotating deformable surface. {\bf Top row:} Without rotation
handling, tracking with the original GBDF descriptors eventually breaks down. {\bf Bottom row:} Our modified GBDF features can track the whole sequence with up to $360^o$ of rotation.}
\label{fig:rotation_handling}
\vspace{-0.4cm}
\end{figure*}
\section{Conclusion}
We have presented a framework for tracking both well textured and sparsely textured deforming surfaces in videos in the presence of occlusions. Our framework computes a relevancy score for each pixel, which is then used to weight the influence of the image information from that pixel in the image energy cost function. The presented method favorably compares to standard cost functions used for handling occlusion, such as Mutual Information and M-estimators.
\subsection{Template Matching}
We assume we are given both a template image $T$ and the rest shape of the corresponding deformable surface, which is a triangular mesh defined by a vector of $N_v$ vertex coordinates in 3D, $\mathbf{V}_T\in\R^{N_v \times 3}$.
To recover the shape of the deformed surface in an input image {\it I}, the vertex coordinates $\mathbf{V}_T$ of the 3D reference shape must be adjusted so that their projection onto the image plane aligns with {\it I}.
We assume the internal parameters of the camera are known and, without loss of generality, that the world reference system coincides with the one of the camera.\comment{All RGB color images are converted to grayscale before feature descriptors are extracted.}
In order to register each input image, a pixel-wise correspondence is sought between the template and the input image. Each pixel $\mathbf{x}\in\R^2$ on the template corresponds to a point $\mathbf{p}\in\R^3$ on the 3D surface. This 3D point is represented by fixed barycentric coordinates which are computed by backprojecting the image location $\mathbf{x}$ onto the 3D reference shape.
The camera projection defines an image warping function $\mathbf{W}: \R^2\times \R^{3\times N_v}\rightarrow \R^2$ which sends pixel $\mathbf{x}$ to a new image location based on the current surface mesh $\mathbf{V}$ as illustrated in Fig.~\ref{fig:warp_function}.
The optimal warping function should minimize the difference between $T(\mathbf{x})$ and $I(\mathbf{W(x;V)})$, according to some measurement of pixel similarity. Traditionally, image intensity has been used, but more robust pixel feature descriptors $\phi_I(\mathbf{x})$ will lead to more meaningful comparisons, as discussed in Section \ref{features}.
\begin{figure}
\vspace{-0mm}
\centering
\includegraphics[height=3.4cm]{fig/warp_function}
\caption{An image warping function maps a pixel from the template image onto the deforming surface in the input image.}
\label{fig:warp_function}
\vspace{-3mm}
\end{figure}
The image energy cost function is a comparison between $\phi_T(\mathbf{x})$ and $\phi_I(\mathbf{W(x;V)})$ at every image point $\mathbf{x}$ defining the quality of their alignment
\begin{equation}
E_{\text{image}}(\V) = \sum_{\mathbf{x}} d\big(\phi_T(\mathbf{x}),
\phi_I(\mathbf{W(x;V)}) \big) .
\label{eq:minImageEnergy}
\end{equation}
There are many possible choices for the function $d$ comparing the descriptor vectors, such as \ac{SSD}, \ac{NCC}, \ac{MI}, and others. We will discuss more in detail about the choice of $d$ in Section \ref{sec_sim_fcn}.
Since monocular 3D surface reconstruction is an under-constrained problem and there are multiple 3D shapes having the same reprojection on the image plane, minimizing the image energy in Eq.~\eqref{eq:minImageEnergy} alone is ill-posed. Additional constraints must be added, such as isometric deformation constraints enforcing that the surface should not stretch or shrink. A change in the length between vertex $\mathbf{v}_i$ and vertex $\mathbf{v}_j$ as compared to the template rest length $l_{ij}$ from $\mathbf{V}_T$ is penalized as
\begin{equation}
E_{\textrm{length}}(\V) = \sum_{i,j}(\left \| \mathbf{v}_i - \mathbf{v}_j \right \| - l_{ij})^2 .
\label{eq:lengthEnergy}
\end{equation}
To encourage physically plausible deformations, the Laplacian mesh smoothing proposed in \cite{Ostlund12} is used. This rotation-invariant curvature-preserving regularization term based on Laplacian smoothing matrix $\mathbf{A}$ penalizes non-rigid deformations away from the reference shape, based on the preservation of affine combinations of neighboring vertices.
\begin{equation}
E_{\text{smooth}}(\V) = \left \| \mathbf{AV} \right \|^2 .
\label{eq:smoothEnergy}
\end{equation}
To reconstruct the surface, we therefore seek the mesh configuration $\V$ that minimizes the following total energy:
\begin{equation}
\underset{\mathbf{V}}{\arg\min} \ \ E_{\text{image}}(\V) + \lambda_{\text{L}} E_{\text{length}}(\V) + \lambda_{\text{S}} E_{\text{smooth}}(\V) ,
\label{eq:totalEnergy}
\end{equation}
for relative weighting parameters $\lambda_{\text{L}}$ and $\lambda_{\text{S}}$.
\section{Experiments and Results}
\label{sec:experiments}
3D surface reconstructions are computed with and without occlusion on both well and poorly textured deforming surfaces. We compare recent methods described in Section \ref{sec_related_work}, which are representative of the current state-of-the-art, against our dense template matching-based reconstruction methods using the various similarity measures and occlusion handling techniques described in Section \ref{sec_framework}.
In particular, we report detailed results of comparisons with the following methods: ``Bartoli12'' \cite{Bartoli12b}, that reconstructs the surface by analytically solving a system of PDEs starting from an estimated 2D parametric warp between images; ``Chhatkuli14'' \cite{Chhatkuli14}, that infers the surface shape exploiting the depth gradient non-holomonic solution of a PDE; ``Brunet14'' \cite{Brunet14}, that reconstructs a smooth surface imposing soft differential constraints of isometric deformation; ``Ostlund12'' \cite{Ostlund12}, that introduces the Laplacian mesh smoothing we employ; and ``Salzmann11'' \cite{Salzmann11a}, that uses pre-learned linear local deformation models.\footnote{Code provided by the authors of these papers was used for all comparisons.}
As for pixel-based template matching techniques, comparing pixel intensity values ``Intensity'' and gradient direction values ``GD'' are done using \ac{SSD}. We also compare standard ``NCC'' and ``MI'' over intensity values. The ``GBDF'' features are compared using \ac{SSD}, and were seen to be the strongest feature descriptor, so it is these values that we test in the M-estimator framework using the ``Huber'' and ``Tukey'' loss functions. Our proposed framework from Section \ref{sec:newMethod} is labeled ``GBDF+Oc'' in the figures. We see that it achieves state-of-the-art performances on a standard, well-textured dataset, and it achieves optimal reconstruction performance in all datasets with occlusions and low texture.
Image sequences were acquired using a Kinect camera, and ground truth surfaces were generated from the depth information. The template is constructed from the first frame, and 3D reconstruction is performed for the rest of the sequence using the image information alone. The initial mesh coordinates for each frame are set to the locations of the final reconstruction of the previous frame in the sequence.
We consider two different metrics to define the reconstruction accuracy. Many previous methods compare the average distance of the reconstructed 3D mesh vertices to their closest projections onto the depth images. This metric ignores the correspondences between the mesh points and the point cloud. As a more meaningful metric, we use the Kinect point cloud to build ground truth meshes, and compute the average vertex-to-vertex distance from the reconstructed mesh to the ground truth mesh. This metric is used for the paper itself. Results using the vertex-to-point-cloud distance are provided in the supplementary material.
\input{parameters}
Our approach relies on frame-to-frame tracking and thus requires a
sufficiently good initialization. However, because the method has a wide basin
of convergence, a rough initialization suffices. Our method can fail when the
initialization is too far from the solution, when frame-to-frame deformations
are so large that the relevancy scores stop being reliable, or when large
changes in surface appearance and severe occlusions cause the image energy
term to become uninformative. If the tracking is lost, it must be
reinitialized, for example by using a feature point based method. However,
this did not prove to be necessary to obtain any of the results shown below.
\input{basin_convergence}
\input{capture_paper_results}
\input{minimal_textured_results}
\input{applications}
\section{Proposed Framework} \label{sec_framework}
In this work, we demonstrate that a carefully designed dense template matching framework can lead to state-of-the-art results in monocular reconstruction of deformable surfaces\comment{, without the need for any statistical learning step}. In this section we describe our framework, based on a recently introduced gradient-based pixel descriptors \cite{Crivellaro14} for robust template matching and the computation of a relevancy score for outlier rejection.
\input{densematching}
\input{robust_optim}
\input{new_method}
\input{minimal_textures}
\section{Introduction}
Being able to recover the 3D shape of deformable surfaces from ordinary images
will make it possible to field reconstruction systems that require only a single
video camera, such as those that now equip most mobile devices. It will also
allow 3D shape recovery in more specialized contexts, such as when performing
endoscopic surgery or using a fast camera to capture the deformations of a
rapidly moving object. Depth ambiguities make such monocular shape recovery
highly under-constrained. Moreover, when the surface is partially
occluded or has minimal texture, the problem
becomes even more challenging because there is little or no useful information
about large parts of it.
Arguably, these ambiguities could be resolved by using a
depth-camera, such as the popular Kinect sensor~\cite{Shotton11}.
However, such depth-cameras are more difficult to fit into a
cell-phone or an endoscope and have limited range. In this work, we focus on
3D shape recovery given a reference image and a single corresponding 3D
template shape known {\it a priori}.
When the surface is well-textured, correspondence-based methods have proved
effective at solving this problem, even in the presence of occlusions
\cite{Bartoli12b,Bronte14,Brunet14,Chhatkuli14,Perriollat11,Pizarro12,Vicente12}.
In contrast, when the surface lacks texture, dense pixel-level template
matching should be used instead. Unfortunately, many methods such
as~\cite{Malti11,Salzmann08a} either are hampered by a narrow basin of
attraction, which means they must be initialized from interest points
correspondences, or require supervised learning to enhance robustness. Using
Mutual Information has often been claimed~\cite{Dame12,Dowson06,Panin08,Viola97}
to be effective at handling these difficulties but our experiments do not bear
this out. Instead, we advocate template matching over robust dense features that
relies on a pixel-wise relevancy score pre-computed for each frame, as shown in
Fig.~\ref{fig_relevancy_scores_low_texture}. Our approach can handle occlusions
and lack of texture simultaneously. Moreover, no training step is required as in
\cite{Salzmann08a}, which we consider to be an advantage because this obligates
either collecting training data or having sufficient knowledge of the surface
properties, neither of which may be forthcoming.
\begin{figure}[t]
\vspace{-0mm}
\centering
\newcommand{2.5cm}{2.5cm}
\newcommand{2.1cm}{3.45cm}
{\subfigure[ ] { \includegraphics[height=2.5cm,width=2.1cm]{fig/white_paper/model}}} \hspace{0.3cm} \vspace{-0.2cm}
{\subfigure[ ] { \includegraphics[height=2.5cm,width=2.1cm]{fig/white_paper/frame_226} }} \\
{\subfigure[ ] { \includegraphics[height=2.5cm,width=3.9cm]{fig/white_paper/relevancy_score} }} \hspace{-0.2cm} \vspace{-0.1cm}
{\subfigure[ ] { \includegraphics[height=2.5cm,width=2.1cm]{fig/white_paper/our_method_f226} }}
\caption{Tracking a sparsely textured surface in the presence of occlusion:
(a) template image, (b) input image, (c) relevancy score, (d) surface tracking
result with proposed framework. All figures in this paper are best viewed in
color.}
\label{fig_relevancy_scores_low_texture}
\vspace{-5mm}
\end{figure}
Our main contribution is therefore a robust framework for image
registration and monocular 3D reconstruction of deformable surfaces in the
presence of occlusions and minimal texture. A main ingredient is the pixel-wise
relevancy score we use to achieve the robustness. We will make the code
publicly available, and release the dataset we used to validate our
approach, which contains challenging sequences of sparsely-textured deforming
surfaces and the corresponding ground truth.
\subsection{Sparsely Textured Surfaces}
To understand the performance of the methods in a realistic, sparsely textured setting, a different dataset is required. A less textured paper dataset exists, as published in \cite{Salzmann08a}, but no ground truth information is available for this dataset in 3D, and so it is not suitable for numerical comparisons. Nevertheless, for qualitative comparison purposes, we ran the proposed framework on this dataset, and our reconstructions align very well to the image information. Example frames are provided in Fig.~\ref{fig:Salz_video_frames}, and the entire video is provided as supplementary material. The best known published results on this dataset are found in \cite{Salzmann12a}, which uses an algorithm that requires training data in addition to explicitly delineating the edges of the surface. Our proposed framework is seen to perform as well as this previous method, qualitatively, while requiring no learning.
In order to be able to perform more meaningful numerical comparisons, we constructed a new dataset along with ground truth in 3D using a Kinect sensor, example images are provided in Fig.~\ref{fig:white_paper_failure}. \remove{This new dataset, along with ground truth, will be made available for public use. The} This new sparsely textured paper dataset contains various deformations and large lighting changes along with occlusions.
\begin{figure}
\centering
\includegraphics[height=1.6cm]{fig/salzmann/a1.png}
\includegraphics[height=1.6cm]{fig/salzmann/a2.png}
\includegraphics[height=1.6cm]{fig/salzmann/a4.png} \\
\includegraphics[height=1.6cm]{fig/salzmann/s1.png}
\includegraphics[height=1.6cm]{fig/salzmann/s2.png}
\includegraphics[height=1.6cm]{fig/salzmann/s4.png}
\caption{Sample reconstructions from the \cite{Salzmann08a} dataset. While no ground truth is available in 3D, our results (top row) are qualitatively observed to be very accurate; the best published results on this dataset are \cite{Salzmann08a} (bottom row), which has to extract the image edges explicitly, and involves learning, while our method does not. We do not have access to a reference image where the surface is in its planar rest shape, as our mesh assumes, causing some misalignment at the surface boundary.}
\label{fig:Salz_video_frames}
\vspace{-2mm}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth,height=3cm]{fig/white_paper/error_graph}
\caption{Reconstruction results on the sparsely textured paper dataset. Feature-based methods fail to reconstruct plausible surfaces, as indicated by the out-of-range error bars on the left.}
\label{fig:white_paper_errors}
\vspace{-0.5cm}
\end{figure}
Quantitative results using the new dataset are presented in Fig.~\ref{fig:white_paper_errors}.
Feature-based methods that fail to reconstruct plausible surface shapes are indicated by high error bars that exceed y-axis range. Fig.~\ref{fig:white_paper_failure} provides a representative reconstruction on a single frame. \ac{NCC} and \ac{MI} can track the surface fairly accurately, however they fail to capture fine details at the surface boundaries and hence the recovered depths in 3D are not very precise. Without occlusion handling, dense matching with gradient-based descriptors often fails near occlusions. The M-estimators are inconsistent near occlusions. However, the proposed framework is seen to be able to accurately track the surface throughout the entire sequence.
It is interesting to note that while the occlusions are cleanly delineated in the relevancy score over a textured surface, they are much less obviously visible in the relevancy score over a sparsely textured surface. This is expected, because well-textured un-occluded regions have consistently high correlation values with the template, and so it is only the occluded regions that are assigned low relevancy scores. However, image regions of little texture have low and noisy correlation with a template, so occluded regions of similarly low correlation are assigned similarly low relevancy scores, and an occluded region is not as obviously distinct from the low textured regions in the relevancy score map. This is one of the strengths of the proposed framework, because only the truly meaningful image regions are allowed to strongly influence the image energy cost.
\begin{figure}
\vspace{-3mm}
\centering
{\subfigure[GBDF+Oc]
{\includegraphics[height=1.8cm, width=2.3cm]{fig/white_paper/our_method_f226}
\hspace{.1cm}}}
{\subfigure[NCC]
{\includegraphics[height=1.8cm, width=2.3cm]{fig/white_paper/ncc_f226}
\hspace{.1cm}}}
{\subfigure[MI]
{\includegraphics[height=1.8cm, width=2.3cm]{fig/white_paper/mutualinfo_f226}
\hspace{.1cm}}} \\ \vspace{-0.2cm}
{\subfigure[GBDF alone]
{\includegraphics[height=1.8cm, width=2.3cm]{fig/white_paper/MC_f226}
\hspace{.1cm}}}
{\subfigure[Tukey]
{\includegraphics[height=1.8cm, width=2.3cm]{fig/white_paper/tukey_f226}
\hspace{.1cm}}}
{\subfigure[Huber]
{\includegraphics[height=1.8cm, width=2.3cm]{fig/white_paper/huber_f226}
\hspace{.1cm}}}
\caption{Reconstruction results on the same single frame. The proposed framework can track the whole sequence accurately, while other methods are seen to have trouble handling occlusions on top of the sparsely textured surface.}
\label{fig:white_paper_failure}
\vspace{-0.5cm}
\end{figure}
\subsection{Handling Sparsely Textured Surfaces}
The relevancy score described in Section~\ref{sec:newMethod} is also able to handle sparsely textured surfaces. Image regions containing little or no texture have low relevancy scores, so these pixels will not negatively influence the image alignment. For example, see Fig.~\ref{fig_relevancy_scores_low_texture}. Using the proposed relevancy score to weight the utility of the image information coming from each pixel in the image allows the optimization to be driven by the most meaningful available information.
\subsection{Handling Occlusions with a Relevancy Score}
\label{sec:newMethod}
\begin{figure}
\vspace{-0.2cm}
\centering
\newcommand{2.5cm}{1.3cm}
{\subfigure[]
{\includegraphics[height=2.5cm]{fig/tshirt/frame_70.png}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=2.5cm]{fig/tshirt/score_intensity_f70}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=2.5cm]{fig/tshirt/score_MC_f70}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=2.5cm]{fig/tshirt/score_graddirect_f70}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=2.5cm]{fig/tshirt/score_intensity_mc_f70}
\includegraphics[height=2.5cm]{fig/colorbar}
\hspace{-.1cm}}} \\
\vspace{-2mm}
\newcommand{\figureHeightb}{1.12cm}
{\subfigure[]
{\includegraphics[height=\figureHeightb]{fig/white_paper/frame_226}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=\figureHeightb]{fig/white_paper/score_intensity_f226}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=\figureHeightb]{fig/white_paper/score_mc_f226}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=\figureHeightb]{fig/white_paper/score_graddirect_f226}
\hspace{-.1cm}}}
{\subfigure[]
{\includegraphics[height=\figureHeightb]{fig/white_paper/score_intensity_mc_f226}
\includegraphics[height=\figureHeightb]{fig/colorbar}
\hspace{-.1cm}}}
\caption{The relevancy score results using various methods on the (a) cloth dataset and (f) sparsely textured paper dataset using (b)(g) Intensity (c)(h) GBDF (d)(i) Gradient direction (e)(j) GBDF+Intensity. GBDF gives a better relevancy map than intensity and gradient direction on the first dataset while intensity is better than GBDF and gradient direction on the second. We therefore combine both intensity and GBDF in our proposed relevancy score.}
\label{fig:exp_OCS}
\vspace{-0.4cm}
\end{figure}
Our experiments suggest that selecting a robust similarity function is not enough to deal with the occlusions and image variability encountered when attempting to track a deforming surface in real-world imagery.
Inspired by the effectiveness of the occlusion masks developed in works such as \cite{Strecha06,Wang09c,Zhou09}, we derive a more robust method to handle occlusions by pre-computing a relevancy score for each pixel of the current frame, which is then used to weight the pixels during the alignment. Since we would like to handle occlusions and sparsely textured surfaces together, rather than designing a binary occlusion prediction mask, we develop a continuous-valued score that will raise or lower the importance of pixels depending on their relevancy. This pre-processing step can greatly improve the quality of the image information handed to the cost function in Eq.~\eqref{eq:totalEnergy}.
Given the estimated configuration $\mathbf{V}^*_{t-1}$ of the deformable surface from the previous frame, a thin-plate spline-based warping function \cite{Bookstein89} is used to un-warp image $I_{t}$ to closely align with the template $T$. A relevancy score is then computed between each pixel $\mathbf{x}$ on the synthetically back-warped image $\hat{I}_t$ and the same pixel on the template $T$, with a sliding-window approach.
It has been verified repeatedly in the literature that NCC is a reliable choice for measuring patch-based image similarity, and so we compute the NCC over the images as an efficient prediction of relevancy.
In one such approach \cite{Klein07}, local image patches are affinely warped based on the predicted camera pose, and sliding NCC windows are then used to look for correspondences of map points in the input image. Our approach is somewhat different, as we use sliding NCC to measure the relevancy of template pixels on the input image.
We average the NCC of both the image intensity and the GBDF features, as it was found that both descriptors provide relevant and often complementary information at this predictor stage, (see Fig.~\ref{fig:exp_OCS} for a qualitative comparison).
The sliding relevancy score is computed as the maximum NCC value over a range of patches near image location $\mathbf{x}$:
\begin{equation}
\omega(\mathbf{x}) = \max_{\mathbf{\delta}} \ \text{NCC}(\patchT(\mathbf{x}), \mathcal{P}_{\hat{I}}(\mathbf{x} + \mathbf{\delta})),
\end{equation}
where $\patchT(\mathbf{x})$ and $\mathcal{P}_{\hat{I}}(\mathbf{x})$ are patches of size $26 \times 26$ centered at $\mathbf{x}$, $\delta = [ \delta_x, \delta_y ]^T $, and $\delta_x, \delta_y$ vary over $[-30,30]$ in all our experiments. Allowing the patch to be compared to nearby patches accounts for some of the variability between the surface position $\mathbf{V}^*_{t-1}$ and the desired position $\mathbf{V}^*_{t}$ to be recovered.
The similarity scores are then normalized to lie in $[0,1]$, and outlier data is also limited at this stage in a process similar to an M-estimator. The mean $\mu$ and standard deviation $\sigma$ of the NCC scores are found for each frame, and all values further than $3\sigma$ from the mean are clamped to the interval $\mu \pm 3\sigma$. These values are then linearly rescaled to lie between 0 and 1, and the normalized weights $\hat{\omega}$ are applied to the data in the image energy term of Equation \eqref{eq:totalEnergy}:
\begin{equation}
E_{\text{image}}(\V) = \sum_{\mathbf{x}} \hat{\omega}(\mathbf{x}) d\big(\phi_T(\mathbf{x}),
\phi_I(\mathbf{W(x;V)}) \big) ,
\label{eq:energy}
\end{equation}
where the sum here is extended to all the pixels of the template. Relevancy scores for the well-textured paper dataset are shown in Fig.~\ref{fig:comp_ncc}.
\begin{figure}
\centering
\newcommand{2.1cm}{1.6cm}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/seq/frame_21}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/seq/frame_37}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/seq/frame_57}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/seq/frame_64}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/seq/frame_76}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/map/frame_21}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/map/frame_37}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/map/frame_57}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/map/frame_64}
\includegraphics[width=2.1cm]{fig/capture_paper_occlusion/map/frame_76}
\vspace{0.1cm}
\caption{Relevancy scores for the well-textured paper dataset.}
\vspace{-0.5cm}
\label{fig:comp_ncc}
\end{figure}
\section{Related Work}
\label{sec_related_work}
The main approaches for deformable surface reconstruction either require 2D tracking throughout a batch of images or a video sequence \cite{Akhter11,Garg13,Russell12a} \remove{(a problem referred to as \textit{Non-Rigid Structure from Motion}), }or they assume a reference template and corresponding 3D shape is known. In this work, we focus on the second approach, which we refer to as \textit{template-based reconstruction}.
The most successful current approaches generally rely on finding feature point
correspondences
\cite{Ostlund12,Bartoli12b,Bronte14,Chhatkuli14,Perriollat11,Pizarro12,
Vicente12}, because they are robust to occlusions. Unfortunately, as shown by
our experimental results, these methods tend to break down when attempting to
reconstruct sparsely or repetitively textured surfaces, since they rely on a
fairly high number of correct matches.
Pixel-based techniques are able to overcome some limitations of local feature matching, since they reconstruct surfaces based on a global, dense comparison of images. On the other hand, some precautions must be taken to handle occlusions, lighting changes, and noise. \cite{Malti11} estimates a visibility mask on the reconstructed surface, but unlike us, only textured surfaces and self-occlusions are handled. \cite{GayBellile10} registers images of deformable surfaces in 2D and shrinks the image warps in self-occluded areas. \cite{Bartoli12b} proves that an analytical solution to the 3D surface shape can be derived from this 2D warp. However, the surface shape in self-occluded areas is undefined.
\cite{Collins14b} registers local image patches of feature point correspondences to estimate their depths, and geometric constraints are imposed to classify incorrect feature point correspondences. In contrast to these local depth estimations, our method reconstructs surfaces globally in order to be more robust to noise and outliers.
Other recent approaches employ supervised learning for enhancing performance \cite{Salzmann13,Varol12a}. In \cite{Salzmann08a} strong results are achieved with poorly textured surfaces and occlusion by employing trained local deformation models, a dense template matching framework using \ac{NCC} \cite{Scandaroli12} and contour detection. Our proposed framework manages to achieve similar performance without requiring any supervised learning step, while the use of robust, gradient-based dense descriptors recently proposed in \cite{Crivellaro14} avoids the need to explicitly detect contours.
Other techniques employed for dealing with occlusions and noise, such as \ac{MI} \cite{Damen12,Dowson06,Panin08,Viola97} and robust M-estimators \cite{Arya07} are studied explicitly in our context, and found to be successful only up to a point.
Our method is similar to that of~\cite{Pilet07}, where a template matching approach is employed and a visibility mask is computed for the pixels lying on the surface, but in this work a very good initialization from a feature point-based method is required in order for its EM algorithm to converge.
In addition to the geometrical degrees of freedom of the surface, local illumination parameters are explicitly estimated in \cite{Hager98,Silveira07}. This requires a reduced deformation model for the surface to keep the size of the problem reasonable.
In the proposed framework, we achieve good performance without the need to explicitly estimate any illumination model, so that an accurate geometric model for the surface can be employed. Furthermore, rather than estimating a simple visibility mask as is often done in many domains such as stereo vision \cite{Strecha06}, face recognition \cite{Zhou09}, or pedestrian detection \cite{Wang09c}, we employ a real-valued pixel-wise relevancy score, penalizing at the same time pixels with unreliable information originating both from occluded and low-textured regions. Our method has a much wider basin of convergence and we can track both well and poorly textured surfaces without requiring initialization by a feature point-based method.
\subsection{Robust Optimization}
\label{sec_features}
\subsubsection{Optimization Scheme}
To make the optimization more robust to noise and wide pose changes, we employ a
multi-scale approach, iteratively minimizing $E^\sigma =
E_{\text{image}}^\sigma + \lambda_{\text{L}} E_{\text{length}} +
\lambda_{\text{S}} E_{\text{smooth}}$ for decreasing values of a scale parameter
$\sigma$, with:
\begin{equation}
E_{\text{image}}^\sigma = \sum_{\mathbf{x}} d\left(G^\sigma * \phi_T(\mathbf{x}),
G^\sigma * \phi_I(\mathbf{W(x;V)}) \right),
\label{eq:minImageEnergySmooth}
\end{equation}
where $G^\sigma$ is a low-pass Gaussian filter of variance $\sigma^2$.
In our experiments we solve the alignment at three scales, using the final
result of each coarser scale to initialize the next set of iterations, and
initializing the coarsest scale with the final position found for the previous
frame. The first frame of each image sequence is taken as the template, and we
employ a standard Gauss-Newton algorithm for minimization. \remove{Our
method uses the forward additive registration algorithm, which requires the
computation of the Hessian and Jacobian matrices at every iteration. However,
we could use the inverse compositional framework algorithm to speed things
up by pre-computing them only once as in \cite{Brunet11}.}
\subsubsection{Feature Selection}
\label{features}
The image information compared in Eq.~\eqref{eq:minImageEnergy} comes from
pixel-based image features. Previous approaches
\cite{Malti11,Pilet07,Salzmann08a} employ image intensity as a local descriptor,
$\phi_I(\mathbf{x}) = I(\mathbf{x})$. More robust results can be obtained with other features,
such as the lighting-insensitive image gradient direction (GD) \cite{Gopalan10},
where $\phi_I(\mathbf{x}) = \tan ^{-1} \frac{I_y(\mathbf{x})}{I_x(\mathbf{x})}$ with $\mod 2\pi$
differencing. Based on its strong previous performance we also consider the
Gradient Based Descriptor Fields (GBDF) recently proposed in
\cite{Crivellaro14}:
\begin{equation}
\phi_I(\mathbf{x}) = \big[ [ \frac{\partial I }{\partial x}(\mathbf{x})]^+,\ [\frac{\partial I} {\partial x}(\mathbf{x})]^-,\ [\frac{\partial I }{\partial y}(\mathbf{x})]^+,\ [\frac{\partial I} {\partial y}(\mathbf{x})]^- \big]^\top \> , \label{MC_eqn}
\end{equation}
where the $[\cdot]^+$ and $[\cdot]^-$ operations respectively keep the positive
and negative values of a real-valued signal. These descriptors are robust under
light changes, and remain discriminative after the Gaussian smoothing employed
in Eq.~\eqref{eq:minImageEnergySmooth}; however, as originally proposed in
\cite{Crivellaro14}, they are not rotation invariant. To achieve in-plane
rotation invariance, in our final framework we employ a modified version of
GBDF. In order to compare pixel descriptors in the same, unrotated coordinate
system, the reconstruction of the previous frame is used to establish a local
coordinate system for each mesh facet. Each pixel descriptor on the template is
then rotated in accordance with its corresponding mesh facet, to be directly
comparable to the points in the input image. We show in Section
\ref{sec:experiments} that this modification indeed increases registration
accuracy by being able to successfully track a rotating deformable surface.
\subsubsection{Similarity Function Selection} \label{sec_sim_fcn}
Choosing the correct comparison function $d$ for Eq.~\eqref{eq:minImageEnergy} also significantly affects the robustness of the tracking.
Common choices include the \ac{SSD} of the descriptors, and the \ac{NCC} of image intensities \cite{Lewis95}, which is invariant under affine changes in lighting.
\input{mutualinformation}
\input{robust_stats}
\subsection{Tracking a Rotating Object} \label{rotation_results}
To demonstrate that the proposed rotation handling technique can track a
rotating deformable object, we ran our tracking algorithm with and without
rotation handling on the well-textured paper dataset with images
rotated 5 degree every frame. This rotation in addition to original surface
deformations introduce large frame-to-frame motions.
Fig.~\ref{fig:rotation_handling} shows that without rotation handling, the original gradient-based descriptors can only track up to 50 degrees of rotation, while the proposed rotation handling technique can track the whole sequence including a full 360 degrees of rotation.
\begin{figure*}[t]
\centering
\newcommand{2.1cm}{2.1cm}
\begin{tabular}{cccccccc}
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_149} &
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_151} &
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_153} &
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_156} &
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_159} &
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_161} &
\includegraphics[width=2.1cm]{fig/handle_rotation/without/frame_167} \\
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_149} &
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_156} &
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_163} &
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_175} &
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_115} &
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_125} &
\includegraphics[width=2.1cm]{fig/handle_rotation/with/frame_141}
\end{tabular}
\caption{Tracking a rotating deformable surface. {\bf Top row:} Without rotation
handling, tracking with the original gradient-based descriptors eventually breaks down. {\bf Bottom row:} The proposed rotation handling technique can track the whole sequence with up to 360 degrees of rotation.}
\label{fig:rotation_handling}
\end{figure*}
\subsection{Rotation Handling}
The original gradient-based descriptors are not rotation invariant therefore
break down when tracking a rotating deformable object. We propose a modification
of these descriptors that can be used in the presence of rotation while still
maintaining their robust convergence properties. Original gradient-based
descriptors of a pixel are computed in the global image coordinates. When the
surface is deformed or rotated, its descriptors also change. To be able to track
a rotating surface, the descriptors of a pixel must be somewhat rotation
invariant, and thus must be computed in its local image patch coordinates. If the
surface is rotated, the local image patch coordinates of a pixel must be rotated
accordingly. In the tracking context, the surface shape in the current image is
very similar to the one in the previous image. Therefore, we can use the
reconstruction in the previous frame to correct the orientation for pixel local
image patch coordinates. Each pixel on the template is associated with its
corresponding mesh facet. When the facet changes its orientation, the local
coordinates of pixel is rotated with the same angle. We compute descriptors of
input image pixels in the global image coordinates, the descriptors for template
image pixels will be rotated with the angle as the change in the orientation of
its corresponding facet. We experimentally show that our proposed method is able
to track a rotating deformable object.
\subsection{Self Occlusions}
\draft{
Our tracking framework can be further extended to handle self-occlusions by
incorporating a z-buffer. Based on the reconstruction from the previous frame or
a motion model, self-occluded template pixels can be detected and excluded from
the image energy computation. We used a z-buffer to densely register images of
the self-occlusion dataset in \cite{GayBellile10} and reconstruct the surface in
the presence of large self-occlusions. The qualitative results are provided in
Fig.~\ref{fig:selfocclusion}. With no additional modifications, our method can
reliably predict the shape of the surface in the areas of self-occlusion.
\begin{figure}
\centering
\includegraphics[height=3.6cm]{fig/selfocclusion/frame_120.pdf} \hfill
\includegraphics[height=3.6cm]{fig/selfocclusion/view.png}
\caption{A z-buffer is incorporated into our method to handle self-occlusions. Left: Image registration is still successful in presence of 50\% self-occlusions. Right: the recovered 3D surface shape.}
\label{fig:selfocclusion}
\end{figure}
}
\section{Old Abstract}
\vspace{-3mm}
\input{introduction}
\input{related}
\input{framework}
\input{experiments}
\input{conclusion}
\clearpage
{\small
\bibliographystyle{ieee}
\subsection{Cushion Case}
First, we provide results on a cloth surface undergoing a different type of
deformation than studied in the paper datasets. We created a new dataset along
with ground truth in 3D using a Kinect sensor, as before, to which artificial
occlusions were added, example images and our reconstructions are shown in
Fig.~\ref{fig:tshirt_dataset}. Quantitative results are presented in
Fig.~\ref{fig:tshirt_errors}.
\begin{figure} [b]
\vspace{-0.3cm}
\includegraphics[height=1.52cm]{fig/tshirt/frame_87.pdf}
\includegraphics[height=1.52cm]{fig/tshirt/frame_131.pdf}
\includegraphics[height=1.52cm]{fig/tshirt/frame_176.pdf}
\includegraphics[height=1.52cm]{fig/tshirt/frame_181.pdf}
\caption{Our representative reconstructions on the T-shirt dataset with artificial occlusions added. Rightmost: a tracking failure case when occlusions appear at areas with large deformations.}
\label{fig:tshirt_dataset}
\vspace{-0.2cm}
\end{figure}
\begin{figure}
\vspace{-0.25cm}
\centering
\includegraphics[width=0.48\textwidth,height=3cm]{fig/tshirt/error_graph}
\caption{Reconstruction results on the T-shirt dataset.}
\label{fig:tshirt_errors}
\vspace{-0.25cm}
\end{figure} |
\section{Introduction}
\label{introduccion}
Ratios of equivalent widths or central depths of absorption lines with
different excitation potentials have been widely
used as temperature indicators in different kind of stars including solar-type
\citep[e.g.][]{1991PASP..103..439G,1994PASP..106.1248G,2002A&A...386..286C,2003A&A...411..559K,
2007AN....328..938B,
2007HiA....14..598M,2010A&A...512A..13S,2012MNRAS.426..484D,2014MNRAS.439.1028D,2014arXiv1412.8168D},
giant stars \citep[e.g][]{1989ApJ...347.1021G,2000AN....321..277S},
and supergiants \citep[e.g.][]{2000A&A...358..587K}.
To the best of our knowledge, this technique has however not been extended
to the low-mass stars regime most likely due to the difficulties in analysing
their optical spectrum, mainly covered by molecular
bands (in particular TiO and water) which blend or hide
most of the atomic lines commonly used in the spectral
analysis of solar-type stars.
Furthermore, M dwarfs are intrinsically faint in
the optical.
The accurate determination of stellar parameters for M dwarfs has
proven to be a difficult task.
Regarding stellar metallicity, some photometric calibrations based on optical and near-infrared photometry exist,
a technique pioneered by \cite{2005A&A...442..635B} and updated
by \cite{2009ApJ...699..933J,2010A&A...519A.105S} and more recently by
\citet[][hereafter NE12]{2012A&A...538A..25N},
although they require accurate parallaxes
and magnitudes which are
usually available only for nearby and bright stars.
Another common technique to characterise M dwarf metallicities
is based on the use of spectroscopic indices which measure the
relative strength of the TiO molecular band
with respect to the CaH molecular bands near 7000\AA \space
\citep{2007ApJ...669.1235L,2012AJ....143...67D,2013AJ....145..102L}.
Since the M dwarfs spectral energy
distribution peaks at infrared wavelengths,
some previous works have performed a search for spectral features and indices in
this spectral region.
In particular, \cite{2012ApJ...748...93R}
use the equivalent width of the Na~{\sc i} and Ca~{\sc i} triplet and the
H$_{\rm 2}$O-K2 index in the K band of the spectra.
This methodology has been also applied by
\cite{2012ApJ...747L..38T} and \cite{2013AJ....145...52M} providing calibrations for
H and J/optical spectral bands, respectively.
Large samples of M dwarfs have been characterised by means
of near-infrared indices in the recent works by
\cite{2014AJ....147...20N} and \cite{2014MNRAS.443.2561G}.
On the other hand, spectral synthesis has been tested in several works,
usually on small number of stars focusing mainly on strong atomic lines
or on spectral windows known to be less affected by molecular lines
\citep{2005MNRAS.356..963W,2006ApJ...653L..65B,2007PASP..119...90M,
2012A&A...542A..33O,2014A&A...564A..90R}.
Concerning the effective temperature, very few M dwarfs are bright enough
for a direct measurement of their radii
\citep[e.g.][hereafter BO12]{2012ApJ...757..112B}, a technique
pioneered by \cite{2003A&A...397L...5S}.
The most common technique for determining the effective temperature
of an M dwarf is the comparison
of observed spectra
with models atmosphere \citep[e.g.][]{2013AJ....145..102L,2014MNRAS.443.2561G}.
\citet[][hereafter CA08]{2008MNRAS.389..585C}
provides optical/near-infrared photometric calibrations
based on an extension of the infrared flux method (IRFM)
for FGK dwarfs from \cite{2006MNRAS.373...13C} to M dwarfs.
However, significant systematic differences between temperatures
based on CA08 calibrations and temperatures based on interferometric
radii measurements have been recently noted by \cite{2015arXiv150101635M}.
Despite the intrinsic difficulties in their characterisation, low-mass
stars are nowadays at the centre
in the search for small, rocky planets with the potential capability
of hosting life
\citep[e.g.][]{2013ApJ...767...95D,2013EPJWC..4703006S}
In particular, the radial velocity searches currently ongoing with HARPS
at La Silla,
and in the framework of the {\it Global Architecture of Planetary Systems}
project\footnote{http://www.oact.inaf.it/exoit/EXO-IT/Projects/Entries/2011/12/\\27\_GAPS.html}
\citep[GAPS;][]{2013A&A...554A..28C}
at the Telescopio Nazionale Galileo (TNG) with HARPS-N,
are producing a large quantity
of high resolution and high signal-to-noise ratio spectra.
Exoplanet searches would certainly
benefit from a methodology to determine accurate stellar parameters using
the same spectra that are being used for the radial velocity determinations,
i.e., without the necessity of observing at infrared facilities
(usually from space)
or relying on
atmospheric models.
Following this line of reasoning,
a methodology to characterise M dwarfs from high resolution optical spectra by using
pseudo equivalent widths of features has been presented in a recent work
by \citet[][hereafter NE14]{2014A&A...568A.121N}.
The idea of pseudo equivalent width can be further exploited
in order to calibrate empirical relationships for M dwarfs.
This is the goal of this paper, in which we present a large database of
empirical calibrations of spectral features, ratios of features and combinations
of both with the aim of deriving T$_{\rm eff}$, spectral type, and metallicity,
for early M dwarf stars by using optical
HARPS and HARPS-N spectra (wavelength range 383-693 nm).
Unlike NE14 we take as reference temperature scale the one provided
by BO12 and not CA08. We also apply our method to derive spectral types.
Furthermore, we provide relations for stellar masses, radii, and surface gravities
so these quantities can be obtained without using parallaxes or photometry.
We use our methods to characterise
in an homogeneous and coherent way a sample of the M dwarfs which
are currently being monitored in the HARPS GTO radial velocity programme
\citep{2013A&A...549A.109B}.
Late M stars are excluded from this study since exoplanet search around
these stars is difficult at optical wavelengths.
This paper is organised as follows.
Section~\ref{spectroscopic_data} describes the spectroscopic
data used in this work.
Section~\ref{methodology}
describes our methodology and how empirical calibrations for
the main stellar parameters are built. These calibrations are then
applied to a large sample of stars and results are compared
with other techniques in Section~\ref{comparison}.
The derived temperatures and metallicities are used to build
empirical calibrations for stellar masses, radii, and gravities
in Section~\ref{mass_radii_gravity}.
Our conclusions follow in Section~\ref{summary}.
\section{Spectroscopic data}
\label{spectroscopic_data}
This work makes use of HARPS \citep{2003Msngr.114...20M} and
HARPS-N \citep{2012SPIE.8446E..1VC}
spectra mostly taken
from archive. Specifically, the data is taken from:
{\it i)} The ESO pipeline processed FEROS and HARPS archive
\footnote{http://archive.eso.org/wdb/wdb/eso/repro/form};
{\it ii)} The ESO Science Data products Archive
\footnote{http://archive.eso.org/wdb/wdb/adp/phase3\_spectral/form?phase3\_collection
=HARPS}; and
{\it iii)} The TNG Archive
\footnote{http://ia2.oats.inaf.it/archives/tng}.
The corresponding ESO and TNG programme IDs are
listed in the footnote to the paper title.
In addition to the data from archive,
some HARPS-N spectra have been provided by the GAPS team.
The instrumental setup of HARPS and HARPS-N is almost identical so
data from both spectrographs can be used together.
The spectra cover the range 383-693 nm (HARPS-N), and
378-691 nm (HARPS). Both instruments provide a resolving power
of R$\sim$ 115000. The spectra are provided already reduced
using ESO/HARPS-N standard calibration pipelines.
Typical values of the signal-to-noise ratio are between
50 and 90 (measured at $\sim$ 5500 \space \AA). Wavelengths are on air.
The spectra were corrected from radial
velocity shifts by using the IRAF
\footnote{IRAF is distributed by the National Optical Astronomy Observatory, which
is operated by the Association of Universities for Research in Astronomy, Inc., under
contract with the National Science Foundation.}
task {\it dopcor}. For this purpose, we used the accurate radial velocities (measured by applying
the cross-correlation technique)
which are provided with
the reduced spectra.
\section{Methodology}
\label{methodology}
The optical spectrum of an M dwarf is a forest of lines
and molecular bands usually heavily blended in which identifying
individual lines (of moderate strength) or measure equivalent widths
is a difficult task.
In order to overcome this limitation we follow the idea depicted
by NE14 and instead of considering equivalent
widths or depths of lines, we consider pseudo equivalent
widths (hereafter EWs) of features. A feature can be a line or a blend
of lines. The pseudo equivalent width is defined as the traditional
equivalent width, with the
difference that it is not measured with respect to a continuum
normalised to unit,
but to the value of the flux between the peaks of the feature
at each wavelength:
\begin{equation}
EW = \sum\frac{F_{pp}-F_{\lambda}}{F_{pp}}\Delta \lambda
\end{equation}
\noindent where F$_{\lambda}$ is the stellar flux, F$_{\rm pp}$ denotes the
value of flux between the peaks of
the feature at each integration step, and $\Delta\lambda$
is the spectra wavelength's step.
An estimate of the uncertainty
in the measured EWs is given by:
\begin{equation}
\sigma_{EW} = \frac{\Delta \lambda}{<F_{pp}>}\sigma_{F_{pp}}
\end{equation}
\noindent where $<F_{\rm pp}>$ is the mean value of the flux between the peaks of
the feature, and $\sigma_{F_{pp}}$ its corresponding standard deviation.
Figure~\ref{examples_pseudo_ews_measurements} illustrates how the EWs are measured.
An initial list of 4224 features was built taking as reference the
spectra of the star Gl 49, a M1.5V low-mass dwarf.
Spectroscopic observations of this star
were carried out within the framework of the GAPS project with HARPS-N.
Only the red region of the spectra (5300 - 6900\AA) was
considered since the blue part of the optical spectrum of an M dwarf usually
suffers from
lower signal-to-noise ratio. Regions of the spectra affected by chromospheric
activity and atmospheric absorption were avoided.
Figure~\ref{effect_of_teff_and_metal} shows the effects of effective temperature
and metallicity on the EWs measurement.
In the left panel a portion of the stellar spectra is shown
for stars with similar metallicities but different
T$_{\rm eff}$. The same portion of the spectra is shown
in the right panel, this time for stars of similar
T$_{\rm eff}$ but different metallicities.
A list of calibrators was built for each
of the basic stellar parameters considered in the
present work
(T$_{\rm eff}$, spectral type, and metallicity), following
different criteria
as explained in the next subsections.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.45]{example_pseudo_ews_measurements.ps}
\caption{Example of the measurement of EWs.
The figure shows a portion of the spectrum of the star Gl 49.
The flux between the peaks of the features is measured over the blue lines.
}
\label{examples_pseudo_ews_measurements}
\end{figure}
\begin{figure*}[!htb]
\centering
\begin{minipage}{0.48\linewidth}
\includegraphics[angle=270,scale=0.55]{effect_of_teff_on_ews_measurements.ps}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\includegraphics[angle=270,scale=0.55]{effect_of_metal_on_ews_measurements.ps}
\end{minipage}
\caption{
Left: Portion of the stellar spectrum of several T$_{\rm eff}$ calibrators
(see text for details)
with similar metallicity
(from -0.07 to + 0.15 dex) but different effective temperature:
3850 K (purple), 3701 K (blue), 3646 K (green), 3520 K (orange), and 3176 K (red).
Right: Portion of the stellar spectrum of several metallicity calibrators
(see text for details)
with similar effective
temperature
(3450 - 3550 K) but different metallicity:
+0.06 dex (purple), +0.01 dex (blue), -0.11 dex (green), -0.20 dex (orange), and -0.37 dex (red).
For the sake of clarity, an offset of -0.40 in flux was applied between the spectra.}
\label{effect_of_teff_and_metal}
\end{figure*}
\subsection{Effective temperature}
\begin{figure*}[!htb]
\centering
\begin{minipage}{0.48\linewidth}
\includegraphics[angle=270,scale=0.55]{example_teff_calibrations1_24feb15.ps}
\end{minipage}
\begin{minipage}{0.48\linewidth}
\includegraphics[angle=270,scale=0.55]{example_teff_calibrations2_24feb15.ps}
\end{minipage}
\caption{
Examples of ratios of some features identified to be
sensitive to T$_{\rm eff}$ in early-M dwarfs.
Stars are plotted using different colours according to their metallicity
(using 0.10 dex length bins).
Median uncertainties are shown in the left upper corner of each plot.
The corresponding fits are also plotted.
The features' central wavelengths as well as the rms standard
deviation of the residuals are given in each plot.
}
\label{examples_of_teff_ratios}
\end{figure*}
The accuracy of the T$_{\rm eff}$ derived from temperature sensitive ratios is intimately
tied to the accuracy of the temperature of the stars used as calibrators.
We used as calibrators the sample of early M dwarfs with angular sizes
obtained with long-baseline interferometry to better than
5\% given in BO12.
BO12
list 22 low-mass stars with spectral-types
equal to or cooler than K5V, spanning a range of T$_{\rm eff}$ between
3000 and 4500 K.
HARPS spectra were obtained for seven stars from the ESO archive,
whilst HARPS-N spectra were obtained for
three stars from the TNG archive.
HARPS-N spectra for two further stars have been provided
by the GAPS team.
To these stars, we added two more from the recent work
by \cite{2014MNRAS.438.2413V} who analyze their stars in the same way as BO12
\footnote{The authors list three M dwarfs but we were unable
to obtain a HARPS or HARPS-N spectra of Gl 649.}.
\cite{2013ApJ...779..188M} revised the temperature scale of BO12 by noticing
an issue regarding the determination of the bolometric flux of the stars.
Although the temperature differences are relatively small, we use the
set of updated temperatures. For the two stars taken from
\cite{2014MNRAS.438.2413V}, updated T$_{\rm eff}$ values computed in the
same way as in \cite{2013ApJ...779..188M} are provided in
\cite{2014arXiv1412.2758N}.
The final list of T$_{\rm eff}$ calibrators amounts to 14 stars whose
parameters are listed in Table~\ref{calibration_sample}.
\begin{table}
\centering
\caption{The effective temperature calibration sample.}
\label{calibration_sample}
\begin{tabular}{lcc}
\hline\noalign{\smallskip}
Star & SpT & T$_{\rm eff}$ \\
& & (K) \\
\hline
GJ338A & M0V & 3953 $\pm$ 41 \\
GJ205 & M1.5V & 3850 $\pm$ 22 \\
GJ880 & M1.5V & 3731 $\pm$ 16 \\
GJ176 & M2.5V & 3701 $\pm$ 90 \\
GJ887 & M0.5V & 3695 $\pm$ 35 \\
GJ526 & M1.5V & 3646 $\pm$ 34 \\
GJ15A & M1.5V & 3602 $\pm$ 13 \\
GJ412A & M1V & 3537 $\pm$ 41 \\
GJ436 & M3V & 3520 $\pm$ 66 \\
GJ581 & M2.5V & 3487 $\pm$ 62 \\
GJ725A & M3V & 3417 $\pm$ 17 \\
GJ699 & M4V & 3238 $\pm$ 11 \\
GJ876 & M5V & 3176 $\pm$ 20 \\
GJ725B & M3.5V & 3142 $\pm$ 29 \\
\hline
\end{tabular}
\tablefoot{
Effective temperatures are from \cite{2013ApJ...779..188M}, and spectral types
for BO12. For GJ176 and GJ876, T$_{\rm eff}$ values are from \cite{2014arXiv1412.2758N},
and spectral types from the GJ catalogue \citep{1991adc..rept.....G}.
}
\end{table}
Starting from our initial list of 4224 identified features, the
EW of all features were measured in all calibration stars.
In order to avoid possible dependencies on microturbulence, rotation,
or stellar metallicity, we rejected features with
EW $<$ 20 m\AA \space or EW $>$ 120 m\AA \space in any of the
calibration stars, thus excluding too weak and too strong features.
We also rejected lines with relative errors
($\frac{\sigma_{EW}}{EW}$) larger than 2\%.
For every possible ratio of features a Spearman's
correlation test was computed to check whether the ratio is temperature
sensitive or not.
The ratios were selected with the only condition that the features
are separated by no more than 15\AA.
This limit was set in order to
avoid problems with scattered light correction or continuum normalisation.
All ratios with a probability of correlation by chance
lower than 2\% were considered for further study
\footnote{
Ideally, for a ratio of lines to be temperature-sensitive the excitation
potential of the lines, $\chi$, must differ as much
as possible. This is because the EWs of lines with higher $\chi$
change with T$_{\rm eff}$ faster than those of lines
with lower $\chi$ values \citep{1994PASP..106.1248G}.}.
Following \cite{2003A&A...411..559K} for each considered EW ratio we
fitted the T$_{\rm eff}$-ratio relationship to several functions:
a Hoerl function (T$_{\rm eff}$ = ab$^{\rm r}$$\times$r$^{\rm c}$), a modified
Hoerl function (T$_{\rm eff}$ = ab$^{\rm 1/r}$$\times$r$^{\rm c}$),
a power-law function (T$_{\rm eff}$ = a$\times$r$^{\rm b}$),
an exponential law function T$_{\rm eff}$ = a$\times$b$^{\rm r}$, and a
logarithmic function T$_{\rm eff}$ = a + b$\times$$\ln$(r), where
r = EW$_{\rm 1}$/EW$_{\rm 2}$ is the ratio between the EW of two features.
All fits were performed using a nonlinear least-squares fitting routine in IDL
\citep[MPFIT;][]{2009ASPC..411..251M} taking into account the uncertainties
in T$_{\rm eff}$.
For each calibration we selected the function that produced the
smallest standard deviation, retaining only those calibrations with
a standard deviation smaller than 75 K.
The number of selected
temperature-sensitive ratios amounts to 112.
Given our relatively low number of calibrators we performed an additional
check to ensure that the selected ratios are not correlated with T$_{\rm eff}$
simply by coincidence. We created 1000 series of simulated random effective
temperatures and errors, keeping the media and the standard deviation of the
original data. For each series of simulated data we repeted our analysis
and computed the number of ``suitable'' calibrations.
The results show that in 98\% of the simulations our methodology does not recover any suitable
T$_{\rm eff}$-calibration, whilst only in 0.8\% of the simulations the number
of obtained calibrations is larger than 10.
We therefore conclude that it is very unlikely
that our obtained T$_{\rm eff}$-ratios are correlated with T$_{\rm eff}$
just by chance.
Some examples of the selected
temperature-sensitive ratios are shown in Figure~\ref{examples_of_teff_ratios},
whilst full details regarding the calibrations for the same examples
can be found in Table~\ref{calibrations_data}.
\begin{table}
\centering
\caption{
Coefficients of our feature ratio-temperature relations. Columns (1) to (4) provide information
about the features involved in the ratio (central wavelength and width in \AA),
while columns (5) to (9) show the coefficients of the best-fitting relationships, their functional
form, and the corresponding standard deviation of the T$_{\rm eff}$ calibration.
Only four examples are shown here. The same examples are shown in
Figure~\ref{examples_of_teff_ratios}.
}
\label{calibrations_data}
\begin{scriptsize}
\begin{tabular}{ccccccccc}
\hline\noalign{\smallskip}
$\lambda_{\rm 1}$ & $\Delta\lambda_{\rm 1}$ & $\lambda_{\rm 2}$ & $\Delta\lambda_{\rm 2}$ &
a & b & c & func.$^{\dag}$ & $\sigma$(K) \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) \\
\hline
6435.00 & 0.38 & 6432.01 & 0.34 & 4529.63 & 0.73 & -0.015 & MH & 49 \\
5720.46 & 0.33 & 5718.32 & 0.28 & 4842.67 & 0.66 & 0.89 & H & 53 \\
6226.67 & 0.35 & 6213.44 & 0.29 & 4350.09 & 0.86 & - & E & 54 \\
5708.61 & 0.30 & 5718.32 & 0.28 & 5122.92 & 0.634 & 0.92 & H & 54 \\
\hline
\multicolumn{9}{l}{$^{\dag}$: H: Hoerl; MH: modified Hoerl; PL: Power-law; E: Exponential; L: Logarithmic}\\
\end{tabular}
\end{scriptsize}
\end{table}
\subsection{Spectral Types}
\label{spectraltypes}
A similar approach was followed to derive spectral types.
HARPS or HARPS-N spectra were obtained for
a sample of 33 stars with homogeneously derived spectral
types in \cite{1991ApJS...77..417K} and \cite{1994AJ....108.1437H}.
The sample contains stars with spectral types between K7V and M4.5V.
These stars are listed in Table~\ref{sptype_calibration_sample}.
\begin{table}
\centering
\caption{Spectral Type calibration sample.}
\label{sptype_calibration_sample}
\begin{tabular}{lll|lll}
\hline\noalign{\smallskip}
GJ & SpT & Ref$^{\dag}$ & GJ & SpT & Ref$^{\dag}$ \\
\hline
185 & K7V & h & 408 & M2V & h \\
686 & M0V & h & 250B & M2.5V & k \\
701 & M0V & h & 581 & M2.5V & h \\
846 & M0.5V & k & 352A & M3V & k \\
720A & M0.5V & k & 436 & M3V & k \\
229 & M1V & k & 752A & M3V & k \\
412A & M1V & h & 569A & M3V & k \\
514 & M1V & h & 273 & M3.5V & k \\
570B & M1V & h & 643 & M3.5V & k \\
908 & M1V & h & 734B & M3.5V & k \\
205 & M1.5V & k & 213 & M4V & k \\
15A & M1.5V & h & 402 & M4V & k \\
526 & M1.5V & h & 699 & M4V & k \\
625 & M1.5V & h & 83.1 & M4.5 & k \\
220 & M2V & k & 166C & M4.5 & k \\
382 & M2V & k & 234A & M4.5 & k \\
393 & M2V & h & & & \\
\hline
\end{tabular}
\tablefoot{$^{\dag}$
k: standard from \cite{1991ApJS...77..417K};
h: standard from \cite{1994AJ....108.1437H}.}
\end{table}
Spectral type-sensitive ratios were identified by means of a
Spearman's correlation test.
For each of them a third order polynomial
fit was performed between the numerical spectral-type index (with value 0.0 for
M0; 0.5 for M0.5 and so on) and the ratio. A negative value implies
that the star is a late-K dwarf instead of an M star
being the index value for K7 equal to -1.0
(K7 is the subtype preceding M0).
The third order polynomial
fit was preferred amongst other functional fits since we found it to give
the lowest rms standard deviation.
Our ``final'' selection of spectral type-sensitive ratios includes 82
ratios of features with a standard deviation lower than 0.5 spectral subtypes.
The derived mean numerical spectral types are rounded to the nearest half integer.
A couple of examples are shown in Figure~\ref{examples_of_sptype_ratios}.
Full details for some examples are given in Table~\ref{calibrations_sptype}.
As for T$_{\rm eff}$, we performed 1000 simulations with random
spectral-type values. The results show that we are not able to find
any suitable calibration in any of the simulations when random spectral-types are
used.
\begin{figure}
\centering
\includegraphics[angle=270,scale=0.55]{example_sptype_calibrations_v23feb15.ps}
\caption{
Spectral type as a function of two different spectral-type sensitive ratios.
Stars are plotted using different colours according to their metallicity
(using 0.10 dex length bins).
A third order polynomial fit is shown. The features' central wavelengths as well as the rms standard
deviation on spectral type of the residuals are given in each plot.
}
\label{examples_of_sptype_ratios}
\end{figure}
\begin{table}
\centering
\caption{
Coefficients of our spectral type calibrations. Columns (1) and (2) show the wavelength of the
corresponding features, columns (3) to (6) the coefficients of the fit,
while column (7) gives the standard deviation of the calibration of spectral types.
Only four examples are shown here.}
\label{calibrations_sptype}
\begin{tabular}{ccccccc}
\hline\noalign{\smallskip}
$\lambda_{\rm 1}$ & $\lambda_{\rm 2}$ & a$_{0}$ & a$_{1}$ & a$_{2}$ & a$_{3}$ & $\sigma$ \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) \\
\hline
5468.06 & 5481.92 & -3.88 & 10.79 & -4.25 & 0.48 & 0.49 \\
5627.63 & 5667.53 & 5.11 & -0.75 & -2.18 & 0.50 & 0.41 \\
5467.06 & 5481.92 & -3.80 & 8.18 & -0.56 & -0.77& 0.48 \\
5467.06 & 5512.54 & -5.09 & 10.67 & -4.13 & 0.53 & 0.45 \\
\hline
\end{tabular}
\end{table}
\subsection{Metallicity}
\label{metallicity}
A common approach to find metallicity calibrators for low-mass stars relies
on the search of M dwarfs in common proper motion pairs orbiting around
a solar-type star with accurate spectroscopic metallicity determinations.
The basic assumption is that both stars are coeval and born in the same
protostellar cloud so the
metallicity of the secondary M dwarf is the same as the one of the primary
star \citep[e.g.][]{2005A&A...442..635B}.
Until very recently, only the most nearby and bright M stars
have been searched for planets by means of the Doppler technique.
As a consequence there is a lack of HARPS and HARPS-N spectra for most of the identified
M dwarfs in binary systems around solar-type stars.
To overcome this difficulty we built a list of 47 metallicity calibrators with available HARPS spectra,
known parallaxes, and magnitudes by using the most
recent photometric calibration provided in NE12.
This calibration is in turn based on metallicity determinations from FGK
primaries with an M dwarf secondary.
The sample of metallicity calibrators covers a wide range in metallicity from -0.54 to +0.24 dex
with
typical error bars of the order of
$\lesssim$ 0.05 dex. These stars are listed in Table~\ref{metallicity_calibration_sample}.
We caution that the uncertainties
reported in Table~\ref{metallicity_calibration_sample}
do not take into account the scatter in the NE12 calibration, which is
of the order of $\sim$ 0.17 dex.
The effects of metallicity on the EW of the features are
entangled with the effects of T$_{\rm eff}$,
with T$_{\rm eff}$ as the primary driver of changes
in the EW.
This can be easily seen in the histograms
in Figure~\ref{histograma_partial}. They show the distribution of the Spearman's rank correlation factor
of the EW with the stellar metallicity and with
T$_{\rm eff}$.
The figure shows that a significant fraction of the features, $\sim$ 43\%, shows a high
(Spearman's correlation factor $<$ -0.80)
anticorrelation with T$_{\rm eff}$, while only a relatively small fraction
($\sim$ 3\%) shows a significant positive correlation.
The correlation between stellar metallicity and EWs is generally less
significant.
The distribution of Spearman's coefficients for metallicity
shows a clear peak at +0.25 which drops almost to zero at +0.50, while at negative values it has
a smooth tail down to -0.80.
Effects of metallicity and effective temperature
should therefore be considered simultaneously.
We
searched for empirical relationships for metallicity as a function
of features and ratios of features (i.e., an indicator of temperature)
with the following analytical form:
\begin{equation}
[Fe/H] = (A\times EW) + (B\times r) + C
\end{equation}
\noindent where $EW$ is the EW of a feature
showing a strong-metallicity correlation, $r$ is a temperature-sensitive
ratio of features, and $A$, $B$, $C$ are independent coefficients.
We considered every combination of features and ratios satisfying the condition that
the correlation of $EW$ with metallicity, and the correlation of $r$ with T$_{\rm eff}$ show at least a 98\%
of significance.
Our final selection consists of 696 calibrations with standard
deviation values between 0.07 and $\sim$ 0.10 dex.
We point out that these uncertainties should be considered as
lower limits since
they do not take into account possible
systematic errors in the underlying NE12 calibration.
As before, we performed a series of simulations
using random metallicities and errors. In 84\% of the simulations we do not find
any suitable metallicity calibration, although in 7.5\% of the cases the simulation
finds a large number of calibrations (larger than 348).
Some examples of our obtained metallicity calibrations are
provided in Table~\ref{calibrations_metallicity}.
\begin{table}
\centering
\caption{Metallicity calibration sample. [Fe/H] are computed
from V magnitudes and parallaxes taken from the compilation of \cite{2013A&A...549A.109B}
and 2MASS magnitudes \citep{2003yCat.2246....0C}. The photometric calibration by NE12 is used.
Errors are computed by propagating the uncertainties in the parallaxes and the
photometry.
}
\label{metallicity_calibration_sample}
\begin{tabular}{lr|lr}
\hline\noalign{\smallskip}
Star & [Fe/H] & Star & [Fe/H] \\
\hline
GJ1 & -0.37 $\pm$ 0.04 & GJ551 & 0.13 $\pm$ 0.04 \\
GJ105B & -0.13 $\pm$ 0.03 & GJ555 & 0.10 $\pm$ 0.04 \\
GJ176 & 0.02 $\pm$ 0.04 & GJ569A & 0.02 $\pm$ 0.03 \\
GJ205 & -0.03 $\pm$ 0.19 & GJ588 & 0.05 $\pm$ 0.03 \\
GJ2066 & -0.10 $\pm$ 0.03 & GJ618A & -0.06 $\pm$ 0.04 \\
GJ213 & -0.24 $\pm$ 0.04 & GJ628 & -0.05 $\pm$ 0.03 \\
GJ229 & -0.02 $\pm$ 0.17 & GJ674 & -0.20 $\pm$ 0.03 \\
GJ250B & -0.10 $\pm$ 0.04 & GJ678.1A& -0.14 $\pm$ 0.04 \\
GJ273 & -0.11 $\pm$ 0.03 & GJ680 & -0.05 $\pm$ 0.04 \\
GJ300 & 0.06 $\pm$ 0.03 & GJ682 & 0.10 $\pm$ 0.03 \\
GJ357 & -0.32 $\pm$ 0.03 & GJ686 & -0.30 $\pm$ 0.03 \\
GJ358 & 0.05 $\pm$ 0.03 & GJ693 & -0.29 $\pm$ 0.03 \\
GJ367 & -0.05 $\pm$ 0.04 & GJ701 & -0.19 $\pm$ 0.03 \\
GJ382 & 0.05 $\pm$ 0.03 & GJ752A & 0.01 $\pm$ 0.03 \\
GJ393 & -0.11 $\pm$ 0.04 & GJ832 & -0.17 $\pm$ 0.04 \\
GJ413.1 & -0.11 $\pm$ 0.04 & GJ846 & -0.07 $\pm$ 0.04 \\
GJ438 & -0.51 $\pm$ 0.07 & GJ849 & 0.24 $\pm$ 0.04 \\
GJ447 & -0.26 $\pm$ 0.03 & GJ876 & 0.13 $\pm$ 0.03 \\
GJ465 & -0.54 $\pm$ 0.04 & GJ877 & -0.02 $\pm$ 0.03 \\
GJ479 & 0.05 $\pm$ 0.04 & GJ887 & -0.35 $\pm$ 0.14 \\
GJ514 & -0.11 $\pm$ 0.04 & GJ908 & -0.39 $\pm$ 0.03 \\
GJ526 & -0.16 $\pm$ 0.03 & HIP31293& 0.03 $\pm$ 0.04 \\
GJ536 & -0.15 $\pm$ 0.04 & LTT9759 & 0.14 $\pm$ 0.04 \\
GJ54.1 & -0.47 $\pm$ 0.05 & & \\
\hline
\end{tabular}
\end{table}
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.45]{histograma_partial_correlations_v19febrero15.ps}
\caption{
Spearman's correlation factor distribution
of the EW with the stellar metallicity (blue histogram) and
with the T$_{\rm eff}$ (red histogram). The sample of metallicity calibrators
and the initial list of 4224 features is considered.
Metallicities are computed using NE12 whilst effective temperatures
are obtained with our methodology.}
\label{histograma_partial}
\end{figure}
\begin{table}
\centering
\caption{
Coefficients of our metallicity calibrations. Column (1) shows the wavelength of the
corresponding feature, column (2) the temperature-sensitive ratio, columns (3) to (5)
the coefficients $A$, $B$, and $C$,
while column (6) gives the standard deviation of the calibration.
Only five examples are shown here.}
\label{calibrations_metallicity}
\begin{scriptsize}
\begin{tabular}{cccccc}
\hline\noalign{\smallskip}
$\lambda_{1}$ & $\lambda_{2}$/ $\lambda_{3}$ & $A$ & $B$ & $C$ & $\sigma$ (dex) \\
(1) & (2) & (3) & (4) & (5) & (6) \\
\hline
6785.77 & 6785.38/6799.25 & -0.0354 & -1.876 & 1.753 & 0.07 \\
6785.77 & 6799.25/6785.38 & -0.0349 & 0.227 & 0.416 & 0.07 \\
6785.77 & 6785.38/6790.93 & -0.0359 & -0.930 & 1.633 & 0.07 \\
6785.77 & 6785.38/6788.76 & -0.0375 & -1.163 & 1.581 & 0.07 \\
6785.77 & 6790.93/6785.38 & -0.0356 & 0.263 & 0.613 & 0.07 \\
\hline
\end{tabular}
\end{scriptsize}
\end{table}
\section{Comparison with other methods}
\label{comparison}
Our calibrations have been applied to a sample of 53 M dwarfs
from the HARPS GTO M dwarf sample \citep{2013A&A...549A.109B}
for which HARPS data have been obtained from the ESO Science
Data Products Archive\footnote{archive.eso.org/wdb/wdb/adp/phase3\_spectral/form?phase3\_collection\\=HARPS}.
Only spectra with a median signal-to-noise ratio of at least 25 were considered.
For stars with more than one spectrum available we took the one with the highest signal-to-noise ratio.
No further restrictions were applied.
The sample is composed of nearby (distance $<$ 11 pc),
bright (V $<$ 12, K$_{\rm S}$ $<$ 7), early-type M dwarfs
(spectral types M0V-M4.5V).
Our methods were applied to compute effective temperatures,
stellar metallicities and to derive spectral-types. ``Final'' values
for these parameters are the mean of the individual values from
all the calibrations.
All these quantities are provided in Table~\ref{parameters_table_full},
which is available in the online version of this paper.
Our results are compared with:
{\it i)} A photometric scale, namely CA08 for T$_{\rm eff}$ and
NE12 for [Fe/H];
{\it ii)} The recent work by \citet[][hereafter GA14]{2014MNRAS.443.2561G};
and {\it iii)} The values obtained with the methodology
developed by NE14.
\subsection{Comparison of effective temperatures}
\label{subseccomparteff}
Photometric effective temperatures are derived from
V magnitudes from the compilation of \cite{2013A&A...549A.109B}
and 2MASS \citep{2003yCat.2246....0C} photometry using
the calibration provided by CA08.
Computed errors take into account the propagation of the uncertainties
of the 2MASS magnitudes as well as the accuracy of the CA08 calibrations.
The comparison between the photometrically derived temperatures and our
spectroscopic ones is illustrated in Figure~\ref{comparison_temperatures}.
There is a clear offset between our spectroscopic estimates and the
photometric temperatures, being the latter
cooler than ours (the median
difference $\Delta$T$_{\rm eff}$ = T$^{\rm phot}_{\rm eff}$ - T$^{\rm spec}_{\rm eff}$
is -198 K with a rms standard deviation of 176 K).
Our temperatures can be converted into the CA08 scale by a
linear transformation:
T$_{\rm eff}$[CA08 scale] = (1.29 $\pm$ 0.02)$\times$T$_{\rm eff}$[this work]
+ (-1271 $\pm$ 89) K (dashed grey line).
The reason for this discrepancy is
the choice of \cite{2013ApJ...779..188M} temperatures as our reference temperature scale.
In order to test this, photometric T$_{\rm eff}$ values were computed for our
sample of temperature calibrators (Table~\ref{calibration_sample}) following CA08.
The comparison between \cite{2013ApJ...779..188M} and CA08 temperatures is shown
in Figure~\ref{comparison_boya_casagrande}.
It can be seen from this figure that CA08 temperatures tend to be systematically
lower than those provided by \cite{2013ApJ...779..188M}.
The discrepancy between CA08 values and interferometry-based temperatures
has also been noted in a recent work by \cite{2015arXiv150101635M}.
The difference $\Delta$T$_{\rm eff}$ between CA08 and interferometric-based
temperatures noted by these authors is 160 K (the CA08 temperatures being
cooler) in agreement with our results.
CA08 temperatures are obtained by extending the IRFM
for FGK dwarfs from \cite{2006MNRAS.373...13C} to M dwarfs.
One of the assumptions of the IRFM is that a star can be
approximated as a blackbody for wavelengths beyond $\approx$ 2 $\mu$m.
\cite{2015arXiv150101635M} argue that whilst this assumption is reasonable
for warmer stars, it does not suit M dwarfs, which have significantly
more flux in the near-infrared than predicted by a blackbody.
As a consequence CA08 temperatures tend to be systematically lower, with increasing
disparity at cooler temperatures where stars deviate more from
the blackbody emission (see Figures~\ref{comparison_temperatures} and
\ref{comparison_boya_casagrande}).
\cite{2015arXiv150101635M} also note that the temperature scale of the old
version of the PHOENIX models used in CA08
differs from interferometric-based temperatures.
We have also compared our temperatures with the values
given by NE14.
Since NE14
method is calibrated using the CA08 photometric relationship,
the comparison of our temperatures with NE14 shows results similar to the comparison
with CA08 (Figure~\ref{comparison_temperatures}).
Our sample contains 51 stars in common with the sample of GA14
who determine effective temperatures by comparing low-medium resolution
spectra with the BT-SETTL version of the PHOENIX model atmospheres
\citep{2012RSPTA.370.2765A,2012EAS....57....3A}.
Their procedure was calibrated using the
stars listed in BO12, although with the stellar bolometric fluxes
computed as in \cite{2013ApJ...779..188M}.
As can be seen in Figure~\ref{comparison_temperatures}, GA14 temperatures
tend to be slightly hotter than ours, specially for T$_{\rm eff}$ $>$ 3400 K.
For the coolest dwarfs in this sample, GA14 temperatures depart from ours and tend
to be smaller than ours.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{harps_m_gto_comparison_temperatures_ver20febrero15.ps}
\caption{
T$_{\rm eff}$ values from the literature estimates versus the values obtained in this work.
The upper
panel shows the differences between the temperatures
given in the literature and the values derived in this work.
Median uncertainties in the derived temperatures are also shown.
The symbol $<>$ in the legend represents the median difference.
The black continuous line represents the 1:1 relation whilst
the grey dashed one
represents the best linear fit between our estimates and those obtained using
CA08 relationship (see text in Section~\ref{subseccomparteff}).}
\label{comparison_temperatures}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{teff_calibrators_comparison_boyajian_casagrande_v20febrero.ps}
\caption{
T$_{\rm eff}$ estimates based on
the calibration provided by CA08,
versus values from \cite{2013ApJ...779..188M} and \cite{2014arXiv1412.2758N}.
The upper
panel shows the differences.
The symbol $<>$ in the legend represents the median difference.
The black continuous line represents the 1:1 relation whilst
the grey dashed one
represents the best linear fit.}
\label{comparison_boya_casagrande}
\end{figure}
\subsection{Comparison of metallicities}
\label{subseccomparmetal}
We also compare our metallicities with those reported previously in the literature.
Values for the comparison are taken
from the photometric calibration by NE12;
from GA14
who determine metallicities
following the method of \cite{2013ApJ...779..188M} based on empirical calibrations
between the strength of atomic and molecular spectroscopic features and
stellar metallicity; and from NE14.
The comparison is shown in Figure~\ref{comparison_metallicity}.
The comparison reveals an overall good agreement between our
metallicity estimates and those by NE12, GA14, and NE14.
The median differences with these works are consistent with
zero and the scatter although somewhat large is consistent
within the (also large) error bars.
A linear fit between our metallicities and those obtained
using NE12 relationship provides a slope slightly larger than
one and a small difference in the zero point:
[Fe/H][NE12 scale] = (1.22 $\pm$ 0.04)$\times$[Fe/H][this work]
+ (0.02 $\pm$ 0.01) dex (dashed grey line).
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{harps_m_gto_comparison_metallicities_ver20febrero15.ps}
\caption{
[Fe/H] values from the literature estimates versus the values obtained in this work.
The upper
panel shows the differences between the metallicities
given in the literature and those derived in this work.
Median uncertainties in the derived metallicities are also shown.
The symbol $<>$ in the legend represents the median difference.
The black continuous line represents the 1:1 relation whilst
the grey dashed one
represents the best linear fit between our estimates and those obtained using
NE12 relationship (see text in Section~\ref{subseccomparmetal}).}
\label{comparison_metallicity}
\end{figure}
\subsection{Comparison of spectral types}
We finally compare the spectral types derived by us
with those obtained by using
the automatic procedure of the HAMMER spectral code
\citep[][hereafter CO07]{2007AJ....134.2398C}.
The CO07 code was designed to classify stars in the Sloan
Digital Sky Survey Spectroscopic database, therefore before using it
our spectra were degraded to a resolution R $\sim$
2000 by convolving them with a gaussian profile.
We also caution that roughly half of the spectral-type sensitive
band indices defined in CO07 are outside the HARPS
spectral coverage.
The comparison is shown in Figure~\ref{comparison_sptypes}.
It can be seen that there seems to be no significant differences
between our estimates and those by CO07, with an
overall good agreement
within $\pm$ 1 spectral subtype (dashed lines in
Figure~\ref{comparison_sptypes}).
\cite{2011AJ....141...97W} and \cite{2013AJ....145..102L} found the automatic spectral types given by the Hammer
code
to be about one subtype earlier than the manual classification.
While this effect is not evident in our comparison it can not
be ruled out either.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{harps_m_gto_comparison_sptypes_vers16febrero.ps}
\caption{
M subtype values obtained by using the HAMMER code
versus the values obtained in this work.
The upper
panel shows the differences with
the values given in the literature.
Random values
between $\pm$0.2 have been added to the
Hammer values to help in the comparison.
The symbol $<>$ in the legend represents the median difference.
The black continuous line represents the 1:1 relation whilst
the dashed ones correspond to $\pm$ 1 spectral subtype. }
\label{comparison_sptypes}
\end{figure}
Figure~\ref{sptype_teff} shows our derived effective temperatures as a function
of the spectral type. For comparison data from \citet[][Table~A.5]{1995ApJS..101..117K}
is overplotted (red circles). The median T$_{\rm eff}$-spectral type sequence from
\citet{2013AJ....145..102L} is also shown (green squares).
It can be seen that except for the presence of some outliers,
effective temperatures and spectral types are well correlated following
the expected trend. In other words, our spectral types appear
to be consistent with our temperature scale. Unlike \cite{2013AJ....145..102L},
our data do not suggest the presence of a T$_{\rm eff}$ plateau
in the spectral range M1-M3 although we note that our sample is relatively
small in comparison with the one in \cite{2013AJ....145..102L}.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{harps-gto-M-temperature_sptype_relationship_v20febrero15.ps}
\caption{
Effective temperature as a function of the spectral type.
For clarity
the spectral types are not rounded.
A second order polynomial fit is shown.
Possible outliers are removed by using a 2.5$\sigma$ clipping
procedure.
The coefficients of the fit
as well as the rms standard deviation are given in the plot.
Data from \citet[][Table~A.5]{1995ApJS..101..117K}
is overplotted using red filled circles, whilst green squares
represent the median T$_{\rm eff}$-spectral type sequence from
\cite{2013AJ....145..102L}.}
\label{sptype_teff}
\end{figure}
We conclude that
our metallicities and spectral types agree reasonably well with other
literature estimates.
Regarding effective temperatures,
there is a clear offset between BO12 and CA08 scales, as explained.
In summary, our methodology can be confidently used to characterise
large samples of stars in an homogeneous way.
\section{Empirical relationships for stellar mass, radius, and gravity}
\label{mass_radii_gravity}
We made use of the temperature and metallicity values derived with our
method to search for
empirical relationships with the stellar evolutionary parameters
namely, stellar mass, radius, and surface gravity.
We derived our own mass-radius relationship by combining the
stars with known interferometric radius from BO12 and \cite{2014MNRAS.438.2413V}
with data from low-mass eclipsing binaries provided in the compilation
by \citet[][Table~5]{2014arXiv1408.1758H}. A 3$\sigma$ clipping procedure was
used to remove potential outliers. Our derived calibration is as
follows:
\begin{equation}
\label{eq_masa_radio}
R = 0.0753 + 0.7009\times M + 0.2356 \times M^2
\end{equation}
\noindent where radius and masses are given in solar units and the rms
standard deviation of the calibration is 0.02 R$_{\odot}$.
The radius-mass plane is shown in Figure~\ref{mass_radio}.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{masa_radio_relationship_v19feb15.ps}
\caption{
Stellar radius as a function of the stellar mass. The sample includes
stars with interferometric measurements of their radii from
BO12 and \cite{2014MNRAS.438.2413V} as well as low-mass eclipsing
binaries from \citet[][Table~5]{2014arXiv1408.1758H}.
The best fit is also shown. A 3$\sigma$ clipping procedure was
used to remove outliers. The upper panel shows the differences
between the radius derived with our fit and the radius given
in the literature. Median errors in radius (not shown in the plot)
are of the order of 0.006 R$_{\odot}$.}
\label{mass_radio}
\end{figure}
Values of stellar
masses were obtained for each of our target stars
following the relations based on near infrared
photometry by \cite{1993AJ....106..773H}.
We chose this calibration since it is the same used
by BO12.
These calibrations are provided in the CIT photometric system therefore,
before applying them, 2MASS magnitudes
were converted into CIT magnitudes following the transformations
provided by \cite{2001AJ....121.2851C}.
Once the stellar masses were computed, values of the radius
were derived using Equation~\ref{eq_masa_radio}. Finally,
surface gravities, $\log g$, were derived from masses and
radii.
\begin{figure*}[!htb]
\centering
\begin{minipage}{0.47\linewidth}
\includegraphics[angle=270,scale=0.45]{stellar_mass_as_teff_metal_harps_m_gto_v26feb15.ps}
\end{minipage}
\begin{minipage}{0.47\linewidth}
\includegraphics[angle=270,scale=0.45]{stellar_radii_as_teff_metal_harps_m_gto_v26feb15.ps}
\end{minipage}
\caption{
Stellar mass (left panel), and radius (right panel),
as a function of the effective temperature. Stars are plotted
using different colours and symbols according to their metallicity. Several fits for fixed
metallicity values are plotted: +0.15 (dashed line), +0.00 (solid line),
-0.15 (dash-dotted line), and -0.30 (dotted line).
The upper left panel shows the differences between the mass
obtained from Equation~\ref{eqn:m} and those derived by using
\cite{1993AJ....106..773H} relationship.
The upper right panel shows the differences between the radius
derived from Equation~\ref{eqn:r} and by using
Equation~\ref{eq_masa_radio}.}
\label{mrg_teff_metal_plot}
\end{figure*}
We first investigated the correlation of M$_{\star}$, R$_{\star}$, and
$\log g$ with the effective temperature and with the stellar metallicity by using the Spearman
correlation test. Results are given in Table~\ref{mrg_teff_metal}.
It can be seen that although the main dependence of the evolutionary parameters
is on the effective temperature, they also show a moderate but significant
dependence on the stellar metallicity.
We also evaluated the significance of the correlations by a bootstrapp
Monte Carlo (MC) test plus a gaussian, random shift of each
data-point within its error bars. For each pair of variables 10000 random
datasets were created, determining the coefficient of
correlation, $\rho$, and its corresponding z-score each time.
The tests were done using the code
{\sc MCSpearman}\footnote{https://github.com/PACurran/MCSpearman/}
by \citet{2014arXiv1411.3816C} which might be consulted for
further details on this method.
Mean values
of $\rho$ and z-score are
given in Table~\ref{mrg_teff_metal}.
We note that the results from the MC method do not exclude a metallicity dependence but suggest
that any possible correlation is relatively weak.
\begin{table}
\centering
\caption{Results from the Spearman's correlation test and the MC method showing the dependence of the
evolutionary parameters on the effective temperature, and on the stellar metallicity.}
\label{mrg_teff_metal}
\begin{tabular}{lcccc}
\hline\noalign{\smallskip}
& \multicolumn{4}{c}{T$_{\rm eff}$} \\
& \multicolumn{2}{c}{Spearman's test} & \multicolumn{2}{c}{MC method} \\
& \multicolumn{2}{c}{\hrulefill} & \multicolumn{2}{c}{\hrulefill} \\
& $\rho$ & prob. & $\rho$ & z-score \\
\hline
Mass & 0.72 & 2*10$^{\rm -9}$ & 0.47 & 1.55 \\
Radius & 0.72 & 2*10$^{\rm -9}$ & 0.47 & 1.55 \\
$\log g$ & -0.73 & 2*10$^{\rm -9}$ & -0.30 & -0.93 \\
\hline
& \multicolumn{4}{c}{[Fe/H]} \\
& \multicolumn{2}{c}{Spearman's test} & \multicolumn{2}{c}{MC method} \\
& \multicolumn{2}{c}{\hrulefill} & \multicolumn{2}{c}{\hrulefill} \\
& $\rho$ & prob. & $\rho$ & z-score \\
\hline
Mass & 0.34 & 0.02 & 0.20 & 0.60 \\
Radius & 0.34 & 0.02 & 0.20 & 0.60 \\
$\log g$ & -0.33 & 0.02 & -0.12 & -0.40 \\
\hline
\end{tabular}
\end{table}
A dependence of the radius on stellar metallicity
is expected from model predictions
\citep{1998A&A...337..403B,2008ApJS..178...89D} however
BO12 find the interferometry-based radius rather insensitive to metallicity.
Furthermore, while fitting
M$_{\star}$, R$_{\star}$, and L$_{\star}$ as a function
of the effective temperature \cite{2013ApJ...779..188M}
find that adding the stellar
metallicity as a parameter does not improve the fits.
However, in a more recent work, \cite{2015arXiv150101635M}
do find a significant effect of the metallicity on
the T$_{\rm eff}$-R$_{\star}$ relation. The authors
point towards small sample sizes and a sparser sampling
on [Fe/H] as the reasons why the effect of [Fe/H]
was not noticed in their previous studies.
We therefore performed two kinds of fit, one using only
the effective temperature and another one adding the
stellar metallicity as a parameter.
The extra sum-of-squares F test \citep[e.g.][]{1993stp..book.....L}
was used to test whether the addition of the metallicity
to the functional form of the calibrations provides
any improvement or not.
The test returns
a measure of the likelihood ($p$-value)
that the simpler model (the one with fewer parameters)
provides a better representation than the more complicated one.
The resulting values\footnote{The tests were performed using
the {\sc MPFTEST} IDL routine
included in the {\sc MPFIT} package \citep[][]{2009ASPC..411..251M}
and available at http://cow.physics.wisc.edu/~craigm/idl/idl.html},
$p$-M$_{\star}$ $\sim$ 2$\times$10$^{\rm -7}$,
$p$-R$_{\star}$ $\sim$ 7$\times$10$^{\rm -8}$,
$p$-$\log g$ $\sim$ 2$\times$10$^{\rm -7}$,
indicate that by including the metallicity there is improvement in the fits
in line with \cite{2015arXiv150101635M}.
The relationships we obtain are the following:
\begin{eqnarray}
\label{eqn:m} \nonumber M_{\star} (M_{\odot}) = -171.616 + 0.139\times T_{\rm eff} - 3.776\times10^{\rm -5}T_{\rm eff}^{\rm 2}\\+ 3.419\times10^{\rm -9}T_{\rm eff}^{\rm 3} + 0.382 \times [Fe/H] \\
\label{eqn:r} \nonumber R_{\star} (R_{\odot}) = -159.857 + 0.130\times T_{\rm eff} - 3.534\times10^{\rm -5}T_{\rm eff}^{\rm 2}\\+ 3.208\times10^{\rm -9}T_{\rm eff}^{\rm 3} + 0.347 \times [Fe/H] \\
\label{eqn:g} \nonumber \log g (cgs) = 174.462 -0.138\times T_{\rm eff} + 3.728\times10^{\rm -5}T_{\rm eff}^{\rm 2}\\- 3.376\times10^{\rm -9}T_{\rm eff}^{\rm 3} - 0.332 \times [Fe/H]
\end{eqnarray}
\noindent where the rms standard deviations of the calibrations are $\sigma_{\rm M\star}$ = 0.02 M$_{\odot}$,
$\sigma_{\rm R\star}$ = 0.02 R$_{\odot}$ , and $\sigma_{\log g}$ = 0.02 (cgs).
The calibrations are valid for 3340 K $<$ T$_{\rm eff}$ $<$ 3840 K,
and -0.40 $<$ [Fe/H] $<$ +0.16 dex.
Empirical relationships for stellar luminosity are not provided since they
can be easily obtained from T$_{\rm eff}$ and R$_{\star}$ just applying
Stefan-Boltzmann's law.
All these quantities (M$_{\star}$, R$_{\star}$, $\log g$, and $\log (L_{\star}/L_{\odot})$ )
for the stars analyzed in this work are provided in Table~\ref{parameters_table_full}.
Typical uncertainties are in the order of 13.1\% for the stellar mass,
11.8\% for the radius, 25\% for
luminosities, and 0.05 dex for $\log g$.
We note that these uncertainties are computed by taking into account the
$\sigma$ of the corresponding calibration and the propagation of the
errors in T$_{\rm eff}$ and [Fe/H].
A word of caution should be given regarding the relative errors in mass
since they tend to increase towards lower masses.
Relative errors in mass might be larger than 20\% for stars with
M$_{\star}$ $<$ 0.35 M$_{\odot}$ and reach up to more than 40\% for
the few stars with M$_{\star}$ $<$ 0.25 M$_{\odot}$. In a similar
way, relative errors in radius can be larger than 20\%
for stars with R$_{\star}$ $<$ 0.35 R$_{\odot}$.
Relative errors in luminosities are also larger for low-luminosity
stars, being significantly high (larger than 70\%) for those
stars with $\log (L_{\star}/L_{\odot})$ $<$ -2.
We point as a possible explanation the fact that relative errors
in masses obtained from \cite{1993AJ....106..773H} relationship
tend to be larger at lower masses.
The M$_{\star}$ and R$_{\star}$
versus temperature planes are shown in
Figure~\ref{mrg_teff_metal_plot} where the stars are plotted with different colours
according to their metallicities.
It can be seen that for a given effective temperature, larger
stellar metallicities predict larger masses and radii.
Regarding surface gravity, see Figure~\ref{logg_teff_metal_plot},
the effect of metallicity tends to be the opposite
with lower gravities in stars with higher metallicity content.
\begin{figure}[!htb]
\centering
\includegraphics[angle=270,scale=0.50]{stellar_gravity_as_teff_metal_harps_m_gto_v26feb15.ps}
\caption{
Logarithmic surface
gravity
as a function of the effective temperature. Stars are plotted
using different colours and symbols according to their metallicity. Several fits for fixed
metallicity values are plotted: +0.15 (dashed line), +0.00 (solid line),
-0.15 (dash-dotted line), and -0.30 (dotted line).
The upper panel shows the differences between the $\log g$ obtained with our calibrations
(Equation~\ref{eqn:g}) and those derived from masses and radius.
Median errors in $\log g$ estimates
are 0.18 dex (mass-radius derived values)
and 0.05 dex (values from Equation~\ref{eqn:g}).}
\label{logg_teff_metal_plot}
\end{figure}
\section{Summary}
\label{summary}
The determination of accurate stellar parameters of low-mass stars is certainly
a major topic in nowadays astrophysics.
This is in part because of their advantages
with respect to solar-type stars in the search for small, rocky,
potentially habitable planets.
This fact motivated us to develop a methodology to
characterise early M dwarfs using the high-resolution spectra that are being
obtained in the current radial velocity exoplanet programmes.
We made use of ratios of features as a method to determine effective temperatures,
and combinations of features and temperature-sensitive ratios to determine metallicities.
This technique largely applied to solar-type, subgiant, and giant stars had not been extended before
to the low-mass stars regime, probably because of the difficulty in identifying spectral lines in their
optical spectra. We also provide empirical calibrations for masses, radii, and gravities
as a function of effective temperature and metallicity.
Our main results are as follows:
\begin{itemize}
\item The behaviour of the EW of features
was studied as a function of the effective temperature and the metallicity.
The results show that for a significant fraction of the features,
$\sim$ 50\%, the EW shows a high anticorrelation with T$_{\rm eff}$, whilst
the correlations between EW and metallicity are in general weaker.
\item Empirical calibrations for the effective temperature were obtained
using stars with interferometric measurement of their radii from BO12
as calibrators. 112 ratios of features sensitive to the temperature were calibrated
providing effective temperatures with
typical uncertainties of the order of 70 K.
\item In the same way 82 ratios of features were calibrated to derive spectral types.
\item Stellar metallicities were obtained from 696 combinations of EW of
individual features
and temperature-sensitive ratios, with estimated uncertainties in the range of
0.07-0.10 dex.
\item We made use of our technique to characterise 53 early M dwarfs which are currently
being monitored in the HARPS exoplanet search programme.
Photometric estimates
of stellar mass, radius, and surface gravity were used
to search for possible correlations of these parameters on T$_{\rm eff}$ and
[Fe/H].
\item We found stellar masses, radii, and surface gravities
to have a moderate but statistically significant
correlation with the stellar metallicity, in the sense that at a given effective temperature
larger metallicities predict slightly larger masses and radii
whereas, larger gravities are found in stars with
lower metallicity content.
\end{itemize}
Although high-resolution HARPS and HARPS-N optical spectra were used for this work, a similar
methodology can be used to derive T$_{\rm eff}$ and [Fe/H] for other instruments/spectral
ranges.
\begin{acknowledgements}
This work was supported by the Italian Ministry of Education,
University, and Research through the
\emph{Premiale HARPS-N} research project under grant
\emph{Ricerca di pianeti intorno a stelle di piccola massa}.
M. P. and I. R. acknowledge financial support from the Spanish Ministry
of Economy and Competitiveness (MINECO) and the \emph{Fondo Europeo de
Desarrollo Regional} (FEDER) through grants AYA2012-39612-C03-01 and
ESP2013-48391-C4-1-R. We sincerely appreciate the careful
reading of the manuscript and the constructive comments of an anonymous
referee.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction} \label{intro}
Protoplanetary disk observations provide snapshots of the planetary formation processes that help us understand the physical characteristics of such disks. However, understanding planetary formation requires modeling the evolution and composition of protoplanetary disks. Considering the different physical states of the various components of a protoplanetary disk is therefore necessary to constrain favorable scenarios for planetary formation, growth, and migration, and solve the apparent inconsistency between the formation and migration timescales.
Since the discovery of the first exoplanet \citep{mayor95}, planetary formation scenarios have frequently been revisited. Even before \citet{charbonneau00} showed the gaseous giant nature of the observed exoplanets, \citet{pollack96} described how gaseous planets could form by accretion of gas on previously accumulated solid cores of a few times the mass of the Earth and estimated a typical planetary formation timescale around $10^{6-7}$ years (compatible with the typical disk lifetime of a few million years as inferred from observations by \citet{beckwith96} and \citet{hartmann98}). \citet{papaloizou99} also noted that hot Jupiters were unlikely to have formed in situ, therefore requiring some sort of planetary migration. Type I inward migration due to Lindblad resonances with the planet are well known from \citet{goldreich79}, \citet{arty93}, and \citet{ward97}, who estimated a migration timescale of about $10^{5}$ years for a typical Earth-mass planet in a minimum mass solar nebula \citep{weiden77, hayashi81}. \citet{kory93} noted that planets should therefore be lost by spiraling into their host star before they could actually grow. Early on, \citet{ward91} detailed the analytical expression of the corotation torque, before \citet{tanaka02} added a 3D analytical expression of the vortensity gradient horseshoe drag torque (tested numerically by \citet{bate03} and \citet{dangelo03}), and \citet{baruteau08} later completed this with the entropy gradient horseshoe drag torque. In the meantime, \citet{alibert05} and \citet{ida08} investigated the inconsistency between the timescales of planet formation and migration and noted that the mass-distance distribution of the exoplanets is inconsistent with the rapid inward migration of planetary cores. They noted that the type I inward migration should be slowed down by at least an order of magnitude to allow planets to form and avoid falling into the star. Part of the solution, investigated by \citet{hellary12}, would consist in accelerating the planet formation by considering a proper temperature treatment and a 3D disk model. However, this is not sufficient, and most of the efforts have been concentrated on slowing down the inward migration. Various models have been tested to achieve that goal. One of them is the model reported by \citet{terquem03},
who studied the effects of a regular magnetic field, while \citet{nelson04} focused on turbulent magnetism. \citet{kuchner02}, on the other hand, studied the consequences of an inner cavity and determined that such a hole could halt inward migration. \citet{jc05} found that self-shadowing in the disk can slightly decrease the migration rate. \citet{masset06b} described that a sharp positive surface mass density gradient could have similar consequences for the creation of planetary traps at the zero-torque radii. Subsequent efforts from \citet{paarm06,paarm08}, \citet{paarp08}, \citet{baruteau08}, \citet{kley08}, \citet{kley09}, and \citet{ayliffe10} investigated the possibility of slowing down and reversing the inward migration by considering more complete models involving proper radiative transfer and thermal consideration: this appeared to be possible for sufficiently low mass planets ($M_{P} \, <\, 40\, M_{\oplus}$). \citet{menou04} studied the effects of opacity transitions on the migration rate and showed that if some specific conditions are met, the migration could be stopped in a steady-state $\alpha$ disk. \citet{paarp09a} estimated that almost any positive surface mass density gradient could act as a protoplanet trap. Zero-torque radii were later analyzed by \citet{lyra10}, \citet{bitsch11}, and \citet{hasegawa112}: they appear to be very important in preventing the fall of the planets into their star, but also in creating a favorable zone for the interactions between planetesimals where they are able to combine. \citet{horn12} estimated that giant planet cores could form at convergent zero-torque radius for sub-$M_{\oplus}$ planets in 2-3 Myr.
There appears to be a strong correlation between the migration rate and the surface mass density and mid-plane temperature gradients that also strongly depend on the disk composition (gas-to-dust ratio, chemical abundances). Although the influence of the dust composition is not commonly used in recent numerical simulations of evolving protoplanetary disks, \citet{helling00} and \citet{semenov03} studied how opacities are affected by temperature. It therefore appears to be necessary to consider how the dust main component phases change in order to estimate the temperature more accurately. From reliable evolved disk radial profiles, we skip the planetary core formation stage and consider how a formed core dynamically interacts with the disk. The resonant torques that a planet exerts on a disk can be calculated from \citet{goldreich79}, \citet{ward88}, and \citet{arty93}, and we considered the refinements from \citet{paar11} to more accurately calculate the various contributions of the corotation torques. In their simulations, \citet{bitsch11} found a possible equilibrium radius of a planet with 20 Earth masses around 12.5 AU. \citet{hasegawa112} also analyzed the influence of the heat transition barrier on the migration torques. Whereas most of the previous work has applied the torque formulae to simplified semi-analytical density and temperature profiles \citep{hasegawa111,paar11}, to simple density prescriptions with a self-consistent 2D-temperature structure \citep{bitsch11,bitsch13}, or to steady-state accretion disk models \citep{bitsch14}, we intend to apply similar reasoning to more realistic disks obtained from numerically simulated viscous evolution rather than analytical steady state disks.
\citet{baillie14} (referred to hereafter as \citetalias{baillie14}) have shown that some of the features of viscous $\alpha$-model protoplanetary disks that are observed around forming stars can be retrieved numerically using a viscous evolution hydrodynamical code. These simulations confirmed the importance of jointly considering the dynamics, thermodynamics, and geometry of the disk. However, as the temperature affects the gas-to-dust ratio of the main components of the disk, the opacity of the disk is affected as well. Therefore, it is important to take into account a consistent composition of the disk when calculating its temperature. We here improve the numerical model of BC14 to consider these changes in the phases of the disk components. The thermal model includes both viscous heating and irradiation heating. We follow the evolution of an initial minimum mass solar nebula. The geometry (including the delimitations of shadowed regions) is calculated self-consistently, with the thermal structure obtained by semi-analytical radiative transfer calculations. The obtained density and temperature radial profiles show discontinuities compared to previous results from \citetalias{baillie14}. Density bumps mainly result in temperature irregularities such as bumps, troughs, and plateaux. The snow or sublimation lines are also enlarged. Using our density and temperature profiles, we compute the torque that the disk would exert on a planet embedded in the disk. The total torques (resulting from the Lindblad and corotation resonances) provide the direction of migration of the planetary core within the disk, allowing us to identify convergence and divergence regions, which are also called planetary traps and deserts
Section \ref{methods} details how the hydrodynamical code of \citet{baillie14} is upgraded to consider variations of the dust composition with temperature. The density radial profiles resulting from the viscous evolution and the calculated thermodynamical and geometrical profiles of the disk are shown in Sect. \ref{res}. These radial profiles, and especially the mid-plane temperature, are analyzed in Sect. \ref{disc}, where we also discuss the influence of the composition of the disk and reconsider the definitions of snowline and sublimation line. Finally, Sect. \ref{torque} calculates the resonant torques of potential planetary embryos in the disk and follows their migration, defining planetary traps and deserts.
\section{Methods} \label{methods}
\subsection{Disk model}
We consider the same model of a viscous $\alpha$ disk \citep{shakura73} as was used in \citetalias{baillie14} and use the same terminology as in that paper. The turbulent viscosity is set to $\alpha_{\mathrm{visc}}= 10^{-2}$ as is commonly accepted for T Tauri star protoplanetary disks without deadzones \citep{fromang06}. The time evolution of the surface mass density is given by Eq. \ref{lb74} from \citet{lyndenbellpringle74},
\begin{equation}
\frac{\partial \Sigma(r,t)}{\partial t} = \frac{3}{r} \, \frac{\partial}{\partial r}\left(\sqrt{r} \, \frac{\partial}{\partial r} \left( \nu(r,t) \, \Sigma(r,t) \, \sqrt{r}\right) \right)
\label{lb74}
.\end{equation}
Similarly to BC14, we applied Eq. \ref{lb74} to a one-dimensional grid of masses that are logarithmically distributed in radius between $R_{*}$ and 1000 AU: each of these masses represents an annulus in the disk. We imposed that the flux at the inner edge cannot be directed outward. The inner mass flux gives the mass accretion rate of the disk. The temperature in the mid-plane, $T_{m}(r)$, results from the combination of viscous heating, stellar irradiation heating, and radiative cooling in the mid-plane.
The grazing angle $\alpha_{gr}(r)$ defines the angle at which the star sees the photosphere at a given radial location $r$. Comparisons between an imposed geometry following that prescription and a free geometry calculated along with a consistent temperature are shown in \citetalias{baillie14}, along with a discussion of the necessity of these geometric refinements. The grazing angle, which controls the amount of energy that the star provides to the disk photosphere, is related to the photosphere height $H_{ph}(r)$ through Eq. \ref{alphagr}:
\begin{equation}
\alpha_{gr}(r) = \arctan\left(\frac{dH_{ph}}{dr}(r)\right) - \arctan\left(\frac{H_{ph}(r)-0.4 R_{*}}{r}\right)
\label{alphagr}
.\end{equation}
A positive grazing angle at a given radius results in an irradiated photosphere at that location, while regions that are not irradiated are shadowed by inner regions. Using Eq. 18 from \citet{calvet91}, we calculated the temperature in the mid-plane, accounting for viscous heating, stellar irradiation, and radiative cooling. The viscous contribution depends on both the surface mass density obtained after temporal evolution and on the mid-plane temperature itself (through the viscosity):
\begin{equation}
F_{v}(r) = \frac{1}{2} \Sigma(r) \nu(r) \left( R \frac{\mathrm{d}\Omega}{\mathrm{d}r} \right)^{2} = \frac{9}{4} \Sigma(r) \nu(r) \Omega^{2}(r)
.\end{equation}
Therefore, we solved Eq. 18 from \citet{calvet91} numerically as an implicit equation on the mid-plane temperature. Considering a hydrostatic equilibrium, the vertical density distribution follows a Gaussian, and we can use Eq. A9 from \citet{dullemond01} to calculate the ratio $\chi$ of the photosphere height to the pressure scale height. The disk vertical density profile is therefore assumed to be the same as an isothermal vertical structure at the mid-plane temperature. This approximation is reasonable below a few pressure scale heights, where most of the disk mass is located. The opacities are also functions of the temperature, as detailed in Sect. \ref{realopa}. We can then estimate the corresponding presumed photosphere height $H_{ph}$ at each radial location and therefore access $\frac{dH_{ph}}{dr}$. Applying Eq. \ref{alphagr}, we can verify whether the presumed grazing angle has the required precision or if we should iterate on it. The impossibility of solving that problem for any positive value of the grazing angle results in a disk column that is not directly irradiated by the star, and therefore we removed the irradiation heating term from the mid-plane temperature equation.
Thus, the geometrical structure (photosphere and pressure heights) was determined jointly with the temperature by iterating numerically on the grazing angle value: the algorithm is thoroughly described in \citetalias{baillie14}.
\subsection{Realistic opacities}
\label{realopa}
While \citetalias{baillie14} did take into account both sources of heating, they only partially considered the thermal variations of the physical composition of the disk. Indeed, \citetalias{baillie14} considered that the dust grains invariably have the same opacity for any temperature below 1500 Kelvin (considered here to be the sublimation temperature of the silicate dust); and another constant opacity (a hundred times lower) for temperatures above 1500 K. In the present paper, we refine the model by considering a more elaborate model of opacities based on optical constants measured in laboratory experiments, accounting for the variations of dust composition as a function of the local temperature. The computation of the Rosseland mean opacities follows the procedure described by \citet{helling00} and \citet{semenov03}. We assumed that the dust grains are composed of a mixture of different elements: olivine silicate, iron, pyroxene, troilite, refractory and volatile organics, and water ice, with initial abundances (before volatiles sublimation) given in Table \ref{tempchgt}. The relative abundances of the opacity model are updated depending on the temperature of the medium and the sublimation temperatures (Table \ref{tempchgt}). We assumed for the dust grains a modified MRN size distribution \citep{pollack85, helling00,semenov03}, that is, grain sizes varying from 0.005 to 5 $\mu$m with a $-3.5$ power-law exponent size distribution. The absorption coefficients of the composite dust grains were computed using the Maxwell-Garnett mixing rule followed by the Mie theory. The tabulated values of the gas opacity computation are taken from \citet{helling00}.
Figure \ref{opa} presents the Rosseland ($\chi_{R}$) and Planck ($\kappa_{P}$) opacity variations with temperature. The Planck mean opacities at stellar effective temperature $T_{*} = 4000$ K in extinction ($\chi_{P}^{*}$) and absorption ($\kappa_{P}^{*}$) are also shown. These opacities are necessary in particular for calculating the irradiation heating as described in \citet{calvet91}, \citet{jc04}, and \citet{jc08}.
It appears that these opacities vary by several orders of magnitude over the concerned temperature range and that these variations are particularly steep near the sublimation temperature of the major disk gas components. These elements and their sublimation temperatures are presented in Table \ref{tempchgt}; sublimation temperatures are given by \citet{pollack94}, corresponding to gas densities of about $10^{-10} \, \mathrm{g.cm^{-3}}$.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{opacity_table.eps}
\caption{Mean-opacity variations with local temperature. Black: Rosseland mean opacity in extinction. Red: Planck mean opacity in absorption. Yellow: Planck mean opacity in extinction at stellar irradiation temperature. Blue: Planck mean opacity in absorption at stellar irradiation temperature.}
\label{opa}
\end{figure}
\begin{table}
\begin{center}
$\begin{array}{c|c|c}
\mathrm{Elements} & \mathrm{Sublimation} & \mathrm{Relative}\\
& \mathrm{Temperature} & \mathrm{Abundances}\\
\hline
\mathrm{Water\, ice} & \mathrm{160\, K} & 59.46 \, \% \\
\mathrm{Volatile\, Organics} & \mathrm{275\, K} & 5.93 \, \% \\
\mathrm{Refractory\, Organics} & \mathrm{425\, K} & 23.20 \, \% \\
\mathrm{Troilite\, (FeS)} & \mathrm{680\, K} & 1.57 \, \% \\
\mathrm{Olivine} & \mathrm{1500\, K} & 7.46 \, \% \\
\mathrm{Pyroxene} & \mathrm{1500\, K} & 2.23 \, \% \\
\mathrm{Iron} & \mathrm{1500\, K} & 0.16 \, \% \\
\end{array}$
\end{center}
\caption{Sublimation temperatures and relative abundances that affect the disk gas opacity.}
\label{tempchgt}
\end{table}
\subsection{Testing the radial resolution}
Our evolution code requires a double check both on the temporal and on the spatial resolution that the structure of our code (density time evolution precedes geometry and temperature calculation and so on for each iteration) allows us to treat sequentially. First, the timestep is adjusted to limit the mass transfers to 1\% of the available mass in each bin of the simulation, and then the radial resolution can be controlled by ensuring that the surface mass density profiles are consistent over 1 million years for different radial resolutions. Figure \ref{rescomp} shows the radial density profiles between 1 and 100 AU after 100,000 years and 1 million years for a range of radial resolutions extending from 5 points per decade to 50 points per decade. Stronger differences around 1 AU are due to the boundary conditions of the test numerical simulations for which the inner edge of the disk was considered to be at 1 AU for the purpose of the comparison: as the computation time increases exponentially, simulations using the highest number of points per decade could only be run over a few decades. We note that the surface mass density profiles are very similar until 100,000 years, after which the differences are more obvious. However, the density profiles remain quite close and the only consequence of the difference seems to be a delay in the viscous evolution for the simulations with 30 and more points per decade.
\begin{figure}
\begin{center} $
\begin{array}{c}
\includegraphics[width=8cm, clip=true]{yr_100000.eps}\\
\includegraphics[width=8cm, clip=true]{yr_1000000.eps}\\
\end{array} $
\end{center}
\caption{Compared evolution of the surface-mass density after 100,000 years (upper panel) and 1 million years of evolution (lower panel) of an initial MMSN for various resolutions from 5 to 50 points per decade.}
\label{rescomp}
\end{figure}
Therefore, we estimate that a disk evolution with a good radial resolution (e.g., 40 points per decade as in the rest of the paper) can be approximated from interpolating the density radial profile of a disk generated by a less radially resolved simulation (5 points per decade) for an accessible computation time. The thermal and geometric structure can then be calculated based on the resampled density structure of the disk.
Assuming a temperature profile following a power law in $T \propto r^{-0.5}$ and a temperature resolution of 0.01 K, we can analytically estimate the highest possible radial resolution for two consecutive bins to have temperatures that differ by at least that temperature precision: this optimal resolution is around 100 points per decade. To keep a safety margin (in particular to allow for the temperature fluctuations due to the opacities), we chose to resample our disks with 40 points per decade, which provides a good radial precision and allows a good thermal and geometrical precision.
\section{Results} \label{res}
We followed the evolution of an initial minimum mass solar nebula \citep{weiden77} with the scaling from \citet{hayashi81}:
\begin{equation}
\label{eqmmsn}
\Sigma (r) = 17,000 \left(\frac{r}{1 \, \mathrm{AU}}\right)^{-3/2} \mathrm{kg\cdot m^{-2}}
.\end{equation}
The central star was a classical T Tauri type young star with constant $M_{*} = 1 \, M_{\odot}$, $R_{*} = 3 \, R_{\odot}$, $T_{*} = 4000 \, \mathrm{K}$ and $\mathcal{L}_{*} = 4 \pi \, R_{*}^{2} \, \sigma_{B} \, T_{*}^{4}$ throughout the simulation.
The choice of the minimum mass solar nebula (MMSN) is motivated by several arguments. First, \citetalias{baillie14} investigated a diversity of initial disk conditions (total disk mass and radial distribution of the initial angular momentum) and found that these numerical simulations converged to similar steady states (characterized by a uniform mass flux and an asymptotic surface mass density distribution in $r^{-1}$), although in slightly different timescales. In addition, \citet{vorobyov07} showed that the MMSN density profile (following a power-law in $r^{-1.5}$) was consistent with an intermediate stage of an evolving protoplanetary disk under self-regulated gravitational accretion. Therefore, the MMSN profile makes an initial profile as reasonable as any other snapshot that could have been taken in the disk evolution. Finally, this fiducial case makes sense because we can better compare our results with previous studies.
\subsection{Time evolution}
\label{evol}
Figure \ref{sigma} presents the evolution of such a disk over 10 million years. Although the protoplanetary disk gas is believed to photo-evaporate in a few million years \citep{font04, alexander07, alexander09, owen10}, we let our numerical simulations extend over longer times to reach a stationary steady-state, which is
characterized by a uniform mass flux (within one order of magnitude maximum), as seen in Fig. \ref{fluxb}.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{sigmam_MMSNmultit_1000.eps}
\caption{Surface mass density radial profile evolution for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities. Simulations are resampled to 40 points per decade.}
\label{sigma}
\end{figure}
As \citetalias{baillie14} detailed, the disk starts to spread viscously outward before beginning to accrete onto the central star after a few thousand years. Figure \ref{sigma} shows that the surface mass density radial profile can be modeled as a power
law by segments. While two segments seem to be sufficient until 10,000 years, longer evolution times require at least one more segment. In the early times, the single connexion point between the two power-laws connects an inner region that has already evolved toward a steady-state power-law radial profile for this region, and an outer region that is much closer to the initial state. We therefore call that radial location the relaxation radius. After 10,000 years, the (inner) relaxation radius is located between 7 and 10 AU. After 1 million years, a secondary "knee" appears around 250 AU. In the intermediate region (10-250 AU), the viscous evolution appears weaker than in the simulations
of \citetalias{baillie14} with a simpler opacity model. At 1 Myr, the surface mass density profile can be modeled as a power
law between 10 and 250 AU with $\Sigma(r) \propto r^{-1.2}$, and at 10 Myr, we can approximate the density profile by $\Sigma(r) \propto r^{-1.1}$. In this region, and evolved disks, the power-law indices are comparable with those from \citetalias{baillie14} and with those observed by \citet{isella09} in the Taurus region and \citet{andrews09} and \citet{andrews10} in the Ophiuchus region.
Figure \ref{fluxb} shows the time evolution of the mass flux as a function of the radial distance. We note that these radial profiles present bumps that were not visible in \citetalias{baillie14}. Although the directions and amplitudes remain globally identical, some irregularities appear; for example, in the 7-10 AU region where the 1 Myr flux profile is briefly inverted and the 10 Myr profile is no longer uniform over that range (the flux varies by almost a factor 10).
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{fluxb_MMSNmultit_1000.eps}
\caption{Mass flux radial profile evolution for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities.}
\label{fluxb}
\end{figure}
\subsection{Thermal evolution}
\label{Tevol}
The thermal evolution of the disk is presented in Fig. \ref{Tm}. The mid-plane temperature is calculated in a similar way to \citetalias{baillie14}. However, the calculation of the mid-plane temperature takes into account a composition of the dust that is consistent with the temperature (i.e., where sublimated species have been removed from the opacity calculation according to Table \ref{tempchgt}) while iterating over the possible grazing angles and mid-plane temperatures: the temperature obtained after the algorithm converged therefore validates the consistent composition. As described in Sect. \ref{methods}, the geometric and thermal structure are recalculated over 40 points per decade.
As in \citetalias{baillie14}, the mid-plane temperature presents irregular features that did not appear in the radial profiles obtained with the simpler opacity model of \citetalias{baillie14}. In particular, we note the temperature plateaux at the change of phase temperatures of the dust components (see Table \ref{tempchgt}). As the disk evolves, the sublimation plateaux drift inward: the silicate sublimation plateau is inside 0.2 AU after 1 Myr while it was originally between 0.3 and 1.3 AU. In addition to these plateaux, we note the troughs around 10 AU and 250 AU, which drift inward as the disk evolves. These features are analyzed more thoroughly in Sect. \ref{tempprofile}. However, it appears that the temperature in the region between these features can be modeled as a power law, with $T(r) \propto r^{-0.47}$ beyond 1 Myr, which recalls the usual approximation of \citet{chiang97} or \citet{dullemond01}, according to which $T(r) \propto r^{-1/2}$.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{Tm_MMSNmultit_1000.eps}
\caption{Evolution of the mid-plane temperature as a function of the radial distance for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities.}
\label{Tm}
\end{figure}
\subsection{Geometry evolution}
\label{geoevol}
As the geometry of the disk is calculated jointly and self-consistently with the thermodynamical structure, irregular features can also be expected in the height radial profiles, as shown in Fig. \ref{hcgm}: the pressure scale height shows radial gradient discontinuities similar to those observed in Fig. \ref{Tm}. Similarly, it is possible to approximate the pressure scale height by an almost constant power law between 20 AU and 250 AU: $h_{\mathrm{pressure}} = r^{1.26}$ after 1 Myr and $h_{\mathrm{pressure}} = r^{1.28}$ at 10 Myr. These index values are very close to the $9/7$ index suggested by \citet{chiang97} in the case of a passive disk for which the photosphere height (height at which the line-of-sight optical depth equals 1) would be directly proportional to the pressure scale height. It appears then that the approximation
of $h_{\mathrm{pressure}} \propto r^{9/7}$ can only be valid in the region between 10 and 250 AU, where the disk photosphere is irradiated and the opacity varies smoothly with the temperature (i.e., where the temperature is lower than the sublimation temperature of the main components, as listed in Table \ref{tempchgt}).
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{hcgm_MMSNmultit_1000.eps}
\caption{Evolution of the pressure scale height as a function of the radial distance for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities.}
\label{hcgm}
\end{figure}
The radial profile of the photosphere height (Fig. \ref{Hm}) reveals a zone inside of 10 AU where the disk photosphere is irregularly irradiated, while it is entirely irradiated in the outer regions. Between 20 and 250 AU, the pressure scale height can be approximated by a power law: $H_{\mathrm{photo}} \propto r^{1.11}$ after 100,000 yr and $h_{\mathrm{pressure}} \propto r^{1.14}$ at 1 Myr and 10 Myr. These index values are reminiscent of those expected by \citet{kenyon87} around $9/8$ and obtained by the numerical simulations of \citetalias{baillie14}.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{Hm_MMSNmultit_1000.eps}
\caption{Evolution of the photosphere height as a function of the radial distance for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities.}
\label{Hm}
\end{figure}
The evolution of the photosphere height over pressure scale height ratio $\chi$ is displayed in Fig. \ref{khim}. As predicted by \citet{dullemond01}, the values are in the range from 1.5 to 6. However, unlike the usual approximation ($\chi = 4$) suggested by \citet{chiang97}, this ratio is neither constant nor uniform. This $\chi$ profile, like the grazing angle $\alpha$ profile (Fig. \ref{alphagrm}) show similar gradient discontinuities as in the surface mass density or temperature radial profiles. The non-irradiated zones also appear quite obvious in these figures because they correspond to locations where $\chi$ and the grazing angle are not defined.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{khim_MMSNmultit_1000.eps}
\caption{Evolution of the pressure scale height to photosphere height ratio as a function of the radial distance for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities. Missing points are due to the regions of the disk that are not directly in the stellar line of sight at a given location and evolution time.}
\label{khim}
\end{figure}
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{alphagrm_MMSNmultit_1000.eps}
\caption{Evolution of the grazing angle as a function of the radial distance for an initial minimum mass solar nebula in the case of a self-consistently calculated geometry with a full continuous model of opacities. Missing points are due to the regions of the disk that are not directly in the stellar line of sight at a given location and evolution time.}
\label{alphagrm}
\end{figure}
\section{Discussion} \label{disc}
\subsection{Influence of the disk composition on density and temperature local gradients}
By comparing these results with the simulations of \citetalias{baillie14}, we note how important it is to take the physical composition of the dust into account by using a complete model of opacities that varies with temperature. The main difference in the surface mass density radial profiles resides in the secondary knee that appears at 1 Myr around 250 AU. Earlier conclusions from \citetalias{baillie14} remain valid: the surface-mass density profile evolves toward a shallower profile, first in the inner regions and then increasingly farther away from the star. However, when the relaxation radius reaches 250 AU, the profile is then fragmented into three power-law segments. This creates a discontinuity in the density gradient that may strongly influence the exchange of angular momentum with possible planetary cores within the disk (see Sect. \ref{torque}). In addition to that, we note that the mass flux may reverse in the middle of the disk until as late as 1 Myr. These density bumps obviously have consequences for the thermodynamical and geometrical structures. The mid-plane temperature and grazing angle radial profiles do in fact present new features, particularly at the radii of these discontinuities, when the physical composition of the disk is properly treated. Temperature plateaux are other important consequences as they may affect, for instance, the snowline position (see Sect. \ref{snowline}).
\subsection{Temperature radial profile analysis}
\label{tempprofile}
Figure \ref{profilT} details the temperature structure in the mid-plane of the protoplanetary disk after 1 Myr of evolution. As well as the mid-plane temperature, Fig. \ref{profilT} also displays the grazing angle and optical depth radial profiles, and the distribution of the heating sources.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{disk_1000000sav_tau.eps}
\caption{Mid-plane temperature radial profile (black) after 1 million years of evolution of a minimum mass solar nebula with a self-consistently calculated geometry and a full continuous model of opacities. Disk shadowed regions are displayed in gray. The ratio of the viscous heating contribution over the total heating (viscous heating rate) is presented in red, the grazing angle radial profile in yellow, and the optical depth radial profile in blue.}
\label{profilT}
\end{figure}
We first note is that the black curve shows temperature plateaux that coincide with the sublimation temperatures listed in Table \ref{tempchgt}. These plateaux are not flat, but present variations of a few Kelvin over radial regions that may be up to 1 AU wide. As Fig. \ref{opa} showed, opacities may vary quite strongly with temperature around these changes of phases. Thus, the temperature drop induced by a local surface mass density variation can be compensated for by an opacity variation over a few Kelvin to provide the equivalent heat and maintain a quasi-constant temperature. In these regions, the irradiation heating effectively has a more important contribution than outside, where it tends to become negligible compared to the viscous heating. This is confirmed by the red curve, which represents the ratio of the viscous contribution over the total heating received by the disk (viscous heating rate): the troughs at 0.1, 0.4, 0.7, 1.2, and 2 AU coincide with the temperature plateaux. This transition is called heat transition barrier by \citet{hasegawa112}. In addition, we note that the silicate sublimation plateau migrates inward as the disk evolves, and after 1 Myr, it is located at the disk inner edge. We interpret the inner disk physical structure as follows:
\begin{itemize}
\item The surface mass density is generally a decreasing function of the radial distance.
\item When a phase-transition temperature is reached, the opacity suddenly increases because of new condensed species, even though the temperature remains stable over a few fractions of an AU.
\item The viscosity $\nu = \alpha_{\mathrm{visc}} \, c_{s} \, h_{\mathrm{pressure}}$ therefore follows the temperature stability. As the density and the angular velocity decrease outward, the quantity of viscous heating decreases as well.
\item To maintain the temperature, the irradiation heating tries to compensate for the viscous heating loss, looking for an optimal grazing angle that can maximize the irradiation heating from the star.
\item At some radii, the disk photosphere geometry is such that the irradiation heating can no longer compensate for the loss in viscous heating. The grazing angle then drops until the disk photosphere is not irradiated anymore. The viscous heating remains the only source of heating, resulting in a shadowed region.
\item The temperature decreases again because of the lack of irradiation heating. However, the opacity variation is smoother now that we are no longer at the phase transition temperature. The viscosity now decreases, reinforcing the decrease in viscous heating. In the meantime, the opacity is now much higher than it was before the change of phase.
\end{itemize}
In addition to these plateaux, the mid-plane temperature radial profile shows two important troughs with an amplitude drop of about 10 Kelvin. The first one, located around 10 AU, coincides with the limit where the dominating source of heating changes: viscous heating is clearly stronger below 10 AU, while stellar irradiation heating dominates outside that limit (red curve from Fig. \ref{profilT}). The second trough, around 250 AU, generates a drop in temperature from 20 to 8 Kelvin. At these temperatures, the opacity varies by two orders of magnitude over a temperature range of 20 Kelvin (Fig. \ref{opa}): the grazing angle drops and the geometry tends to shadowing. The temperature bump disappears in simpler opacity models \citepalias{baillie14}. In addition, it remains present (although less pronounced) in earlier evolution ages.
Furthermore, we note that in the irradiation-dominated region, the disk is permanently irradiated, which is not the case in the inner regions (see gray bands in Fig. \ref{profilT}). These shadowed regions coincide with a drop in the grazing angle (see yellow curve in Fig. \ref{profilT}). Moreover, all the temperature plateaux (all are located below 10 AU) correspond to irradiated zones, while the outer edges of the plateaux (where the temperature starts dropping again) trigger shadowed regions, as detailed above. The region between 20 and 250 AU matches the two-layer model of \citet{chiang97} well because the temperature is clearly below the dust sublimation temperatures, and therefore in a temperature domain where the opacity varies so smoothly that it can approximated by a constant, allowing the retrieval of the simulation results from \citetalias{baillie14}.
Finally, the stellar line-of-sight optical depth radial profile (blue curve in Fig. \ref{profilT}) shows that the disk becomes optically thin beyond 20 AU. However, the optical thickness drops again at the heat transition barrier (around 10 AU at 1 Myr)
\subsection{Snow region}
\label{snowline}
While the snowline is defined as the radius at which the temperature is equal to the water ice condensation temperature (160 K), the fact that we now observe a plateau around that temperature shows that we should rather talk about a snow region than a snowline. Figure \ref{snowzone} presents the evolution of the inner and outer edges of the plateau around the water ice condensation temperature. Given the precision on the temperature, and in the opacity model, we chose to define the plateau temperature range as $160 \, \mathrm{K} \, \pm \, 2 \, \mathrm{K}$. The snow region can be as wide as 1 AU at early ages and migrates inward until it stabilizes below 2 AU in a few Myr
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{snowzone2.eps}
\caption{Time evolution of the snow region (mid-plane radial location for which the temperature coincides with the water-ice condensation temperature $\pm$ 2 K).}
\label{snowzone}
\end{figure}
Similarly, we define the silicate sublimation region in place of the silicate sublimation line at $1500 \, \mathrm{K} \, \pm \, 20 \, \mathrm{K,}$ and we present the evolution of its edges in Fig. \ref{sublzone}. The sublimation region initially spreads from 0.3 to 1.3 AU. It also migrates inward and seems to reach the inner edge of the disk in a few Myr.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{subzone2.eps}
\caption{Time evolution of the silicates sublimation zone (mid-plane radial location for which the temperature coincides with the silicate sublimation temperature $\pm$ 20 K).}
\label{sublzone}
\end{figure}
It appears that the zone of the mid-plane in which the sublimation of an element can occur is AU-wide instead of the expected sharp frontier usually called snow or sublimation line. At this temperature, the corresponding element is sublimated, which affects the medium opacity and therefore the heating received by the disk at this location. This heating variation is compensated for by a grazing angle variation that maintains a smoother continuity of the temperature and gas-to-dust ratio. The origin of the plateau is therefore related to the partial and gradual sublimation of the element that is sublimated at this temperature: the gas-to-dust ratio of that element is close to 0 at the outer edge of the plateau while it increases inward up to 1 at the inner edge. The other elements gas-to-dust ratios are mostly unaffected at this temperature. Partial sublimation could occur layer by layer on dust particles or on a vertical scale (the column height of the sublimated material increasing inward).
The changes of phase can now occur in much wider regions than the previous snowlines. This may favor smooth variations in physical compositions across the disk as the planetary embryo migrates inward across the snow or sublimation region. This may influence the chemical models, for example, those that try to explain the abundance of carbon in the so-called carbon-rich planets: \citet{fortney10} and \citet{madhu11} suggested that, at equilibrium, all the available O should go into organics, whereas \citet{venot12} described a non-equilibrium chemistry to model the C/O ratio. Therefore, the shallow variations of the temperature profile may favor equilibrium chemistry.
\section{Consequences on planet traps and deserts}
\label{torque}
The results reported above clearly show that rapid variations of temperature and density occur in the disk in the transition region. It is thought that such a transition may potentially create planets traps and deserts. This has been explored previously in static-disk models, but never in dynamically evolving disks. Thus in the following we compute the torque that the disk would exert on a putative planet. In particular we explore how planet traps appear, move, and disappear in the disk.
Assuming that it has already formed, a planetary embryo exchanges angular momentum with the disk \citep{goldreich79,ward88,arty93,jc05}. These exchanges are due to the resonances excited by the planetary embryo in the disk (Lindblad resonances caused by the action of the induced spiral arms, and corotation resonances). Thus, the planet exerts a torque on the disk and therefore the disk exerts an opposite torque on the planet. We calculate these torques in the case of the evolved disk described in the previous section. We then study their effects on potential planetary embryos. We assume here that the disk structure is not modified by the planet.
\subsection{Lindblad torques}
A given perturber, such as the planet in our disk, excites Lindblad resonances of multiple orders \citep{goldreich79,goldreich80}. Using a two-dimensional approximation, considering laminar disks, a planet on a circular orbit, ignoring the disk self-gravity and assuming thermal equilibrium, \citet{paarp08} were able to derive the following formula for the total Lindblad torque exerted by the disk over the planet:
\begin{equation}
\label{gammalin}
\Gamma_{\mathrm{Lindblad}} = - \frac{\Gamma_{0}(r_{P})}{\gamma} \left(2.5 \, - 1.7 \frac{\partial \ln T}{\partial \ln r}(r_{P}) + 0.1 \frac{\partial \ln \Sigma}{\partial \ln r}(r_{P}) \right)
,\end{equation}
with $\gamma = 1.4$, the adiabatic index,\\
$\Gamma_{0}(r_{P}) = \left(\frac{q}{h}\right)^{2} \, \Sigma(r_{P}) \, r_{P}^{4} \, \left(\Omega(r_{P})\right)^{2}$,\\
$h=\frac{h_{\mathrm{press}}(r_{P})}{r_{P}}$,\\
and $\Omega(r_{P})$ the Keplerian angular velocity at the planet position in the disk.
Although we just showed in the previous sections that the disk can present radial variation in density and temperature gradients, the torque expressions are still evaluated at the planet radial location. \citet{ward97}, \citet{hasegawa111,hasegawa112}, and \citet{masset11} (Appendix B) developed fully analytic torque expressions that could account for the local gradients at each resonance location rather than just considering the gradients at the planet radius. However, these formulas could not retrieve the amplitude of the expression derived in \citet{paarp08}, which was benchmarked against numerical simulations. In addition, more subtle expressions involving the second derivative of the temperature derived by \citet{masset11} (Eq. 79) were only tested in an isothermal disk. To be consistent with the corotation torque expressions from \citet{paar11}, we therefore estimated the total Lindblad torque using the expression of \citet{paarp08} (our Eq. \ref{gammalin}).
We note that the Lindblad torque presents a stronger dependence on the temperature gradient than on the density gradient, which reflects the effect of the pressure buffer described in \citet{ward97}. Following a similar process as in \citet{bitsch11} and \citet{bitsch13,bitsch14}, we can use here the results of Sect. \ref{res}, which provide the density and temperature of an evolved disk at a given date, therefore gaining in consistency as the temperature is not set to a power law and is calculated jointly with the geometry of the disk from the density resulting of the viscous evolution.
\subsection{Corotation torques}
Corotation resonances are known to exert complicated torques that include linear and nonlinear parts. \citet{paarp09b} showed that the corotation torques are generally nonlinear in the usual range of viscosity ($\alpha_{\mathrm{visc}} < 0.1$). The nonlinear contribution, due to the horseshoe drag \citep{ward91} caused by the interaction between the planet and the fluid element moving in its vicinity, is also known for having two possible origins: barotropic, intially formalized by \citet{tanaka02}, and entropic, detailed by \citet{baruteau08}.
Concerning the horseshoe drag, \citet{paar11} described the density perturbation generated by the corotation resonances and provided expressions for both the entropy and vortensity (or barotropic) contributions. Assuming a gravitational softening $b=0.4 h_{\mathrm{press}}$, \citet{bitsch11} and \citet{bitsch14} summarized these expressions to obtain the following contributing torques:
\begin{eqnarray}
\label{gammahsentro} \Gamma_{\mathrm{hs,entro}} &=& - \frac{\Gamma_{0}(r_{P})}{\gamma^{2}} \, 7.9 \, \left(-\frac{\partial \ln T}{\partial \ln r}(r_{P}) + (\gamma-1) \frac{\partial \ln \Sigma}{\partial \ln r}(r_{P}) \right)\\
\label{gammahsbaro} \Gamma_{\mathrm{hs,baro}} &=& - \frac{\Gamma_{0}(r_{P})}{\gamma} 1.1 \left(\frac{\partial \ln \Sigma}{\partial \ln r}(r_{P}) + \frac{3}{2}\right)
.\end{eqnarray}
\citet{paar11} showed that in the absence of saturation of the linear contributions, the total corotation torque can be defined as
\begin{equation}
\label{gammacor}
\Gamma_{\mathrm{corotation}} = \Gamma_{\mathrm{hs,baro}} + \Gamma_{\mathrm{hs,entro}}
.\end{equation}
It appears that this unsaturated corotation torque strongly depends on the temperature and surface mass density gradients. It also
scales with $M_{\mathrm{P}}^2$, as does the Lindblad torque. However, \citet{paarp09a} showed that given the viscous, diffusive, and libration timescales, the linear effects of the corotation torques can be saturated for some viscosities and some planet masses. For our disk that evolved for 1 Myr, the viscosity range compared to Fig. 14 from \citet{paarp09a} suggests that saturation cannot be neglected for planetary masses higher than $6 M_{\mathrm{\oplus}}$. \citet{paar11} defined weight functions for the partial saturation of the corotation torque. These functions vary with the half-width of the horseshoe, which depends on the mass of the planet. Appendix A of \citet{bitsch11} summarized this method and added correcting factors. We used a similar torque calculation, which is necessary to take into account the variations with the planet mass.
\subsection{Planetary core migration}
\subsubsection{Migration direction}
In the present exercise, we set the planet mass to a typical giant planet core mass $M_{P} = 10 \, M_{\mathrm{\oplus}}$ and varied $r_{P}$, the initial planet location. The total torque exerted by the disk over the planet is given by
\begin{equation}
\label{gammatot}
\Gamma_{\mathrm{tot}} = \Gamma_{\mathrm{Lindblad}} + \Gamma_{\mathrm{corotation}}
.\end{equation}
A positive torque exerted by the disk on the planet means that the disk gives angular momentum to the planet and therefore that the planet migrates outward. In contrast, a negative total torque results in the disk gaining angular moment from the planet, and the planet migrating inward. Thus, at the locations where the total torque sign changes, we can define two physical radii:
\begin{itemize}
\item If the total torque is negative inside and positive outside that line, the potential planets are driven away from that location, therefore defining a depleted region in planetary cores, corresponding to the planet deserts of \citet{hasegawa12}.
\item If the total torque is positive inside and negative outside, the potential planets converge toward this location, which we thus call a planetary trap. Such a zero-torque radius was called by \citet{lyra10} the equilibrium radius. \citet{hasegawa111} studied the vertical effects of the planet mass, but only considered fixed surface mass density radial profiles, while \citet{hasegawa112} estimated planet trap locations due to Lindblad torques alone.
\end{itemize}
\subsubsection{Density and temperature features at the origins of the traps and deserts}
Figure \ref{multitorque1M} presents the radial profiles of the Lindblad, corotation and total torques after 1 million years of evolution, when the planetary cores can already exist but the gas disk is not yet completely photo-evaporated. We chose to focus on the planetary formation region and therefore to limit the radial extent of our migration investigation up to 20 AU. The torques are normalized by $\Gamma_{0}$. A left-pointing arrow shows a negative torque and therefore an inward migration, while a right-pointing arrow shows a positive torque or outward migration.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{resample1000000_xlog.eps}
\caption{Radial radial profile of the migration torques (black arrows showing the direction of migration: outward for positive torques and inward for negative torques) after 1 Myr of evolution. Lindblad torques (blue), corotation torques (red), and shadowed regions (gray) are also represented.}
\label{multitorque1M}
\end{figure}
The Lindblad torque is mostly negative except around 10 AU, where the temperature gradient is reversed. The corotation torque is mainly positive, but it can become negative in the temperature plateaux regions where the temperature gradient is shallower (Fig. \ref{tottorque1M}). The orders of magnitude of the two contributions are comparable, and the resulting total torque alternates between positive (outward) regions and negative (inward) regions. After 1 million years of evolution, outward migration could occur below 0.33 AU, between 0.41 and 0.60 AU, between 0.85 and 1.7 AU, and between 2.2 and 8.0 AU. As a consequence, planets could accumulate at the traps located around 0.33, 0.60, 1.7, and 8.0 AU and the regions around 0.41, 0.85, and 2.2 AU should be strongly depleted in planets. The snowline (dotted line at 1.8 AU in Fig. \ref{multitorque1M}) seems to be closely related to the nearby trap at 1.7 AU: the inner edge of the 160 K-plateau coincides with a trap, while there is a planetary desert at the outer edge of the plateau, as shown in Fig. \ref{tottorque1M},
which presents the total torque together with the surface mass density and temperature radial profiles. As a comparison, \citet{bitsch11} found a possible equilibrium radius around 12.5 AU for a 20 Earth-masses planet, which could be consistent with our 8 AU-trap.
As the total torque strongly depends on the surface mass density and temperature gradients (see Eqs. \ref{gammalin}-\ref{gammatot}), we can here relate the torque radial profile with the bumps in surface mass density and troughs in temperature. We note here that the density bump around 6 AU seems to result in a stronger positive torque gradient, in agreement with the estimate of \citet{masset06b} that sharp surface mass density gradients are required to slow-down type I migration. On the other hand, the temperature irregularity located between 9 and 10 AU seems to be at the origin of the torque inversion in this region. The density profile there does not seem to affect the total torque trend. This sharp temperature gradient appears to be correlated with an inner shadowed region, as expected by \citet{kretke12}, who described a model in which the absence of stellar irradiation could lead to an outward migration, and therefore estimated that sustaining an outward migration in an irradiated and active disk would be difficult without any shadowing effects. This correlation confirms the importance of properly taking into account the geometry and photosphere irradiation. \citet{kretke12} found traps inside 1 AU for the most massive embryos and estimated that a mass accretion rate $\dot{M}\, > \, 10^{-8}\, M_{\odot}.\mathrm{yr}^{-1}$ was required to trap planetary cores at radii compatible with known gas giant planet radii. This is compatible with our present study and the mass accretion rates obtained by \citetalias{baillie14} for evolution ages younger than 1 Myr.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{resample1000000_lincor_TSxlog.eps}
\caption{Radial profile of the total torque (black arrows showing the direction of migration) after 1 Myr of evolution. Temperature radial profile (blue curve), surface mass density radial profile (red curve), and shadowed regions (gray regions) are also represented.}
\label{tottorque1M}
\end{figure}
\subsubsection{Earlier time evolution}
To follow the evolution of the trap and desert locations, Fig. \ref{torquemultit} presents the total torque profiles as functions of the radial distance for different ages. After 1000 years, we can only find two traps around 0.22 and 11 AU and one desert located at 7 AU. The trap is correlated with the shadow region just outside the temperature plateau at the water ice sublimation temperature (160 K).
After 10,000 years, we now have two traps (4 and 13 AU) and three deserts (3.5 and 6 AU), while the water ice sublimation plateau appears to be located around the 5 AU. The outer trap may be due to the temperature gradient inversion there.
After 100,000 years of evolution, we now have four traps (0.45, 0.75, 2, and 10 AU) and four deserts (0.23, 0.6, 1, and 3 AU). The traps appear to be located in the outer parts of the shadowed regions, while three of the deserts are at the inner edges of the shadowed zones. In addition, the water ice line coincides with the trap at 2 AU.
Traps and deserts seem to relate better to the variation in the temperature gradient than to the surface mass density gradient trends, as expected from the coefficients in Eq. \ref{gammalin}.
\begin{figure}[htbp!]
\begin{center} $
\begin{array}{cc}
\includegraphics[width=\hsize, clip=true]{resample1000_lincor_TSxlog.eps}\\
\includegraphics[width=\hsize, clip=true]{resample10000_lincor_TSxlog.eps}\\
\includegraphics[width=\hsize, clip=true]{resample100000_lincor_TSxlog.eps}\\
\end{array} $
\end{center}
\caption{Radial profiles of the total torque (black arrows showing the direction of migration) after 1000 years (upper panel), 10,000 years (middle panel), and 100,000 years (lower panel) of evolution. Temperature radial profiles (blue curve), surface mass density radial profiles (red curve), and shadowed regions (gray regions) are also represented.
\label{torquemultit}
\end{figure}
\subsubsection{Trap and desert migration}
To better follow the evolution of the traps and deserts, Fig. \ref{traptime} presents the variations in number and position of these traps and deserts.
\begin{figure}[htbp!]
\center
\includegraphics[width=\hsize]{rmulti_1000_lc_log.eps}
\caption{Time evolution of the migration traps (blue "+") and deserts (red "x"). The snowline position (dotted line) and the heat transition radius (dashed line) are also represented.}
\label{traptime}
\end{figure}
Figure \ref{traptime} shows the evolution of the traps and deserts. We first note that in the early evolution of the disk, it seems to be possible to create traps and deserts below 0.3 AU. However, as these traps disappear in a few thousand years (much before a steady-state is reached), they probably only exist as long as the disk "remembers" the initial condition of the simulation. Second, it appears that there is a permanent trap correlated with the heat transition barrier (dashed line): this trap is systematically slightly (less than 1 AU) inward of that viscous-irradiation transition (consistent with the planetary trap position of a 20 Earth-masses planet estimated by \citet{bitsch11}), recalling
the estimate of \citet{hasegawa112} that the heat transition barrier was at the origin of trapping regions. Between 10,000 and 100,000 years, the heat transition barrier moves slightly outward, which may coincide with the outward drift of the relaxation radius that crosses the planetary formation region at this age (see Sect. \ref{sigma}). That trapping location is the outer boundary of an outward-migrating region that initially spreads from around 6.5 AU (about 1 AU outer to the water ice sublimation line) up to $\sim$ 13 AU. After 10,000 years of evolution, that outward-migrating region is divided into two (sometimes three) sequences of (outward-migrating + inward-migrating) regions. A depleted zone in planets can be found $\sim$ 1 AU outside of the water ice sublimation line at any time.
After 10,000 years, we also note traps and deserts following the main dust component sublimation lines:
\begin{itemize}
\item the water ice sublimation line coincides with a systematic trap,
\item the volatile organics sublimation line is closely accompanied by an inner trap and an outer desert,
\item the refractory organics and troilite sublimation lines are similarly bordered, if not as closely as the volatile organics line,
\item the silicate sublimation line is closely followed by an outer desert region.
\end{itemize}
From Fig. \ref{tottorque1M}, we can infer that the space between two sublimation regions is a zone of outward migration: therefore the chances of a planet to become trapped or significantly slowed down in the inner disk are not negligible. Once trapped, the forming planet is very likely to follow the trap migration. In addition, most of the planet traps (those that are correlated with the heat-transition barrier or the ice lines) seem quite sustainable: these traps and deserts slowly migrate inward with the sublimation lines as the disk ages and cools down. Only a few isolated transient traps and deserts appear between the water-ice line and the heat-transition barrier. However, it is very likely that a planet in such a disappearing trap would be trapped again in the next outer trap.
\citet{kley09} suggested that a possible scenario to address the inconsistency between the formation and migration timescales could be the retention of the icy cores at the snowline. We extend
this: it is possible to trap planets at any main dust component sublimation line. Dust sublimation is a physical mechanism sufficiently efficient to trap planets along the ice lines. In addition, as suggested by the correlation of the trap or desert locations with respect to the shadowed regions (Fig. \ref{tottorque1M}), the heating sources play an important role in trapping planets. This is reinforced by the presence of traps along the heat-transition barrier.
\subsubsection{Perspectives}
Although multiple planet-systems seem to be the most common configuration, we here only studied the interaction of a single planet with the disk. \citet{cossou13} showed that multiple planetary cores tend to be trapped in chains of mean motion resonances. Therefore, a proper treatment of multiple planet interaction with the disk requires considering the dynamics of these planets in addition to the disk dynamics. Coupling our hydrodynamical approach with an N-body code requires a more thorough study and will be the object of a future paper. In addition, as we have detailed, our current model does not take into account the feedback of the planet on the disk, which is currently not affected by the planet. This might also be necessary for future studies.
We leave the consideration of dead zones for
a future paper because they probably generate density and temperature irregularities that have been shown to be favorable for creating planetary traps and deserts. Although the torque expressions detailed above show a limited dependency on the planet mass ($\propto M_{\mathrm{P}}^{2})$), this is kept for a subsequent review, along with the dependency on the disk mass and the mass accretion rate.
\section{Conclusions}
We have studied the viscous evolution of a protoplanetary disk, using an $\alpha$ prescription for its viscosity. Starting with an initial minimum mass solar nebula, we self-consistently calculated the geometry, thermodynamics, and density distribution evolution of such a disk, taking into account the various sources of heating (viscous and irradiation), the shadowing effects, and the physical composition. The physical phases of the various dust components have a dramatic influence on the opacity of the disk and therefore reflect on the mid-plane temperature. The evolved density radial profiles present a few irregularities compared to the previous
results of \citetalias{baillie14}. These density bumps mainly result in temperature features (bumps, troughs, plateaux) and also widen the snowline and sublimation line regions. Our main result is that snowlines may better be called snow regions, AU-wide smooth variations, rather than sharp frontiers in the disk. Close to the sublimation temperature of the main dust elements, the disk mid-plane temperature becomes constant. This could result in a less sudden and strong change of phase than initially thought, involving partial sublimation. Given an element in Table \ref{tempchgt}, and if we focus on the disk mid-plane region for which the temperature corresponds to that element sublimation temperature within a few Kelvin, the gas-to-dust ratio of that element would vary from $\sim 0$ at the outer edge of the temperature plateau to $\sim 1$ at the inner edge (the rest of the material is unaffected at these temperatures). A planetesimal crossing such a region inward might then be able to melt layer by layer, with possible consequences on its mineral petrology in the case of a future recondensation. Another way to model the partial sublimation would be to consider a vertical gradient of gas-to-dust ratio for the element in question: the element would be sublimated from the mid-plane up to a given altitude that would vary with the radial distance in the plateau. The farther into the plateau, the higher that limit, and the greater the proportion of the material of the column that is sublimated. Such a refinement would affect the dust vertical transport as studied by \citet{ciesla10b}, for example. Possible combinations of vertical and radial gradients of gas-to-dust ratio could also reflect the physics of the temperature plateaux.
From these density and temperature radial profiles, we estimated the resonant torques that potential planetary cores would exert on the disk. From the total torque sign (resulting from the Lindblad and corotation resonances), we derived the locations of planetary traps and deserts and followed their own migration as the disk ages: after 1 Myr, four planetary traps were identified around 0.33, 0.60, 1.7, and 8.0 AU, with three planetary deserts around 0.41, 0.85, and 2.2 AU.
As stated by \citet{masset06b}, density and temperature discontinuities appear to be the key characteristics leading to the generation of planetary traps: temperature troughs, density bumps, changes of phases, and heat transitions clearly affect the potential planet migration. Previous works indicated that density and temperature irregularities can slow down type I migration to allow planets to grow in a reasonable time. Our evolution code shows that such structures naturally arise during the disk evolution when the coupling between the disk dynamics, thermodynamics, and geometry is taken into account: planets can be trapped at the main dust component sublimation lines. Additional simulations are still required to estimate by how much planetary embryos can be slowed down by being trapped, and whether they can actually migrate along with the planetary traps. It appears necessary to synchronize the planet and trap migration rates, and this study requires a more thorough calculation of the planet feedback on the disk. Other sources of discontinuities, such as deadzones, may affect these migration torques and create more planetary traps or deserts. This will be treated in a separate paper.
\begin{acknowledgements}
We thank Esther Taillifet for enlightening discussions. We are also indebted to Bertram Bitsch and Dominic Macdonald for valuable suggestions that improved the quality of the manuscript significantly. We also thank the referee for detailed and constructive comments that improved the quality of the paper. This work was supported by IDEX Sorbonne Paris Cit\'e. We acknowledge the financial support from the UnivEarthS Labex program of Sorbonne Paris Cit\'e (ANR-10-LABX-0023 and ANR-11-IDEX-0005-02).
\end{acknowledgements}
|
\section{Introduction}
A contractive map of a group has a unitary dilation if and only if it is completely positive definite, in the sense that certain operator matrices are positive. Consequently, for a semigroup $P$ contained in a group $G$, a contractive representation of $P$ has a unitary dilation if and only if it can be extended to a completely positive definite map on $G$. Introduced in \cite{DFK2014}, such representations on a semigroup are called completely positive definite. In particular, when the group is lattice-ordered, a representation is called regular if a certain natural extension to the group is completely positive definite.
Nica \cite{Nica1992} introduced the study of isometric representations of quasi-lattice ordered semigroups. This generalized the notion of doubly commuting representations of semigroups with nice generators. Laca and Raeburn \cite{Laca1996} developed the theory, and showed there is a universal $C^*$-algebra for isometric Nica covariant representations. This field has also been explored in \cite{Solel2008}.
Davidson, Fuller, and Kakariadis \cite{Fuller2013, DFK2014} defined and studied contractive Nica-covariant representation on lattice ordered semigroups. The regularity of such representations was seen as a critical property in describing the $C^*$-envelope of semicrossed products. They posed a question \cite[Question 2.5.11]{DFK2014} of whether regularity is automatic for Nica-covariant representations. Fuller \cite{Fuller2013} established this for certain abelian semigroups.
This paper answers this question affirmatively by establishing a necessary and sufficient condition for a representation of a lattice ordered semigroup to be regular. This condition generalizes a result of Brehmer \cite{Brehmer1961}, where he gave a necessary and sufficient condition for a representation of $\mathbb{Z}_+^\Omega$ to be regular. As an immediate consequence of Brehmer's condition, it is known that doubly commuting representations and commuting column contractions are both regular \cite[Proposition I.9.2]{SFBook}. This paper generalizes both results in the lattice ordered group settings. We first show that a Nica-covariant representation, which is an analog of a doubly commuting representation, is regular. We then introduce an analog of commuting column contractions, which is shown to be regular as well.
\section{Preliminaries}
Let $G$ be a group. A unital semigroup $P\subseteq G$ is called a cone. A cone $P$ is \emph{spanning} if $PP^{-1}=G$, and is \emph{positive} when $P\bigcap P^{-1}=\{e\}$. A positive cone $P$ defines a partial order on $G$ via $x\leq y$ when $x^{-1}y\in P$. $(G,P)$ is called \emph{totally ordered} if $G=P\bigcup P^{-1}$, in which case the partial order on $G$ is a total order. If any finite subset of $G$ with a upper bound in $P$ also has a least upper bound in $P$, the pair $(G,P)$ is called a \emph{quasi-lattice ordered group}. We call this partial order \emph{compatible with the group} if for any $x\leq y$ and $g\in G$, we always have $gx\leq gy$ and $xg\leq yg$. Equivalently, the corresponding positive cone satisfies a normality condition that $gPg^{-1}\subseteq P$ for any $g\in G$, and thus $x\leq y$ whenever $yx^{-1}\in P$ as well. When $P$ is a positive spanning cone of $G$ whose partial order is compatible with the group, if every two elements $x,y\in G$ have a least upper bound (denoted by $x\vee y$) and a greatest lower bound (denoted by $x\wedge y$), the pair $(G,P)$ is called a \emph{lattice ordered group}. It is immediate that a lattice ordered group is also a quasi-lattice ordered group.
\begin{example} (Examples of Lattice Ordered Groups)
\begin{enumerate}
\item $(\mathbb{Z},\mathbb{Z}_{\geq 0})$ is a lattice ordered group. In fact, this partial order is also a total order. More generally, any totally ordered group $(G,P)$ is also a lattice ordered group.
\item If $(G_i, P_i)_{i\in I}$ is a family of lattice ordered group, their direct product $(\prod G_i, \prod P_i)$ is also a lattice ordered group.
\item Let $G=C_\mathbb{R}[0,1]$, the set of all continuous functions on $[0,1]$. Let $P$ be the set of all non-negative functions in $G$. Then $(G,P)$ is a lattice ordered group.
\item Let $\mathcal{T}$ be a totally ordered set. A permutation $\alpha$ on $\mathcal{T}$ is called order preserving if for any $p,q\in\mathcal{T}$, $p\leq q$, we also have $\alpha(p)\leq\alpha(q)$. Let $G$ be the set of all order preserving permutations, which is clearly a group under composition. Let $P=\{\alpha\in G: \alpha(t)\geq t,\mbox{ for all }t\in\mathcal{T}\}$. Then $(G,P)$ is a non-abelian lattice ordered group \cite{LatticeOrderBookIntro}.
\item Let $\mathbb{F}_n$ be the free group of $n$ generators, and $\mathbb{F}_n^+$ be the semigroup generated by the $n$-generators. Then $(\mathbb{F}_n,\mathbb{F}_n^+)$ defines a quasi-lattice ordered group \cite[Examples 2.3]{Nica1992}. However, the induced partial order is not compatible with the group and the pair is not a lattice ordered group.
\end{enumerate}
\end{example}
For any element $g\in G$ of a lattice ordered group $(G,P)$, $g$ can be written uniquely as $g=g_+g_-^{-1}$ where $g_+,g_-\in P$, and $g_+\wedge g_-=e$. In fact, $g_+=g\vee e$ and $g_-=g^{-1}\vee e$. Here are some important properties of a lattice ordered group:
\begin{lemma}\label{lm.LatticeBasic} Let $(G,P)$ be a lattice order group, and $a,b,c\in G$.
\begin{enumerate}
\item $a(b\vee c)=(ab)\vee(ac)$ and $(b\vee c)a=(ba)\vee(ca)$. A similar distributive law holds for $\wedge$.
\item $(a\wedge b)^{-1}=a^{-1}\vee b^{-1}$ and similarly $(a\vee b)^{-1}=a^{-1}\wedge b^{-1}$.
\item $a\geq b$ if and only if $a^{-1}\leq b^{-1}$.
\item $a(a\wedge b)^{-1} b=a\vee b$. In particular, when $a\wedge b=e$, $ab=ba=a\vee b$.
\item If $a,b,c\in P$, then $a\wedge(bc)\leq (a\wedge b)(a\wedge c)$.
\end{enumerate}
\end{lemma}
One may refer to \cite{LatticeOrderBook} for a detailed discussion of this subject. Notice by statement (4) of Lemma \ref{lm.LatticeBasic} $g_+,g_-$ commute and thus $g=g_+g_-^{-1}=g_-^{-1} g_+$.
For a group $G$, a unital map $S:G\to\bh{H}$ is called \emph{completely positive definite} if for any $g_1,g_2,\cdots,g_n\in G$
$$\big[S(g_i^{-1} g_j)\big]_{1\leq i,j\leq n}\geq 0.$$
Here, $i$ denotes the row index and $j$ the column index, and we shall follow this convention throughout this paper. A well known result (\cite{Naimark1943}, see also \cite[Proposition I.7.1]{SFBook}) stated that a completely positive definite map of $G$ has a unitary dilation. The converse is elementary.
\begin{theorem}\label{thm.NagyFoias} If $S:G\to\bh{H}$ is a unital completely positive definite map. Then there exists a unitary representation $U:G\to\bh{K}$ where $\mathcal{H}$ is a subspace of $\mathcal{K}$, and that $P_\mathcal{H} U(g)|_\mathcal{H} = S(g)$. Moreover, this unitary representation can be chosen to be minimal in the sense of $\mathcal{K}=\bigvee_{g\in G} U(g)\mathcal{H}$.
\end{theorem}
When $(G,P)$ is a lattice ordered group, we may simultaneously increase or decrease $g_i$ so that it would suffices to take $g_i\in P$:
\begin{lemma} Let $S:G\to\bh{H}$ be a map, then the following are equivalent:
\begin{enumerate}
\item $\big[S(g_i^{-1} g_j)\big]_{1\leq i,j\leq n} \geq 0$ for any $g_1,g_2,\cdots,g_n\in G$;
\item $\big[S(g_i g_j^{-1})\big]_{1\leq i,j\leq n} \geq 0$ for any $g_1,g_2,\cdots,g_n\in G$;
\item $\big[S(p_i^{-1} p_j)\big]_{1\leq i,j\leq n} \geq 0$ for any $p_1,p_2,\cdots,p_n\in P$;
\item $\big[S(p_i p_j^{-1})\big]_{1\leq i,j\leq n} \geq 0$ for any $p_1,p_2,\cdots,p_n\in P$.
\end{enumerate}
\end{lemma}
\begin{proof}
Since $G$ is a group, by considering $g_i$ and $g_i^{-1}$, it is clear that (1) and (2) are equivalent. Statement (1) clearly implies statement (3), and conversely when statement (3) holds true, for any $g_1,\cdots,g_n\in G$, take $g=\vee_{i=1}^n \left(g_i\right)_-$. Denote $p_i=g\cdot g_i$ and notice that from our choice of $g$, $g\geq \left(g_i\right)_-$. Hence, $$p_i=g\cdot \left(g_i\right)_-^{-1} \left(g_i\right)_+\in P.$$
But notice that for each $i,j$, $p_i^{-1} p_j=g_i^{-1} g^{-1} gg_j=g_i^{-1}g_j$. Therefore, $$\big[S(g_i^{-1} g_j)\big]_{1\leq i,j\leq n}=\big[S(p_i^{-1} p_j)\big]_{1\leq i,j\leq n}\geq 0.$$
Similarly, statements (2) and (4) are equivalent. \end{proof}
For the convenience of computation, when $(G,P)$ is a lattice ordered group, $S:G\to\bh{H}$ is called completely positive definite when $$\big[S(p_i p_j^{-1})\big]_{1\leq i,j\leq n} \geq 0.$$
For a spanning cone $P\subset G$, a contractive representation $T:P\to\bh{H}$ is called \emph{completely positive definite} when it can be extended to some completely positive definite map on $G$. There is a well-known result due to Sz.Nagy that every contraction has a unitary dilation, and therefore, every contractive representation of $\mathbb{Z}_+$ is completely positive definite. Ando \cite{Ando1963} further showed that every contractive representation of $\mathbb{Z}_+^2$ is completely positive definite. However, Parrott \cite{Parrott1970} provided an counterexample where a contractive representations on $\mathbb{Z}_+^3$ is not completely positive definite.
For a completely positive definite representation $T$ on a lattice ordered semigroup, one might wonder what its extension looks like. In a lattice ordered group $(G,P)$, any element $g\in G$ can be uniquely written as $g=g_+g_-^{-1}$ where $g_\pm\in P$ and $g_+\wedge g_-=e$. Suppose $U:G\to\bh{K}$ is a unitary dilation of $T$, we can make the following observation.
\begin{eqnarray*}
\tilde{T}(g) &=& P_\mathcal{H} U(g)\big|_\mathcal{H} \\
&=& P_\mathcal{H} U(g_-)^*U(g_+)\big|_\mathcal{H}.
\end{eqnarray*}
This motivates the question of whether the extension $\tilde{T}(g)=T(g_-)^*T(g_+)$ is completely positive definite. We call a contractive representation $T$ \emph{right regular} whenever $\tilde{T}$ defined in such way is completely positive definite. There is a dual definition that call $T$ \emph{left regular} (or \emph{$*$-regular}) if $\overline{T}(g)=T(g_+)T(g_-)^*$ is completely positive definite.
When $(G,P)$ is a lattice ordered group, $(G,P^{-1})$ is also a lattice ordered group. A representation $T:P\to\bh{H}$ give raise to a dual representation $T^*:P^{-1}\to\bh{H}$ where $T^*(p^{-1})=T(p)^*$. Consider $g=g_+g_-^{-1}=g_-^{-1} \left(g_+^{-1}\right)^{-1}$, we have $$\tilde{T}(g)=T(g_-)^* T(g_+)=T^*(g_-^{-1}) T^*(g_+^{-1})^*=\overline{T^*}(g).$$ Hence, $\tilde{T}$ agrees with $\overline{T^*}$ on $G$. Therefore, we obtain the following Proposition.
\begin{proposition} Let $(G,P)$ be a lattice ordered group, and $T:P\to\bh{H}$ be a representation and $T^*$ defined as above. Then the following are equivalent
\begin{enumerate}
\item $T$ is right regular.
\item $T^*$ is left regular.
\item For any $p_1,\cdots,p_n\in P$, $[\tilde{T}(p_ip_j^{-1})]\geq 0$ (equivalently, $[\overline{T^*}(p_ip_j^{-1})]\geq 0$).
\end{enumerate}
\end{proposition}
Due to this equivalence, we shall focus on the right regularity and call a representation \emph{regular} when it is right regular. Regular dilations were first studied by Brehmer \cite{Brehmer1961}, and they were also studied in \cite{Nagy1961, Halperin1962}. A necessary and sufficient condition for regularity for the abelian group $\mathbb{Z}^\Omega$ was proven by Brehmer \cite[Theorem I.9.1]{SFBook}.
\begin{theorem}[Brehmer]\label{thm.Brehmer} Let $\Omega$ be a set, and denote $\mathbb{Z}^\Omega$ to be the set of $(t_\omega)_{\omega\in\Omega}$ where $t_\omega\in\mathbb{Z}$ and $t_\omega=0$ except for finitely many $\omega$. Also, for a finite set $V\subset\Omega$, denote $e_V\in\mathbb{Z}^\Omega$ to be $1$ at those $\omega\in V$ and $0$ elsewhere. If $\{T_\omega\}_{\omega\in\Omega}$ is a family of commuting contractions, we may define a contractive representation $T:\mathbb{Z}_+^\Omega\to\bh{H}$ by $$T(t_\omega)_{\omega\in\Omega}=\prod_{\omega\in\Omega} T_\omega^{t_\omega}.$$
Then $T$ is right regular if and only if for any finite $U\subseteq\Omega$, the operator $$\sum_{V\subseteq U} (-1)^{|V|} T(e_V)^* T(e_V)\geq 0.$$
\end{theorem}
It turns out that not all completely positive definite representations are regular.
\begin{example} It follows from Brehmer's theorem that a representation $T$ on $\mathbb{Z}_+^2$ is regular if and only if $T_1=T(e_1), T_2=T(e_2)$ are contractions that satisfy $$I-T_1^* T_1-T_2^* T_2+(T_1 T_2)^* T_1T_2\geq 0.$$
Take $T_1=T_2=\left[ \begin{array}{cc}
0 & 1 \\
0 & 0 \\ \end{array} \right]$ and notice, $$I-T_1^* T_1-T_2^* T_2+(T_1 T_2)^* T_1T_2=\left[ \begin{array}{cc}
1 & 0 \\
0 & -1 \\ \end{array} \right].$$
Brehmer's result implies that $T$ is not regular. However, from Ando's theorem \cite{Ando1963}, any contractive representation on $\mathbb{Z}_+^2$ has a unitary dilation and thus is completely definite definite.
\end{example}
Isometric Nica-covariant representations on quasi-lattice ordered groups were first introduced by Nica \cite{Nica1992}: an isometric representation $W:G\to\bh{H}$ is Nica-covariant if for any $x,y$ with an upper bound, $W_x W_x^* W_y W_y^*=W_{x\vee y} W_{x\vee y}^*$. When the order is a lattice order, it is equivalent to the property that $W_s,W_t^*$ commute whenever $s\wedge t=e$. Therefore, the notion of Nica-covariant is extended to abelian lattice ordered groups in \cite{DFK2014}, and we shall further extend such definition to non-abelian lattice ordered groups and call a representation $T:P\to\bh{H}$ \emph{Nica-covariant} if $T(s)T(t)^*=T(t)^*T(s)$ whenever $s\wedge t=e$. For a Nica-covariant representation $T$, since $T(g^+)$ commutes with $T(g^-)^*$ for any $g\in G$, there is no difference between left and right regularity. It observed in \cite{DFK2014} that Nica-covariant representations are regular in many cases.
\begin{example} (Examples of Nica covariant representations)
\begin{enumerate}
\item On $(\mathbb{Z},\mathbb{Z}_+)$, a contractive representation $T$ on $\mathbb{Z}_+$ only depends on $T_1=T(1)$ since $T(n)=T_1^n$. This representation is always Nica-covariant since for any $s,t\geq 0$, $s\wedge t=0$ if and only if one of $s,t$ is $0$. A well known result due to Sz.Nagy shows that its extension to $\mathbb{Z}$ by $\tilde{T}(-n)=T^{*n}$ is completely positive definite and thus $T$ is regular.
\item Similarly, any contractive representation of a totally ordered group $(G,P)$ is Nica-covariant. A theorem of Mlak \cite{Mlak1966} shows that such representations are regular.
\item $(\mathbb{Z}^n,\mathbb{Z}_+^n)$, the finite Cartesian product of $(\mathbb{Z},\mathbb{Z}_+)$ is a lattice ordered group. A representation $T$ on $\mathbb{Z}_+^n$ depends on $n$ contractions $T_1=T(1,0,\cdots,0)$, $T_2=T(0,1,0,\cdots,0)$,$\cdots$, $T_n=T(0,\cdots,0,1)$. Notice $T$ is Nica covariant if and only if $T_i, T_j$ $*$-commute whenever $i\neq j$. Hence Nica covariant representations are equivalent to doubly commuting. It is known \cite[Section I.9]{SFBook} that doubly commuting contractive representations are regular.
\item For a lattice ordered group made from a direct product of totally ordered groups, Fuller \cite{Fuller2013} showed that their contractive Nica-covariant representations are regular.
\end{enumerate}
\end{example}
A question posed in \cite[Question 2.5.11]{DFK2014} asks whether contractive Nica-covariant representations on abelian lattice ordered groups are regular in general. For example, for $G=C_\mathbb{R}[0,1]$ and $P$ equal to the set of non-negative continuous functions, there are no known results on whether contractive Nica-covariant representations are regular on such semigroup. Little is known for the non-abelian lattice ordered groups as well. In this paper, we establish that all Nica-covariant representations of lattice ordered semigroups are regular.
Let $(G,P)$ be a lattice-ordered group, not necessarily abelian. Recall that the regularity conditions require a matrix involving entries in the form of $\tilde{T}(pq^{-1})$ to be positive, where $p,q\in P$. We start by investigating this quantity of $pq^{-1}$.
\begin{lemma}\label{lm.1} Let $p,q\in P$. Then,
\begin{eqnarray*}
(pq^{-1})_+ &=& p(p\wedge q)^{-1}\mbox{ and,}\\
(pq^{-1})_- &=& q(p\wedge q)^{-1}.
\end{eqnarray*} \end{lemma}
\begin{proof}
By property (1) and (2) in Lemma \ref{lm.LatticeBasic},
\begin{eqnarray*}
(pq^{-1})_+ &=& (pq^{-1} \vee e) \\
&=& p(q^{-1} \vee p^{-1}) \\
&=& p(p\wedge q)^{-1}.
\end{eqnarray*}
Similarly, $(pq^{-1})_-=q(p\wedge q)^{-1}.$ \end{proof}
\begin{lemma}\label{lm.1b} Let $p,q,g\in P$ such that $g\wedge q=e$. Then $(pg)\wedge q=p\wedge q$.
\end{lemma}
\begin{proof}
By the property (5) of Lemma \ref{lm.LatticeBasic}, we have that
$$(pg)\wedge q\leq (p\wedge q)(g\wedge q)=p\wedge q.$$
On the other hand, $p\wedge q$ is clearly a lower bound for both $p\leq pg$ and $q$, and hence $p\wedge q\leq (pg)\wedge q$. This proves the equality. \end{proof}
\begin{lemma}\label{lm.2} Let $p,q\in P$. If $g\in P$ is another element where $g\wedge q=0$, then
\begin{eqnarray*}
(pgq^{-1})_- &=& (pq^{-1})_-\mbox{ and,} \\
(pgq^{-1})_+ &=& (pq^{-1})_+g.
\end{eqnarray*}
In particular, if $0\leq g\leq p$, then
\begin{eqnarray*}
(pg^{-1}q^{-1})_- &=& (pq^{-1})_-\mbox{ and,} \\
(pg^{-1}q^{-1})_+ &=& (pq^{-1})_+g^{-1}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
By Lemma \ref{lm.1}, we get $(pgq^{-1})_+=pg(q\wedge pg)^{-1}$. Apply Lemma \ref{lm.1b} to get
$$(q\wedge pg)^{-1} = (q\wedge p)^{-1}.$$
Now $g\wedge (p\wedge q)=e$ and thus $g$ commutes with $p\wedge q$ by property (4) of Lemma \ref{lm.LatticeBasic}. Therefore,
\begin{eqnarray*}
(pgq^{-1})_+ &=& pg(q\wedge pg)^{-1} \\
&=& p(q\wedge p)^{-1}g \\
&=& (pq^{-1})_+ g.
\end{eqnarray*}
The statement $(pgq^{-1})_- = (pq^{-1})_-g$ can be proven in a similar way.
Finally, for the case where $0\leq g\leq p$, it follows immediately by considering $p'=pg^{-1}$ and thus $p=p'g$. \end{proof}
\begin{lemma}\label{lm.order} If $p_1,p_2,\cdots,p_n\in P$ and $g_1,\cdots,g_n\in P$ be such that $g_i\leq p_i$ for all $i=1,2,\cdots,n$. Then $\wedge_{i=1}^n p_i g_i^{-1} \leq \wedge_{i=1}^n p_i$. In particular, when $\wedge_{i=1}^n p_i=e$, we have $\wedge_{i=1}^n p_i g_i^{-1}=e$.
\end{lemma}
\begin{proof} It is clear that $e\leq p_i g_i^{-1}\leq p_i$, and thus $$e\leq \wedge_{i=1}^n p_i g_i^{-1} \leq \wedge_{i=1}^n p_i.$$ Therefore, the equality holds when the later is $e$. \end{proof}
\section{A Necessary and Sufficient Condition For Regularity}
When $T:P\to\bh{H}$ is a representation of lattice ordered semigroup, we denote $\tilde{T}(g)=T(g^-)^* T(g^+)$. Recall that $T$ is \emph{regular} if $\tilde{T}$ is completely positive definite. The main result is the following necessary and sufficient condition for regularity:
\begin{theorem}\label{thm.main} Let $(G,P)$ be a lattice ordered group and $T:P\to\bh{H}$ be a contractive representation. Then $T$ is regular if and only if for any $p_1,\cdots,p_n\in P$ and $g\in P$ where $g\wedge p_i=e$ for all $i=1,2,\cdots,n$, we have
\begin{equation}\label{eq.right}\left[T(g)^* \tilde{T}(p_i p_j^{-1}) T(g)\right]\leq \left[\tilde{T}(p_i p_j^{-1})\right].\tag{$\star$}\end{equation}
\end{theorem}
\begin{remark}\label{rm.pos} If we denote $$X=\left[\tilde{T}(p_i p_j^{-1})\right]$$ and $D=
\diag(T(g),T(g),\cdots,T(g))$, Condition (\ref{eq.right}) is equivalent of saying $D^* X D\leq X$. Notice that we made no assumption on $X\geq 0$. Indeed, it follows from the main result that Condition (\ref{eq.right}) is equivalent of saying the representation $T$ is regular, which in turn implies $X\geq 0$. Therefore, when checking the Condition (\ref{eq.right}), we may assume $X\geq 0$.
\end{remark}
\begin{remark} By setting $p_1=e$ and picking any $g\in P$, Condition (\ref{eq.right}) implies that $T(g)^* T(g)\leq I$, and thus $T$ must be contractive.
\end{remark}
The following Lemma is taken from \cite[Lemma 14.13]{NestAlgebra}.
\begin{lemma}\label{lm.Davidson} If $A,X,D$ are operators in $\bh{H}$ where $A\geq 0$. Then a matrix of the form $\begin{bmatrix} A & A^{1/2}X \\ X^* A^{1/2} & D\end{bmatrix}$ is positive if and only if $D\geq X^* X$.
\end{lemma}
Condition (\ref{eq.right}) can thus be interpreted in the following equivalent definition.
\begin{lemma}\label{lm.equiv} Let $p_1,\cdots,p_n\in P$ and $g\in P$ with $g\wedge p_i=e$ for all $1\leq i\leq n$. Denote $q_1=p_1g,\cdots,q_n=p_ng$ and $q_{n+1}=p_1,\cdots,q_{2n}=p_n$. Then Condition (\ref{eq.right}) is equivalent to $\left[\tilde{T}(q_i q_j^{-1})\right]\geq 0$.
\end{lemma}
\begin{proof}
Let $X=[\tilde{T}(p_i p_j^{-1})]\geq 0$ and $D=\diag(T(g),T(g),\cdots,T(g))$. Notice by Lemma \ref{lm.2} that
\begin{eqnarray*}
(p_i g p_j^{-1})_+&=&(p_i p_j^{-1})_+ g\\
(p_i g p_j^{-1})_-&=&(p_i p_j^{-1})_-,
\end{eqnarray*}
and thus $\tilde{T}(p_i g p_j^{-1})=\tilde{T}(p_i p_j^{-1}) T(g)$. Therefore,
$$\left[\tilde{T}(q_i q_j^{-1})\right]=\begin{bmatrix} X & XD \\ D^* X & X\end{bmatrix}.$$ Lemma \ref{lm.Davidson} implies that this matrix is positive if and only if $D^* X D\leq X$, which is Condition \ref{eq.right}.\end{proof}
We shall first show that $\left[\tilde{T}(p_i p_j^{-1})\right]\geq 0$ given $p_i\wedge p_j=e$ and Condition (\ref{eq.right}). This will serve as a base case in the proof of the main result.
\begin{lemma}\label{lm.base} Let $(G,P)$ be a lattice ordered group, and $T$ be a representation on $P$ that satisfies Condition (\ref{eq.right}). If $p_i\wedge p_j=e$ for all $i\neq j$, then $[\tilde{T}(p_i p_j^{-1})]\geq 0$.
\end{lemma}
\begin{proof}
Let $q_1=e,q_2=p_1$ and for each $1<m\leq n$, recursively define $q_{2^{m-1}+k}=p_m q_k$ where $1\leq k\leq 2^{m-1}$. Since $T$ is contractive,
$$[\tilde{T}(q_i q_j^{-1})]_{1\leq i,j\leq 2}=\begin{bmatrix} I & \tilde{T}(q_1 q_2^{-1}) \\ \tilde{T}(q_2q_1^{-1}) & I \end{bmatrix}\geq 0.$$
By Lemma \ref{lm.equiv}, for each $m$, $[\tilde{T}(q_i q_j^{-1})]_{1\leq i,j\leq 2^m}\geq 0$. Notice that $q_{2^{m-1}}=p_m$ for each $1\leq m\leq n$. Therefore, $[\tilde{T}(p_i p_j^{-1})]$ is a corner of $[\tilde{T}(q_i q_j^{-1})]\geq 0$, and thus must be positive.\end{proof}
For arbitrary choices of $p_1,\cdots,p_n\in P$, the goal is to reduce it to the case where $p_i\wedge p_j=e$. The following lemma does the reduction.
\begin{lemma}\label{lm.reduce} Let $(G,P)$ be a lattice ordered group. Assuming $T$ is a representation that satisfies Condition (\ref{eq.right}).
Assume there exists $2\leq k<n$ where for each $J\subset\{1,2,\cdots,n\}$ with $|J|>k$, $\wedge_{j\in J} p_j=e$. Then let $g=\wedge_{j=1}^{k} p_j$ and $q_1=p_1 g^{-1},\cdots,q_k=p_k g^{-1}$, and $q_{k+1}=p_{k+1},\cdots,q_n=p_n$.
Then $[\tilde{T}(p_i p_j^{-1})]\geq 0$ if $[\tilde{T}(q_i q_j^{-1})]\geq 0$.
\end{lemma}
\begin{proof}
Let us denote $X=[\tilde{T}(q_jq_i^{-1})]\geq 0$ and its lower right $(n-k)\times (n-k)$ corner to be $Y$. Notice first of all, when $i,j\in\{1,2,\cdots,k\}$, $$q_i q_{j}^{-1}=p_i g^{-1} g p_{j}^{-1}=p_ip_{j}^{-1}.$$
So the upper left $k\times k$ corner of $[\tilde{T}(q_i q_j^{-1})]$ and the lower right $(n-k)\times(n-k)$ corner of $X$ are both the same as those in $[\tilde{T}(p_i p_j^{-1})]$.
Now consider $i\in\{1,2,\cdots,k\}$ and $j\in\{k+1,\cdots,n\}$. It follows from the assumption that $g\wedge p_j = \left(\wedge_{s=1}^k p_s\right)\wedge p_j=e$ and $g\leq p_i$. Therefore, we can apply Lemma \ref{lm.2} to get
\begin{eqnarray*}
(p_i g^{-1}p_j^{-1})_- &=& (p_ip_j^{-1})_- \\
(p_i g^{-1}p_j^{-1})_+ &=& (p_ip_j^{-1})_+g^{-1}.
\end{eqnarray*}
Now $g\in P$, so that $T((q_i q_j^{-1})_+)T(g)=T((p_i p_j^{-1})_+)$ and $T((q_i q_j^{-1})_-)=T((p_i p_j^{-1})_-)$. Hence, $$\tilde{T}(q_i q_j^{-1}) T(g)=\tilde{T}(p_i p_j^{-1}).$$
Similarly, for $i\in\{k+1,\cdots,n\}$, $j\in\{1,2,\cdots,k\}$, we have $\tilde{T}(p_i p_j^{-1})=T(g)^*\tilde{T}(q_j q_i^{-1})$. Now define $D=\operatorname{diag}(I,\cdots, I, T(g), \cdots, T(g))$ be the block diagonal matrix with $k$ copies of $I$ followed by $n-k$ copies of $T(g)$. Consider $D X D^*$: it follows immediately from the assumption that $D^* X D\geq 0$. We have,
$$D^*[\tilde{T}(q_i q_j^{-1})] D= \renewcommand\arraystretch{1.4}\left[ \begin{array}{ccc|c}
\cdots & \cdots & \cdots & \vdots \\
\cdots & \tilde{T}(p_i p_j^{-1}) & \cdots & \tilde{T}(q_i q_j^{-1}) T(g) \\
\cdots & \cdots & \cdots & \vdots \\ \hline
\cdots & T(g)^* \tilde{T}(q_i q_j^{-1}) & \cdots & [T(g)^* \tilde{T}(p_i p_j^{-1}) T(g)] \\ \end{array} \right]\geq 0.$$
It follows from our previous computation that each entry in the lower left $(n-k)\times k$ corner and upper right $k\times (n-k)$ corner are the same as those in $[\tilde{T}(p_i p_j^{-1})]$. Hence, $D X D^*$ only differs from $[\tilde{T}(p_ip_j^{-1})]$ on the lower right $(n-k)\times (n-k)$ corner. It follows from Condition (\ref{eq.right}) that
$$[T(g)^* \tilde{T}(p_i p_j^{-1}) T(g)]\leq [\tilde{T}(p_i p_j^{-1})].$$
Hence, the matrix remains positive when the lower right corner $[T(g)^* \tilde{T}(p_i p_j^{-1}) T(g)]$ in $D^* X D$ is replaced by $[\tilde{T}(p_i p_{j}^{-1})]$. The resulting matrix is exactly $[\tilde{T}(p_ip_j^{-1})]$, which must be positive.
\end{proof}
Now the main result (Theorem \ref{thm.main}) can be deduced inductively:
\begin{proof}
First assume that $T:P\to\bh{H}$ is a representation that satisfies Condition (\ref{eq.right}), which has to be contractive. The goal is to show for any $n$ elements $p_1,p_2,\cdots,p_n\in P$, the operator matrix $[\tilde{T}(p_i p_j^{-1})]\geq 0$ and thus $T$ is regular. We proceed by induction on $n$.
For $n=1$, $\tilde{T}(p_1 p_1^{-1})=I\geq 0$.
For $n=2$, we have,
$$[\tilde{T}(p_i p_j^{-1})]= \left[ \begin{array}{cc}
I & \tilde{T}(p_1 p_2^{-1}) \\
\tilde{T}(p_2 p_1^{-1}) & I \end{array} \right].$$
Here, $\tilde{T}(p_2 p_1^{-1}) = \tilde{T}(p_1 p_2^{-1})^*$, and they are contractions since $T$ is contractive. Therefore, this $2\times 2$ operator matrix is positive.
Now assume that there is an $N$ such that for any $n<N$, we have $[\tilde{T}(p_i p_j^{-1})]\geq 0$ for any $p_1,p_2,\cdots,p_n\in P$. Consider the case when $n=N$:
For arbitrary choices $p_1,\cdots,p_N\in P$, let $g=\wedge_{i=1}^N p_i$, and replace $p_i$ by $p_i g^{-1}$. By doing so, $ p_i g^{-1} \left(p_j g^{-1}\right)^{-1}=p_i p_j^{-1}$, and thus they give the same matrix $[\tilde{T}(p_i p_j^{-1})]$. Moreover, $\wedge_{i=1}^n p_ig^{-1}=(\wedge_{i=1}^N p_i)g^{-1}=e$. Hence, without loss of generality, we may assume $\wedge_{i=1}^N p_i=e$.
Let $m$ be the smallest integer such that for all $J\subseteq \{1,2,\cdots, N\}$ and $|J|>m$, we have $\wedge_{j\in J} p_j=e$. It is clear that $m\leq N-1$. Now do induction on $m$:
For the base case when $m=1$, we have $p_i\wedge p_j=e$ for all $i\neq j$. Lemma \ref{lm.base} tells that Condition (\ref{eq.right}) implies $[\tilde{T}(p_i p_j^{-1})]\geq 0$.
Now assume $[\tilde{T}(p_i p_j^{-1})]\geq 0$ whenever $m\leq M-1<N-1$ and consider the case when $m=M$: For a subset $J\subseteq \{1,2,\cdots,n\}$ with $|J|=M$, let $g=\wedge_{j\in J} p_j$ and set $q_j=p_j g^{-1}$ for all $j\in J$, and $q_j=p_j$ otherwise. Lemma \ref{lm.reduce} concluded that $[\tilde{T}(p_i p_j^{-1})]\geq 0$ whenever $[\tilde{T}(q_i q_j^{-1})]\geq 0$ and the sub-matrix $[\tilde{T}(p_i p_j^{-1})]_{i,j\notin J}\geq 0$.
Since $|\{1,2,\cdots,N\}\backslash J|=N-M<N$, the induction hypothesis on $n$ implies that $[\tilde{T}(p_i p_j^{-1})]_{i,j\notin J}\geq 0$. Therefore, $[\tilde{T}(p_ip_j^{-1})]\geq 0$ whenever $[\tilde{T}(q_i q_j^{-1})]\geq 0$, and by dropping from $p_i$ to $q_i$, we may, without loss of generality, assume that $\wedge_{j\in J} p_j=e$. Repeat this process for all subsets $J\subset\{1,2,\cdots,n\}$ where $|J|=M$, and with Lemma \ref{lm.order}, we eventually reach a state when $\wedge_{j\in J} p_j=e$ for all $J\subseteq \{1,2,\cdots,N\}$, $|J|=M$. But in such case, for all $|J|\geq M$, we have $\wedge_{j\in J} p_j=e$. Therefore, we are in a situation where $m\leq M-1$. The result follows from the induction hypothesis on $m$.
Conversely, suppose that $T$ is regular. Fix $g\in P$ and $p_1,p_2,\cdots,p_k\in P$ where $g\wedge p_i=e$ for all $i=1,2,\cdots,k$. Denote $q_1=p_1g,q_2=p_2g,\cdots,q_k=p_kg$, and $q_{k+1}=p_1,q_{k+2}=p_2,\cdots,q_{2k}=p_k$. It follows from regularity that $[\tilde{T}(q_iq_j^{-1})]\geq 0$, which is equivalent to Condition (\ref{eq.right}) by Lemma \ref{lm.equiv}. \end{proof}
As an immediate consequence of Theorem \ref{thm.main}, we can show that isometric representations on any lattice ordered group must be regular.
\begin{corollary}\label{cor.iso} Let $T:P\to\bh{H}$ be an isometric representation of a lattice ordered semigroup. Then $T$ is regular.
\end{corollary}
\begin{proof}
Take $p_1,\cdots,p_n\in P$ and $g\in P$ with $g\wedge p_i=e$. It is clear that $g\wedge\left(p_i p_j^{-1}\right)_\pm =e$ and therefore $g$ commutes with each $\left(p_i p_j^{-1}\right)_\pm$. Hence,
\begin{eqnarray*}
T(g)^* \tilde{T}(p_ip_j^{-1}) T(g) &=& T(g)^* T((p_ip_j^{-1})_-)^* T((p_ip_j^{-1})_+) T(g) \\
&=& T((p_ip_j^{-1})_-)^*T(g)^* T(g) T((p_ip_j^{-1})_+) \\
&=& T((p_ip_j^{-1})_-)^* T((p_ip_j^{-1})_+) =\tilde{T}(p_ip_j^{-1}).
\end{eqnarray*}
Therefore, $[T(g)^* \tilde{T}(p_ip_j^{-1}) T(g)]=[\tilde{T}(p_ip_j^{-1})]$ and Condition (\ref{eq.right}) is satisfied. \end{proof}
For a contractive representation $T$, it would suffice to dilate it to an isometric representation. This provides an analog of \cite[Proposition 2.5.4]{DFK2014} on non-abelian lattice ordered groups.
\begin{corollary} Let $T:P\to\bh{H}$ be a contractive representation. Then $T$ is completely positive definite if and only if there exists an isometric representation $V:P\to\bh{K}$ such that $P_\mathcal{H} V(p)\big|_\mathcal{H}=T(p)$ for all $p\in P$. Such $V$ can be taken to be minimal in the sense that $\mathcal{K}=\bigvee_{p\in P} V(p)\mathcal{H}$.
In particular, $T$ is regular if and only if there exists such isometric dilation $V$ and in addition, $P_\mathcal{H} V(p)^* V(q)\big|_\mathcal{H}=T(p)^* T(q)$ for all $p,q\in P$ with $p\wedge q=e$.
\end{corollary}
\begin{proof}
When $T:P\to\bh{H}$ is completely positive definite and its extension $S$ to $G$ has minimal unitary dilation $U:G\to\bh{L}$, let $\mathcal{K}=\bigvee_{p\in P} U(p)\mathcal{H}$. It is clear that $\mathcal{K}$ is invariant for any $U(p)$, $p\in P$. Define a map $V:P\to\bh{K}$ via $V(p)=P_{\mathcal{K}} U(p)\big|_{\mathcal{K}}$, which must be isometric due to the invariance of $\mathcal{K}$. $V$ is an isometric dilation of $T$ that satisfies $P_\mathcal{H} V(p)|_\mathcal{H} = T(p)$, and $\mathcal{K}=\bigvee_{p\in P} V(p)\mathcal{H}$. In other words, $V$ is a minimal isometric dilation of $T$. In particular, when $T$ is regular, for any $p,q\in P$ with $p\wedge q=e$
\begin{eqnarray*}
T(p)^* T(q) &=& P_\mathcal{H} U(p)^* U(q)\big|_{\mathcal{H}}\\
&=& P_\mathcal{H} P_\mathcal{K} U(p)^* U(q)\big|_{\mathcal{K}}\big|_{\mathcal{H}} \\
&=& P_\mathcal{H} V(p)^* V(q)\big|_\mathcal{H}.
\end{eqnarray*}
Conversely, when $V:P\to\bh{K}$ is a minimal isometric dilation of $T$, Corollary \ref{cor.iso} implies that $V$ is regular and thus completely positive definite. There exists a unitary dilation $U:G\to\bh{L}$ where $P_\mathcal{K} U(p)\big|_{\mathcal{K}}=V(p)$. Therefore,
\begin{eqnarray*}
P_\mathcal{H} U(p)\big|_{\mathcal{H}} &=& P_\mathcal{H}P_\mathcal{K} U(p)\big|_{\mathcal{H}} \\
&=& P_\mathcal{H} V(p) \big|_{\mathcal{H}} = T(p).
\end{eqnarray*}
Hence, $U$ is also a unitary dilation of $T$ and thus $T$ is completely positive definite. Moreover, when $P_\mathcal{H} V(p)^* V(q)\big|_\mathcal{H}=T(p)^* T(q)$ for all $p,q\in P$ with $p\wedge q=e$, by the regularity of $V$,
$$P_\mathcal{H} U(p)^* U(q)\big|_\mathcal{H}=P_\mathcal{H} P_\mathcal{K}U(p)^* U(q)\big|_\mathcal{K}\big|_\mathcal{H}=T(p)^* T(q).$$
Therefore, $\tilde{T}(g)=T(g_-)^* T(g_+)$ is completely positive definite and $T$ is regular.
\end{proof}
\section{Nica-covariant Representations}
In this section, we answer the question of whether contractive Nica-covariant representations are regular. It suffices to show contractive Nica-covariant representations on lattice ordered groups satisfy Condition (\ref{eq.right}).
\begin{theorem}\label{thm.nc} A contractive Nica-covariant representation on a lattice ordered group is regular.
\end{theorem}
\begin{proof}
Let $p_1,\cdots,p_k\in P$ and $g\in P$ with $g\wedge p_i=e$ for all $i=1,2,\cdots,k$. $X=[\tilde{T}(p_i p_j^{-1})]$ and $D=\operatorname{diag}(T(g),T(g),\cdots,T(g))$. By Remark \ref{rm.pos}, we may assume $X\geq 0$.
Since for each $p_i,p_j\in P$, $\tilde{T}(p_ip_j^{-1})=T(p_{i,j}^-)^* T(p_{i,j}^+)$ where $e\leq p_{i,j}^\pm\leq p_i,p_j$. Hence, $g\wedge p_{i,j}^\pm=e$ and thus $g$ commutes with $p_{i,j}^\pm$. Therefore $T(g)$ commutes with $T(p_{i,j}^+)$ because $T$ is a representation and it also commutes with $T(p_{i,j}^-)^*$ by the Nica-covariant condition. As a result, $T(g)$ commutes with each entry in $X$, and thus $D$ commutes with $X$. Similarly, $D^*$ commutes with $X$ as well.
By continuous functional calculus, since $X\geq 0$, we know $D,D^*$ also commutes with $X^{1/2}$. Hence, in such case,
$$D^*XD=D^*X^{1/2}X^{1/2} D=X^{1/2} D^* D X^{1/2}\leq X. \qedhere$$ \end{proof}
It was shown in \cite[Proposition 2.5.10]{DFK2014} that a contractive Nica-covariant representation on abelian lattice ordered groups can be dilated to an isometric Nica-covariant representation. Here, we shall extend this result to non-abelian case.
\begin{corollary}\label{cor.NicaIso} Any minimal isometric dilation $V:P\to\bh{K}$ of a contractive Nica-covariant representation $T:P\to\bh{H}$ is also Nica-covariant.
\end{corollary}
\begin{proof} Let $T:P\to\bh{H}$ be a contractive Nica-covariant representation. Theorem \ref{thm.nc} implies that $T$ is regular, and thus by Theorem \ref{thm.NagyFoias}, it has a minimal unitary dilation $U:G\to\bh{L}$, which gives rise to a minimal isometric dilation $V:P\to\bh{K}$. Here $\mathcal{K}=\bigvee_{p\in P} V(p)\mathcal{H}$ and $V(p)=P_\mathcal{K} U(p)|_\mathcal{K}$. Notice that $\mathcal{K}$ is invariant for $U$ and therefore, $P_\mathcal{K}U(p)^* U(q)|_\mathcal{K}=V(p)^* V(q)$ for any $p,q\in P$. In particular, if $p\wedge q=e$, $p,q\in P$, we have from the regularity that
\begin{eqnarray*}
T(p)^* T(q) &=& P_\mathcal{H} U(p)^* U(q)|_\mathcal{H} \\
&=& P_\mathcal{H} (P_\mathcal{K}U(p)^* U(q)|_\mathcal{K}) |_\mathcal{H} \\
&=& P_\mathcal{H} V(p)^* V(q) |_\mathcal{H}.
\end{eqnarray*}
Now let $s,t\in P$ be such that $s\wedge t=e$. First, we shall prove $V(s)^* V(t)|_\mathcal{H}=V(t) V(s)^*|_\mathcal{H}$: Since $\{V(p) h: p\in P, h\in\mathcal{H}\}$ is dense in $\mathcal{K}$, it suffices to show for any $h,k\in\mathcal{H}$ and $p\in P$, $$\left\langle V(s)^* V(t) h, V(p) k\right\rangle=\left\langle V(t) V(s)^* h, V(p) k\right\rangle.$$
Start from the left,
\begin{eqnarray*}
& & \left\langle V(s)^* V(t) h, V(p) k\right\rangle \\
&=& \left\langle V(p)^* V(s)^* V(t) h, k\right\rangle = \left\langle V(sp)^* V(t) h, k\right\rangle \\
&=& \left\langle V((sp\wedge t)^{-1} sp)^* V(sp\wedge t)^* V(sp\wedge t) V((sp\wedge t)^{-1}t) h,k\right\rangle \\
&=& \left\langle V((sp\wedge t)^{-1} sp)^* V((sp\wedge t)^{-1}t) h, k\right\rangle \\
&=& \left\langle T((sp\wedge t)^{-1} sp)^* T((sp\wedge t)^{-1}t) h, k\right\rangle.
\end{eqnarray*}
The last equality follows from $\left((sp\wedge t)^{-1} sp\right)\wedge\left((sp\wedge t)^{-1}t\right)=e$ and thus,
$$T((sp\wedge t)^{-1} sp)^* T((sp\wedge t)^{-1}t) = P_\mathcal{H} V((sp\wedge t)^{-1} sp)^* V((sp\wedge t)^{-1}t) |_\mathcal{H}.$$
Since $s\wedge t=e$, Lemma \ref{lm.1b} implies that $sp\wedge t=p\wedge t$. Notice $(p\wedge t)\wedge s\leq t\wedge s=e$, and thus by Property (4) of Lemma \ref{lm.LatticeBasic}, $s$ commutes with $p\wedge t$. By the Nica-covariance of $T$, this also implies $T(s)^*$ commutes with $T((p\wedge t)^{-1}t)$. Put all these back to the equation:
\begin{eqnarray*}
& & \left\langle T((sp\wedge t)^{-1} sp)^* T((sp\wedge t)^{-1}t) h, k\right\rangle \\
&=& \left\langle T( s(p\wedge t)^{-1}p)^*T((p\wedge t)^{-1}t) h, k\right\rangle \\
&=& \left\langle T((p\wedge t)^{-1}p)^*T(s)^* T((p\wedge t)^{-1}t) h, k\right\rangle \\
&=& \left\langle T((p\wedge t)^{-1}p)^* T((p\wedge t)^{-1}t) \left(T(s)^* h\right), k\right\rangle \\
&=& \left\langle V((p\wedge t)^{-1}p)^* V((p\wedge t)^{-1}t) \left(T(s)^* h\right), k\right\rangle \\
&=& \left\langle V((p\wedge t)^{-1}p)^* V((p\wedge t)^{-1}t) \left(V(s)^* h\right), k\right\rangle \\
&=& \left\langle V(p)^* V(t) \left(V(s)^* h\right), k\right\rangle =\left\langle V(t) V(s)^* h, V(p) k\right\rangle.
\end{eqnarray*}
Here we used the fact that $P_\mathcal{H} V(p)^* V(q) |_\mathcal{H}=T(p)^* T(q)$ whenever $p\wedge q=e$. Also, that $\mathcal{H}$ is invariant under $V(s)^*$, so that $T(s)^* h\in\mathcal{K}$ is the same as $V(s)^* h$.
Now to show $V(s)^* V(t)=V(t) V(s)^*$ in general, it suffices to show for every $p\in P$, $V(s)^* V(t) V(p)|_\mathcal{H}=V(t) V(s)^* V(p)|_\mathcal{H}$. Start with the left hand side and repeatedly use similar argument as above,
\begin{eqnarray*}
& & V(s)^* V(t) V(p)|_\mathcal{H} \\
&=& V(s)^* V_{tp}|_\mathcal{H} = V((s\wedge tp)^{-1}s)^* V((s\wedge tp)^{-1}tp)|_\mathcal{H} \\
&=& V(t(s\wedge p)^{-1}p) V((s\wedge p)^{-1}s)^* |_\mathcal{H} \\
&=& V(t (s\wedge p)^{-1}p) V((s\wedge p)^{-1}s)^* |_\mathcal{H} \\
&=& V(t) V((s\wedge p)^{-1}s)^* V((s\wedge p)^{-1}p) |_\mathcal{H} = V(t) V(s)^* V(p)|_\mathcal{H}.
\end{eqnarray*}
This finishes the proof. \end{proof}
\section{Row and Column Contractions}
A commuting $n$-tuple $(T_1,\cdots,T_n)$ where each $T_i\in\bh{H}$ is called a \emph{row contraction} if $\sum_{i=1}^n T_i T_i^* \leq I$. Equivalently, the operator $[T_1,T_2,\cdots,T_n]\in\mathcal{B}(\mathcal{H}^n,\mathcal{H})$ is contractive. It can be naturally associated with a contractive representation $T:\mathbb{Z}_+^n\to\bh{H}$ that sends the $i$-th generator $e_i$ to $T_i$. There is a dual definition called column contractions, when $T_i$ satisfies $\sum_{i=1}^n T_i^* T_i \leq I$. It is clear that $T$ is a row contraction if and only if $T^*$ is a column contraction.
As an immediate corollary to Brehmer's theorem (Theorem \ref{thm.Brehmer}), a column contraction $T$ is always right regular \cite[Proposition I.9.2]{SFBook}, and therefore a row contraction $T$ is always left regular. This section generalizes the notion of row contraction to arbitrary lattice ordered groups and establishes a similar result.
\begin{definition}\label{def.rc} Let $T:P\to\bh{H}$ be a contractive representation of a lattice ordered group $(G,P)$. $T$ is called \emph{row contractive} if for any $p_1,\cdots,p_n\in P$ where $p_i\neq e$ and $p_i\wedge p_j=e$ for all $i\neq j$, $$\sum_{i=1}^n T(p_i)T(p_i)^*\leq I.$$
Dually, $T$ is called \emph{column contractive} if for such $p_i$, $$\sum_{i=1}^n T(p_i)^*T(p_i)\leq I.$$
\end{definition}
\begin{remark} Definition \ref{def.rc} indeed generalized the notion of commuting row contractions: when the group is $(\mathbb{Z}^\Omega,\mathbb{Z}_+^\Omega)$ where $\Omega$ is countable, a representation $T:\mathbb{Z}_+^\Omega\to\bh{H}$ is uniquely determined by its value on the generators $T_\omega=T(e_\omega)$. $T$ is called commuting row contraction when $\sum_{\omega\in\Omega} T_\omega T_\omega^* \leq I$. For any $p_1,\cdots,p_k\in\mathbb{Z}_+^\Omega$ where $p_i\wedge p_j=0$ for all $i\neq j$ and $p_i\neq 0$, each $p_i$ can be seen as a function from $\Omega$ to $\mathbb{Z}_+$ with finite support. Let $S_i\subseteq\Omega$ be the support of $p_i$, which is non-empty since $p_i\neq 0$. We have $S_i\bigcap S_j=\emptyset$ since $p_i\wedge p_j=0$. Therefore, pick any $\omega_i\in S_i$ and by $T$ contractive, $T(\omega_i) T(\omega_i)^*\geq T(p_i) T(p_i)^*$. Since $S_i$ are pairwise-disjoint, $\omega_i$ are distinct. Therefore, we get that
$$\sum_{i=1}^n T(p_i) T(p_i)^* \leq \sum_{i=1}^n T(\omega_i) T(\omega_i)^* \leq I.$$
and thus $T$ satisfies the Definition \ref{def.rc}. Hence, two definitions coincides on $(\mathbb{Z}^\Omega,\mathbb{Z}_+^\Omega)$.
\end{remark}
Our goal is to prove the following result:
\begin{theorem} A column contractive representation is right regular. Therefore, a row contractive representation is left regular.
\end{theorem}
We shall proceed with a method similar to the proof of Theorem \ref{thm.main}.
\begin{lemma}\label{lm.rcbase} Let $T$ be a column contractive representation. Let $p_1,\cdots,p_n\in P$ and $g_1,\cdots,g_k\in P$ where $p_i\wedge p_{i'}=p_i\wedge g_j=g_j\wedge g_{j'}=e$ for all $1\leq i\neq i'\leq n$ and $1\leq j\neq j'\leq k$. Moreover, assume that $g_i\neq e$. Denote $X=[\tilde{T}(p_i p_j^{-1})]$ and $D_i=\diag(T(g_i),\cdots,T(g_i))$. Then, $$\sum_{i=1}^k D_i^* X D_i\leq X.$$
\end{lemma}
\begin{proof}
The statement is clearly true for all $k$ when $n=1$. Now assuming it is true for all $k$ whenever $n<N$, and consider the case when $n=N$:
It is clear that when all of the $p_i$ are equal to $e$, then $X-\sum_{i=1}^k D_i^* X D_i$ is a $n\times n$ matrix whose entries are all equal to $I-\sum_{i=1}^k T(g_i)^* T(g_i)\geq 0$, and thus the statement is true. Otherwise, we may assume without loss of generality that $p_1\neq e$. Let $q_1=e$ and $q_2=p_2,\cdots,q_n=p_n$. Denote $X_0=[\tilde{T}(q_i q_j^{-1})]$ and $E=\diag(I,T(p_1),\cdots,T(p_1))$ be a $n\times n$ block diagonal matrix.
Denote $Y=[\tilde{T}(p_i p_j^{-1})]_{2\leq i,j\leq n}$ and set $E_i=\diag(T(g_i),\cdots,T(g_i))$ be a $(n-1)\times(n-1)$ block diagonal matrix. Finally, set $E_{k+1}=\diag(T(p_1),\cdots,T(p_1))$ be a $(n-1)\times(n-1)$ block diagonal matrix.
From the proof of Theorem \ref{thm.main}, $$X=E^* X_0 E +\begin{bmatrix} 0 & 0 \\ 0 & Y-E_{k+1}^* Y E_{k+1} \end{bmatrix}.$$
Now $Y$ is a matrix of smaller size and thus by induction hypothesis, $$\sum_{i=1}^{k+1} E_i^* Y E_i\leq Y.$$
Hence,
\begin{eqnarray*}
Y-E_{k+1}^* Y E_{k+1} &\geq& \sum_{i=1}^k E_i^* Y E_i \\
&\geq& \sum_{i=1}^k E_i^* (Y-E_{k+1}^* Y E_{k+1}) E_i.
\end{eqnarray*}
Also notice that $E$ commutes with $D_i$ and therefore, if $\sum_{i=1}^k D_i^* X_0 D_i\leq X_0$, we have
\begin{eqnarray*}
& &\sum_{i=1}^k D_i^* X D_i \\
&=& E^*\left(\sum_{i=1}^k D_i^* X_0 D_i\right)E+\begin{bmatrix} 0 & 0 \\ 0 & \sum_{i=1}^k E_i^*(Y-E_{k+1}^* Y E_{k+1})E_i \end{bmatrix} \\
&\leq& E^* X_0 E +\begin{bmatrix} 0 & 0 \\ 0 & Y-E_{k+1}^* Y E_{k+1} \end{bmatrix} = X.
\end{eqnarray*}
Hence, $\sum_{i=1}^k D_i^* X D_i\leq X$ if $\sum_{i=1}^k D_i^* X_0 D_i\leq X_0$. This reduction from $X$ to $X_0$ changes one $p_i\neq e$ to $e$, and therefore by repeating this process, we eventually reach a state where all $p_i=e$. \end{proof}
The main result can be deduced immediately from the following Proposition:
\begin{proposition} Let $T$ be a column contractive representation on a lattice ordered semigroup $P$. Let $p_1,\cdots,p_n\in P$ and $g_1,\cdots,g_k\in P$ where $g_i\wedge p_j=e$ and $g_i\wedge g_l=e$ for all $i\leq l$. Assuming $g_i\neq e$ and denote $X=[\tilde{T}(p_i p_j^{-1})]$ and $D_i=\diag(T(g_i),\cdots,T(g_i))$. Then $$\sum_{i=1}^k D_i^* X D_i\leq X.$$ In particular, Condition (\ref{eq.right}) is satisfied when $k=1$.
\end{proposition}
\begin{proof}
The statement is clear when $n=1$. Assuming it's true for $n<N$, and consider the case when $n=N$: Let $m$ be the smallest integer such that for all $J\subseteq \{1,2,\cdots, N\}$ and $|J|>m$, $\wedge_{j\in J} p_j=e$. It was observed in the proof of Theorem \ref{thm.main} that $m\leq N-1$. Proceed by induction on $m$:
In the base case when $m=1$, $p_i\wedge p_j=e$ for all $i\neq j$, the statement is shown in Lemma \ref{lm.rcbase}. Assuming the statement is true for $m<M-1<N-1$ and consider the case when $m=M$. For each $J\subseteq\{1,2,\cdots,N\}$ with $|J|=M$ and $\wedge_{j=1}^M p_j=g\neq e$, denote $q_i=p_i$ when $i\notin J$ and $q_i= q_i g^{-1}$ when $i\in J$. Let $X_0=[\tilde{T}(q_i q_j^{-1})]$ and $E$ be a block diagonal matrix whose $i$-th diagonal entry is $I$ when $i\notin J$ and $T(g)$ otherwise. Denote $Y=[\tilde{T}(q_i q_j^{-1})]_{i,j\notin J}$ and $E_i=\diag(T(g_i),\cdots,T(g_i))$ with $N-M$ copies of $T(g_i)$. Finally, let $E_{k+1}=\diag(T(g),\cdots,T(g))$ with $N-M$ copies of $T(g)$.
From the proof of Theorem \ref{thm.main}, by assuming without loss of generality that $J=\{1,2,\cdots,M\}$, we have
$$X=E^* X_0 E+\begin{bmatrix} 0 & 0 \\ 0 & Y-E_{k+1}^* Y E_{k+1} \end{bmatrix}.$$
Now $Y$ has a smaller size and thus by induction hypothesis on $n$,
$$\sum_{i=1}^{k+1} E_i^* Y E_i\leq Y.$$
and thus
\begin{eqnarray*}
Y-E_{k+1}^* Y E_{k+1} &\geq& \sum_{i=1}^k E_i^* Y E_i \\
&\geq& \sum_{i=1}^k E_i^* (Y-E_{k+1}^* Y E_{k+1}) E_i.
\end{eqnarray*}
Therefore, if $\sum_{i=1}^k D_i^* X_0 D_i\leq X_0$,
\begin{eqnarray*}
& & \sum_{i=1}^k D_i^* X D_i \\
&=& E^*\left(\sum_{i=1}^k D_i^* X_0 D_i\right)E+\begin{bmatrix} 0 & 0 \\ 0 & \sum_{i=1}^k E_i^*(Y-E_{k+1}^* Y E_{k+1})E_i \end{bmatrix} \\
&\leq& E^* X_0 E +\begin{bmatrix} 0 & 0 \\ 0 & Y-E_{k+1}^* Y E_{k+1} \end{bmatrix} = X.
\end{eqnarray*}
Hence, the statement is true for $p_i$ if it is true for $q_i$, where $\wedge_{j\in J} q_j=e$. Repeat the process until all such $|J|=M$ has $\wedge_{j\in J} p_j=e$, which reduces to a case where $m<M$. This finishes the induction. Notice Condition (\ref{eq.right}) is clearly true when $g=e$, and when $g\neq e$, it is shown by the case when $m=1$. This finishes the proof. \end{proof}
\section{Brehmer's Condition}
Brehmer \cite{Brehmer1961} established a necessary and sufficient condition for a representation on $P=\mathbb{Z}_+^\Omega$ to be regular (see Theorem \ref{thm.Brehmer}). This section explores how Brehmer's result relates to Condition (\ref{eq.right}) without invoking their equivalence to regularity. In particular, we show that Brehmer's condition allows us to decompose certain $X=[\tilde{T}(p_i-p_j)]$ as a product $R^* R$, where $R$ is an upper triangular matrix.
Let $\{T_\omega\}_{\omega\in\Omega}$ be a family of commuting contractions, which leads to a contractive representation on $\mathbb{Z}_+^\Omega$ by sending each $e_\omega$ to $T_\omega$. For each $U\subseteq\Omega$, denote
$$Z_U=\sum_{V\subseteq U} (-1)^{|V|} T(e_V)^* T(e_V).$$
For example,
\begin{eqnarray*}
Z_{\emptyset} &=& I \\
Z_{\{1\}} &=& I-T_1^* T_1 \\
Z_{\{1,2\}} &=& Z_{\{1\}}-T_2^* Z_{\{1\}} T_2 = I-T_1^*T_1-T_2^*T_2+T_2^*T_1^*T_1T_2 \\
&\vdots&
\end{eqnarray*}
Brehmer's theorem stated that $T$ is regular if and only if $Z_U\geq 0$ for any finite subset $U\subseteq\Omega$. We shall first transform Brehmer's condition into an equivalent form.
\begin{lemma}\label{lm.brehmer1} $Z_U\geq 0$ for each finite subset $U\subseteq \Omega$ if and only if for any finite set $J\subseteq\Omega$ and $\omega\in\Omega$, $\omega\notin J$, $$T_\omega^* Z_J T_\omega\leq Z_J.$$
\end{lemma}
\begin{proof}
Take any finite subset $J\subseteq\Omega$ and $\omega\in\Omega$, $\omega\notin J$.
\begin{eqnarray*}
& & Z_J-T_\omega^* Z_J T_\omega\\
&=& \sum_{V\subseteq J} (-1)^{|V|} T(e_V)^* T(e_V) + \sum_{V\subseteq J} (-1)^{|V|+1} T_\omega^*T(e_V)^* T(e_V)T_\omega \\
&=& \sum_{V\subseteq \{\omega\}\bigcup J, \omega\notin V} (-1)^{|V|} T(e_V)^* T(e_V) + \sum_{V\subseteq \{\omega\}\bigcup J, \omega\in V} (-1)^{|V|} T(e_V)^* T(e_V) \\
&=& Z_{\{\omega\}\bigcup J}.
\end{eqnarray*}
Therefore, $T_\omega^* Z_J T_\omega\leq Z_J$ if and only if $Z_{\{\omega\}\bigcup J}\geq 0$. This finishes the proof.\end{proof}
A major tool is the following version of Douglas Lemma \cite{DouglasLemma}:
\begin{lemma}[Douglas]\label{lm.Douglas} For $A,B\in\bh{H}$, $A^*A\leq B^*B$ if and only if there exists a contraction $C$ such that $A=CB$.
\end{lemma}
As an immediate consequence of Lemma \ref{lm.Douglas}, $T_\omega^* Z_J T_\omega\leq Z_J$ is satisfied if and only if there is a contraction $W_{\omega,J}$ such that $Z_J^{1/2} T_\omega=W_{\omega,J} Z_J^{1/2}$. Therefore, it would suffices to find such contraction $W_{\omega,J}$ for each finite subset $J\subseteq\Omega$ and $\omega\in\Omega$, $\omega\notin J$. By symmetry, it would suffices to do so for each $J_n=\{1,2,\cdots,n\}$ and $\omega_n=n+1$. Without loss of generality, we shall assume that $\Omega=\mathbb{N}$.
Consider $\mathcal{P}(J_n)=\{U\subseteq J_n\}$, and denote $p_U=\sum_{i\in U} e_i\in\mathbb{Z}_+^\Omega$. Denote $X_n=[\tilde{T}(p_U-p_V)]$ where $U$ is the row index and $V$ is the column index.
\begin{lemma}\label{lm.telescope} Assuming $Z_J\geq 0$ for all $J\subseteq J_n$. Then for a fixed $F\subseteq J_n$, we have, $$\sum_{U\subseteq F} T_U^* Z_{F\backslash U} T_U = I.$$
\end{lemma}
\begin{proof}
We first notice that by definition, $Z_J=\sum_{U\subseteq J} (-1)^{|U|} T_U^* T_U$. Therefore,
$$\sum_{U\subseteq F} T_U^* Z_{F\backslash U} T_U = \sum_{U\subseteq F} \sum_{V\subseteq F\backslash U} (-1)^{|V|} T_{U\bigcup V}^* T_{U\bigcup V}.$$
For a fixed set $W\subseteq F$, consider the coefficient of $T_W^* T_W$ in the double summation. It appears in the expansion of every $T_U^* Z_{F\backslash U} T_U$, where $U\subseteq W$, and its coefficient in the expansion of such term is equal to $(-1)^{|W\backslash U|}$. Therefore, the coefficient of $T_W^* T_W$ is equal to $$\sum_{U\subseteq W} (-1)^{|W\backslash U|}=\sum_{i=0}^{|W|} {|W|\choose i} (-1)^i.$$
This evaluates to $0$ when $|W|>0$ and $1$ when $|W|=0$, in which case, $W=\emptyset$ and $T_W=I$. \end{proof}
Now can now decompose $X_n=R_n^* R_n$ explicitly.
\begin{proposition}\label{prop.decomp} Assuming $Z_J\geq 0$ for all $J\subseteq J_n$. Define a block matrix $R_n$, whose rows and columns are indexed by $\mathcal{P}(J_n)$, by $R_n(U,V)=Z_{J_n\backslash U}^{1/2} T_{U\backslash V}$ whenever $V\subseteq U$ and $0$ otherwise. Then $X_n=R_n^* R_n$
\end{proposition}
\begin{proof} Fix $U,V\subseteq J_n$, the $(U,V)$-entry in $X_n$ is $\tilde{T}(p_U-p_V)=T_{V\backslash U}^* T_{U\backslash V}$. Now the $(U,V)$-entry in $R_n^* R_n$ is equal to
$$\sum_{W\subseteq J_n} R_n(W,U)^* R_n(W,V).$$
It follows from the definition that $R_n(W,U)^* R_n(W,V)=0$ unless $U,V\subseteq W$, and thus $U\bigcup V\subseteq W$. Hence,
\begin{eqnarray*}
& &\sum_{W\in\mathcal{P}(J_n)} R_n(W,U)^* R_n(W,V) \\
&=& \sum_{U\bigcup V\subseteq W} T_{W\backslash U}^* Z_{J_n\backslash W} T_{W\backslash V} \\
&=& \sum_{U\bigcup V\subseteq W} T_{V\backslash U}^* T_{W\backslash(U\bigcup V)}^* Z_{J_n\backslash W} T_{W\backslash(U\bigcup V)} T_{W\backslash U} \\
&=& T_{V\backslash U}^* \left(\sum_{U\bigcup V\subseteq W} T_{W\backslash(U\bigcup V)}^* Z_{J_n\backslash W} T_{W\backslash(U\bigcup V)}\right) T_{W\backslash U}.
\end{eqnarray*}
If we denote $F=J_n\backslash(U\bigcup V)$ and $W'=W\backslash (U\bigcup V)$, since $U\bigcup V\subseteq W$, we have $J_n\backslash W=F\backslash W'$. Hence the summation becomes $$\sum_{U\bigcup V\subseteq W} T_{W\backslash(U\bigcup V)}^* Z_{J_n\backslash W} T_{W\backslash(U\bigcup V)}=\sum_{W'\subseteq F} T_{W'}^* Z_{F\backslash W'} T_{W'},$$ which by Lemma \ref{lm.telescope} is equal to $I$. Therefore, the $(U,V)$-entry in $R_n^* R_n$ is equal to $T_{V\backslash U}^* T_{W\backslash U}$ and $X_n=R_n^* R_n$
\end{proof}
\begin{remark} If we order the subsets of $J_n$ by cardinality and put larger sets first, then since $R_n(U,V)\neq 0$ only when $V\subseteq U$, $R_n$ becomes a lower triangular matrix. In particular, the row of $\emptyset$ contains exactly one non-zero entry, which is $Z_{J_n}^{1/2}$ at $(\emptyset,\emptyset)$.
\end{remark}
\begin{example} Let us consider the case when $n=2$, and $J_2$ has 4 subsets $\{1,2\}$, $\{2\}$,$\{1\}$,$\emptyset$. Under this ordering, $$X_n=\begin{bmatrix} I & T_1 & T_2 & T_1T_2 \\ T_1^* & I & T_1^* T_2 & T_2 \\ T_2^* & T_2^* T_1 & I & T_1 \\ T_1^* T_2^* & T_2^* & T_1^* & I \end{bmatrix}.$$ Proposition \ref{prop.decomp} gives that $$R_n=\begin{bmatrix} I & T_1 & T_2 & T_1 T_2 \\ 0 & Z_1^{1/2} & 0 & Z_1^{1/2} T_2 \\ 0 & 0 & Z_2^{1/2} & Z_2^{1/2} T_1 \\ 0 & 0 & 0 & Z_{1,2}^{1/2} \end{bmatrix}$$ satisfies $R_n^* R_n=X_n$.
\end{example}
We can now prove Brehmer's condition from Condition (\ref{eq.right}) without invoking their equivalence to regularity.
\begin{proposition} In the case of $T:\mathbb{Z}_+^\Omega\to\bh{H}$, Condition (\ref{eq.right}) implies Brehmer's condition.
\end{proposition}
\begin{proof} Without loss of generality, we may assume $\Omega=\mathbb{N}$. We shall proceed by induction on the size of $J\subseteq \mathbb{N}$.
For $|J|=1$ (i.e. $J=\{\omega\}$), Condition (\ref{eq.right}) implies $T$ is contractive. Hence, $Z_{J}=I-T_\omega^* T_\omega\geq 0$. Assuming $Z_{J}\geq 0$ for all $|J|\leq n$, and consider the case when $|J|=n+1$. By symmetry, it would suffices to show this for $J=J_{n+1}=\{1,2,\cdots,n+1\}$.
By Proposition \ref{prop.decomp}, $X_n=R_n^* R_n$ where the $(\emptyset,\emptyset)$-entry of $R_n$ is equal to $Z_{J_n}^{1/2}$. Let $D_n$ be a block diagonal matrix with $2^n$ copies of $T_{n+1}$ along the diagonal. Condition (\ref{eq.right}) implies that $$D_n^* X_n D_n =D_n^* R_n^* R_n D_n\leq X_n=R_n^* R_n.$$ Therefore, by Lemma \ref{lm.Douglas}, there exists a contraction $W_n$ such that $W_n R_n=R_n D_n$. By comparing the $(\emptyset,\emptyset)$-entry on both sides, there exists $C_n$ such that $C_n Z_{J_n}^{1/2}=Z_{J_n}^{1/2} T_{n+1}$, where $C_n$ is the $(\emptyset,\emptyset)$-entry of $W_n$, which must be contractive as well. Hence, by Lemma \ref{lm.brehmer1} and \ref{lm.Douglas}, $$Z_{J_{n+1}}=Z_{J_n}-T_{n+1}^* Z_{J_n}T_{n+1}\geq 0.$$ This finishes the proof.
\end{proof}
\section{Covariant Representations}
The semicrossed products of a dynamical system by Nica-covariant representations was discussed in \cite{Fuller2013, DFK2014}, where its regularity is seen as a key to many results. Our result on the regularity of Nica-covariant representations (Theorem \ref{thm.nc} and Corollary \ref{cor.NicaIso}) allows us to generalize some of the results to arbitrary lattice ordered abelian groups.
\begin{definition} A $C^*$-dynamical system is a triple $(A,\alpha,P)$ where
\begin{enumerate}
\item $A$ is a $C^*$-algebra;
\item $\alpha:P\to\operatorname{End}(A)$ maps each $p\in P$ to a $*$-endomorphism on $A$;
\item $P$ is a spanning cone of some group $G$.
\end{enumerate}
\end{definition}
\begin{definition} A pair $(\pi,T)$ is called a \emph{covariant pair} for a $C^*$-dynamical system if
\begin{enumerate}
\item $\pi: A\to\bh{H}$ is a $*$-representation;
\item $T:P\to\bh{H}$ is a contractive representation of $P$;
\item $\pi(a) T(s)=T(s) \pi(\alpha_s(a))$ for all $s\in P$ and $a\in A$.
\end{enumerate}
In particular, a covariant pair $(\pi,T)$ is called Nica-covariant/isometric, if $T$ is Nica-covariant/isometric.
\end{definition}
The main goal is to prove that Nica-covariant pairs on $C^*$-dynamical systems can be lifted to isometric Nica-covariant pairs. This can be seen from \cite[Theorem 4.1.2]{DFK2014} and Corollary \ref{cor.NicaIso}. However, we shall present a slightly different approach by taking the advantage of the structure of lattice ordered abelian group.
\begin{theorem}\label{thm.semi} Let $(A,\alpha,P)$ be a $C^*$-dynamical system over a positive cone $P$ of a lattice ordered abelian group $G$. Let $\pi:A\to\bh{H}$ and $T:P\to\bh{H}$ form a Nica-covariant pair $(\pi,T)$ for this $C^*$-dynamical system. If $V:P\to\mathcal{K}$ is a minimal isometric dilation of $T$, then there is an isometric Nica-covariant pair $(\rho, V)$ such that for all $a\in A$,
$$P_\mathcal{H}\rho(a)\big|_{\mathcal{H}}=\pi(a).$$
Moreover, $\mathcal{H}$ is invariant for $\rho(a)$.
\end{theorem}
\begin{proof}
Fix a minimal dilation $V$ of $T$ and consider any $h\in\mathcal{H}$, $p\in P$, and $a\in A$: define $$\rho(a) V(p) h = V(p) \pi(\alpha_p(a)) h$$ We shall first show that this is a well defined map. First of all, since $V$ is a minimal isometric dilation, the set $\{V(p) h\}$ is dense in $\mathcal{K}$. Suppose $V(p) h_1=V(s) h_2$ for some $p,s\in P$ and $h_1,h_2\in\mathcal{H}$. It suffices to show that for any $t\in P$ and $h\in\mathcal{H}$, we have
\begin{equation}
\left\langle V(p) \pi(\alpha_p(a)) h_1, V(t) h\right\rangle=\left\langle V(s) \pi(\alpha_s(a)) h_2, V(t) h\right\rangle.
\end{equation}
Since $A$ is a $C^*$-dynamical system, it follows from the covariant condition $\pi(a) T(s)=T(s)\pi(\alpha_s(a))$ that $T(s)^* \pi(a)=\pi(\alpha_s(a)) T(s)^*$. Hence,
\begin{eqnarray*}
& & \left\langle V(p) \pi(\alpha_p(a)) h_1, V(t) h\right\rangle \\
&=& \left\langle V(t)^*V(p) \pi(\alpha_p(a)) h_1, h\right\rangle \\
&=& \left\langle V(t-t\wedge p)^*V(p-t\wedge p) \pi(\alpha_p(a)) h_1, h\right\rangle \\
&=& \left\langle T(t-t\wedge p)^*T(p-t\wedge p) \pi(\alpha_p(a)) h_1, h\right\rangle \\
&=& \left\langle \pi(\alpha_{p-(p-t\wedge p)+(t-t\wedge p)}(a)) T(t-t\wedge p)^* T(p-t\wedge p)h_1, h\right\rangle \\
&=& \left\langle \pi(\alpha_{t}(a)) T(t-t\wedge p)^* T(p-t\wedge p)h_1, h\right\rangle.
\end{eqnarray*}
Here we used that fact that $V$ is regular and thus $$P_\mathcal{H} V(t-t\wedge p)^* V(p-t\wedge p)\big|_\mathcal{H}=T(t-t\wedge p)^* T(p-t\wedge p).$$ Now notice that
\begin{eqnarray*}
T(t-t\wedge p)^* T(p-t\wedge p)h_1 &=& P_\mathcal{H} V(t-t\wedge p)^* V(p-t\wedge p) h_1 \\
&=& P_\mathcal{H} V(t)^* V(p) h_1.
\end{eqnarray*}
Similarly,
$$\left\langle V(s) \pi(\alpha_s(a)) h_2, V(t) h\right\rangle= \left\langle \pi(\alpha_{t}(a)) T(t-t\wedge s)^* T(s-t\wedge s)h_2, h\right\rangle,$$
where $$T(t-t\wedge s)^* T(s-t\wedge s)h_2= P_\mathcal{H} V(t)^* V(s) h_2=P_\mathcal{H} V(t)^* V(p) h_1.$$
Therefore, $\rho$ is well defined on the dense subset $\{V(p) h\}$.
Since $V(p)$ is isometric and $\pi,\alpha$ are completely contractive, $$\|V(p) \pi(\alpha_p(a)) h\|=\|\pi(\alpha_p(a)) h\|\leq\|h\|=\|V(p) h\|,$$
and thus $\rho(a)$ is contractive on $\{V(p) h\}$. Hence, $\rho(a)$ can be extended to a contractive map on $\mathcal{K}$. Moreover, for any $h\in\mathcal{H}$ and $a\in A$, we have $\rho(a) h = \pi(a) h\in\mathcal{H}$, and thus $\mathcal{H}$ is invariant for $\rho$. For any $a,b\in A$, $p\in P$, and $h\in\mathcal{H}$,
\begin{eqnarray*}
\rho(a)\rho(b) V(p) h &=& V(p) \pi(\alpha_p(a))\pi(\alpha_p(b)) h \\
&=& V(p) \pi(\alpha_p(ab)) h \\
&=& \rho(ab) V(p) h.
\end{eqnarray*}
Therefore, $\rho$ is a contractive representation of $A$ and thus a $*$-representation. Now for any $p,t\in P$ and $h\in\mathcal{H}$,
\begin{eqnarray*}
\rho(a)V(p) V(t)h &=& V(p+t) \pi(\alpha_{p+t}(a))h \\
&=& V(p) V(t) \rho(\alpha_{p+t}(a))h\\
&=& V(p) \rho(\alpha_p(a)) V(t) h.
\end{eqnarray*}
Hence, $(\rho,V)$ is an isometric Nica-covariant pair. \end{proof}
This lifting of contractive Nica-covariant pairs to isometric Nica-covariant pairs has significant implication in its associated semi-crossed product. A family of covariant pairs gives rise to a semi-crossed product algebra in the following way \cite{Fuller2013, DFK2014}. For a $C^*$-dynamical system $(A,\alpha,P)$, denote $\mathcal{P}(A,P)$ be the algebra of all formal polynomials $q$ of the form
$$q=\sum_{i=1}^n e_{p_i} a_{p_i},$$
where $p_i\in P$ and $a_{p_i}\in A$. The multiplication on such polynomials follows the rule that $a e_s = e_s \alpha(a)$ and $e_p e_q=e_{pq}$. For a covariant pair $(\sigma,T)$ on this dynamical system, define a representation of $\mathcal{P}(A,P)$ by $$(\sigma\times T)\left(\sum_{i=1}^n e_{p_i} a_{p_i}\right)=\sum_{i=1}^n T(p_i) \sigma(a_{p_i}).$$
Now let $\mathcal{F}$ be a family of covariant pairs on this dynamical system. We may define a norm on $\mathcal{P}(A,S)$ by
$$\|p\|_\mathcal{F}=\sup\{(\sigma\times T)(p):(\sigma,T)\in\mathcal{F}\},$$
and the semi-crossed product algebra is defined as
$$A\times_\alpha^\mathcal{F} P=\overline{\mathcal{P}(A,S)}^{\|\cdot\|_\mathcal{F}}.$$
In particular, $A \times_\alpha^{nc} P$ is determined by the Nica-covariant representations, and $A \times_\alpha^{nc,iso} P$ is determined by the isometric Nica-covariant representation. As an immediate corollary from Theorem \ref{thm.main} and \ref{thm.semi},
\begin{corollary} For a $C^*$-dynamical system $(A,\alpha,P)$, the semi-crossed product algebra given by Nica-covariant pairs agrees with that given by isometric Nica-covariant pairs. In other words, $$A \times_\alpha^{nc} P \cong A \times_\alpha^{nc,iso} P.$$
\end{corollary}
\section*{Acknowledgements}
I would like to thank Professor Ken Davidson for pointing out this area of research and giving me directions. I would also like to thank Adam Fuller and Evgenios Kakariadis for many valuable comments.
\bibliographystyle{plain}
|
\section{Introduction}
\label{intro}
For a pair of points in a smooth convex planar curve, its mid-line is the line that passes through its mid-point and the intersection of the corres\-pon\-ding tangent lines. The envelope of the $2$-parameter family of mid-lines is an important affine invariant symmetry set associated with the curve. It has applications in computer graphics and has been studied by many authors
(\cite{Sapiro},\cite{Giblin},\cite{Giblin2},\cite{Holtom}). It is well-known that the limit of the mid-lines when one point tends to the other is the affine normal and
the limit of the envelope of mid-lines is its affine evolute.
The mid-plane is a natural generalization of the concept of mid-line to a smooth locally convex surface $S$: For $p_1,p_2\in S$, the mid-plane is the plane that passes through the mid-point $M$ of $p_1$ and $p_2$ and the intersection line of the tangent planes at $p_1$ and $p_2$. We verify that if we fix a tangent vector $T$ and make $p_2\to p_1$ along this direction, the mid-plane converges to the Transon plane of the surface at $p_1$ in the direction $T$ (\cite{Transon}).
What can we say about the envelope of the mid-planes of a surface $S$ in $3$-space? In this paper we are interested in understanding this set when
$p_2\to p_1$. It turns out that, when $p_2\to p_1$ along the direction $T$, the limit of this envelope consists of the solution of a system
of $4$ equations. The first equation of this system defines the Transon plane of $T$ at $p_1$, the first and second define the line of the cone of B.Su of $T$ at $p_1$, and the first, second and third equations define the center of the Moutard quadric of $T$ at $p_1$. The notions of Transon Plane, cone of B.Su and Moutard's quadric are classical, but there are some modern references (\cite{Juttler},\cite{Buchin}). The close connection between these classical notions of affine differential geometry and the apparently unrelated concept of envelope of mid-planes is a main point of this paper.
But what happens with the fourth equation? We verify that the tangent directions that leads to some solution of this system of $4$ equations are zeros of a polynomial equation of degree $6$, thus they are at most $6$. The set of solutions of this system is an affine invariant set that, up to our knowledge, has not yet been considered. We call it the {\it Affine Mid-Planes Evolute}. We verify that under some generic conditions, the affine mid-planes evolute is a regular surface in $3$-space.
This work is part of the doctoral thesis of the first author under the supervision of the second author.
\section{Some classical concepts in affine differential geometry}
In this section, we recall the notions of Transon plane, B.Su's cone and Moutard's quadric of a given direction at a point of a smooth locally convex surface.
\subsection{Preliminaries}
Consider a smooth convex surface $S$ and a point $p_0\in S$. Assume that $p_0=(0,0,0)$ and that the tangent plane at $p_0$ is $z=0$. Assume also that the axes $x$ and $y$ are orthogonal in the Blaschke metric of $S$. Then, close to $p_0$, $S$ is the graph of a function $f$ that can be written as
\begin{equation}\label{eq:GraphS}
f(x,y)=\frac{1}{2}(x^2+y^2)+f_3(x,y)+ f_4(x,y) +O(5), %
\end{equation}
where
\begin{equation}\label{eq:GraphS1}
f_3(x,y)=\sum_{i=0}^3f_{3-i,i}x^{3-i}y^{i},\ \ f_4(x,y)=\sum_{i=0}^{4}f_{4-i,i}x^{4-i}y^i
\end{equation}
are homogeneous polynomials of degree $k$, $k=3,4$, and $O(n)$ denotes some expression of order $\geq n$ in $(x,y)$.
We may also assume that the affine normal vector of $S$ at the origin is $(0,0,1)$. In this case, the apolarity condition implies
that $3f_{3,0}+f_{1,2}=0$ and $3f_{0,3}+f_{2,1}=0$ (see \cite{Nomizu}). Thus we can write
\begin{equation}\label{eq:CubicaNormalizada}
f_3(x,y)=a(x^3-3xy^2)+b(y^3-3yx^2),
\end{equation}
where $6a=C(e_1,e_1,e_1)$, $6b=C(e_2,e_2,e_2)$, $e_1=(1,0)$, $e_2=(0,1)$ and $C$ denotes the cubic form at the origin.
\begin{comment}
Any other $h$-orthonormal basis, $\{\bar{U},\bar{V}\}$ is related to $\{U,V\}$ by
\begin{equation}\label{eq:MudaBase}
\begin{array}{c}
\bar{U}=\cos(\theta)U-\sin(\theta)V\\
\bar{V}=\sin(\theta)U+\cos(\theta)V,
\end{array}
\end{equation}
for some $\theta\in\mathbb{R}$. Then the corresponding cubic form is given by
\begin{equation}
\begin{array}{c}
a=\cos(3\theta)\bar{a}-\sin(3\theta)\bar{b}\\
b=\sin(3\theta)\bar{a}+\cos(3\theta)\bar{b}.
\end{array}
\end{equation}
\end{comment}
\subsection{Transon planes, cone of B.Su and Moutard's quadric}\label{sec:MedialCurve}
We begin with a two-hundred years old result of A.Transon (\cite{Transon}). For a mo\-dern reference, see \cite{Juttler}.
For the sake of completeness, we give a proof below.
\begin{proposition}\label{prop:Transon}
Consider a smooth convex surface $S\subset\mathbb{R}^3$, $p_0\in S$ and $T\in T_{p_0}S$. Then the affine normal lines of the planar curves obtained as the intersection of $S$ with planes containing $T$ form a plane, which is called the {\it Transon plane} of the tangent $T$ at $p_0$.
\end{proposition}
\begin{proof}
We may assume that $T=(1,0)$. Take then a plane of the form $y=\lambda z$, $\lambda\in\mathbb{R}$. Using formulas \eqref{eq:GraphS} and \eqref{eq:GraphS1}, the projection of the corresponding section of $S$
on the $xz$ plane is given by
\begin{equation}\label{eq:proj}
z=\frac{1}{2}x^2+f_{3,0}x^3+\left( \frac{\lambda^2}{8}+\frac{\lambda}{2}f_{2,1}+f_{4,0} \right) x^4+ O(5).
\end{equation}
From appendix \ref{app1}, the affine normal direction of this projection is $(-2f_{3,0},1)$. Thus the affine normal of the section is contained in the
plane $x+2f_{3,0}z=0$, which is independent of $\lambda$.
\end{proof}
Similar calculations show that when $S$ is the graph of a function given by \eqref{eq:GraphS}, the Transon plane at the origin in a direction $T=(\xi,\eta)$ is given by $G(\xi,\eta,X)=0$, where $X=(x,y,z)$ and
\begin{equation}\label{planodetranson}
G(\xi,\eta,X)=\frac{\xi}{2}(\xi^2+\eta^2)x+\frac{\eta}{2}(\xi^2+\eta^2)y+f_3(\xi,\eta)z,
\end{equation}
(see also \cite{Juttler}).
\begin{figure}[ht]
\centering
\includegraphics[scale=0.65]{plano_de_transon.eps}\\
\caption{The Transon plane of the tangent $T$: Formed by the affine normal lines of the planar sections that contain $T$.}
\end{figure}
Consider now the family of all Transon planes obtained as the direction of $T$ varies. The envelope of this family is called {\it cone of B.Su} and is
obtained by solving the equations $G=G_{\xi}=G_{\eta}=0$, where
\begin{equation}
\left\{
\begin{array}{c}
G_{\xi}=\frac{1}{2}(3\xi^2+\eta^2)x+(\xi\eta) y+\left((f_3)_{\xi}\right)z\\
G_{\eta}=(\xi\eta) x+\frac{1}{2}(\xi^2+3\eta^2)y+\left((f_3)_{\eta}\right)z.
\end{array}
\right.
\end{equation}
Since $G$ is homogeneous of degree $3$, Euler's relation says that
$$3G=\eta G_{\eta}+\xi G_{\xi}.$$
Thus the cone of B.Su is obtained form the equations
$G_{\xi}=G_{\eta}=0$. We shall denote a vector in the direction of this line by $s(T)=s(p,T)$.
We can thus calculate $s(T)$ as the vector product of the normal vectors of $G_{\xi}=0$ and $G_{\eta}=0$.
In particular, for $(\xi,\eta)=(1,0)$ we obtain
\begin{equation}\label{eq:ConeSu10}
s(1,0)=\left( -2f_{3,0}, -2f_{2,1}, 1\right).
\end{equation}
We shall consider also the osculating conics of all planar sections obtained from planes containing $T$. The following proposition is an old result of T.Moutard (\cite{Juttler},\cite{Moutard}). We give a proof for the sake of completeness.
\begin{proposition}
The union of the osculating conics of all planar sections containing $T$ form a quadric, which is called the {\it Moutard's quadric} of the tangent $T$.
\end{proposition}
\begin{proof}
Assume that $S$ is the graph of $f$ given by equation \eqref{eq:GraphS}, $p_0=(0,0,0)$ and $T=(1,0)$. Then the projection of the section of $S$ by the plane $y=\lambda z$
is given by equation \eqref{eq:proj}. By appendix \ref{app1}, the affine curvature is given by
$$
\mu=\lambda^2+4\lambda f_{2,1}+8f_{4,0}-20f_{3,0}^2.
$$
The projection of the osculating conic of this section in the plane $xz$ is given by
$$
(x+2f_{3,0}z)^2+\mu z^2-2z=0.
$$
Substituting $\lambda=y/z$ in the above equation, after some calculations we obtain that the osculating conic is contained in
\begin{equation}
z=\frac{1}{2}(x^2+y^2)+2f_{2,1}yz+2f_{3,0}xz+4(f_{4,0}-2f_{3,0}^2)z^2,
\end{equation}
thus proving the lemma. For later reference, we remark that the center of this Moutard quadric is
\begin{equation}\label{eq:CenterMoutard}
X=\dfrac{1}{4(2f_{4,0}-5f_{3,0}^2-f_{2,1}^2)} \left(-2f_{3,0},-2f_{2,1},1 \right).
\end{equation}
\end{proof}
\begin{proposition}\label{prop:centro_quadrica_moutard}
The center of the Moutard's quadric of a tangent $T$ coincides with the center of affine curvature of the intersection $\gamma$ of the plane
generated by $T$ and $s(T)$ with the surface.
\end{proposition}
\begin{proof}
We may assume that $T=(1,0)$.
The plane generated by $T$ and $s(T)$ has equation $y+2f_{2,1}z=0$. Thus, it follows from equation \eqref{eq:proj} that the
projection of $\gamma$ in the $xz$ plane is given by
$$
z= \dfrac{1}{2}x^2+f_{3,0}x^3+\left(f_{4,0}-\dfrac{f_{2,1}^2}{2}\right)x^4+O(5).
$$
From appendix \ref{app1}, the affine curvature is given
$\mu=-4(5f_{3,0}^2-2f_{4,0}+f_{2,1}^2)$.
Thus the affine center of curvature at the origin is given by
$$
\dfrac{1}{4(2f_{4,0}-5f_{3,0}^2-f_{2,1}^2)} \left(-2f_{3,0},-2f_{2,1},1 \right),
$$
which coincides with formula \eqref{eq:CenterMoutard}.
\end{proof}
\section{Mid-planes}
Let $S$ be a regular surface and $p_1,p_2\in S$. Denote by $M(p_1,p_2)$ the mid-point and by $C(p_1,p_2)$ the mid-chord of $p_1$ and $p_2$, i.e.,
$$
M(p_1,p_2)=\dfrac{p_1+p_2}{2},\ \ C(p_1,p_2)=\dfrac{p_1-p_2}{2}.
$$
The mid-plane of $(p_1,p_2)$ is the plane that contains $M(p_1,p_2)$ and the line of intersection of the tangent planes at $p_1$ and $p_2$.
Consider $F:S\times S\times\mathbb{R}^3\to\mathbb{R}$ given by
\begin{equation}\label{eq:Normais}
F(p_1,p_2,X)=\left[ (N_1\cdot C)N_2+(N_2\cdot C)N_1\right]\cdot (X-M),
\end{equation}
where $\cdot$ denotes the canonical inner product and $N_i$ is any functional that is zero in $T_{p_i}S$, $i=1,2$.
It is not difficult to verify that the equation of the mid-plane of $(p_1,p_2)$ is $F=0$.
Along this section, we shall assume that $S$ is the graph of a function $f$ defined by equations \eqref{eq:GraphS} and \eqref{eq:GraphS1}, and take
$$
N_i=(-f_x,-f_{y},1)(u_i,v_i)
$$
as normal vectors, where $p_i=(u_i,v_i,f(u_i,v_i))$, $i=1,2$. Write $\Delta u=u_1-u_2$, $\Delta v=v_1-v_2$.
\subsection{Limit of mid-planes}
Denote by $O(n)$ terms of order $n$ in $(u_1,u_2,v_1,v_2)$. Next lemma is the main tool to establish the connection between the envelope of mid-planes and the concepts
of Transon planes and B.Su's cone:
\begin{lemma}\label{Lemma:Main3}
We have that
\begin{equation}\label{eq:Main3}
F(u_1,v_1,u_2,v_2,X)=G(\Delta u,\Delta v,X)+O(4).
\end{equation}
where $G$ is given by equation \eqref{planodetranson}.
\end{lemma}
We can prove this lemma through long but straightforward calculations, the reader can find it in appendix \ref{app2}.
This lemma has some important consequences, one of them is the following corollary:
\begin{corollary}\label{transon}
Take $p_1=(u_1,v_1,f(u_1,v_1))$ and $p_2=(u_2,v_2,f(u_2,v_2))$ such that $(\Delta u,\Delta v)=t(\xi,\eta)$, $t\in\mathbb{R}$. Then
the limit of the mid-planes of $(p_1,p_2)$ when $t$ goes to $0$ is the Transon plane of $(\xi,\eta)$ at $p_1$.
\end{corollary}
\begin{proof}
It follows from Lemma \ref{Lemma:Main3} that
$t^3 G(\xi,\eta,X)=O(t^4)$.
When $t\to 0$, this equation is converging to $G(\xi,\eta,X)=0$, which is the equation of the Transon plane.
\end{proof}
\subsection{Limit of envelope of mid-planes-I}
We have to consider the system
$$
\left\{
\begin{array}{ll}
F(u_1,v_1,u_2,v_2,X)=0 \\
F_{u_1}(u_1,v_1,u_2,v_2,X)=0 \\
F_{v_1}(u_1,v_1,u_2,v_2,X)=0 \\
F_{u_2}(u_1,v_1,u_2,v_2,X)=0 \\
F_{v_2}(u_1,v_1,u_2,v_2,X)=0 \\
\end{array}
\right.,
$$
which is equivalent to
\begin{equation}\label{sistema}
\left\{
\begin{array}{ll}
F(u_1,v_1,u_2,v_2,X)=0 \\
F_{u_1}(u_1,v_1,u_2,v_2,X)-F_{u_2}(u_1,v_1,u_2,v_2,X)=0 \\
F_{v_1}(u_1,v_1,u_2,v_2,X)-F_{v_2}(u_1,v_1,u_2,v_2,X)=0 \\
F_{u_1}(u_1,v_1,u_2,v_2,X)+F_{u_2}(u_1,v_1,u_2,v_2,X)=0 \\
F_{v_1}(u_1,v_1,u_2,v_2,X)+F_{v_2}(u_1,v_1,u_2,v_2,X)=0 \\
\end{array}
\right. \ .
\end{equation}
\begin{corollary}
We have that
\begin{equation}
\left\{
\begin{array}{ll}
F_{u_1}(u_1,v_1,u_2,v_2,X)-F_{u_2}(u_1,v_1,u_2,v_2,X)=2G_{\xi}(\Delta u,\Delta v)+O(3),\\
F_{v_1}(u_1,v_1,u_2,v_2,X)-F_{v_2}(u_1,v_1,u_2,v_2,X)=2G_{\eta}(\Delta u,\Delta v)+O(3)
\end{array}
\right. \ .
\end{equation}
\end{corollary}
\begin{proof}
Differentiate formula \eqref{eq:Main3}.
\end{proof}
\begin{corollary}\label{Su}
Take $p_1=(u_1,v_1,f(u_1,v_1))$ and $p_2=(u_2,v_2,f(u_2,v_2))$ such that $(\Delta u,\Delta v)=t(\xi,\eta)$. Then
the limit of the second and third equations of system \eqref{sistema} when $t$ goes to $0$ are $G_{\xi}=G_{\eta}=0$. As a consequence,
$X$ satisfies the first, second and third equations of the limit of system \eqref{sistema} if and only if $X$ belongs to the cone of B.Su at the origin
in the direction $(\xi,\eta)$.
\end{corollary}
\begin{proof}
The second equation of system \eqref{sistema} can be written as $G_{\xi}(\Delta u,\Delta v)=O(3)$. Taking pairs $(p_1,p_2)$
such that $(\Delta u,\Delta v)=t(\xi,\eta,X)$, the second equation of system \eqref{sistema} becomes $t^2G_{\xi}(\xi,\eta,X)=O(t^3)$.
When $t\to 0$, this equation is converging to $G_{\xi}(\xi,\eta,X)=0$. The same reasoning applied to the third equation of system \eqref{sistema} leads to
the equation $G_{\eta}(\xi,\eta,X)=0$.
\end{proof}
\subsection{Limit of envelope of mid-planes-II}
Define
$$
H_1(\xi,\eta,X)=H_{11}x+H_{12}y+H_{13}z-H_{14},\ $$$$H_2(\xi,\eta,X)=H_{21}x+H_{22}y+H_{23}z-H_{24},
$$
where
$$
H_{11} = \dfrac{1}{2}a\left(5\xi^3+3\eta^2\xi\right)-b\left(3\xi^2\eta+2\eta^3\right), \ \
H_{12} =-3a\eta^3 -\dfrac{3}{2}b\left(\xi^3+3\eta^2\xi\right),
$$
$$
H_{13} = 2f_{4,0}\xi^3 +\dfrac{3}{2}f_{3,1}\eta\xi^2+f_{2,2}\eta^2\xi+\dfrac{1}{2}f_{1,3}\eta^3, \ \
H_{14} = \dfrac{1}{4}(\xi^2+\eta^2)\xi,
$$
$$
H_{21} = -\dfrac{3}{2}a\left(3\xi^2\eta+\eta^3\right)-3b\xi^3, \ \
H_{22} = -a\left(2\xi^3+3\xi\eta^2\right)+\dfrac{1}{2}b\left(3\xi^2\eta+5\eta^3\right),
$$
$$
H_{23} = \dfrac{1}{2}f_{3,1}\xi^3+f_{2,2}\xi^2\eta+\dfrac{3}{2}f_{1,3}\xi\eta^2+2f_{0,4}\eta^3, \ \
H_{24} = \dfrac{1}{4}(\xi^2+\eta^2)\eta.
$$
Next lemma is the main tool to connect the concept of Moutard's quadric and the envelope of mid-planes:
\begin{lemma}\label{Lemma:Main4}
We have that $F(u_1,v_1,u_2,v_2,X)$ is given by
\begin{equation}\label{eq:Main4}
F=G(\Delta u,\Delta v,X)+H_1(\Delta u,\Delta v,X) \sigma u+H_2(\Delta u,\Delta v,X) \sigma v +O(5),
\end{equation}
where $\sigma u=u_1+u_2$ and $\sigma v=v_1+v_2$.
\end{lemma}
The proof of this lemma are long but straightforward calculations, the reader can find it in appendix \ref{app2}.
\begin{corollary}
We have that
\begin{equation}
\left\{
\begin{array}{ll}
F_{u_1}(u_1,v_1,u_2,v_2,X)+F_{u_2}(u_1,v_1,u_2,v_2,X)=2H_1(\Delta u,\Delta v)+O(4),\\
F_{v_1}(u_1,v_1,u_2,v_2,X)+F_{v_2}(u_1,v_1,u_2,v_2,X)=2H_2(\Delta u,\Delta v)+O(4)
\end{array}
\right. \ .
\end{equation}
\end{corollary}
\begin{proof}
Differentiate equation \eqref{eq:Main4}.
\end{proof}
\begin{corollary}
The limit of the system \eqref{sistema} defining the envelope of mid-planes is the system
\begin{equation}\label{eq:SystemEvolute}
G=G_{\xi}=G_{\eta}=H_1=H_2=0.
\end{equation}
\end{corollary}
\begin{proof}
Take pairs $(p_1,p_2)$ such that $(\Delta u,\Delta v)=t(\xi,\eta)$.
Then the limit of the fourth and fifth equations when $t$ goes to zero are $H_1(\xi,\eta,X)=0$
and $H_2(\xi,\eta,X)=0$, respectively.
\end{proof}
\begin{proposition}\label{moutard}
If system \eqref{eq:SystemEvolute} admits a solution, then this solution is the center of the Moutard's quadric of $T=(\xi,\eta)$.
\end{proposition}
\begin{proof}
We may assume that the tangent $(\xi,\eta)$ is $(1,0)$.
Taking $\xi=1, \eta=0$ in the system $G=G_{\eta}=H_1=0$ we get
$$
\left\{
\begin{array}{ccc}
\dfrac{1}{2}x+az & = & 0 \\
\dfrac{1}{2}y-3bz & = & 0 \\
\dfrac{5}{2}ax-\dfrac{3}{2}by+2f_{4,0}z & = & \dfrac{1}{4} \\
\end{array}
\right. \ .
$$
The solution of this system is
$$
(x,y,z)= \dfrac{1}{4(2f_{4,0}-5a^2-9b^2)} \left( -2a, 6b, 1 \right),
$$
which is exactly the center of the Moutard's quadric of $T=(1,0)$ at the origin (see formula \eqref{eq:CenterMoutard}).
\end{proof}
\section{Affine mid-planes evolute}
Any direction in the tangent plane of $S$ at $p$ can be represented by a vector $T$ with $h(T)=1$, where $h$
denotes the Blaschke metric. Thus we can write $(p,T)\in T^1S$, where $T^1S$ denotes the unit tangent bundle of $S$.
Since $G=G_{\xi}=G_{\eta}=0$ are linearly dependent, we shall discard the equation $G=0$.
Denote by $D:T^1S\longrightarrow \mathbb{R}$ the determinant of the extended matrix of
$$
G_{\xi}=G_{\eta}=H_1=H_2=0.
$$
Then the system \eqref{eq:SystemEvolute} admits a solution if and only if $D(p,T)=0$.
\begin{lemma}\label{lemma:q}
For each $p\in S$, the system \eqref{eq:SystemEvolute} admit solutions if and only if $T=(\xi,\eta)$ is a root of a homogeneous polynomial of degree $6$.
\end{lemma}
\begin{proof}
We may assume $p=(0,0,0)$, $T=(\xi,\eta)$ and $S$ is the graph of $f$ given by equation \eqref{eq:CubicaNormalizada}. Straightforward calculations show that
$$
D(p,T) = \frac{3}{32}(\xi^2+\eta^2)^2 \cdot q(\xi,\eta),
$$
where $q(\xi,\eta)=12q_3(\xi,\eta)+q_4(\xi,\eta)$ is a homogeneous polynomial of degree $6$, $q_3$ is given by
$$
ab\xi^6+3(a^2-b^2)\xi^5\eta-15ab\xi^4\eta^2+10(b^2-a^2)\xi^3\eta^3+15ab\xi^2\eta^4+3(a^2-b^2)\xi\eta^5-ab\eta^6
$$
and
$$
q_4=-f_{3,1}\xi^6+(4f_{4,0}-2f_{2,2})\xi^5\eta+(2f_{3,1}-3f_{1,3})\xi^4\eta^2+4(f_{4,0}-f_{0,4})\xi^3\eta^3
$$
$$
+(3f_{3,1}-2f_{1,3})\xi^2\eta^4+(2f_{2,2}-4f_{0,4})\xi\eta^5+f_{1,3}\eta^6.
$$
Thus $D=0$ admit at most $6$ solutions.
\end{proof}
\begin{corollary}
For each $p\in S$, the system \eqref{eq:SystemEvolute} admits a solution for at most $6$ values of the direction $T$.
\end{corollary}
\begin{example}
Consider a point $p\in S$ with $f_4=0$. We may assume, by a rotation of the tangent plane, that $b=0$. Then
$$
q_3=3a^2\xi^5\eta-10a^2\xi^3\eta^3+3a^2\xi\eta^5.
$$
This polynomial has exactly six roots, $(\xi,\eta)=(\cos\frac{k\pi}{6},\sin\frac{k\pi}{6})$, $0\leq k\leq 5$.
\end{example}
For each $p\in S$ and $T_i(p)$, $1\leq i\leq 6$, given by the above corollary, define $X_i(p)$ as the solution of the system \eqref{eq:SystemEvolute} and write $E(p)=\cup_{i=1}^6 X_i(p)$. Up to our knowledge, the set $E(p)$ has not yet been considered. We call the set $E(p)$, $p\in S$, the {\it Affine Mid-Planes Evolute} of $S$.
\section{Regularity of the Affine Mid-Planes Evolute}
In this section we study the regularity of the branches of the Affine Mid-Planes Evolute. We begin by showing that, under certain conditions, the vector fields $T_i(p)$, $1\leq i\leq 6$, and the corresponding map $X_i(p)$ are smooth.
\begin{lemma}
Let $p_0\in S$ and take $(\xi_0,\eta_0)\in T_{p_0}S$ with $q(p_0)(\xi_0,\eta_0)=0$ where $q$ is defined in Lemma \ref{lemma:q}. Assume that $(\xi_0,\eta_0)$ is a simple root of $q(p_0)$.
Then there exists a neighborhood $U$ of $p_0$ and a map $\left(Id,(\xi,\eta)\right):U\to T^1S$ such that $(\xi,\eta)(p_0)=(\xi_0,\eta_0)$ and $(\xi,\eta)(p)$
is a simple root of $q(p)$. Denoting by $X(p)$ the center of the Moutard quadric of $(\xi,\eta)(p)$, the map $X:U\to\mathbb{R}^3$ is differentiable.
\end{lemma}
\begin{proof}
The first claim follows from the implicit function theorem and the second one from the formula of the center of the Moutard quadric.
\end{proof}
From now on, we shall assume that, for a given branch of the Affine Mid-Planes Evolute, $T=T(p)$ and $X=X(p)$ are smooth functions of $p\in U$.
We now look for conditions under which this branch of the Affine Mid-Planes Evolute is a regular surface.
\bigskip\bigskip
Assume that the frame $\{T_1(p),T_2(p)\}$ is $h$-orthonormal and that $T_1(p)$ vanishes the cubic form at each $p$. Then $S\times S$
is locally parameterized by
$$
(p,x,y)\to \left( p, (xT_1(p)+yT_2(p)+f(p,x,y)\zeta(p)) \right),
$$
where $\zeta(p)$ is the affine normal vector at $p$. We may write
$$
f(p,x,y)=\frac{1}{2}(x^2+y^2)+b(p)(y^3-3yx^2)+O(4)(p)(x,y),
$$
where $b^2(p)$ is the Pick invariant of $S$ at $p$. Consider $\mathbf{G}:T^1S\to \mathbb{R}^3$ given by
$$
\mathbf{G}(p,\xi,\eta)=\left( \frac{\xi}{2}(\xi^2+\eta^2),\frac{\eta}{2}(\xi^2+\eta^2),b(p)(\eta^3-3\eta\xi^2) \right).
$$
Then, for fixed $(p,\xi,\eta)$, denoting $X=(x,y,z)$, $\mathbf{G}(p,\xi,\eta)\cdot X=0$ is the equation of the Transon plane at $p$ in the direction $(\xi,\eta)$.
Consider a curve $p=p(t)$, $p(0)=p_0$, along the surface $S$ and denote by $T(t)=(\xi(t),\eta(t))$ and $X(t)$ the corresponding values of $T$ and $X$
along the branch.
\begin{lemma}\label{lemma:Transon}
Assume that $b'(0)\neq 0$. Then $X'(0)$ belong to the Transon plane if and only if $(\xi,\eta)(0)$ vanishes the cubic form at $p$. Moreover, if $(\xi,\eta)(0)$ vanishes the cubic form at $p$, then $X'(0)$ is not in the cone of B.Su.
\end{lemma}
\begin{proof}
Write $\mathbf{G}(t)=\mathbf{G}(p(t),\xi(t),\eta(t))$. Differentiating $\mathbf{G}(t)\cdot X=0$ at $t=0$ we obtain
$$
\mathbf{G}_{\xi}\cdot X\ \xi'+\mathbf{G}_{\eta}\cdot X\ \eta'+\mathbf{G}\cdot X'+b'(\eta^3-3\eta\xi^2)z=0.
$$
For $X=(x,y,z)$ center of Moutard's quadric, $\mathbf{G}_{\xi}\cdot X=\mathbf{G}_{\eta}\cdot X=0$. Thus $\mathbf{G}\cdot X'=0$ if and only
if $\eta^3-3\eta\xi^2=0$. We conclude that $X'$ belongs to the Transon plane if and only if $\eta^3-3\eta\xi^2=0$, thus proving the first claim of the lemma.
For the second claim, we may assume that $(\xi(0),\eta(0))=(1,0)$. Thus, at $t=0$, $X$ is given by equation \eqref{eq:CenterMoutard} with $f_{3,0}=0$, $f_{2,1}=-3b(0)$, i.e.,
$$
(x,y,z)(0)=-\frac{1}{4(9b^2-2f_{4,0})}\left(0, 6b(0),1 \right).
$$
Observe that
$$
\mathbf{G}_{\eta}(p,\xi,\eta)=\left( \xi\eta,\frac{1}{2}(\xi^2+3\eta^2),3b(p)(\eta^2-\xi^2) \right).
$$
Differentiating $\mathbf{G}_{\eta}(t)\cdot X(t)=0$ we obtain
$$
\mathbf{G}_{\eta\xi}\cdot X\ \xi'+\mathbf{G}_{\eta\eta}\cdot X\ \eta'+\mathbf{G}_{\eta}\cdot X' +3b'(\eta^2-\xi^2)z=0.
$$
Now take $t=0$ and use $\xi(0)=1$, $\eta(0)=0$ to obtain
$$
\left(0,1,-6b(0)\right)\cdot X\ \xi'+\left( 1, 0, 0\right)\cdot X\ \eta'+\mathbf{G}_{\eta}\cdot (X')-3b'(0)z=0.
$$
Using the above formula for $(x,y,z)$ we get
$$
\mathbf{G}_{\eta}\cdot X'+\frac{3b'(0)}{4(9b^2-2f_{4,0})}=0.
$$
We conclude that $G_{\eta}(X')\neq 0$, which implies that $X'$ is not on the cone of B.Su.
\end{proof}
\begin{lemma}\label{lemma:Su1}
Assume $p'(0)=T$ and denote by $\gamma$ the intersection of $S$ with the plane generated by $T$ and $s(T)$. If $\mu_\gamma'(0)\neq 0$, where $\mu_\gamma$ denotes the affine curvature of $\gamma$, $X'(0)$ is a non-zero vector in the direction of the cone of B.Su.
\end{lemma}
\begin{proof}
Since $X$ is a point of the affine evolute of $\gamma$ and $T$ is tangent to this curve, $X'(0)$ is in the direction of the its affine normal, which belongs to the cone of B.Su
(see Proposition \ref{prop:centro_quadrica_moutard}).
\end{proof}
If we assume that $S$ is given by \eqref{eq:CubicaNormalizada}, $p=(0,0)$ and $T(p)=(1,0)$, the condition $\mu_\gamma'(0)\neq 0$ can be explicitly
described. Expanding $f$ until order $5$, it is straightforward to verify that the projection of $\gamma$
in the $xz$ plane is given by
$$
z=\frac{1}{2}x^2+ax^3+\left(f_{4,0}-\frac{9}{2}b^2\right)x^4+\left(-27ab^2+3bf_{3,1}+f_{5,0}\right)x^5 +O(6).
$$
From appendix \ref{app1}, it follows that
$$
\mu_\gamma'(0)=\frac{1}{27}\left( 9a_5+40a_3^3-45a_3a_4\right),
$$
where $a_3=6a$, $a_4=24\left(f_{4,0}-\frac{9}{2}b^2\right)$ and
$a_5=120\left(-27ab^2+3bf_{3,1}+f_{5,0}\right)$.
\begin{proposition}
Assume that $p$ is not critical for the Pick invariant and that $\mu_\gamma'(0)\neq 0$.
Then the corresponding branch of the Affine Mid-Planes Evolute is smooth at $p$.
\end{proposition}
\begin{proof}
From lemma \ref{lemma:Su1}, $X_T$ is a non-zero vector in the direction of the cone of B.Su. Take a direction $W\neq T$
such that the derivative of $b$ in this direction in non-zero. From lemma \ref{lemma:Transon}, $X_W$ is not a multiple of $X_T$,
thus proving the proposition.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction}
Photoinduced changes in electronic properties of correlated electron and electron-phonon systems have been investigated extensively so far.\cite{koshigono_jpsj06,yonemitsu_pr08} Development in experimental techniques has allowed for progressively shortening optical pulses and consequently improving time resolutions. At the same time, the field amplitude that can be applied to materials has significantly increased for different excitation energies. This fact enhances the possibility of controlling the electronic properties.
As for events that occur with an oscillating electric field of large amplitude, dynamical localization is known in the continuous-wave case.\cite{dunlap_prb86,grossmann_prl91} When the frequency of the field is comparable to or higher than that of electron transfers determined by the transfer integral, the slow component of the electron dynamics is governed by the time average of the transfer integral with the Peierls-phase factor.\cite{kayanuma_pra94} The time average for a sinusoidally oscillating field gives an effective transfer integral modulated by the zeroth-order Bessel function whose argument is proportional to the ratio of the amplitude to the frequency of the field. Thus, the effective transfer integral can vanish to cause dynamical localization or can even change from positive (negative) to negative (positive).
Applying such a strong field to interacting electron systems, Tsuji {\it et al.} have shown that the sign inversion of the transfer integral and the consequent inversion of the band structure, a ``negative-temperature'' state, are regarded as the sign inversion of the interaction in the half-filled Hubbard model, using the dynamical mean-field theory, by which a superconducting state is possibly induced.\cite{tsuji_prl11} In addition, they have shown that such a repulsion-to-attraction conversion is realized not only by continuous waves but also by half-cycle pulses and asymmetric monocycle pulses.\cite{tsuji_prb12} In the limit of the vanishing pulse width, the irradiation causes a sudden change in the Peierls phase of the transfer integral, and the dynamical phase shift is equivalent to the crystal momentum shift. In this limit, however, a symmetric monocycle pulse causes nothing. Indeed, the repulsion-to-attraction conversion is hardly realized by symmetric monocycle pulses with various trial time profiles of finite pulse widths.\cite{tsuji_prb12}
The effective transfer integral modulated by the application of a continuous wave and that by the application of a pulse are not always the same. It depends on the system and how the system is excited. The quasi-two-dimensional metal complex Et$_2$Me$_2$Sb[Pd(dmit)$_2$]$_2$ (dmit = 1,3-dithiol-2-thione-4,5-dithiolate) is known to show a photoinduced transition from a charge-ordered-insulator phase to a Mott-insulator phase.\cite{ishikawa_prb09,nishioka_jpsj13a,nishioka_jpsj13b} In this particular case, it has been demonstrated that continuous-wave lasers modulate the effective transfer integrals between the dimers, and pulsed lasers modulate that within the dimer, both through the zeroth-order Bessel function.\cite{nishioka_jpsj14} On the basis of this fact, the dynamical localization is transiently expected for pulsed lasers if the field amplitude is sufficiently large. Indeed, optical freezing of charge motion has been found in an organic conductor, $\alpha$-(BEDT-TTF)$_2$I$_3$ (BEDT-TTF = bis[ethylenedithio]-tetrathiafulvalene).\cite{ishikawa_ncomms14} Although its mechanism is not theoretically understood yet, it is regarded as due to collaboration with the dynamical localization and intermolecular charge correlations.
The realization of a negative-temperature state or modulation of transfer integrals and interaction parameters requires an oscillating electric field of large amplitude. In general, it is much larger than what is needed for some phase transition to be simply induced. In any case, the fact that transfer integrals are effectively modulated by pulsed lasers is encouraging because the field amplitude can be increased by shortening the pulse width. In the actual case of $\alpha$-(BEDT-TTF)$_2$I$_3$ above,\cite{ishikawa_ncomms14} the dynamical phase shift is known to be much smaller than $\pi$. On the basis of these findings, we study the effect of symmetric monocycle electric-field pulses whose time integrals are zero, i.e., whose dynamical phase shifts are zero. It was expected to be difficult to form a negative-temperature state or to invert interaction parameters by such symmetric pulses.\cite{tsuji_prb12} The model systems we treat are similar to that used in Ref.~\citen{nishioka_jpsj14}, but they are simplified to use the exact diagonalization method: one-dimensional three-quarter-filled strongly dimerized extended Peierls-Hubbard and Holstein models. Their band-insulator phases consist of almost half-filled dimers alternating with almost completely filled dimers. Charge transfers take place between these dimers, but the optical gap is determined by the electronic transition within the almost half-filled dimer, as in Et$_2$Me$_2$Sb[Pd(dmit)$_2$]$_2$.
In this paper, we show that, even by a symmetric monocycle electric-field pulse, a negative-temperature state is produced, and interactions are effectively inverted when we interpret the dependences of the time-averaged correlation functions on the interactions. Namely, the stronger on-site repulsion causes the larger space-time-averaged double occupancy, the stronger nearest-neighbor repulsion causes the larger averaged nearest-neighbor charge-density correlation, and so on. Such a negative-temperature state appears when the total-energy increment is maximized as a function of the electric field amplitude.
\section{Strongly Dimerized Models with Charge-Ordered Ground States}
We study the electron-phonon dynamics after strong photoexcitation in a one-dimensional three-quarter-filled strongly dimerized model with on-site and nearest-neighbor repulsions and with the Peierls and Holstein types of electron-phonon couplings,
\begin{eqnarray}
H & = & -\sum_{j=1}^{N/2} (t_0-u_j) \sum_\sigma \left(
c_{2j-1,\sigma}^\dagger c_{2j,\sigma} + c_{2j,\sigma}^\dagger c_{2j-1,\sigma}
\right) \nonumber \\ & &
-\sum_{j=1}^{N/2} t_{\rm inter} \sum_\sigma \left(
c_{2j,\sigma}^\dagger c_{2j+1,\sigma} + c_{2j+1,\sigma}^\dagger c_{2j,\sigma}
\right) \nonumber \\ & &
+ U \sum_{i=1}^{N} n_{i,\uparrow} n_{i,\downarrow}
+ V \sum_{i=1}^{N} n_{i} n_{i+1}
+ \sum_{j=1}^{N/2} \left(
\frac{u_j^2}{2S} + \frac{\dot{u}_j^2}{2S\omega_P^2}
\right) \nonumber \\ & &
+ g \sum_{i=1}^{N} (b_i+b_i^\dagger) \left( n_i - \frac{3}{2} \right)
+ \omega_H \sum_{i=1}^{N} b_i^\dagger b_i
\;, \label{eq:hamiltonian}
\end{eqnarray}
where $ c_{i,\sigma}^\dagger $ creates an electron with spin $ \sigma $ at site $ i $, $ n_{i,\sigma} = c_{i,\sigma}^\dagger c_{i,\sigma} $, and $ n_i = \sum_\sigma n_{i,\sigma} $. The transfer integral within the $ j $-th dimer (i.e., between sites $ 2j-1 $ and $ 2j $) is modulated by the lattice displacement $ u_j $: $ t_0-u_j $. The transfer integral between the dimers is denoted by $ t_{\rm inter} $. They are shown schematically in Fig.~\ref{fig:negativeT}(a).
\begin{figure}
\includegraphics[height=9cm]{66977Fig1.eps}
\caption
(a) Schematic representation of transfer integrals and distances between neighboring sites.
(b) Half-filled dimer whose bonding orbitals are occupied and completely filled dimer in ground state.
(c) Half-filled dimer whose antibonding orbitals are occupied and completely filled dimer in negative-temperature state.
\label{fig:negativeT}}
\end{figure}
The parameter $ U $ represents the onsite repulsion strength, and $ V $ represents the nearest-neighbor repulsion strength. The lattice displacement $ u_j $ has a bare frequency $ \omega_P $ and is coupled with the bond density between sites $ 2j-1 $ and $ 2j $ with strength $ S $. It is noted that $ u $'s have a dimension of energy. The operator $ b_i^\dagger $ creates a phonon, which has a bare frequency $ \omega_H $ and is coupled with the charge density relative to its average value at site $ i $ with strength $ g $.
The periodic boundary condition is imposed. For $ g $=0, we use $ N $=16 sites. For $ g\neq 0 $, we use $ N $=8 sites and limit the number of phonons up to a maximum of two on every site. The number of electrons is 24 for $ N $=16 and 12 for $ N $=8 at three-quarter filling. The electron-quantum-phonon system is treated by the exact diagonalization method. The lattice system is treated classically.
This is because the lattice vibration is caused by the intermolecular restoring force and consequently is slow compared with other dynamics. Meanwhile, the molecular vibration is caused by the intramolecular restoring force and its energy is often comparable to those of the intermolecular charge transfers. Its quantum nature is important and essential for the polaron effect.
Before the photoexcitation, the system is in the lowest-energy state so that the lattice displacements are determined by
\begin{equation}
u_j = -S \sum_\sigma \left(
\langle c_{2j-1,\sigma}^\dagger c_{2j,\sigma} \rangle +
\langle c_{2j,\sigma}^\dagger c_{2j-1,\sigma} \rangle
\right)
\;. \label{eq:initial_displacement}
\end{equation}
They spontaneously alternate large and weak distortions, to which $ u_1 $ and $ u_2 $ are assigned [Fig.~\ref{fig:negativeT}(a)]. The electron-quantum-phonon ground state is in the band-insulator phase, which is schematically shown in Fig.~\ref{fig:negativeT}(b) with $ t_1 > t_2 > 0 $: it consists of strongly ($ u_1 < 0 $, large $ \mid u_1 \mid $) distorted dimers with the bonding orbitals occupied and weakly ($ u_2 < 0 $, small $ \mid u_2 \mid $) distorted dimers with both the bonding and antibonding orbitals occupied.
Photoexcitation is introduced through the Peierls phase
\begin{equation}
c_{i,\sigma}^\dagger c_{j,\sigma} \rightarrow
\exp \left[
\frac{ie}{\hbar c} r_{ij} A(t)
\right] c_{i,\sigma}^\dagger c_{j,\sigma}
\;, \label{eq:photo_excitation}
\end{equation}
where $ r_{ij}=r_j-r_i $ and $ r_i $ being the location of the $ i $-th site, which is substituted into Eq.~(\ref{eq:hamiltonian}). For symmetric monocycle electric-field pulses, we use the time-dependent vector potential,
\begin{equation}
A(t) = \frac{cF}{\omega} \left[ \cos (\omega t)-1 \right]
\theta (t) \theta \left( \frac{2\pi}{\omega}-t \right)
\;, \label{eq:monocycle_pulse}
\end{equation}
where $ F $ is the field amplitude and the central frequency $ \omega $ is chosen to be nearly resonant with the optical gap. The time-dependent Schr\"odinger equation for the exact many-electron-quantum-phonon wave function is numerically solved by expanding the exponential evolution operator with a time slice $ dt $=0.02 to the 15th order and by checking the conservation of the norm. \cite{yonemitsu_prb09,matsubara_prb14} We use the leapfrog method to solve the classical equations for the lattice displacements,
\begin{equation}
\frac{1}{S\omega_P^2} \frac{d^2 u_j}{dt^2} = F_j(\left\{ u \right\})
\;, \label{eq:lattice_motion}
\end{equation}
where the force $ F_j(\left\{ u \right\}) $ is obtained with the aid of the Hellmann-Feynman theorem:
\begin{equation}
F_j(\left\{ u \right\}) =
-\sum_\sigma \left[
e^{\frac{ie}{\hbar c}r_{2j-1,2j}A(t)}
\langle c_{2j-1,\sigma}^\dagger c_{2j,\sigma} \rangle +
e^{\frac{ie}{\hbar c}r_{2j,2j-1}A(t)}
\langle c_{2j,\sigma}^\dagger c_{2j-1,\sigma} \rangle
\right] -\frac{u_j}{S}
\;. \label{eq:force}
\end{equation}
For the bare transfer integrals, we use $ t_0 $=1 and $ t_{\rm inter} $=0.1, which correspond to an extrinsically and strongly dimerized system. The interaction parameters $ U $, $ V $, $ g $, and $ S $ are varied. The Holstein coupling strength $ g $ is chosen to be so weak that the number of phonons is always much smaller than one on every site, $ \langle \Psi (t) \mid b_i^\dagger b_i \mid \Psi (t) \rangle \ll 1 $. The bare phonon frequency $ \omega_H $, corresponding to a molecular vibration frequency, is set to be $ \omega_H=0.2 $. The bare phonon frequency $ \omega_P $, corresponding to an intermolecular oscillation frequency, is set to be $ \omega_P=0.05 $ if only $ S $ is nonzero among the couplings $ U $, $ V $, $ g $, and $ S $ (i.e., in the Peierls model) and $ \omega_P=0 $ with $ S=0.1 $ otherwise (i.e., in the extended Hubbard and Holstein models with frozen lattice displacements). The intermolecular distances $ r_{ij} $ are fixed for simplicity and so chosen that the Bessel functions $ J_0(er_{ij}F/\hbar\omega) $ modulating the transfers between sites $ i $ and $ j=i+1 $ ($ i $=1, 3 for the two nonequivalent intradimer transfers and $ i $=2, 4 for the interdimer transfers) are easily distinguishable: $ a_1 \equiv r_{12} = a $, $ a_2 \equiv r_{34} = 1.2a $, and $ a_{\rm inter} \equiv r_{23} = r_{45} = 1.6a $. For $ S=0.1 $, $ t_1=t_0-u_1 \simeq 1.2 $ and $ t_2=t_0-u_2 \simeq 1.0 $ [Fig.~\ref{fig:negativeT}(a)]. The difference between $ t_1 $ and $ t_2 $ becomes smaller for larger $ U $.
As pointed out previously in Appendix of Ref.~\citen{nishioka_jpsj14} for the two-site tight-binding model and for a resonantly excited strongly dimerized model, the total-energy increment $ \Delta E_{\rm tot} $ is roughly proportional to $ J_0(eaF/\hbar\omega) \sin(eaF/\hbar\omega) $ although $ \Delta E_{\rm tot} $ is nonnegative.
For reference, we show again, in Fig.~\ref{fig:nishioka_A13A11}, the energy increment [Eq.~(A$\cdot$11) in Ref.~\citen{nishioka_jpsj14}] and the term proportional to $ J_0(eaF/\hbar\omega)\sin(eaF/\hbar\omega) $ [Eq.~(A$\cdot$13) in Ref.~\citen{nishioka_jpsj14}] for the monocycle nearly resonant ($ \omega=1.05\omega_0 $ with the bonding-to-antibonding transition energy $ \hbar \omega_0=2t $) excitation of the one-electron two-site tight-binding model with transfer integral $ t $ (upper panel of Fig.~A$\cdot$3 in Ref.~\citen{nishioka_jpsj14}).
\begin{figure}
\includegraphics[height=5.6cm]{66977Fig2.eps}
\caption{(Color online)
Energy increment and $ J_0(eaF/\hbar\omega)\sin(eaF/\hbar\omega) $ term for monocycle nearly resonant ($ \omega=1.05\omega_0 $) excitation of one-electron two-site tight-binding model.\cite{nishioka_jpsj14}
\label{fig:nishioka_A13A11}}
\end{figure}
The maximum of $ J_0(eaF/\hbar\omega)\sin(eaF/\hbar\omega) $ appears around $ eaF/\hbar\omega $=1, while the maximum of the energy increment appears at $ eaF/\hbar\omega $ slightly smaller than 1.
\section{States after Symmetric Monocycle Pulse Excitation}
As a typical case, the time evolution of the charge density $ \langle \Psi (t) \mid n_i \mid \Psi (t) \rangle $ and that of the double occupancy $ \langle \Psi (t) \mid n_{i,\uparrow} n_{i,\downarrow} \mid \Psi (t) \rangle $ are shown in Figs.~\ref{fig:t_evol_U0p15f2p4_nd}(a) and \ref{fig:t_evol_U0p15f2p4_nd}(b), respectively, for $ U=0.15 $, $ V=0 $, $ g=0 $, and $ 0 < t < 50 T $.
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig3.eps}
\caption{(Color online)
(a) Charge density $ \langle \Psi (t) \mid n_i \mid \Psi (t) \rangle $ and (b) double occupancy $ \langle \Psi (t) \mid n_{i,\uparrow} n_{i,\downarrow} \mid \Psi (t) \rangle $, at four nonequivalent sites during and after monocycle pulse excitation with $ eaF/\hbar\omega=1.0 $ and $ \omega=2.4 $. The model parameters are $ U=0.15 $, $ V=0 $, $ g=0 $, $ S=0.1 $, and $ \omega_P=0 $. The time averages below are taken for the time region indicated by $ t_w $.
\label{fig:t_evol_U0p15f2p4_nd}}
\end{figure}
The charge density is, initially and a short time after the pulse excitation $ t > 5T $, almost 1, 1, 2, 2 for sites $ i $=1, 2, 3, 4, respectively [see also Fig.~\ref{fig:negativeT}(a), denoted formally by the (1,1,2,2) configuration hereafter]. Because the present system is in the band-insulator phase, the time evolution of the double occupancy is quite similar to that of the charge density. The time-averaged double occupancy below is calculated by
\begin{equation}
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle =
\frac{1}{t_w} \int_{t_s}^{t_s+t_w}
\langle \Psi (t) \mid n_{i,\uparrow} n_{i,\downarrow} \mid \Psi (t) \rangle dt
\;, \label{eq:time_average}
\end{equation}
with $ t_s=5 T $, $ t_w=45 T $, and $ T=2\pi/\omega $ being the period of the oscillating electric field. The time domain of integration is indicated by the two-headed arrow. The other time-averaged quantities are calculated likewise.
\subsection{Effect of on-site repulsion $ U $}
Time-averaged quantities for $ V=0 $ and $ g=0 $ with different values of $ U $ and fixed $ \omega $=2.4 are shown in Fig.~\ref{fig:Fw_d_nn_U_2p5} in the range of $ 0< eaF/\hbar\omega <2.5 $.
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig4.eps}
\caption{(Color online)
(a) Total-energy increment $ \Delta E_{\rm tot} $, (b) time-averaged double occupancy averaged over strongly distorted dimer sites $ (\langle \langle n_{1,\uparrow} n_{1,\downarrow} \rangle \rangle + \langle \langle n_{2,\uparrow} n_{2,\downarrow} \rangle \rangle)/2 $ and that averaged over weakly distorted dimer sites $ (\langle \langle n_{3,\uparrow} n_{3,\downarrow} \rangle \rangle + \langle \langle n_{4,\uparrow} n_{4,\downarrow} \rangle \rangle)/2 $, (c) average double occupancy $ \frac{1}{N} \sum_{i=1}^{N} \langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle $, and (d) average nearest-neighbor charge-density correlation $ \frac{1}{N} \sum_{i=1}^{N} \langle \langle n_i n_{i+1} \rangle \rangle $, as a function of $ eaF/\hbar\omega $, for different values of $ U $. The other parameters are $ V=0 $, $ g=0 $, $ S=0.1 $, $ \omega_P=0 $, and $ \omega=2.4 $. In (c), the vicinity of $ eaF/\hbar\omega $=0.9 is enlarged in the inset.
\label{fig:Fw_d_nn_U_2p5}}
\end{figure}
The total-energy increment $ \Delta E_{\rm tot} $ is almost independent of $ U $ in the range of $ U < 0.23 $ [Fig.~\ref{fig:Fw_d_nn_U_2p5}(a)]. The maximum appears around $ eaF/\hbar\omega $=0.9. As is clear from the comparison between Figs.~\ref{fig:nishioka_A13A11} and \ref{fig:Fw_d_nn_U_2p5}(a), it is governed by the Bessel function $ J_0(eaF/\hbar\omega) $ modulating the transfer between sites 1 and 2 within the strongly distorted dimer: the system is excited by the electronic transitions within the strongly distorted dimers. The maximum value is indeed close to the energy increment expected when all electrons in the strongly distorted dimers transit from the bonding orbitals to the antibonding orbitals [Fig.~\ref{fig:negativeT}(c)]: $ 2 t_1 \times 2 \times $(16 sites/4 sites)$ \simeq 2.3 \times 2 \times 4 \simeq 18$.
As $ eaF/\hbar\omega $ increases from zero to about 0.5, the time-averaged charge density $ \langle \langle n_{i} \rangle \rangle $ changes from the (1,1,2,2) configuration toward a uniform distribution, especially for large $ U $, so that the charge order is weakened and the time-averaged double occupancy $ \langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle $ also changes toward a uniform distribution [Fig.~\ref{fig:Fw_d_nn_U_2p5}(b)]. However, around $ eaF/\hbar\omega $=0.9, where $ \Delta E_{\rm tot} $ reaches the maximum, the configurations of $ \langle \langle n_{i} \rangle \rangle $ and $ \langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle $ almost return to those before the photoexcitation, which correspond to $ eaF/\hbar\omega $=0. This indicates that, in the corresponding state, all electrons in the strongly distorted dimers indeed transit from the bonding orbitals to the antibonding orbitals [Fig.~\ref{fig:negativeT}(c)] and almost no electrons are transferred between dimers. If only these intradimer transitions were realized, the site-diagonal quantities such as $ \langle \langle n_{i} \rangle \rangle $, $ \langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle $, and the nearest-neighbor charge-density correlation $ \langle \langle n_i n_{i+1} \rangle \rangle $ would be unchanged. This very state is called a negative-temperature state.
In the ground state and weakly excited states of small $ eaF/\hbar\omega $, as $ U $ increases, the average double occupancy decreases [Fig.~\ref{fig:Fw_d_nn_U_2p5}(c)]. Around $ eaF/\hbar\omega $=0.9, however, the average double occupancy for $ U > 0 $ is larger than that of the noninteracting ($ U $=0) ground state ($ eaF/\hbar\omega $=0) and increases as $ U $ increases. This behavior with increasing $ U $ corresponds to more attractive on-site interactions for equilibrium: it appears as if the negative-temperature state has an attractive on-site interaction and its strength increases with $ U $ when the time-averaged correlation is interpreted in terms of the correlation in equilibrium. Near the maximum of $ \Delta E_{\rm tot} $, the energy supplied by the photoexcitation goes slightly to the $ U $ term in Eq.~(\ref{eq:hamiltonian}) to enhance the average double occupancy, as discussed later in Sect.~\ref{subsec:flow}.
For small $ eaF/\hbar\omega $, as $ U $ increases, the average nearest-neighbor charge-density correlation increases [Fig.~\ref{fig:Fw_d_nn_U_2p5}(d)]. Around $ eaF/\hbar\omega $=0.9, this correlation is larger than that of the ground state, slightly decreases as $ U $ varies from $ U $=0 to $ U $=0.1, and then increases as $ U $ varies from $ U $=0.1 to $ U $=0.2. This fact cannot be interpreted in terms of an effective on-site interaction for equilibrium.
The same quantities as in Fig.~\ref{fig:Fw_d_nn_U_2p5} are shown in Fig.~\ref{fig:Fw_d_nn_U_3_8} for larger field amplitudes, $ 3< eaF/\hbar\omega <8 $.
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig5.eps}
\caption{(Color online)
Same quantities as in Fig.~\ref{fig:Fw_d_nn_U_2p5}, with the same ordinate axes and for $ 3< eaF/\hbar\omega <8 $.
\label{fig:Fw_d_nn_U_3_8}}
\end{figure}
For comparison, the same ordinate axes are used. Local maxima of the total-energy increment $ \Delta E_{\rm tot} $ appear around $ eaF/\hbar\omega $=4 and $ eaF/\hbar\omega $=7.3 [Fig.~\ref{fig:Fw_d_nn_U_3_8}(a)]. Their values depart from the maximum value with increasing $ eaF/\hbar\omega $. Around $ eaF/\hbar\omega $=4, the weakened charge order recovers to a small extent [Fig.~\ref{fig:Fw_d_nn_U_3_8}(b), compare it with Fig.~\ref{fig:Fw_d_nn_U_2p5}(b)]. Around $ eaF/\hbar\omega $=7.3, the weakened charge order does not recover at all. The realized state differs from the typical negative-temperature state at $ eaF/\hbar\omega $=0.9 because the electrons are transferred from the weakly distorted dimers to the strongly distorted dimers. Thus, the charge order is weaker and $ \Delta E_{\rm tot} $ is smaller than the corresponding quantities at $ eaF/\hbar\omega $=0.9. Around $ eaF/\hbar\omega $=4 and $ eaF/\hbar\omega $=7.3, as $ U $ increases, the average double occupancy decreases except for $ eaF/\hbar\omega $=4 with $ 0 < U < 0.1 $ [Fig.~\ref{fig:Fw_d_nn_U_3_8}(c)]: the effective on-site attraction exists for $ eaF/\hbar\omega $=4 and $ 0 < U < 0.1 $ only. The average nearest-neighbor charge-density correlation indeed increases as $ U $ increases even around them [Fig.~\ref{fig:Fw_d_nn_U_3_8}(d)].
\subsection{Effect of nearest-neighbor repulsion $ V $}
Time-averaged quantities for $ U=0.2 $ and $ g=0 $ with different values of $ V $ and fixed $ \omega=2.4 $ are shown for $ 0 < eaF/\hbar\omega < 2.5 $ in Fig.~\ref{fig:Fw_nn_d_UmV_2p5}.
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig6.eps}
\caption{(Color online)
(a) Total-energy increment $ \Delta E_{\rm tot} $, (b) time-averaged nearest-neighbor charge-density correlation $ \langle \langle n_i n_{i+1} \rangle \rangle $ at interdimer ($ i=2 $) and two nonequivalent intradimer ($ i=1,3 $) bonds, (c) average nearest-neighbor charge-density correlation $ \frac{1}{N} \sum_{i=1}^{N} \langle \langle n_i n_{i+1} \rangle \rangle $, and (d) average double occupancy $ \frac{1}{N} \sum_{i=1}^{N} \langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle $, as a function of $ eaF/\hbar\omega $, for different values of $ V $. The other parameters are $ U=0.2 $, $ g=0 $, $ S=0.1 $, $ \omega_P=0 $, and $ \omega=2.4 $.
\label{fig:Fw_nn_d_UmV_2p5}}
\end{figure}
The total-energy increment $ \Delta E_{\rm tot} $ is almost independent of $ V $ [Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(a)]. The maximum appears around $ eaF/\hbar\omega $=0.9. The maximum value is again close to the energy increment expected when Fig.~\ref{fig:negativeT}(c) is realized.
As $ eaF/\hbar\omega $ increases from zero to about 0.5, $ \langle \langle n_{i} \rangle \rangle $ changes from the (1,1,2,2) configuration toward a uniform distribution so that the charge order is weakened and $ \langle \langle n_i n_{i+1} \rangle \rangle $ also changes toward a uniform distribution [Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(b)]. However, around $ eaF/\hbar\omega $=0.9, the configurations of $ \langle \langle n_{i} \rangle \rangle $ and $ \langle \langle n_i n_{i+1} \rangle \rangle $ almost return to those before the photoexcitation. Namely, the negative-temperature state [Fig.~\ref{fig:negativeT}(c)] is realized.
In the ground state and weakly excited states of small $ eaF/\hbar\omega $, as $ V $ increases, the average nearest-neighbor charge-density correlation decreases [Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(c)] and the average double occupancy increases [Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(d)]. Around $ eaF/\hbar\omega $=0.9, the average nearest-neighbor charge-density correlation is larger than that of the ground state ($ eaF/\hbar\omega $=0) for $ V $=0 and increases as $ V $ increases. At the same place, the average double occupancy decreases as $ V $ increases. These behaviors with increasing $ V $ are consistent with more attractive nearest-neighbor interactions for equilibrium. Near the maximum of $ \Delta E_{\rm tot} $, the supplied energy goes to the $ V $ term in Eq.~(\ref{eq:hamiltonian}) to enhance the average nearest-neighbor charge-density correlation, as discussed later in Sect.~\ref{subsec:flow}.
A typical negative-temperature state does not appear at $ eaF/\hbar\omega $=4 or $ eaF/\hbar\omega $=7.3 (not shown), where a local maximum of $ \Delta E_{\rm tot} $ appears. Around $ eaF/\hbar\omega $=4, the weakened charge order recovers to a small extent. Around $ eaF/\hbar\omega $=7.3, the weakened charge order does not recover at all. In both cases, as $ V $ increases, the average nearest-neighbor charge-density correlation decreases except for $ eaF/\hbar\omega $=4 with $ 0 < V < 0.05 $: the effective nearest-neighbor attraction exists for $ eaF/\hbar\omega $=4 and $ 0 < V < 0.05 $ only. How these states deviate from the typical negative-temperature state is similar to that shown previously in Fig.~\ref{fig:Fw_d_nn_U_3_8}.
\subsection{Effect of Holstein electron-phonon coupling $ g $}
The total-energy increment $ \Delta E_{\rm tot} $ and the time-averaged Holstein electron-phonon correlation $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ for $ U=0 $, $ V=0 $, and $ \omega_H=0.2 $ with different values of $ g $ and fixed $ \omega=2.4 $ are shown for $ 0 < eaF/\hbar\omega < 2.5 $ in Figs.~\ref{fig:Fw_hlst_G_2p5}(a) and \ref{fig:Fw_hlst_G_2p5}(b), respectively.
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig7.eps}
\caption{(Color online)
(a) Total-energy increment $ \Delta E_{\rm tot} $ and (b) time-averaged Holstein electron-phonon correlation averaged over almost singly occupied sites $ (\langle \langle (b_1+b_1^\dagger)(n_1-3/2) \rangle \rangle + \langle \langle (b_2+b_2^\dagger)(n_2-3/2) \rangle \rangle)/2 $ and that averaged over almost doubly occupied sites $ (\langle \langle (b_3+b_3^\dagger)(n_3-3/2) \rangle \rangle + \langle \langle (b_4+b_4^\dagger)(n_4-3/2) \rangle \rangle)/2 $, as a function of $ eaF/\hbar\omega $, for different values of $ g $. The other parameters are $ U=0 $, $ V=0 $, $ S=0.1 $, $ \omega_H=0.2 $, $ \omega_P=0 $, and $ \omega=2.4 $.
\label{fig:Fw_hlst_G_2p5}}
\end{figure}
The maximum of $ \Delta E_{\rm tot} $ appears around $ eaF/\hbar\omega $=0.9 again. The maximum value is close to the energy increment expected when Fig.~\ref{fig:negativeT}(c) is realized: $ 2 t_1 \times 2 \times $(8 sites/4 sites)$ \simeq 2.3 \times 2 \times 2 \simeq 9$.
In the ground state and weakly excited states of small $ eaF/\hbar\omega $, the correlations $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ are negative and their magnitudes increase as $ g $ increases [Fig.~\ref{fig:Fw_hlst_G_2p5}(b)].
Note that, in equilibrium, the magnitude of the correlation $ \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle $ is larger at the almost singly occupied sites even though the intramolecular displacement $ \langle b_i+b_i^\dagger \rangle $ at the almost singly occupied sites and that at the almost doubly occupied sites have opposite signs and the same magnitude, as $ \langle n_i-3/2 \rangle $ at the almost singly occupied sites ($ \simeq -0.5 $) and that at the almost doubly occupied sites ($ \simeq +0.5 $).
As $ eaF/\hbar\omega $ increases from zero to about 0.7, the correlations are weakened so that the magnitudes of $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ become small. Around $ eaF/\hbar\omega $=0.9, $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ at the almost singly occupied sites is positive and increases as $ g $ increases. On the other hand, $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ at the almost doubly occupied sites remains negative, but its magnitude is quite small around $ eaF/\hbar\omega $=0.9. If $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ are spatially averaged over all sites, the average correlation is positive and increases with $ g $ around $ eaF/\hbar\omega $=0.9. This behavior corresponds to inverted electron-phonon couplings (from positive $ g $ to negative $ g $) for equilibrium. The energy supplied by the photoexcitation goes slightly to the $ g $ term in Eq.~(\ref{eq:hamiltonian}) to invert the average Holstein correlation, as discussed later in Sect.~\ref{subsec:flow}.
This fact is consistent with the finding in Ref.~\citen{kawakami_prl10}: at the early stage of the photoinduced melting of charge order in $\alpha$-(BEDT-TTF)$_2$I$_3$, the oscillation of the correlated electrons between the neighboring molecules interferes destructively with intramolecular vibrations. Their coupling is theoretically treated by the Holstein term.\cite{kawakami_prl10} When a sufficient amount of energy goes to the $ g $ term, the phase of the electronic oscillation relative to the phase of the intramolecular vibration becomes opposite to that in equilibrium. It is because, in the ground state, the relative phase is determined in such a way that the interaction energy is lowered. This change in the relative phase is observed as the destructive interference.
A typical negative-temperature state does not appear at $ eaF/\hbar\omega $=4 or $ eaF/\hbar\omega $=7.3 (not shown), where a local maximum of $ \Delta E_{\rm tot} $ appears.
Around $ eaF/\hbar\omega $=4, the magnitudes of $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ are small, but they remain negative. Around $ eaF/\hbar\omega $=7.3, the magnitudes of $ \langle \langle (b_i+b_i^\dagger)(n_i-3/2) \rangle \rangle $ become slightly smaller than the neighborhood, but they are not so small compared with those around $ eaF/\hbar\omega $=4. Namely, the realized state departs from any state with an inverted interaction.
\subsection{Effect of Peierls electron-lattice coupling $ S $}
Time-averaged quantities for $ U=0 $, $ V=0 $, $ g=0 $, and $ \omega_P=0.05 $ with different values of $ S $ and $ \omega $ are shown for $ 0 < eaF/\hbar\omega < 2.5 $ in Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}.
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig8.eps}
\caption{(Color online)
(a) Total-energy increment $ \Delta E_{\rm tot} $, (b) time-averaged bond density $ \langle \langle \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $, (c) time-averaged lattice displacement $ \langle \langle u_1 \rangle \rangle $, and (d) time-averaged Peierls electron-lattice correlation $ \langle \langle u_1 \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $, as a function of $ eaF/\hbar\omega $, for different values of $ S $. The other parameters are $ U=0 $, $ V=0 $, $ g=0 $, and $ \omega_P=0.05 $. The central frequency is set at $ \omega=2.19 $ for $ S=0.02 $, $ \omega=2.25 $ for $ S=0.05 $, $ \omega=2.41 $ for $ S=0.1 $, and $ \omega=2.8 $ for $ S=0.2 $.
\label{fig:Fw_bnd_frc_prl_S_2p5}}
\end{figure}
The optical gap sensitively depends on $ S $, so that the central frequency $ \omega $, which is chosen to be resonant with the optical gap, is varied: $ \omega=2.19 $ for $ S=0.02 $, $ \omega=2.25 $ for $ S=0.05 $, $ \omega=2.41 $ for $ S=0.1 $, and $ \omega=2.8 $ for $ S=0.2 $. As $ S $ increases, the energy difference between the bonding and antibonding orbitals in the strongly distorted dimer increases. Thus, the maximum of $ \Delta E_{\rm tot} $, which appears around $ eaF/\hbar\omega $=0.9, also increases with $ S $ [Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(a)]. The maximum value is again close to the energy increment expected when Fig.~\ref{fig:negativeT}(c) is realized.
As $ eaF/\hbar\omega $ increases, the time-averaged bond density between the almost singly occupied sites $ \langle \langle \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $ decreases and then becomes negative [Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(b)]. Around $ eaF/\hbar\omega $=0.9, its magnitude shows a local maximum. Although its magnitude is smaller than that in the ground state, this corresponds to the negative-temperature state shown in Fig.~\ref{fig:negativeT}(c). Around $ eaF/\hbar\omega $=2, where $ \Delta E_{\rm tot} $ almost vanishes, the time-averaged bond density almost takes the value taken by the ground state, i.e., the time-averaged bond density almost remains the initial value before the photoexcitation. In the ground state and weakly excited states of small $ eaF/\hbar\omega $, the time-averaged lattice displacement between the almost singly occupied sites $ \langle \langle u_1 \rangle \rangle $ is negative and its magnitude increases as $ S $ increases [Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(c)]. As $ eaF/\hbar\omega $ increases, $ \langle \langle u_1 \rangle \rangle $ increases and then becomes positive. Around $ eaF/\hbar\omega $=0.9, $ \langle \langle u_1 \rangle \rangle $ shows a local maximum, which increases with $ S $. The magnitude of $ \langle \langle u_1 \rangle \rangle $ here is close to, but smaller than, that in the ground state.
Actually, around $ eaF/\hbar\omega $=2, all time-averaged quantities almost take the values taken by the ground state in all cases of Figs.~ \ref{fig:Fw_d_nn_U_2p5}, \ref{fig:Fw_nn_d_UmV_2p5}, \ref{fig:Fw_hlst_G_2p5}, and \ref{fig:Fw_bnd_frc_prl_S_2p5}. This situation is similar to the time behavior of the double occupancy in the cases $ eaF/\hbar\omega $=2 and $ eaF/\hbar\omega $=2.5 of Ref.~\citen{tsuji_prl11} after the sudden application of a continuous-wave field to the half-filled Hubbard model in the dynamical mean-field theory. Although the double occupancy is shown for a short time in Ref.~\citen{tsuji_prl11}, its time average would be equal to its initial value for some value of $ eaF/\hbar\omega $ slightly smaller than 2.5. Note that the dynamical localization is expected for $ eaF/\hbar\omega $=2.4 in the continuous-wave case.
In the ground state and weakly excited states of small $ eaF/\hbar\omega $, the time-averaged Peierls electron-lattice correlation between the almost singly occupied sites $ \langle \langle u_1 \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $ is negative and its magnitude increases as $ S $ increases [Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(d)]. As $ eaF/\hbar\omega $ increases from zero to about 0.3, the correlation is weakened so that the magnitude of $ \langle \langle u_1 \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $ becomes small. Around $ eaF/\hbar\omega $=0.5 and $ eaF/\hbar\omega $=1.3, $ \langle \langle u_1 \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $ is quite small but positive for $ S > 0.05 $ and increases as $ S $ increases. The energy used to invert the Peierls correlation is discussed below in Sect.~\ref{subsec:flow}. The field amplitudes $ eaF/\hbar\omega $=0.5 and $ eaF/\hbar\omega $=1.3 correspond to those that weaken the charge order to the largest degree in Figs.~ \ref{fig:Fw_d_nn_U_2p5}(b) and \ref{fig:Fw_nn_d_UmV_2p5}(b). Namely, the electrons are transferred most extensively between dimers. This is the reason why $ \langle \langle u_1 \sum_\sigma \left( c_{1,\sigma}^\dagger c_{2,\sigma} + c_{2,\sigma}^\dagger c_{1,\sigma} \right) \rangle \rangle $ is quite small.
If both the time-averaged bond density and $ \langle \langle u_1 \rangle \rangle $ invert their signs, their correlation should not invert its sign. This happens around $ eaF/\hbar\omega $=0.9. Here also, the magnitude of their correlation increases as $ S $ increases. This corresponds to the negative-temperature state. Namely, the negative-temperature state appears when $ \Delta E_{\rm tot} $ reaches its maximum ($ eaF/\hbar\omega $=0.9), while the inverted electron-lattice coupling [$(t_0-u_j)$ to $(t_0+u_j)$ in Eq.~(\ref{eq:hamiltonian})] appears in different field amplitudes ($ eaF/\hbar\omega $=0.5 and $ eaF/\hbar\omega $=1.3).
\subsection{Energy flows and inverted interactions \label{subsec:flow}}
So far, we mainly show whether the correlation functions merely increase or decrease as a function of the coupling strength or the field amplitude. In this section, we focus on those photoinduced states whose correlation functions behave as if the interactions are inverted. We show how the inverted correlations are related to changes in the total energy and in the interaction energy.
The enhanced average double occupancy is defined by
\begin{equation}
\frac{1}{N} \sum_{i=1}^{N} \Delta_U
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle =
\frac{1}{N} \sum_{i=1}^{N} \left(
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle \mid_{U>0} -
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle \mid_{U=0}
\right)
\;,
\end{equation}
for fixed $ eaF/\hbar\omega $. The regions where the enhanced average double occupancy is positive are shown as the regions of the inverted interaction in Fig.~\ref{fig:UVGS_DeltaCorr}(a).
\begin{figure}
\includegraphics[height=11.2cm]{66977Fig9.eps}
\caption{(Color online)
Regions of inverted interactions, denoted by `II', in planes spanned by $ eaF/\hbar\omega $ and
(a) $ U $ ($ V $=0, $ g $=0, $ S $=0.1, and $ \omega_P $=0),
(b) $ V $ ($ U $=0.2, $ g $=0, $ S $=0.1, and $ \omega_P $=0),
(c) $ g $ ($ U $=0, $ V $=0, $ S $=0.1, $ \omega_H $=0.2, and $ \omega_P $=0), and (d) $ S $ ($ U $=0, $ V $=0, $ g $=0, and $ \omega_P $=0.05),
from data shown in Figs.~ \ref{fig:Fw_d_nn_U_2p5}(c)--\ref{fig:Fw_d_nn_U_3_8}(c), \ref{fig:Fw_nn_d_UmV_2p5}(c), \ref{fig:Fw_hlst_G_2p5}(b), and \ref{fig:Fw_bnd_frc_prl_S_2p5}(d), respectively, for resonant excitations.
\label{fig:UVGS_DeltaCorr}}
\end{figure}
They appear around the local maxima of $ \Delta E_{\rm tot} $. For $ eaF/\hbar\omega $=0.92 and $ U $=0.15, the enhanced average double occupancy is 6$\times$10$^{-3}$ in Fig.~\ref{fig:Fw_d_nn_U_2p5}(c). This value is rather small compared with the $ U $ dependence of the average double occupancy in equilibrium. The energy flowed to enhance the average double occupancy $ \Delta_U E_U $ is defined as
\begin{equation}
\Delta_U E_U = U \sum_{i=1}^{N} \Delta_U
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle
\;.
\end{equation}
In the case above, $ \Delta_U E_U $ is 1.4$\times$10$^{-2}$. We further define $ \Delta_U \Delta E_{\rm tot} $ by
\begin{equation}
\Delta_U \Delta E_{\rm tot} =
\Delta E_{\rm tot} \mid_{U>0} - \Delta E_{\rm tot} \mid_{U=0}
\;,
\end{equation}
for fixed $ eaF/\hbar\omega $. In the case above, $ \Delta_U \Delta E_{\rm tot} $ is 1.8$\times$10$^{-1}$, while $ \Delta E_{\rm tot} $ is 17. The ratio $ \Delta_U E_U / \Delta_U \Delta E_{\rm tot} $ is then 7.8$\times$10$^{-2}$, while the ratio $ \Delta_U E_U / \Delta E_{\rm tot} $ is 8.3$\times$10$^{-4}$. The increased average double occupancy is defined by
\begin{equation}
\frac{1}{N} \sum_{i=1}^{N} \Delta
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle =
\frac{1}{N} \sum_{i=1}^{N} \left(
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle \mid_{F>0} -
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle \mid_{F=0}
\right)
\;,
\end{equation}
for fixed $ U $. Note that this quantity is negative for $ U $=0 and positive for $ U \geq $0.05 in Fig.~\ref{fig:Fw_d_nn_U_2p5}(c). Using this quantity, the interaction-energy increment $ \Delta E_U $ is evaluated as
\begin{equation}
\Delta E_U = U \sum_{i=1}^{N} \Delta
\langle \langle n_{i,\uparrow} n_{i,\downarrow} \rangle \rangle
\;.
\end{equation}
For $ eaF/\hbar\omega $=0.92 and $ U $=0.15, the increased average double occupancy is 3.2$\times$10$^{-2}$ in Fig.~\ref{fig:Fw_d_nn_U_2p5}(c), and $ \Delta E_U $ is 7.7$\times$10$^{-2}$, so that $ \Delta E_U / \Delta E_{\rm tot} $ is 4.4$\times$10$^{-3}$.
The enhanced average nearest-neighbor charge-density correlation is defined by
\begin{equation}
\frac{1}{N} \sum_{i=1}^{N} \Delta_V
\langle \langle n_{i} n_{i+1} \rangle \rangle =
\frac{1}{N} \sum_{i=1}^{N} \left(
\langle \langle n_{i} n_{i+1} \rangle \rangle \mid_{V>0} -
\langle \langle n_{i} n_{i+1} \rangle \rangle \mid_{V=0}
\right)
\;,
\end{equation}
for fixed $ eaF/\hbar\omega $. The regions where the enhanced average nearest-neighbor charge-density correlation is positive are shown as the regions of the inverted interaction in Fig.~\ref{fig:UVGS_DeltaCorr}(b). They appear around the local maxima of $ \Delta E_{\rm tot} $. For $ eaF/\hbar\omega $=0.92 and $ V $=0.2, the enhanced average nearest-neighbor charge-density correlation is 3.4$\times$10$^{-2}$ in Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(c). This value is rather large compared with the $ V $ dependence of the average nearest-neighbor charge-density correlation in equilibrium. The energy flowed to enhance the average nearest-neighbor charge-density correlation is evaluated as
\begin{equation}
\Delta_V E_V = V \sum_{i=1}^{N} \Delta_V
\langle \langle n_{i} n_{i+1} \rangle \rangle
\;.
\end{equation}
In the case above, $ \Delta_V E_V $ is 1.1$\times$10$^{-1}$. The quantity $ \Delta_V \Delta E_{\rm tot} $, defined by
\begin{equation}
\Delta_V \Delta E_{\rm tot} =
\Delta E_{\rm tot} \mid_{V>0} - \Delta E_{\rm tot} \mid_{V=0}
\;,
\end{equation}
is 5.5$\times$10$^{-2}$, while $ \Delta E_{\rm tot} $ is 17. The ratio $ \Delta_V E_V / \Delta_V \Delta E_{\rm tot} $ is then about 2, while the ratio $ \Delta_V E_V / \Delta E_{\rm tot} $ is 6.3$\times$10$^{-3}$. The fact that $ \Delta_V E_V / \Delta_V \Delta E_{\rm tot} $ is larger than 1 is realized by the relation $ \Delta_V E_U < 0 $, as shown in Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(d): the large enhancement of the average nearest-neighbor charge-density correlation is realized by the reduction of the average double occupancy in the system with competing interactions $ U $ and $ V $. The increased average nearest-neighbor charge-density correlation, which is defined by
\begin{equation}
\frac{1}{N} \sum_{i=1}^{N} \Delta
\langle \langle n_{i} n_{i+1} \rangle \rangle =
\frac{1}{N} \sum_{i=1}^{N} \left(
\langle \langle n_{i} n_{i+1} \rangle \rangle \mid_{F>0} -
\langle \langle n_{i} n_{i+1} \rangle \rangle \mid_{F=0}
\right)
\;,
\end{equation}
for fixed $ V $, is 6.5$\times$10$^{-2}$ in Fig.~\ref{fig:Fw_nn_d_UmV_2p5}(c), and the interaction-energy increment $ \Delta E_V $,
\begin{equation}
\Delta E_V = V \sum_{i=1}^{N} \Delta
\langle \langle n_{i} n_{i+1} \rangle \rangle
\;,
\end{equation}
is 2.1$\times$10$^{-1}$, so that $ \Delta E_V / \Delta E_{\rm tot} $ is 1.2$\times$10$^{-2}$ in the case above.
The enhanced average Holstein correlation is given by the average of the quantities for the almost singly occupied and doubly occupied sites shown in Fig.~\ref{fig:Fw_hlst_G_2p5}(b) because it is zero for $ g $=0. The region where the enhanced average Holstein correlation is positive is shown as the region of the inverted interaction in Fig.~\ref{fig:UVGS_DeltaCorr}(c). It appears only around the maximum of $ \Delta E_{\rm tot} $. Because we used small $ g $ for the technical reason mentioned before, every energy scale is small.
The enhanced average Peierls correlation is given by the quantity shown in Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(d) because it is zero for $ S $=0. The regions where the enhanced average Peierls correlation is positive are shown as the regions of the inverted interaction in Fig.~\ref{fig:UVGS_DeltaCorr}(d). They appear on both sides of the local maxima of $ \Delta E_{\rm tot} $. For $ eaF/\hbar\omega $=0.57 and $ S $=0.2, this quantity is 1.2$\times$10$^{-2}$ in Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(d). The energy flowed to enhance the average Peierls correlation $ \Delta_S E_S $,
\begin{equation}
\Delta_S E_S = \sum_{j=1}^{N/2} \langle \langle u_j \sum_\sigma \left(
c_{2j-1,\sigma}^\dagger c_{2j,\sigma} + c_{2j,\sigma}^\dagger c_{2j-1,\sigma}
\right) \rangle \rangle
\;,
\end{equation}
is 5$\times$10$^{-2}$. The quantity $ \Delta_S \Delta E_{\rm tot} $ is defined as before, and $ \Delta_S \Delta E_{\rm tot} $ is 1.8, while $ \Delta E_{\rm tot} $ is 13. The ratio $ \Delta_S E_S / \Delta_S \Delta E_{\rm tot} $ is 2.7$\times$10$^{-2}$, while the ratio $ \Delta_S E_S / \Delta E_{\rm tot} $ is 3.6$\times$10$^{-3}$. The increased average Peierls correlation is 7.6$\times$10$^{-1}$ in Fig.~\ref{fig:Fw_bnd_frc_prl_S_2p5}(d), and the interaction-energy increment $ \Delta E_S $,
\begin{equation}
\Delta E_S = \sum_{j=1}^{N/2} \left[
\langle \langle u_j \sum_\sigma \left(
c_{2j-1,\sigma}^\dagger c_{2j,\sigma} + c_{2j,\sigma}^\dagger c_{2j-1,\sigma}
\right) \rangle \rangle \mid_{F>0} -
\langle \langle u_j \sum_\sigma \left(
c_{2j-1,\sigma}^\dagger c_{2j,\sigma} + c_{2j,\sigma}^\dagger c_{2j-1,\sigma}
\right) \rangle \rangle \mid_{F=0}
\right]
\;,
\end{equation}
is about 3.0, so that $ \Delta E_S / \Delta E_{\rm tot} $ is 0.23. This ratio is relatively large because the ground state is stabilized by the Peierls electron-lattice coupling in the first place.
Comparing Figs.~\ref{fig:UVGS_DeltaCorr}(c) with \ref{fig:UVGS_DeltaCorr}(d), we notice that the region of the inverted interaction is wider for $ S $ than for $ g $. When we see the local maxima of $ \Delta E_{\rm tot} $, there is indeed a difference between them. For $ g \neq 0 $, with increasing $ eaF/\hbar\omega $, the deviation from the typical negative-temperature state becomes large. For $ S \neq 0 $ but $ g=0 $, with increasing $ eaF/\hbar\omega $, the deviation does not seem to become so large at least for $ eaF/\hbar\omega < 10 $. In the latter, the disordering effect seems small. If we increased $ \omega_P $ and treated $ u_j $ quantum-mechanically, the disordering effect might become large to suppress the inversion of the Peierls electron-phonon coupling for large $ eaF/\hbar\omega $.
In all cases above, for the interaction parameter $ X $ ($ X $=$ U $, $ V $, $ g $, or $ S $), the average correlation enhanced by the interaction $ X $ is smaller than the average correlation increased by the field $ F $. Even in the system with competing interactions $ U $ and $ V $, the ratio of the average nearest-neighbor charge-density correlation enhanced by $ V $ to that increased by $ F $ is about 1 to 2. When the interaction-induced change is compared with the field-induced change for the energy ratio, we find $ \Delta_X E_X / \Delta_X \Delta E_{\rm tot} > \Delta E_X / \Delta E_{\rm tot} $ for $ X $=$ U $ and $ V $, while $ \Delta_X E_X / \Delta_X \Delta E_{\rm tot} < \Delta E_X / \Delta E_{\rm tot} $ for $ X $=$ S $. In any case, $ \Delta_X E_X / \Delta E_{\rm tot} $ is less than $ 10^{-2} $ and very small. The efficiency of the inversion of interactions is quite low.
\section{Discussion}
In general, the formation of a negative-temperature state requires suppressing the disordering effect (i.e., thermalization) on increasing the total energy. The present model has a simple structure. The strongly distorted dimer is regarded as a type of two-level system, which is a typical system to achieve a negative-temperature state. The weakly distorted dimers are then regarded as the environment, to which the energy flow from the system is allowed. The interdimer charge transfer is a type of relaxation process. The structure of the present model may be advantageous to forming a negative-temperature state in the sense that the thermalization is suppressed by separating the energy scale of the interdimer electron transfer processes and that of the intradimer process.
In real substances, the network of electron transfers is not always simple as treated in this study. For instance, in the model previously used for Et$_2$Me$_2$Sb[Pd(dmit)$_2$]$_2$,\cite{nishioka_jpsj13a} the weakly distorted dimer has an unoccupied molecular orbital whose energy is not far from those of the antibonding orbitals in the strongly distorted dimer. In such a case, even under optimal conditions (such as $ eaF/\hbar\omega $=0.9 in the present study), the electrons can be easily transferred between dimers so that the negative-temperature state, if realized, can be short-lived, depending on the transfer integral between the corresponding molecular orbitals. We expect that, if the photoinduced melting of charge order (i.e., a high-temperature state) is easily achieved, the negative-temperature state would hardly be achieved, and vice versa.
\section{Conclusions}
Electron-phonon dynamics in the band-insulator phase of one-dimensional three-quarter-filled strongly dimerized extended Peierls-Hubbard and Holstein models after the application of a symmetric monocycle electric-field pulse with central frequency resonant with the optical gap are calculated using the exact diagonalization method. From the field-amplitude and interaction-parameter dependences of time-averaged correlation functions, we discuss the negative-temperature state and the inversion of interactions. The total-energy increment is maximized when the energy gain of an electron transferred to the neighboring site $ eaF $ under the electric field $ F $ is comparable to the photoexcitation energy $ \hbar\omega $ ($ eaF/\hbar\omega $=0.9). The maximum value is close to the energy increment expected when all electrons in the strongly distorted dimers transit from the bonding orbitals to the antibonding orbitals, i.e., the energy difference between the negative-temperature state and the ground state. In this state, the site-diagonal quantities such as the time-averaged charge density, double occupancy, and nearest-neighbor charge-density correlation are close to those in the ground state, and the time-averaged bond density is inverted. Around $ eaF/\hbar\omega $=4 and $ eaF/\hbar\omega $=7.3, the total-energy increment shows a local maximum, but the realized state departs from the typical negative-temperature state at $ eaF/\hbar\omega $=0.9. Compared with the latter, the electrons are transferred between dimers and the charge order is thus weakened.
The interaction-parameter dependences of the time-averaged correlation functions in the negative-temperature state become similar to those in the ground state if the interactions (except the Peierls electron-lattice coupling) are inverted: the repulsive on-site (nearest-neighbor) interaction is converted to an attractive on-site (nearest-neighbor) interaction, and the relative sign between the charge density ($ n_i - 3/2 $) and the intramolecular displacement is inverted. The energy supplied by the photoexcitation goes to these interactions terms. As for the Peierls electron-lattice coupling, the relative sign between the bond density and the intermolecular displacement is inverted for those field amplitudes ($ eaF/\hbar\omega $=0.5 and $ eaF/\hbar\omega $=1.3), which are different from the field amplitude for the negative-temperature state to appear ($ eaF/\hbar\omega $=0.9). The formation of the negative-temperature state and the inversion of interactions have been discussed by Tsuji {\it et al.}\cite{tsuji_prb12} using the dynamical mean-field theory, but they need half-cycle pulses or asymmetric monocycle pulses in the cases studied. In this paper, we demonstrate that they are realized even by symmetric monocycle pulses.
\begin{acknowledgment}
The authors are grateful to Y. Tanaka for various discussions.
This work was supported by Grants-in-Aid for Scientific Research (C) (Grant No. 23540426) and Scientific Research (A) (Grant No. 23244062) from the Ministry of Education, Culture, Sports, Science and Technology of Japan.
\end{acknowledgment}
|
\section{Introduction}\label{sec:introduction}
Heavy-ion fusion reactions at energies around the Coulomb barrier
have been studied for a long time and give interesting phenomena such as
an enhancement of fusion cross sections at energies below the Coulomb barrier.
For a better understanding of fusion dynamics,
it is crucial to consider the couplings between the collective motions
of colliding nuclei and their internal excitations.
To deal with such couplings, coupled-channels models have been developed
and succeeded in reproducing such an enhancement of fusion cross sections at subbarrier energies \cite{balantekin98,dasgupta98,hagino12,back14}.
In those models, unless a colliding system is too heavy or too light,
it has been generally considered that fusion takes place
once colliding nuclei overcome the Coulomb barrier
because of strong absorption inside the Coulomb barrier.
Then, instead of using an imaginary potential inside the Coulomb barrier
with the regular boundary condition at the origin,
an incoming wave boundary condition \cite{dasso83} has often been employed.
The approaching phase leading to overcoming the Coulomb barrier
to make a contact of projectile and target is considered as the capture process.
The property of low-energy fusion dynamics drastically changes when the product of the charge numbers of the projectile and target nuclei, denoted by $Z_PZ_T$, becomes greater than 1600.
In such heavy systems, it has been known that the fusion probability
is extremely hindered around the Coulomb barrier \cite{gaggeler84,sahm84,sahm85,schmidt91}, which is called fusion hindrance\footnote{Notice that this phenomenon is different from fusion hindrance observed at deep subbarrier energies in light- and medium-mass systems \cite{jiang02,jiang04,dasgupta07}.}.
One of the well known experimental observations for exhibiting this fusion hindrance is a direct comparison between the $^{40}$Ar\,+\,$^{180}$Hf ($Z_PZ_T=1296$) and $^{96}$Zr\,+\,$^{124}$Sn ($Z_PZ_T=2000$) systems leading to $^{220}$Th reported in Ref. \cite{sahm85}.
The authors of Ref. \cite{sahm85} measured evaporation-residue cross sections of these systems to deduce their fusion probabilities for central collisions, and then compared them at the Coulomb barrier energy of both reactions estimated by the Bass model \cite{bass74}.
They reported that the fusion probability of the $^{96}$Zr\,+\,$^{124}$Sn system is hindered by two orders of magnitude from that of the $^{40}$Ar\,+\,$^{180}$Hf system and from a simple barrier penetration model and that the apparent fusion barrier for the $^{96}$Zr\,+\,$^{124}$Sn system is shifted by about 26.7\,MeV from the Bass barrier.
Before this observation, Swiatecki and coworkers had predicted that an additional kinetic energy to the Coulomb barrier, called extra-push energy, is needed for fusion to take place
when the so-called fissility parameter of the system (its definition given in Eqs.~(\ref{eq:fissility}) and (\ref{eq:fissility1}) in Sec.~\ref{sec:result:extrapush}) is larger than its threshold value, which approximately corresponds to $Z_PZ_T>1600 - 1800$ \cite{swiatecki81,swiatecki82,bjornholm82,blocki86}.
They considered that, after the contact of colliding nuclei, large Coulomb force might make the colliding system reseparate under the appearance of an additional potential barrier inside the contact configuration.
This indicates the necessity of treating the second step towards fusion, i.e.,
after the capture process as the first step,
the process of shape evolution from the contact configuration of colliding nuclei to the configuration of compound nucleus formation.
Since the projectile and target densities are significantly overlapped
in this formation process, strong energy dissipation is expected to arise from the collective motions to the internal excitations.
Therefore, the potential energy landscape and the mechanism of energy dissipation inside the Coulomb barrier are important quantities to understand the property of the fusion mechanism in heavy systems.
For the analysis of the competition between compound nucleus formation and quasifission in such heavy systems, especially in the context of the synthesis of superheavy elements \cite{hofmann00,oganessian07,morita04,morita12},
dynamical models such as
a diffusion model based on a master equation \cite{adamian97,diaz-torres01,diaz-torres06}
and a macroscopic fluctuation--dissipation model based on a Langevin equation \cite{abe02, aritomo04, zagrebaev05, aritomo12} have been applied.
In the latter model, relevant collective degrees of freedom such as the relative distance and mass partition of colliding nuclei and their deformations are chosen and non-collective degrees of freedom are treated as heat bath.
The inertia parameter, nucleus--nucleus potential, and friction coefficient for the relevant collective degrees of freedom, appearing as transport coefficients in the equation, are important ingredients for such calculations and are mainly determined by macroscopic models.
The main motivation of the present article is to analyze the fusion hindrance
in heavy systems by \emph{microscopic} reaction models.
In this work, we employ the time-dependent Hartree-Fock (TDHF) theory
based on the Skyrme energy density functionals \cite{skyrme56,vautherin72,vautherin73} as a microscopic reaction model.
Energy density functional theory is selfconsistent, fully microscopic and provides a unique tool for describing both the static and dynamical properties of nuclei across the whole nuclear chart in a unified framework \cite{bender03}.
Since the first applications of TDHF simulations to nuclear fusion reactions \cite{bonche76,koonin77,flocard78,negele78,davies78,negele82} were performed about 40 years ago,
much progress of TDHF simulations has been achieved \cite{simenel12}.
Recent three dimensional TDHF calculations are able to include full Skyrme functionals consistent with nuclear structure calculations
and have been applied to fusion reactions \cite{kim97,simenel01, umar06, umar06a, washiyama08,washiyama09}, charge equilibration \cite{iwata10}, transfer reactions \cite{simenel10, scamps13, sekizawa13},
as well as nuclear responses \cite{simenel03, nakatsukasa05, maruhn05,avez08,ebata10,stetcu11,hashimoto12,fracasso12}.
Furthermore, TDHF simulations to reactions with heavy nuclei have been performed \cite{golabek09, kedziora10, umar10, oberacker10, guo12, simenel12, simenel14, oberacker14,wakhle14,scamps15}.
In Refs. \cite{simenel12,guo12}, extra-push energies in several heavy systems have been estimated and the fusion hindrance resulting in finite extra-push energies has been realized, which is consistent with experimental observations.
In Refs. \cite{umar10, oberacker10}, using the so-called density-constrained TDHF (DC-TDHF) method \cite{umar06a}, nucleus--nucleus potential as the minimized energy and excitation energy of a compound system are estimated along a path obtained from TDHF evolutions.
Quasifission process is extensively studied in Refs. \cite{kedziora10,oberacker14,wakhle14}, which have clarified the capability of TDHF simulations for the dynamics of quasifission.
In the present article, we analyze the fusion hindrance in detail from the point of view of energy dissipation and dynamical nucleus--nucleus potential in the entrance channel of fusion reactions in heavy systems.
For this purpose, we employ a method to simultaneously extract dynamical nucleus--nucleus potential and energy dissipation for relevant collective coordinates
from TDHF trajectories, which we proposed in Refs. \cite{washiyama08,washiyama09} and called dissipative-dynamics TDHF (DD-TDHF) method.
The DD-TDHF method relies on the assumption that a complicated microscopic mean-field dynamics in the entrance channel of fusion reactions can be reduced to a
macroscopic equation of motion with a dissipation term.
Then, we apply this method to fusion reactions of heavy, nearly symmetric systems and analyze the property of nucleus--nucleus potential and energy dissipation extracted directly from TDHF simulations.
Using the properties of the extracted potential and dissipation, we perform a systematic study of the analysis of the fusion hindrance in heavy systems.
This article is organized as follows.
In Sec.~\ref{sec:method}, we explain our method of how to analyze the fusion
hindrance by TDHF simulations including how to extract nucleus--nucleus
potential and energy dissipation from TDHF dynamics.
In Sec.~\ref{sec:result}, we present the results of our systematic calculations
for extra-push energies, nucleus--nucleus potentials, energy dissipation in heavy systems, and the analysis of the fusion hindrance.
We give the conclusions of the present article in Sec.~\ref{sec:conclusions}.
We explain some details of our method in the Appendix.
\section{Method}\label{sec:method}
\subsection{Some remarks on TDHF}
In the TDHF theory, quantum many-body dynamics is replaced with
one-body density dynamics, leading to the TDHF equation
\begin{equation}\label{eq:TDHF}
i\hbar\frac{\partial}{\partial t}{\hat\rho}=[{\hat h}[{\hat\rho}],{\hat\rho}],
\end{equation}
where ${\hat\rho}$ and ${\hat h}[{\hat\rho}]$ denote the one-body density
and selfconsistent single-particle Hamiltonian as a functional
of one-body density, respectively. The selfconsistent single-particle
Hamiltonian can be obtained from the energy density functional,
denoted by ${\cal E}[\hat\rho]$, through a variation ${\hat h}[{\hat\rho}]=\partial{\cal E}[\hat\rho]/\partial{\hat\rho}$.
Since the TDHF theory describes collective motions in a semiclassical way,
one can define semiclassical trajectories corresponding to the time evolutions of relevant collective coordinates from one-body density.
In this article, we concentrate on low energy nuclear dynamics,
where the energy of the system is low enough to validate the use of the mean-field dynamics
and is above the Coulomb barrier.
\subsection{DD-TDHF method}\label{sec:method:DD-TDHF}
In this subsection, we briefly recall the DD-TDHF method, proposed in Refs. \cite{washiyama08,washiyama09}.
The method is described as follows.
First, we properly define the collective coordinates to describe fusion reactions.
In this article, we concentrate on central collisions of fusion dynamics
and employ the relative distance $R$ between projectile and target nuclei as a collective coordinate.
An extension to using several collective coordinates is possible.
Then, we construct two-body kinematics of projectile and target nuclei.
In TDHF simulations, we define at each time
a separation plane to separate the total density to the densities of two subsystems. The plane, called neck, is perpendicular to the collision axis, which is always $x$-axis in the case for central collisions.
We will explain how to determine the neck position in Appendix~\ref{sec:appendix:neck}.
The densities of the two subsystems are computed in the coordinate space representation as
\begin{equation}\label{eq:subsystem}
\rho_{1,2}(\boldsymbol{r},t)\equiv \rho(\boldsymbol{r},t)\,\theta[\pm(x-x_0(t))],
\end{equation}
where $\theta(x)$ is the step function and the neck position is at $x=x_0(t)$
depending on time.
Then, we compute the center-of-mass coordinate $R_i$ and associated momentum $P_i$ as well as the mass number $A_i$ for each subsystem $i=1,2$ from the density of each subsystem, $A_i(t)={\rm Tr}(\hat{\rho}_i(t))$, $R_i(t)={\rm Tr}(\hat{x} \hat{\rho}_i(t)) / A_i$, and $P_i(t)={\rm Tr}(\hat{p}_x \hat{\rho}_i(t))$.
The masses of the subsystems are directly computed as
\begin{equation}
m_{i} = P_{i}/\dot{R}_{i},
\end{equation}
for $i=1,2$.
From these quantities, we construct a TDHF trajectory for the relative distance $R$, associated momentum $P$, and reduced mass $\mu$ at each time by
\begin{align}
R& = R_1 - R_2, \\
P& = \frac{m_2 P_1 - m_1 P_2}{m_1 + m_2}, \\
\mu& = \frac{m_1 m_2}{m_1 + m_2}. \label{eq:reduced-mass}
\end{align}
We assume that the time evolutions of $R$ and $P$ obtained from the TDHF evolution of central collisions follow a one-dimensional macroscopic equation of motion with a dissipation term that is assumed to depend on the velocity ${\dot R}$:
\begin{align}
\frac{dR}{dt}& = \frac{P}{\mu}, \label{eq:newton1} \\
\frac{dP}{dt}& = -\frac{dV}{dR} -\frac{d}{dR}\left(\frac{P^2}{2\mu}\right)
-\gamma \frac{dR}{dt}, \label{eq:newton2}
\end{align}
where $V$ and $\gamma$ denote the nucleus--nucleus potential and
friction coefficient, respectively.
This assumption is based on the hypothesis that
a complicated microscopic dynamics obtained from the TDHF evolution
can be reduced to a simple one-dimensional Newton equation.
The aim is to obtain transport coefficients in the equation from TDHF dynamics.
We have already obtained the reduced mass by Eq.~(\ref{eq:reduced-mass}).
To obtain two unknown quantities $dV/dR$ and $\gamma$ as a function of $R$,
we need two equations.
If we assume that the transport coefficients, $\mu$, $dV/dR$, and $\gamma$, at each $R$ do not change by a slight change of energy $\Delta E$,
we can construct a system of two equations for Eq.~(\ref{eq:newton2})
corresponding to two TDHF trajectories at two slightly different energies $E_I$ and $E_{II}=E_I+\Delta E$.
By solving the system of two equations at $R$, we obtain $\gamma(R)$ and $dV/dR$ as
\begin{align}
\gamma(R)&=-\frac{\dot{P}_I-\dot{P}_{II}}{\dot{R}_I-\dot{R}_{II}}+\frac{\dot{R}_{I}+\dot{R}_{II}}{2}\frac{d\mu}{dR}, \label{eq:gamma} \\
\frac{dV(R)}{dR}&=\frac{\dot{R}_I\dot{P}_{II}-\dot{R}_{II}\dot{P}_I}{\dot{R}_I-\dot{R}_{II}}-\frac{\dot{R}_{I}\dot{R}_{II}}{2}\frac{d\mu}{dR},\label{eq:dvdr}
\end{align}
where the subscripts $I$ and $II$ correspond to the trajectories at energies $E_{I}$ and $E_{II}$, respectively,
and we perform all the time derivatives at $R_{I}=R_{II}=R$.
The nucleus--nucleus potential $V$ is obtained from Eq.~(\ref{eq:dvdr}) by integration with the asymptotic form of the Coulomb potential at sufficiently large distances. We employ $R_{\rm max}=20$\,fm in this article.
Therefore, we do not need any normalizations for the potential.
In Refs. \cite{washiyama08,washiyama09}, we applied the DD-TDHF method to fusion reactions in light- and medium-mass systems and gave the following results:
(i) At energies near the Coulomb barrier, dynamical effects significantly decrease the barrier height of the extracted nucleus--nucleus potentials because of reorganization of the density of colliding nuclei, resulting in energy dependence of potential.
(ii) The extracted friction coefficients show a universal property similar to that in one-body dissipation models, and significant energy dependence
related to the picture of the window formula of one-body dissipation mechanism.
From those results, we concluded that the DD-TDHF method is useful tool to investigate the property of nucleus--nucleus potential and one-body energy dissipation in low-energy nuclear reactions.
When we apply the DD-TDHF method to fusion reactions in heavy systems,
we find that, as the overlap of the densities of colliding nuclei becomes substantially large at small relative distances in TDHF simulations,
constructing two-body kinematics of colliding nuclei with a proper determination of the neck position becomes difficult.
Also, we face in some systems the following situations:
(i) the relative velocity or the velocity of a fragment becomes 0,
(ii) the mass of a fragment becomes negative.
We consider that these situations also make it impossible to construct two-body kinematics.
Therefore, we stop performing the extraction once we face these situations or once we can not define the neck position.
In the following, we denote as $R_{\rm min}$ the relative distance at which we stop the extraction.
\subsection{Nucleus--nucleus potential from the frozen density approximation}
\label{sec:method:FD}
As a reference for the nucleus--nucleus potential based on a framework of
the energy density functional, we employ the so-called frozen density approximation \cite{brueckner68,denisov02}.
In this approximation, the energy of the total system at a given relative
distance $R$ between the centers-of-mass of the projectile and target densities, denoted by ${\cal E}[\rho_{P+T}](R)$, is obtained
with keeping their densities frozen to their respective ground-state densities.
Denoting the ground-state densities of the projectile and target nuclei as $\rho_P$ and $\rho_T$, respectively, we obtain the frozen density potential as
\begin{equation}\label{eq:FD}
V^{\rm FD}(R) = {\cal E}[\rho_{P+T}](R) - {\cal E}[\rho_P] - {\cal E}[\rho_T],
\end{equation}
where ${\cal E[\rho]}$ is calculated from the same Skyrme energy density functionals as those used in our TDHF simulations.
In the frozen density potential, effects of possible rearrangement of internal structure during collision are not considered.
Note that we neglect the Pauli effect from the overlap of the projectile and target densities in
estimating ${\cal E}[\rho_{P+T}](R)$. Namely, we approximate $\rho_{P+T}$ as a simple sum of $\rho_P+\rho_T$.
This will be crucial when the overlap becomes large at small relative distances,
leading to overestimation of the attractive potential energy.
Therefore, we use the frozen density potential as a reference
until a relative distance around the Coulomb barrier radius,
where the overlap of projectile and target densities is expected to be small.
\subsection{Definition of fusion in heavy systems}\label{sec:method:definition}
Semiclassical nature of TDHF dynamics enables us to define a threshold energy for classification of nuclear reactions.
In the case of central collisions at low energies,
we can define the fusion threshold energy above which fusion, i.e., the formation of a compound nucleus, takes place.
Following a manner similar to those in previous works \cite{simenel12,guo12},
we define fusion as a reaction where a colliding system keeps
a compact shape for a sufficiently long time ($\sim 2000$\,fm/$c$)
after contact.
From a sufficiently low energy, we perform TDHF calculations systematically by
increasing progressively the energy with a step of 0.5\,MeV until fusion is reached.
We have found that there is an ambiguity to define a compact shape
leading to the formation of the compound nucleus in some systems.
We will come to this point in Appendix~\ref{sec:appendix:threshold},
and will show that this ambiguity does not affect the conclusion of this article.
\subsection{Numerical setups}\label{sec:method:setup}
In this work, we perform only central collisions.
In our practical computations of TDHF evolutions,
we use the TDHF3D code \cite{kim97} developed by P.~Bonche and coworkers
with the SLy4d parametrization \cite{kim97} of the Skyrme energy density functional,
which is slightly modified from the SLy4 parametrization \cite{chabanat97,chabanat98} by neglecting the center-of-mass corrections in the fitting protocol of determining the parameter set.
As the initial conditions, we solve static Hartree-Fock equations by using the computer code ev8 \cite{ev8} with the imaginary time method.
Then, we place the centers of the projectile and target densities on the $x$-axis.
The initial distance is set to be $R_0=28.8$\,fm for all the systems
considered in this article except for the $^{40}$Ca\,+\,$^{40}$Ca system ($R_0=22.4$\,fm).
We assume that the colliding nuclei follow the Rutherford trajectory
before the initial distance.
The step sizes in the coordinate space and time are 0.8\,fm and 0.45\,fm/$c$, respectively, for both static and dynamical computations.
The numbers of the mesh points in our three dimensional numerical box is $64\times28\times14$ along the $x$, $y$, and $z$-axis, respectively, with a symmetry plane at $z=0$.
The pairing correlation is neglected in both static and dynamical calculations, though this has been included in
some of recent applications to nuclear responses and reactions in
either time-dependent Hartree-Fock-Bogoliubov \cite{avez08,stetcu11,hashimoto12}
or an approximated way \cite{ebata10,scamps13}.
\section{Results}\label{sec:result}
To analyze the fusion hindrance, we perform TDHF simulations for central collisions in heavy systems
$^{90}$Zr\,+\,$^{90,92,94,96}$Zr, $^{100}$Mo\,+\,$^{92,100}$Mo,$^{104}$Ru,$^{110}$Pd,
and $^{124}$Sn\,+\,$^{90,92,94,96}$Zr, where the extra-push energies were deduced
from experimental evaporation-residue cross sections \cite{sahm85, schmidt91},
and $^{96}$Zr\,+\,$^{132}$Sn,$^{136}$Xe, $^{70}$Zn\,+\,$^{208}$Pb.
We obtain the ground-state shapes of $^{92}$Zr, $^{92}$Mo, $^{100}$Mo, and $^{124}$Sn to be oblate with the deformation parameters $\beta=0.03$, $0.05$, $0.23$, and $0.11$, respectively.
The ground-state shapes of $^{94}$Zr and $^{136}$Xe are slightly prolate deformation with $\beta=0.04$ and $0.05$, respectively.
The ground-states of $^{104}$Ru and $^{110}$Pd are triaxial with $\beta=0.31$, $\gamma=21.0^\circ$ and $\beta=0.31$, $\gamma=21.6^\circ$, respectively.
The other nuclei are found to be spherical.
Note that the ground-states of $^{92,94}$Zr, $^{92}$Mo, $^{124}$Sn, and $^{136}$Xe
would be spherical if the pairing correlation were included in the calculations.
For collisions with deformed nuclei, we place the nucleus in the box by setting the symmetry axis of the axially deformed nucleus or the longest axis of the triaxial nucleus parallel to the $z$-axis.
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_Q20newZr90Sn124.eps}
\caption{\label{fig:q20Zr90Sn124} (Color online)
Time evolution of mass quadrupole moment $Q_{20}$ of the total system
at different center-of-mass energies for the $^{90}$Zr\,+\,$^{124}$Sn system.
}
\end{figure}
\begin{figure}[tb]
\includegraphics[width=\linewidth,clip]{2DdensitynewZr90Sn124.E2335.eps}
\caption{\label{fig:2DdensityZr90Sn124} (Color online)
Time evolution of density profiles $\rho(x,y,z=0,t)$ at $E_{\rm c.m.}=233.5$\,MeV
for the $^{90}$Zr\,+\,$^{124}$Sn system.
}
\end{figure}
\begin{figure*}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_extrapush.eps}
\caption{\label{fig:extrapush} (Color online)
Extra-push energies extracted from TDHF simulations (red filled circles) and from
experimental data (blue filled triangles) \cite{schmidt91}.
In the $x$-axis, the nuclide of the target is shown, while the nuclide of the projectile is shown at the top.
}
\end{figure*}
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_extrapush.fissility.eps}
\caption{\label{fig:extrapush.fissility} (Color online)
Same as in Fig.~\ref{fig:extrapush}, but as a function of parameter $x_{\rm mean}$ defined by Eq.~(\ref{eq:mean_fissility}).}
\end{figure}
\subsection{Extra-push energy}\label{sec:result:extrapush}
In our TDHF simulations in heavy systems,
we indeed observe that, after contact of colliding nuclei,
a colliding system keeps a configuration of substantial overlap of projectile and target densities for a long time, and then reseparates to two fragments.
We regard this process as quasifission.
This has been observed in previous works \cite{simenel12,kedziora10,guo12}
and we are not aware of such a process in lighter systems.
As an example of quasifission realized in our TDHF simulations,
we show in Fig.~\ref{fig:q20Zr90Sn124} the time evolution of mass quadrupole moment of the $^{90}$Zr\,+\,$^{124}$Sn system defined as
\begin{equation}\label{eq:q20}
Q_{20}(t) = \int d^3r \, (2x^2-y^2-z^2)\, \rho(x,y,z,t),
\end{equation}
where $\rho(x,y,z,t)$ denotes the density of the total system obtained from TDHF calculations at different center-of-mass energies $E_{\rm c.m.}$
In this system, we can identify that the case at $E_{\rm c.m.}=233.5$\,MeV leads to quasifission after a long contact time $\sim 1800$\,fm/$c$.
This energy is well above the barrier height obtained from
the frozen density potential, 217.1\,MeV.
In Fig.~\ref{fig:2DdensityZr90Sn124}, we show the time evolution of density profile for the quasifission process at $E_{\rm c.m.}=233.5$\,MeV.
Large overlap of the densities from $t=450$\,fm/$c$ to $t=2250$\,fm/$c$ and reseparation at $t=3015$\,fm/$c$ can be seen.
We define the extra-push energy $E_{\rm extra}$ estimated from TDHF simulations as
\begin{equation}
E_{\rm extra}= E_{\rm thres} -V_{\rm FD},
\end{equation}
where $E_{\rm thres}$ and $V_{\rm FD}$ denote the fusion threshold energy and
the potential barrier height obtained from the frozen density approximation, respectively.
Figure~\ref{fig:extrapush} shows the extra-push energies obtained from TDHF simulations denoted by the red filled circles for various heavy systems.
We also show by the blue filled triangles experimental extra-push energies taken from Ref. \cite{schmidt91},
which are defined as the difference between the apparent barrier deduced from experimental data of evaporation-residue cross sections and the barrier height obtained from the Bass model \cite{bass74}.
Our TDHF simulations reasonably reproduce the experimental extra-push energies,
except for heavier systems.
We also reproduce the trend that the extra-push energy becomes large as the charge product of the system becomes large.
Figure~\ref{fig:extrapush.fissility} compares our result with the extra-push model \cite{swiatecki81,swiatecki82,bjornholm82,blocki86}, plotted by the dot-dashed line.
In this figure, we plot extra-push energy as a function of parameter $x_{\rm mean}$ defined as
\begin{equation}\label{eq:mean_fissility}
x_{\rm mean}= \frac{1}{3}x_{\rm eff}+\frac{2}{3}x_{\rm fis}
\end{equation}
with
\begin{align}
x_{\rm eff}&=\frac{4Z_PZ_T}{A_P^{1/3}A_T^{1/3}\left(A_P^{1/3}+A_T^{1/3}\right)}
\biggr/\left(\frac{Z^2}{A}\right)_{\rm crit}, \label{eq:fissility} \\
x_{\rm fis}&=\left(Z^2/A\right)\Bigr/\left(Z^2/A\right)_{\rm crit}.
\label{eq:fissility1}
\end{align}
Here, $(Z^2/A)_{\rm crit}= 50.883(1 - 1.7826I^2)$ is used, where $I=(A-2Z)/A$ is the neutron excess of the compound nucleus with $A=A_P+A_T$, $Z=Z_P+Z_T$ being the mass and charge numbers of the compound nucleus, respectively.
The parameter $x_{\rm mean}$ was originally proposed in Ref. \cite{blocki86},
with which the extra-push energy better scales.
Around $x_{\rm mean}=0.74$, our result is better than the extra-push model,
while the extra-push model is better at $x_{\rm mean}\sim 0.77$.
In some systems, we find that our TDHF simulations underestimate the values of the experimental extra-push energy.
We underestimate the extra-push energy by about 12\,MeV in the $^{96}$Zr\,+\,$^{124}$Sn system.
A possible interpretation of this underestimation is as follows.
As we have explained in Sec.~\ref{sec:method:definition},
we define fusion by TDHF as a reaction where a colliding system keeps
a compact shape for a sufficiently long time.
As one can see in Fig.~\ref{fig:q20Zr90Sn124},
such a reaction never reaches to the shape of resulting compound nucleus in TDHF simulations
because of lack of dissipation beyond the mean-field level,
e.g., two-body dissipation arising from direct nucleon--nucleon collisions \cite{wen13,jiang13,wen14}.
Therefore, the cases defined as fusion from TDHF simulations might lead to
reseparation (quasifission) if dissipation mechanisms beyond the mean-field level were included in the calculations.
That is, fusion threshold energies in some systems might be underestimated within the mean-field level.
Other possible interpretation of this underestimation is the neglect of pairing correlations in both static and dynamical calculations. This might induce the discrepancy between TDHF and experimental fusion threshold energies.
Although we have such a limitation,
our TDHF simulations qualitatively reproduce the overall property of
the dependence of observed extra-push energies on system size.
This fact validates our following analysis based on TDHF simulations.
In the following, we will give detailed analyses of nucleus--nucleus
potential and energy dissipation obtained from the DD-TDHF method
and of the origin of the fusion hindrance in heavy systems.
\subsection{Nucleus--nucleus potential in heavy systems}
\label{sec:result:potential}
\subsubsection{Property of extracted potential}
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_potnewZr96Sn124.big.final.eps}
\caption{\label{fig:potZr96Sn124} (Color online)
Nucleus--nucleus potential as a function of relative distance $R$ for the $^{96}$Zr\,+\,$^{124}$Sn system.
The red solid, green dot-dashed, blue dashed, and brown dotted lines denote the potentials extracted from the TDHF trajectories at $E_{\rm c.m.}=300$, 250, 228, and 220\,MeV, respectively.
The filled circles are the potential obtained from the frozen density approximation.}
\end{figure}
\begin{figure}[tb]
\includegraphics[width=0.8\linewidth,clip]{plot_potCa40Ca40.eps}
\caption{\label{fig:potCa40Ca40} (Color online)
Same as in Fig.~\ref{fig:potZr96Sn124}, but for the $^{40}$Ca\,+\,$^{40}$Ca system.
}
\end{figure}
Figure~\ref{fig:potZr96Sn124} shows the nucleus--nucleus potentials
as a function of relative distance $R$
obtained from the DD-TDHF method for the $^{96}$Zr\,+\,$^{124}$Sn system at $E_{\rm c.m.}=300$, 250, 228, and 220\,MeV used in TDHF. The filled circles are the potential obtained by the frozen density approximation.
Note that, when $R \le 11.2$\,fm, the overlap density of the two nuclei at the neck in the frozen density approximation becomes close to the normal density 0.16\,fm$^{-3}$, indicating the violation of the frozen density approximation at $R \le 11.2$\,fm.
The charge product of this system is 2000, and we observe the fusion hindrance in this system as we have shown in Fig.~\ref{fig:extrapush}.
The fusion threshold energy for this system is $E_{\rm c.m.}=228$\,MeV.
At $E_{\rm c.m.}=220$\,MeV, which is higher than the barrier height of the frozen density potential, the system reseparates after slight contact.
Note that, for the cases at $E_{\rm c.m.}=250$, 228, and 220\,MeV, we stop the extraction of the potential at $R_{\rm min}=10.1$, 11.4, and 12.7\,fm, respectively.
The extracted potentials agree with the frozen density potential until $R\sim 13.5$\,fm and start to deviate from the frozen density one as $R$ decreases.
At $R< 13.5$\,fm, the extracted potentials do not have an
ordinary barrier structure, though the frozen density potential has an ordinary barrier at $R\sim 12.8$\,fm.
The extracted potentials monotonically increase after deviating from
the frozen density potential.
At $R > 11.4$\,fm, the extracted potentials agree with each other, while a deviation is seen between the potentials extracted at $E_{\rm c.m.}=300$ and 250\,MeV
at $R < 11.4$\,fm.
To investigate these findings in detail,
we compare them with the potentials extracted in light systems.
Figure~\ref{fig:potCa40Ca40} show the potentials extracted in the $^{40}$Ca\,+\,$^{40}$Ca system.
From the comparison, we find two main differences between the potentials extracted in light and heavy systems as follows:
\begin{itemize}
\item Potentials for the $^{40}$Ca\,+\,$^{40}$Ca system and the frozen
density potential for both systems have an ordinary barrier structure,
while no barrier appears in the potentials extracted in the $^{96}$Zr\,+\,$^{124}$Sn system.
Monotonic increase of the extracted potentials is only observed in the $^{96}$Zr\,+\,$^{124}$Sn system.
\item In the $^{40}$Ca\,+\,$^{40}$Ca system, energy dependence of potential around the Coulomb barrier is significant,
while for the $^{96}$Zr\,+\,$^{124}$Sn system, energy dependence of potential is less pronounced than in the $^{40}$Ca\,+\,$^{40}$Ca system.
\end{itemize}
\begin{figure}[tb]
\includegraphics[width=\linewidth,clip]{plot_potall0.eps}
\caption{\label{fig:potall} (Color online)
Nucleus--nucleus potentials for various heavy systems are displayed
at the region around the barrier of the frozen density potential.
}
\end{figure}
The above two properties are indeed general in heavy systems.
To see this, we show in Fig.~\ref{fig:potall} extracted nucleus--nucleus potentials for the $^{100}$Mo\,+\,$^{104}$Ru ($Z_PZ_T=1848$), $^{100}$Mo\,+\,$^{110}$Pd (1932), $^{90}$Zr\,+\,$^{124}$Sn, $^{96}$Zr\,+\,$^{132}$Sn (2000), $^{96}$Zr\,+\,$^{136}$Xe (2160), $^{70}$Zn\,+\,$^{208}$Pb (2460) systems.
In each system, except for the $^{70}$Zn\,+\,$^{208}$Pb system, we show the potentials extracted at the fusion threshold energy
by the blue dashed line and at higher energy by the red solid line.
In the $^{70}$Zn\,+\,$^{208}$Pb system, we do not observe fusion at energies
between 265 and 400\,MeV.
In any systems, the potentials show monotonic increase as $R$ decreases
and energy dependence of potential is less pronounced.
(At most, the difference is 0.6\% in the $^{90}$Zr\,+\,$^{124}$Sn system, while the difference is 3\% in the $^{40}$Ca\,+\,$^{40}$Ca system.)
By this systematic study of nucleus--nucleus potential,
we conclude that these properties of the extracted potentials are general in heavy systems.
\subsubsection{Discussion}
\begin{figure}[tb]
\includegraphics[width=0.8\linewidth,clip]{plot_momentum_abs.eps}
\caption{\label{fig:momentum} (Color online)
Absolute value of the relative momentum scaled by the initial reduced mass number $A_0$ as a function of relative distance for the $^{96}$Zr\,+\,$^{124}$Sn at $E_{\rm c.m.}=228$\,MeV (red solid line) and the $^{40}$Ca\,+\,$^{40}$Ca at $E_{\rm c.m.}=60$\,MeV (blue dashed line) systems.
The barrier radii of the frozen density potentials for these systems are indicated by arrow.
}
\end{figure}
We have shown that the nucleus--nucleus potentials extracted in heavy systems exhibit monotonic increase as the relative distance decreases and do not have an ordinary barrier structure and this property is seen only in heavy systems.
In the following, we will consider the reason why this property appears only in this case.
Figure~\ref{fig:momentum} shows the relative momentum scaled by the initial reduced mass number $A_0=A_PA_T/(A_P+A_T)$ as a function of relative distance
for the $^{96}$Zr\,+\,$^{124}$Sn (red solid line) and the $^{40}$Ca\,+\,$^{40}$Ca (blue dashed line) systems.
The Coulomb barrier radius of the frozen density potential is indicated by arrow for these systems.
Note that the sign of the relative momentum is negative in the entrance channel of central collisions.
In the $^{40}$Ca\,+\,$^{40}$Ca system, the momentum decreases as $R$ decreases until the Coulomb barrier, and then the momentum starts to increase after the system overcomes the Coulomb barrier.
In the $^{96}$Zr\,+\,$^{124}$Sn system, however, the momentum monotonically decreases as $R$ decreases.
If the potential of this system were the frozen density one, the momentum would increase after passing the barrier. Therefore, the monotonic decrease of the momentum with $R$ decreasing induces the increase in extracted potentials at the inside of the frozen density barrier.
Dissipation also can decrease the momentum.
However, our analysis shows that the extracted dissipation is not enough to explain the property of the time evolution of the momentum in heavy systems and the increase in potential is necessary.
\begin{figure}[tb]
\includegraphics[width=0.8\linewidth,clip]{plot_conditional_saddle.eps}
\caption{\label{fig:conditional_saddle}
Schematic illustration of the appearance of a conditional saddle in heavy systems. See text for details.
}
\end{figure}
We further consider
the meaning of
the increase in potential at small relative distances. To do so, we introduce with Fig.~\ref{fig:conditional_saddle}
an explanation of the appearance of the fusion hindrance in heavy systems
from a simple geometrical feature of potential used in the extra-push model \cite{swiatecki81,swiatecki82,bjornholm82}.
Figure~\ref{fig:conditional_saddle} shows the schematic illustration of an adiabatic potential energy landscape as a function of deformation of a compound nucleus in light and heavy systems.
In both systems, the left region corresponds to the configuration of a compound nucleus formation, while the right corresponds to the quasifission.
The potential barrier, denoted by ``saddle'' in the figure, is usually called conditional saddle and corresponds to the fission barrier under the condition that the mass partition of the system into two fragments is fixed to be the one at the contact configuration in the entrance channel.
If the equilibration at the contact configuration is fast enough, in other words, the transition from two-body regime to one-body regime in the collision dynamics is fast enough, the system will feel this adiabatic potential once two nuclei touch.
In light systems, the contact configuration is located inside the saddle on the potential landscape. Consequently, fusion automatically takes place once the two nuclei touch.
The situation is changed for heavy systems because of large Coulomb energy.
That is, the saddle is shifted to the inside of the contact configuration,
and the system needs to overcome the saddle for fusion, which gives rise to the fusion hindrance in the extra-push model.
In the DD-TDHF method, dynamical effects are automatically included in the extracted potential through TDHF dynamics.
Since the TDHF theory treats the reorganization of single-particle wave functions during collision in a selfconsistent way,
the extracted potential includes effects of smooth transition from two-body regime of collision dynamics to one-body regime of shape evolution at small $R$.
Therefore, we consider that the increase in potential at small $R$ observed in heavy systems can be regarded as a reflection of the appearance of the conditional saddle inside the contact configuration in the schematic picture explained above.
\subsubsection{Isotope dependence of nucleus--nucleus potential}
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_Zr96Sn.eps}
\caption{(Color online)
\label{fig:potZr96Sn}(a)
Comparison between the nucleus--nucleus potentials for the $^{96}$Zr\,+\,$^{124}$Sn (red solid line) and $^{96}$Zr\,+\,$^{132}$Sn (blue dashed line) systems. Both potentials are extracted at $E_{\rm c.m.}=230$\,MeV.
The red filled circles and blue filled triangles are the frozen density potentials for the $^{96}$Zr\,+\,$^{124}$Sn and $^{96}$Zr\,+\,$^{132}$Sn systems, respectively.
\label{fig:transZr96Sn}(b)
Average number $N_{\rm trans}$ of transferred neutrons (red) and protons (blue) from $^{124}\mathrm{Sn}$ (solid line) and $^{132}\textrm{Sn}$ (dashed line) to $^{96}$Zr in the $^{96}$Zr\,+\,$^{124,132}$Sn systems at $E_{\rm c.m.}=230$\,MeV.
(c)\label{fig:NZratioZr96Sn}
Neutron to proton ratio ($N/Z$) of the two subsystems in the $^{96}$Zr\,+\,$^{124}$Sn (solid line) systems and $^{96}$Zr\,+\,$^{132}$Sn (dashed line) systems.
The gray solid and dashed lines indicate the values of the equilibrated $N/Z$ for the two systems.
}
\end{figure}
Next we will analyze the isotope dependence
of nucleus--nucleus potentials in heavy systems,
For this purpose, we employ the $^{96}$Zr\,+\,$^{124}$Sn and $^{96}$Zr\,+\,$^{132}$Sn systems by changing the neutron number of Sn.
In Fig.~\ref{fig:potZr96Sn}(a), we show the nucleus--nucleus potentials
for the $^{96}$Zr\,+\,$^{124}$Sn (red solid line) and $^{96}$Zr\,+\,$^{132}$Sn (blue dashed line) systems.
The potentials in both systems are extracted at $E_{\rm c.m.}=230$\,MeV.
As a reference, the frozen density potentials are plotted by the red filled circles and blue filled triangles for the $^{96}$Zr\,+\,$^{124}$Sn and $^{96}$Zr\,+\,$^{132}$Sn systems, respectively.
In the DD-TDHF method and frozen density approximation, the potential energy for the $^{96}$Zr\,+\,$^{124}$Sn system is slightly larger than that for the $^{96}$Zr\,+\,$^{132}$Sn system.
In the DD-TDHF case, the difference between the potentials is 0.4\,MeV at the barrier radius of the frozen density potential, $R=12.8$\,fm, which is similar to the difference between the frozen density potential barriers of the two systems, 0.5\,MeV.
Then, the difference between the two extracted potentials evolves as the relative distance becomes small, and becomes 1.8\,MeV at $R=11.1$\,fm.
In order to further understand the role of the neutron excess in $^{132}$Sn on the potential, we show in Fig.~\ref{fig:transZr96Sn}(b) the average numbers of transferred nucleons, $N_{\rm trans}$, through the neck from $^{124}$Sn and $^{132}$Sn to $^{96}$Zr in the $^{96}$Zr\,+\,$^{124,132}$Sn systems at $E_{\rm c.m.}=230$\,MeV.
The positive and negative values correspond to transfer from $^{124,132}$Sn to $^{96}$Zr and transfer from $^{96}$Zr to $^{124,132}$Sn, respectively.
We find that the number of transferred neutrons in the $^{96}$Zr\,+\,$^{132}$Sn system is larger than that in the $^{96}$Zr\,+\,$^{124}$Sn system as $R$ decreases.
We also find the larger number of transferred protons from $^{96}$Zr to $^{132}$Sn
than that from $^{96}$Zr to $^{124}$Sn.
Because of larger neutron to proton ($N/Z$) ratio of $^{132}$Sn ($N/Z=1.64$) than that of $^{124}$Sn ($N/Z=1.48$),
larger nucleon transfer towards charge equilibration occurs during the collision in the $^{96}$Zr\,+\,$^{132}$Sn system.
This is illustrated in Fig.~\ref{fig:NZratioZr96Sn}(c), where $N/Z$ of the two subsystems in the $^{96}$Zr\,+\,$^{124}$Sn (solid line) and $^{96}$Zr\,+\,$^{132}$Sn (dashed line) systems are plotted.
As $R$ decreases, larger change in $N/Z$ towards the change equilibration
in the $^{96}$Zr\,+\,$^{132}$Sn system can be seen.
We consider that the net effects make the potential for the $^{96}$Zr\,+\,$^{132}$Sn system lower.
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_ZrSn124.eps}
\caption{\label{fig:ZrSn124}(Color online)
Same as Fig.~\ref{fig:potZr96Sn} but for the $^{90,96}$Zr\,+\,$^{124}$Sn systems
obtained at the fusion threshold energy $E_{\rm thres}=234.5$\,MeV and $E_{\rm thres}=228$\,MeV, respectively.
}
\end{figure}
Similar property is found in the extracted potentials in $^{90,96}$Zr\,+\,$^{124}$Sn systems.
Figure~\ref{fig:ZrSn124}(a) shows the extracted potentials at the fusion threshold energy and the frozen density potentials for these systems. We find that the increase of the extracted potential from the frozen density potential is larger in $^{96}$Zr\,+\,$^{124}$Sn than in $^{90}$Zr\,+\,$^{124}$Sn. The difference is about 4\,MeV in average.
From Fig.~\ref{fig:ZrSn124}(b) showing the number of transferred nucleons
and Fig.~\ref{fig:ZrSn124}(c) showing $N/Z$ ratio,
larger nucleon transfer towards charge equilibration occurs in the $^{90}$Zr\,+\,$^{124}$Sn system due to larger difference between the initial $N/Z$ ratios of the projectile and target.
From these observations,
we conclude that, as nucleon transfer towards charge equilibration during the collision becomes smaller, increase in the extracted potential from the frozen density potential becomes larger.
\subsubsection{Comparison with the DC-TDHF method}\label{sec:result:DCTDHF}
Figure~\ref{fig:potZr96Sn124compare} compares nucleus--nucleus potentials obtained from our method (red solid line) and from the DC-TDHF method (blue dashed line) in Ref. \cite{oberacker10} for the $^{96}$Zr\,+\,$^{124}$Sn system at $E_{\rm c.m.}=230$\,MeV,
which is 2-MeV above the fusion threshold energy.
As a reference, the frozen density potential is plotted by the filled circles.
The DC-TDHF potential has an ordinary barrier at $R\sim 13.1$\,fm,
which is slightly outside the barrier radius of the frozen density potential, $R\sim 12.8$\,fm.
The barrier height of the DC-TDHF potential is about 3.5-MeV lower than that of the frozen density potential.
The difference between DC-TDHF potential and our potential becomes
large as the relative distance becomes small.
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_potZr96Sn124compare.eps}
\caption{\label{fig:potZr96Sn124compare} (Color online)
Comparison between the nucleus--nucleus potential obtained from our method (red solid line) and that from the DC-TDHF method (blue dashed line) taken from Ref. \cite{oberacker10} for the $^{96}$Zr\,+\,$^{124}$Sn system at $E_{\rm c.m.}=230$\,MeV.
As a reference, the frozen density potential of this system is plotted by the filled circles.
}
\end{figure}
\begin{figure}[tb]
\includegraphics[width=0.8\linewidth,clip]{plot_potCa40Ca40compare.eps}
\caption{\label{fig:potCa40Ca40compare} (Color online)
Same as in Fig.~\ref{fig:potZr96Sn124compare} but for the $^{40}$Ca\,+\,$^{40}$Ca system at $E_{\rm c.m.}=55$\,MeV.
DC-TDHF potential is taken from Ref. \cite{umar14}.
}
\end{figure}
Since both methods solve the same TDHF equations,
the time evolutions of the single-particle wave functions in the two methods should be same, except for numerical errors and slightly different Skyrme's parametrizations (SLy4d for DD-TDHF and SLy4 for DC-TDHF).
In the DC-TDHF method, the potential energy is given as the energy density functional of the total system with the minimization under the constraint that the total density matches the TDHF density.
Therefore, the obtained potential can be regarded as the lowest energy along the dynamical path that includes all dynamical effects through the time evolution of the realistic TDHF density distributions.
In the DD-TDHF method,
the potential energy is obtained under the assumption that
complex TDHF dynamics for central collisions is reduced to one-dimensional macroscopic equation of motion.
Effects coming from other collective degrees of freedom than the degree of freedom of the relative motion
are reduced to the transport coefficients of the reduced mass, nucleus--nucleus potential, and dissipation term in the one-dimensional Newton equation.
Although both methods are based on the same TDHF dynamics,
the way of interpretation of nucleus--nucleus potential is totally different from each other, giving different properties of the potential in heavy systems.
Note that the reduced mass extracted from the DD-TDHF method includes dynamical effects, leading to the $R$ dependence of mass.
However, in light systems, we do not see a significant difference between DC-TDHF and DD-TDHF potentials. As an example,
we show in Fig.~\ref{fig:potCa40Ca40compare} the comparison between DC-TDHF (blue dashed line) and our potential (red solid line) for the $^{40}$Ca\,+\,$^{40}$Ca system at $E_{\rm c.m.}=55$\,MeV,
and find no significant difference between the two potentials
(only 0.3\,MeV difference at most at the barrier).
Therefore, the effects mentioned above are significant only in heavy systems.
\subsection{Dissipation}\label{sec:result:dissipation}
The TDHF theory includes one-body dissipation through the selfconsistent mean field.
As we explained in Sec.~\ref{sec:method:DD-TDHF},
we can simultaneously extract the nucleus--nucleus potential and the friction coefficient from the DD-TDHF method.
In this subsection, we discuss the property of extracted friction
coefficients in heavy systems.
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_gammanewZr96Sn124.eps}
\caption{\label{fig:frictionZr96Sn124} (Color online)
Reduced friction coefficient defined as $\gamma(R)/\mu(R)$ as a function of $R$
extracted from TDHF trajectories at $E_{\rm c.m.}=300$\,MeV (red solid line), $E_{\rm c.m.}=250$\,MeV (green dot-dashed line), $E_{\rm c.m.}=228$\,MeV (blue dashed line), and $E_{\rm c.m.}=220$\,MeV (brown dotted line) for the $^{96}$Zr\,+\,$^{124}$Sn system.
}
\end{figure}
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_gammaall.eps}
\caption{\label{fig:friction} (Color online)
Reduced friction coefficient $\gamma(R)/\mu(R)$ as a function of $R$
for selected systems.
}
\end{figure}
In Fig.~\ref{fig:frictionZr96Sn124}, we show the reduced friction coefficients defined as the friction coefficient divided by the reduced mass $\gamma(R)/\mu(R)$ as a function of relative distance in the $^{96}$Zr\,+\,$^{124}$Sn system extracted at the same energies as used in extracting potential in Fig.~\ref{fig:potZr96Sn124}.
The extracted friction coefficient increases as the colliding nuclei approach at $R>13$\,fm.
And then, unlike the extracted potential,
we observe a clear energy dependence of the friction coefficient.
As $E_{\rm c.m.}$ increases, the peak position of the extracted friction coefficient is shifted towards small $R$, which is close to $R_{\rm min}$, the relative distance at which we stop the extraction, at each energy.
(Note again that, for the cases at $E_{\rm c.m.}=250$, 228, and 220\,MeV, $R_{\rm min}=10.1$, 11.4, and 12.7\,fm, respectively.)
The reason of the appearance of the peak near $R_{\rm min}$ is that,
as $R$ becomes small, the relative velocity becomes small and
the internal structure of the colliding nuclei has much time to reorganize,
giving a large value of the friction coefficient.
Figure~\ref{fig:friction} shows the reduced friction coefficients for selected systems.
For these systems, each friction coefficient is extracted at each fusion threshold energy.
The radial dependence of the friction coefficients is similar in these systems:
As $R$ decreases, the friction coefficient increases,
makes two humps at $R\sim12.6$ and 12\,fm, and then has a sharp peak.
Also, the values of their strengths at the humps are similar to each other, $\gamma/\mu \sim 1\times 10^{21}$\,s$^{-1}$,
and their strengths at the sharp peak are of the same order of magnitude, $\gamma/\mu \sim 2 - 6\times 10^{21}$\,s$^{-1}$.
For comparison, we also show the reduced friction coefficient for the
$^{40}$Ca\,+\,$^{40}$Ca system at $E_{\rm c.m.}=55$\,MeV by the open squares.
The radial dependence of the friction coefficient is slightly different
from those in heavy systems: It has only a single peak.
Its strength is however of the same order of magnitude as that at the peak in heavy systems.
\begin{figure}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_EdissnewZr96Sn124.eps}
\caption{\label{fig:Ediss} (Color online)
Accumulated dissipation energy calculated by Eq.~(\ref{eq:Ediss}) as a function of $R$ obtained from TDHF trajectories at $E_{\rm c.m.}=300$, 250, 228, and 220\,MeV for the $^{96}$Zr\,+\,$^{124}$Sn system. A zoom around $R=13$\,fm is inserted.
}
\end{figure}
With the extracted friction coefficient, we can evaluate accumulated dissipation energy from the relative motion to internal excitations by the formula:
\begin{equation}\label{eq:Ediss}
E_{\rm diss}[R(t)] = \int_0^t dt^\prime \gamma[R(t^\prime)] {\dot R}(t^\prime)^2,
\end{equation}
until time $t$ corresponding to the relative distance $R(t)$ in the entrance channel.
Figure~\ref{fig:Ediss} shows the accumulated dissipation energy as a function of $R$ for
the $^{96}$Zr\,+\,$^{124}$Sn system at the same $E_{\rm c.m.}$ as in Fig.~\ref{fig:frictionZr96Sn124}.
A zoom around $R=13$\,fm is inserted in the figure to focus on a low $E_{\rm diss}$ region.
As expected from the property of the extracted friction coefficient, the dissipation energies are zero at large $R$ and then monotonically increase as the colliding nuclei approach and overlap each other.
Irrespective of humps and peaks of the extracted friction coefficients,
the dissipated energies show a smoothed behavior.
Concerning the dependence of dissipation energy on $E_{\rm c.m.}$ of TDHF trajectory,
the dissipation energy becomes larger at larger $E_{\rm c.m.}$
because of larger relative velocity entering in the integral in Eq.~(\ref{eq:Ediss}).
In the following, we will use the dissipation energy $E_{\rm diss}$ accumulated until $R_{\rm min}$ as the physical quantity to analyze the fusion hindrance in the entrance channel of fusion reactions.
\subsection{Origin of extra-push energy}
\begin{figure}[tb]
\includegraphics[width=0.8\linewidth,clip]{plot_pot_schematic.eps}
\caption{\label{fig:pot.schematic} (Color online)
Schematic picture to define the quantities used in the analysis: Frozen density potential barrier $V_{\rm FD}$, fusion threshold energy $E_{\rm thres}$, extra-push energy $E_{\rm extra}=E_{\rm thres}-V_{\rm FD}$, and increase in potential $\Delta V$ for the $^{96}$Zr\,+\,$^{124}$Sn system.
}
\end{figure}
\begin{figure*}[tb]
\includegraphics[width=0.9\linewidth,clip]{plot_dv.ediss.eps}
\caption{\label{fig:dv.ediss} (Color online)
Contributions of the increase in potential $\Delta V$ and of the accumulated dissipation energy $E_{\rm diss}$ to the extra-push energy for the same systems as those in Fig.~\ref{fig:extrapush}.
}
\end{figure*}
In Sec.~\ref{sec:result:extrapush}, we have observed the fusion hindrance phenomenon by TDHF simulations for the entrance channel of central collisions in heavy systems.
We have shown in Fig.~\ref{fig:extrapush} the extra-push energies estimated from TDHF simulations for several heavy systems
and reasonably reproduced the extra-push energies deduced from experimental observations of evaporation-residue cross sections.
In this subsection, using the properties of the extracted potentials and friction coefficients in heavy systems presented in the previous subsections,
we will analyze the origin of the fusion hindrance and appearance of the extra-push energy.
Before showing the result of the analysis,
we first summarize the quantities in a schematic picture in Fig.~\ref{fig:pot.schematic} that will be used in the analysis.
They are the fusion threshold energy $E_{\rm thres}$ and the frozen density potential barrier $V_{\rm FD}$, indicated by the arrows, for the $^{96}$Zr\,+\,$^{124}$Sn system. We have defined the extra-push energy by TDHF as the difference between the fusion threshold energy and the frozen density potential barrier, $E_{\rm extra}=E_{\rm thres}-V_{\rm FD}$.
Here, we introduce a new quantity, increase in potential energy $\Delta V$, that is the difference between the extracted potential at $E_{\rm thres}$ at $R_{\rm min}$ and the frozen density potential barrier.
Since the extracted potential increases monotonically as $R$ decreases in the heavy systems, $\Delta V$ becomes maximum at $R_{\rm min}$.
We will also use in the analysis the accumulated dissipation energy $E_{\rm diss}$ defined in Eq.~(\ref{eq:Ediss}).
This accumulated dissipation energy is computed from the friction coefficient
extracted at $E_{\rm thres}$. As shown in Fig.~\ref{fig:Ediss}, $E_{\rm diss}$ also monotonically increases as $R$ decreases.
From the energy conservation, we have at $R\geq R_{\rm min}$
\begin{equation}\label{eq:energy_conservation}
E_{\rm c.m.} = E_{\rm kin}(R) + V(R) + E_{\rm diss}(R),
\end{equation}
where $E_{\rm kin}=P^2/2\mu(R)$ is the collective kinetic energy
and all the energies in the right-hand side can be computed by TDHF and by the DD-TDHF method and we confirm that the energy conservation holds.
Subtracting $V_{\rm FD}$ from the both sides of Eq.~(\ref{eq:energy_conservation}) and using the quantities $E_{\rm extra}= E_{\rm thres}-V_{\rm FD}$ and $\Delta V=V(R_{\rm min})-V_{\rm FD}$,
we derive the following relation at $E_{\rm c.m.}=E_{\rm thres}$ at $R=R_{\rm min}$:
\begin{equation}
E_{\rm extra}=E_{\rm kin}(R_{\rm min})+ \Delta V(R_{\rm min})+ E_{\rm diss}(R_{\rm min})
\end{equation}
As shown in Fig.~\ref{fig:momentum}, the absolute value of the momentum $|P|$ is expected to monotonically decrease and to be smallest at $R=R_{\rm min}$ in heavy systems,
so is $E_{\rm kin}$.
We have checked that $E_{\rm kin}$ at $R=R_{\rm min}$ in all the systems is less than 0.6\,MeV and small enough, compared with $\Delta V$ and $E_{\rm diss}$.
Therefore, we approximate the extra-push energy as the sum
of $\Delta V$ and $E_{\rm diss}$,
\begin{equation}
E_{\rm extra}\approx\Delta V(R_{\rm min})+ E_{\rm diss}(R_{\rm min}).
\end{equation}
That is, we can identify the origin of the extra-push energy in the entrance channel as $\Delta V$ and/or $E_{\rm diss}$ extracted from the DD-TDHF method.
Figure~\ref{fig:dv.ediss} shows each contribution to the extra-push energy
for the same systems as in Fig.~\ref{fig:extrapush}.
The red filled circles denote the contribution of the increase in potential $\Delta V$ to the extra-push energy, while the blue filled triangles denote the contribution of the accumulated dissipation energy $E_{\rm diss}$.
In the $^{90}$Zr\,+\,$^{90,92,94}$Zr and $^{100}$Mo\,+\,$^{92}$Mo systems, $\Delta V$ is smaller than $E_{\rm diss}$.
In the rest of the systems considered here, $\Delta V$ is greater than $E_{\rm diss}$.
As the charge product of the system increases, the ratio of $\Delta V$ to $E_{\rm diss}$ increases.
From this figure, we can conclude that, as the system becomes heavier,
the contribution of the increase in potential energy
to the extra-push energy in the entrance channel
is more important than that of the accumulated dissipation energy.
Dissipation plays an important role, but is not sufficient to explain the
amount of the extra-push energy in our analysis with the DD-TDHF method.
\section{Conclusions}\label{sec:conclusions}
The synthesis of superheavy elements in laboratories is extremely challenging
because of an extremely low production rate. One of the reason for a low production late
is the presence of the fusion hindrance. This gives strong competition
between compound nucleus formation and quasifission.
In this article we address this point by the selfconsistent microscopic reaction theory, TDHF.
We have shown the applicability of TDHF dynamics for fusion reactions involving heavy nuclei by reproducing the extra-push energies estimated from experimental evaporation-residue cross sections in heavy systems.
Then, we have analyzed the property of the nucleus--nucleus potentials for the entrance channel of fusion reactions
extracted from the DD-TDHF method, which combines TDHF dynamics with a classical equation of motion.
In heavy systems, the extracted nucleus--nucleus potentials show the following properties:
(i) Ordinary Coulomb barrier in the extracted potential disappears and the potential monotonically increases as the relative distance decreases.
(ii) The dependence of potential on the energy used in the TDHF simulations is not pronounced.
These properties are not seen in the potentials in light- and medium-mass systems and also in the DC-TDHF method.
We have shown that the increase in potential is a reflection of the appearance of the conditional saddle inside the contact configuration of colliding nuclei on an adiabatic potential surface.
By analyzing the isotope dependence of the extracted potentials,
we have found that, as nucleon transfers towards charge equilibration during the collision becomes small, increase in the extracted potential from the frozen density potential becomes large.
Extracted friction coefficients however show energy dependence, and have a sharp peak near the relative distance at which the extraction stops. This indicates strong reorganization of the internal structure near that distance.
Using these properties of the extracted nucleus--nucleus potential and energy dissipation, we have analyzed the origin of the extra-push energy in heavy systems.
The analysis has shown that more contribution comes from the increase in potential energy to extra-push energy than that from the accumulated dissipation energy
as the size of the system is heavier.
We have concluded that the increase in potential is more important than dissipation for understanding the origin of the fusion hindrance.
Although the present study is limited to central collisions,
which enable us to perform the DD-TDHF method with one-dimensional equation of motion,
we have shown the feasibility of the present analysis for the fusion hindrance.
Possible extensions to non-central collisions involving finite angular momenta will be interesting to understand the role of angular momentum for the fusion hindrance.
Furthermore, the orientation dependence of the dynamics due to large nuclear deformation will be important for systems with, for example, actinide nuclei, which have been and will be used in the so-called hot fusion reactions in the synthesis of superheavy elements.
The fission process, as an inverted process of fusion,
is one of the future subject to apply the present method to extract
information on energy dissipation.
\begin{acknowledgments}
The author thanks D. Lacroix, T. Nakatsukasa, and K. Sekizawa for providing valuable comments on this article.
The author is supported by the Special Postdoctoral Researcher Program of RIKEN.
\end{acknowledgments}
|
\section{Introduction}\label{s1}
The theory of global attractors for the 2-D Navier--Stokes
system
$$
\begin{cases}
\Dt u+(u,\Nx)u+\Nx p=\nu\Dx u+g,\\
u\big|_{t=0}=u_0,\ \ \divv u=0.
\end{cases}
$$
has been a starting point of the theory of infinite
dimensional dissipative dynamical systems and remains in
the focus of this theory, see~\cite{BV, Ch-V-book,CF88, FMRT,Lad,temam,S-Y}
and the references therein.
In the case of a bounded domain $\Omega$ the corresponding dynamical system
has a global attractor in the appropriate phase space. The attractor
has finite fractal dimension, measured in terms of the dimensionless
number $G$ (the Grashof number), $G=\frac{\|g\||\Omega|}{\nu^2}$.
The best known estimate in the case of the Dirichlet boundary conditions
$u\vert_{\partial\Omega}=0$ is (see~\cite{temam})
\begin{equation}\label{0.est-dir}
\dim_f\mathcal{A}\le c_{\mathrm{D}}G,
\end{equation}
while in the case of a periodic domain $x\in[0,2\pi L]^2$
the estimate can be significantly improved (see \cite{CFT}):
$$
\dim_f\mathcal{A}\le c_{\mathrm{per}}G^{2/3}(\ln(1+G))^{1/3},
$$
and, moreover, this estimate sharp up to a logarithmic correction as
shown in \cite{Liu}, see also \cite{DG} for the alternative proof of the upper bound.
Finding explicit majorants for the constants $c_{\mathrm{D}}$ and
$c_{\mathrm{per}}$ amounts to finding sharp or explicit constants
in certain Sobolev inequalities and spectral Lieb--Thirring
inequalities. For example, $c_{\mathrm{D}}\le (4\pi 3^{1/4})^{-1}$
\cite{I-Stokes}, and the majorant for $c_{\mathrm{per}}$ can
be easily written down using the recent result \cite{BDZ} on the
sharp constant in the logarithmic
Brezis--Gallouet inequality (which is essential for the attractor
dimension estimate in the periodic case).
In unbounded channel-like domains the Navier-Stokes system with Di\-richlet boundary conditions
is still dissipative in view of the Poincar\'e inequality. In particular, if the case of finite
energy solutions is considered, the associated semigroup possesses a compact global attractor
similarly to the case of bounded domains. Up to the moment, there are two alternative ways to
establish this fact. The first one is based on the weighted energy estimates and careful analysis
of the Leray projection in weighted Sobolev spaces, see \cite{Bab,MZ} and references therein
and the second one utilizes the so-called energy method (which will be also used in our paper)
and the energy equality, see \cite{ball,Rosa,temam}.
\par
However, in contrast to the case of bounded domains,
the solution semigroup is no longer compact, but is rather
asymptotically compact and this fact strongly affects the existing upper bounds for
the dimension of the attractor. Indeed, the best known estimate for the case of
channel-like domains
obtained in \cite{Rosa} can be written as follows
$$
\dim_f\mathcal A\le \frac{c_\mathrm{LT}}2\frac{\|g\|^2} {\lambda_1^{2}\nu^{4}}.
$$
Here $\lambda_1>0$ is the bottom of the spectrum of the Stokes operator in the channel,
and ${c_\mathrm{LT}}$ is the universal Lieb--Thirring constant, see \eqref{L-T}.
For example, in a straight channel of width $d$ we have $\lambda_1\ge \pi^2/d^2$,
so that in this case we obtain
$$
\dim_f\mathcal A\le \frac{1}{4\sqrt{3}\pi^4}\frac{d^4\|g\|^2} {\nu^{4}}.
$$
We observe that these estimates are proportional to $\nu^{-4}$ (unlike
\eqref{0.est-dir} which is proportional to $\nu^{-2}$). On the other hand, it is worth
mentioning that, to the best of our knowledge, no growing as $\nu\to0$ lower bounds for the
dimension of the attractor are known for the case of Dirichlet boundary conditions,
regardless whether the underlying domain is bounded or unbounded. So, in contrast to
the case of periodic boundary conditions, the behaviour of the attractor's dimension
as $\nu\to0$ remains unclear for the case of Dirichlet boundary conditions even in
the case of bounded domains.
We mention also that keeping in mind the Poiseulle flows and the mean flux integral,
it seems more natural to consider the {\it infinite} energy solutions for the
Navier-Stokes system in a pipe. In this case, the system remains dissipative
and the existence of the so-called locally compact attractor can be established,
see \cite{AZ, Zel-glas}. The dimension of this attractor may be infinite in
general, but will be finite if the external forces $g(x)$ decay to zero as
$|x|\to\infty$ (no matter how slow this decay is), see \cite{MZ,Zel-glas} and the
references therein.
In the whole $\mathbb{R}^2$ the Laplacian is not positive-definite
and the Navier--Stokes system is not dissipative at least in a usual sense
even in the case of zero external forces and finite energy solutions,
see e.g. \cite{shonbek}, see also \cite{Gal1,Wie} and the references
therein concerning the decay properties of various types of solutions
for the Navier-Stokes problem with zero external forces. The presence
of external forces makes the problem more complicated and usually only
growing in time bounds for the solutions are available. We mention here
only the recent results concerning the polynomial growth in time for
the so-called uniformly local norms of infinite-energy solutions obtained
in \cite{Gal, Zel-inf}, see also the references therein.
\par
Let us now consider the damped and driven Navier--Stokes system
\begin{equation}\label{1.ns-main}
\begin{cases}
\Dt u+(u,\Nx)u+\Nx p+\alpha u=\nu\Dx u+g,\\
u\big|_{t=0}=u_0,\ \ \divv u=0.
\end{cases}
\end{equation}
with additional dissipative term $\alpha u$.
The drag/friction term $\alpha u$, where $\alpha>0$
is the Rayleigh or Ekman friction coefficient (or the Ekman
pumping/dissipation constant),
models the bottom friction in
two-dimensional oceanic models
and is the main energy sink in large scale atmospheric models~\cite{P}.
\par
The analytic properties of system \eqref{1.ns-main} (such as existence of solutions, their uniqueness and regularity, etc.) remain very close to the analogous properties of the classical Navier-Stokes equations or Euler equations if the inviscid case $\nu=0$ is considered. However, the friction term
$\alpha u$ is very essential for the dynamics since it removes the energy which piles up at the large scales and from the mathematical point of view compensates the lack of the Poincare inequality. This
makes the Navier--Stokes system and even the limit Euler system dissipative whatever
the domain is and allows to study its global attractors in various phase spaces. For instance, the so-called weak global attractor for the inviscid case is constructed in~\cite{I91} for the case of finite energy solutions; its compactness in the strong topology related with the $H^1$-norm is verified in \cite{CVZ} and \cite{CZ} for the cases of finite and infinite energy solutions respectively; the inviscid limit $\nu\to0$ is studied in \cite{const} including the absence of the so-called anomalous dissipation of enstrophy; the existence of a locally compact global attractor in the uniformly local phase spaces is established for the viscous case $\nu>0$ in \cite{Zel-inf}; see also~\cite{BCT}, \cite{W}
for the existence and uniqueness results for the stationary problem
and the
stability of stationary solutions for~\eqref{1.ns-main}
with~$\nu=0$.
From the point of view of the attractors
and their dimension the
system~\eqref{1.ns-main}
in the case of the periodic domain $x\in[0,2\pi L]^2$ was studied
in \cite{I-M-T},
where it was shown that the corresponding dynamical system
possesses a global attractor $\mathcal{A}$ (in $L^2$) whose fractal dimension
is finite and satisfies the following estimate
\begin{equation}\label{min}
\dim_f\mathcal{A}\le \min
\left(\sqrt{6}\,
\frac{\|\rot g\|L}{\nu\alpha}\,,\
\frac38\,\frac{\|\rot g\|^2}{\nu\alpha^3}\right),
\end{equation}
where
the values of the constants are updated in accordance with
\cite{I12JST}. The first estimate
is enforced in the regime $\nu/L^2\gg\alpha$, while the second
estimate
is enforced in the opposite regime $\nu/L^2\ll\alpha$. We observe that both
estimates are of the order $1/\nu$ as $\nu\to0^+$ if all the remaining parameters
are fixed. It was also shown in \cite{I-M-T} that this rate
of growth of the dimension is sharp, and the upper bounds were
supplemented with a lower bound of the order $1/\nu$, based on
the instability analysis of generalized Kolmogorov flows. The finite-dimensionality of the global attractor in the uniformly local phase spaces under the assumption that the external forces $g(x)$ decay to zero as $|x|\to\infty$ has been proved recently in \cite{Pen}, but no explicit upper bounds for the dimension was given there.
We would also like to point out that starting from the
paper~\cite{Lieb} the Lieb--Thirring inequalities are
an essential analytical tool in the estimates
of global Lyapynov exponents for the Navier--Stokes equations.
This fully applies to our case.
We now observe that the first estimate in~\eqref{min} blows up as
the size of the periodic domain $L\to \infty$. On the other hand,
the second estimate survives (the homogeneous $H^1$-norm is
scale invariant in two dimensions). Therefore, one might expect
that this estimate holds for $L=\infty$, that is, for
$x\in\mathbb{R}^2$, and a motivation of the present work
is to show that this is indeed the case.
In this paper we study the damped and driven Navier--Stokes system
\eqref{1.ns-main} in $\mathbb{R}^2$ in the class of {\it finite} energy
solutions and our main aim is to obtain explicit upper bounds for the attractor's dimension in terms of the parameters $\nu$ and $\alpha$ and various norms of the external forces $g$.
\par
In section~\ref{s2} we recall for the reader convenience the proof
of the well-posedness and derive the energy equality. Then
using the energy equality method \cite{rosa,Rosa} we establish the
asymptotic compactness of the solution semigroup and, hence,
the existence of the global attractor~$\mathcal{A}$.
In section~\ref{s3} we consider the case when the right-hand
side $g$ belongs to the scale of homogeneous Sobolev spaces
$\dot H^s$, $s\in[-1,1]$ and derive the following estimate(s)
for the fractal dimension of the attractor $\mathcal{A}$:
$$
\dim_f\mathcal A\le \frac{1-s^2}{64\sqrt{3}}
\(\frac{1+|s|}{1-|s|}\)^{|s|}\frac1{\alpha^{2+s}\,\nu^{2-s}}\|g\|^2_{\dot H^{s}},
\quad s\in[-1,1].
$$
In particular, for $s=1$ we obtain
$$
\dim_f\mathcal A\le \frac{1}{16\sqrt{3}}
\frac{\|\rot g\|^2}{\nu\,\alpha^{3}},
$$
which up to a constant coincides with the second estimate
in \eqref{min}, proving thereby our expectation.
In this paper we use standard notation.
The $L_2$-norm and the corresponding scalar product
are denoted by $\|\cdot\|$ and $(\cdot,\cdot)$.
\section{A priori estimates, well-posedness and
asymptotic compactness}\label{s2}
We study the damped and driven Navier-Stokes system \eqref{1.ns-main}
in $\mathbb R^2$.
Here $u(t,x)=(u^1,u^2)$ is the unknown velocity vector field, $p$ is
the unknown pressure, $g(x)=(g^1,g^2)$ is the given external force
(and without loss of generality we can and shall assume that
$\divv g=0$),
the advection term is
$$
(u,\Nx)v=\sum_{i=1}^2u^i\partial_{x_i}v,
$$
and
$\alpha>0$, $\nu>0$ are given parameters. We restrict ourselves
to considering only {\it finite} energy solutions, so we assume that
$$
g\in [L^2(\mathbb R^2)]^2,\ \ u_0\in \mathcal H:=\{u_0\in[L^2(\mathbb R^2)]^2,\ \divv u_0=0\},
$$
and by definition $u=u(t,x)$ is a weak solution of
\eqref{1.ns-main} if
\begin{equation}\label{1.def}
u\in C(0,T;\mathcal H)\cap L^2(0,T; [H^1(\mathbb R^2)]^2),\ \ T>0,
\end{equation}
and \eqref{1.ns-main} is satisfied in the sense of distributions.
This means that for every divergence free test function
$\varphi(t,x)=(\varphi^1,\varphi^2)$,
$\divv \varphi=0$, $\varphi\in C_0^\infty(R_+\times\mathbb R^2)$,
the following integral identity holds
\begin{multline}\label{1.weak}
-\int_{\mathbb R}(u,\Dt\phi)\,dt+\int_{\mathbb R}((u,\Nx)u,\varphi)\,dt+\alpha\int_{\mathbb R}(u,\varphi)\,dt+\\
+\nu\int_{\mathbb R}(\Nx u,\Nx\varphi)\,dt=\int_{\mathbb R}(g,\varphi)\,dt.
\end{multline}
The following fact concerning the global well-posedness of
the Navier-Stokes equations in 2D is well-known, see, for instance,
\cite{temam,temam1}.
\begin{theorem}\label{Th1.main}
For any $u_0\in\mathcal H$ there exists a unique
solution $u$ of problem \eqref{1.ns-main} satisfying
\eqref{1.def} and \eqref{1.weak}. Moreover, this solution
satisfies the following energy equality for almost all $t\ge0$:
\begin{equation}\label{1.energy}
\frac12 \frac d{dt}\|u(t)\|^2+\nu\|\Nx u(t)\|^2+
\alpha\|u(t)\|^2=(g,u(t)).
\end{equation}
\end{theorem}
\begin{proof} For the convenience of the reader, we remind below the key steps of the proof.
It is strongly based on the so-called Ladyzhenskaya interpolation inequality
\begin{equation}\label{1.lad}
\|u\|_{L^4(\mathbb R^2)}^2\le C\|u\|\|\Nx u\|
\end{equation}
which holds for any $u\in H^1(\mathbb R^2)$. Indeed, this inequality together with \eqref{1.def}
implies that any weak solution $u$ satisfies
$$
u\in L^4(0,T;L^4(\mathbb R^2))
$$
and, integrating by parts and using the Cauchy-Schwartz inequality, we see that
$$
|(u,\Nx)u,\varphi)|=
\bigl|\sum_{i,j=1}^2(u^iu^j,\partial_{x_j}\varphi^i)\bigr|\le C\|u\|^2_{L^4}\|\varphi\|_{H^1}.
$$
Therefore,
$$
\int_{\mathbb R}((u,\Nx u),\varphi)\,dt\le C\|u\|^2_{L^4(0,T;L^4(\mathbb R^2))}\|\varphi\|_{L^2(0,T;H^1(\mathbb R^2))}
$$
and
\begin{equation}\label{1.neg}
(u,\Nx)u\in L^2(0,T;\mathcal H^{-1})=[L^2(0,T;\mathcal H^1)]^*,
\end{equation}
where, as usual, $\mathcal H^1$ is a subspace of $[H^1(\mathbb R^2)]^2$
which consists of divergence free vector fields and
$\mathcal H^{-1}:=[\mathcal H^1]^*$. Thus, from \eqref{1.weak},
we see that $\Dt u\in L^2(0,T;\mathcal H^{-1})$ as well, and all
terms in \eqref{1.weak} make sense for the test function
$\varphi=u\in L^2(0,T;\mathcal H^1)$, so by approximation
arguments (and the fact that the divergent free vector fields
with compact support are dense in $\mathcal H$) we may take
$\varphi=u$ in \eqref{1.weak}. Then, using the well-known
orthogonality relation
\begin{equation}\label{1.null}
((u,\Nx)v,v)=-\frac12(\divv u,|v|^2)=0,
\end{equation}
we end up with the desired energy equality \eqref{1.energy},
see, for instance, \cite{temam} for the details.
\par
The existence of a weak solution can be obtained in a standard way
based on the energy equality \eqref{1.energy} either directly by
the Faedo-Galerkin method or by approximation the problem
\eqref{1.ns-main} in $\mathbb R^2$ by the corresponding problems
in bounded domains (see e.g., \cite{temam,temam1} for the details),
so we leave this proof to the reader and remind below the proof
of the uniqueness. Let $u_1$ and $u_2$ be two weak solutions of
problem \eqref{1.ns-main} and let $v=u_1-u_2$. Then $v$ solves
$$
\Dt v+(u_1,\Nx)v+(v,\Nx)u_2+\alpha v+\Nx \bar p=
\nu\Dx v,\ \ \divv v=0.
$$
Multiplying this equation by $v$ and integrating over
$x\in\mathbb R^2$, analogously to the energy equality, we have
\begin{equation}\label{1.dif-en}
\frac12\frac d{dt}\|v(t)\|^2+\alpha\|v(t)\|^2+
\nu\|\Nx v(t)\|^2=-((u_1,\Nx v),v)-((v,\Nx)u_1,v).
\end{equation}
The first term on the right-hand side of this equality vanishes
in view of~\eqref{1.null} and the second
term can be estimated using the Ladyzhenskaya inequality as follows:
\begin{multline*}
|((v,\Nx)u_2,v)|\le\|v\|^2_{L^4}\|\Nx u_2\|\le
C\|v\|\|\Nx v\|\|\Nx u_2\|\le\\\le \nu\|\Nx v\|^2+
C^2\nu^{-1}\|\Nx u_2\|^2\|v\|^2.
\end{multline*}
Inserting this estimate into the right-hand side of
\eqref{1.dif-en} we obtain
$$
\frac12\frac d{dt}\|v(t)\|^2\le
C^2\nu^{-1}\|\Nx u_2(t)\|^2\|v(t)\|^2,
$$
and the Gronwall inequality gives
$$
\|u_1(t)-u_2(t)\|^2\le
e^{2C^2\nu^{-1}\int_0^t\|\Nx u_2(s)\|^2ds}\|u_1(0)-u_2(0)\|^2.
$$
Thus, the uniqueness is proved and the theorem is also proved.
\end{proof}
\begin{corollary}\label{Cor1.dis} The weak solution $u(t)$ of problem \eqref{1.ns-main}
satisfies the following dissipative estimate:
\begin{equation}\label{1.dis}
\aligned
\|u(t)\|^2&\le\|u_0\|^2e^{-\alpha t}+\alpha^{-2}\|g\|^2,\\
2\nu\int_t^{t+T}\|\Nx u(s)\|^2ds&\le\alpha^{-1}T\|g\|^2+\|u(t)\|^2.
\endaligned
\end{equation}
\end{corollary}
\begin{proof}
Using the Cauchy--Schwartz
inequality, the first
estimate follows from~\eqref{1.energy} by the Gronwall inequality,
and the second estimate is proved by integrating~\eqref{1.energy}.
\end{proof}
Thus, equation \eqref{1.ns-main} defines a solution semigroup $S(t):\mathcal H\to\mathcal H$:
$$
S(t)u_0:=u(t),\ \ u_0\in\mathcal H
$$
where $u(t)$ is a weak solution of equation \eqref{1.ns-main}
with the initial data $u(0)=u_0$. Moreover, this semigroup
is dissipative according to estimate \eqref{1.dis} and is Lipschitz continuous:
for any $u_1,u_2\in\mathcal H$
\begin{equation}\label{1.lipp}
\|S(t)u_1-S(t)u_2\|\le Ce^{Kt}\|u_1-u_2\|,
\end{equation}
where the constants $C$ and $K$ depend only on $\|u_0^i\|$.
Our next aim is to verify the existence of a global attractor for
this semigroup. For the reader convenience, we first remind its
definition, see \cite{BV,Ch-V-book, CF88,temam} for more details.
\begin{definition}\label{Def1.attr} Let $S(t)$, $t\ge0$, be a semigroup acting in a Banach space $\mathcal H$. Then the set $\mathcal A\subset\mathcal H$ is a global attractor of the semigroup $S(t)$ if
\par
1) The set $\mathcal A$ is compact in $\mathcal H$.
\par
2) It is strictly invariant: $S(t)\mathcal A=\mathcal A$.
\par
3) It attracts the images of bounded sets in $\mathcal H$ as $t\to\infty$, i.e.,
for every bounded set $B\subset \mathcal H$ and every neighborhood
$\mathcal O(\mathcal A)$ of the set $\mathcal A$ in $\mathcal H$
there exists $T=T(B,\mathcal O)$ such that
$$
S(t)B\subset\mathcal O(\mathcal A)
$$
for all $t\ge T$.
\end{definition}
To state the abstract theorem on the existence of a global attractor, we need more definitions.
\begin{definition} Let $S(t)$ be a semigroup in a Banach space $\mathcal H$.
Then, a set $\mathcal B\subset\mathcal H$ is an {\it absorbing} set of $S(t)$ if,
for every bounded set $B\subset\mathcal H$, there exists $T=T(B)$ such that
$$
S(t)B\subset\mathcal B.
$$
A semigroup $S(t)$ is {\it asymptotically compact} if for
any bounded sequence $u_0^n$ in $\mathcal H$ and for any
sequence $t_n\to\infty$, the sequence $S(t_n)u_0^n$ is precompact in $\mathcal H$.
\end{definition}
In order to verify the existence of a global attractor we
will use the following criterion, see \cite{BV,Ch-V-book,Lad,S-Y,temam} for its proof.
\begin{proposition}\label{Prop1.attr} Let $S(t)$ be a semigroup
in a Banach space $\mathcal H$. Suppose that
\par
1) $S(t)$ possesses a bounded closed absorbing set $\mathcal B\subset H$;
\par
2) $S(t)$ is asymptotically compact;
\par
3) For every fixed $t\ge0$ the map $S(t):\mathcal B\to\mathcal H$ is continuous.
\par
Then the semigroup $S(t)$ possesses a global attractor
$\mathcal A\subset \mathcal B$. Moreover, the attractor $\mathcal A$ has the following structure:
$$
\mathcal A=\mathcal K\big|_{t=0},
$$
where $\mathcal K\subset L^\infty(\mathbb R,\mathcal H)$ is the set of complete
trajectories $u:\mathbb R\to\mathcal H$ of semigroup $S(t)$ which are defined
for all $t\in\mathbb R$ and bounded.
\end{proposition}
The next theorem which establishes the existence of a global attractor for the solution
semigroup $S(t)$ associated with equation \eqref{1.ns-main} is the main result of this section.
\begin{theorem}\label{Th1.attr}
The solution semigroup $S(t)$ of the damped Navier-Stokes problem \eqref{1.ns-main}
possesses a global attractor $\mathcal A$ in $\mathcal H$.
\end{theorem}
\begin{proof} We will check the assumptions of Proposition
\ref{Prop1.attr}. Indeed, the first assumption is satisfied
due to the dissipative estimate \eqref{1.dis} and the desired
bounded and closed absorbing set can be taken as
\begin{equation}\label{1.abs}
\mathcal B:=\big\{u_0\in\mathcal H\,:\ \|u_0\|_{L^2}^2\le 2\alpha^{-2}\|g\|^2_{L^2}\big\}.
\end{equation}
The third assumption is also satisfied due to estimate \eqref{1.lipp}.
Thus, we only need to check the asymptotic compactness. We will use the
so-called energy method (see \cite{ball,rosa, Rosa}) in order to do this.
\par
Indeed, let $u^0_n\in\mathcal H$ be a bounded sequence of the
initial data. Then, due to estimate \eqref{1.dis}, we may assume
without loss of generality that $u^0_n\in\mathcal B$.
Let $u_n(t)$, $t\ge-t_n$, $t_n\to\infty$, be the sequence of solutions of
\begin{multline*}
\Dt u_n+(u_n,\Nx)u_n+\alpha u_n+\Nx p_n=\nu\Dx u_n+g,\\
\ \divv u_n=0,\ \ u_n\big|_{t=-t_n}=u_n^0.
\end{multline*}
Then, $u_n(0)=S(t_n)u_n^0$ and we only need to verify that the
sequence $\{u_n(0)\}_{n=0}^\infty$ is precompact in $\mathcal H$.
In order to do so, we first verify that there exists a subsequence
(which we also denote by $u_n$ for simplicity) such that
\begin{equation}\label{1.weak-n}
u_n(0)\rightharpoondown u(0)
\end{equation}
converges weakly in $\mathcal H$ to $u(0)$, where
$u(t)$ is a complete bounded
trajectory $u\in \mathcal{K}$. We first note that due to the dissipative estimate \eqref{1.dis},
\begin{equation}\label{1.uniform}
\|u_n\|_{L^\infty(-T,T;\mathcal H)}+
\sup_{t\ge -T}\|u_n\|_{L^2(t,t+1;\mathcal H^1)}\le C,
\end{equation}
where $T\le t_n$ and $C$ is independent of $n$ and $T$. Moreover,
from the Ladyzhenskaya inequality \eqref{1.lad} we conclude also
that $u_n$ is bounded in $L^4(-T,T;L^4)$. Thus, by the
Banach-Alaoglu theorem, we may assume without loss of generality that
\begin{multline*}
u^n\to u\ \text{weak-star in }\ L^\infty(-T,T;\mathcal H) \\
\text{and weakly in\ } L^2(-T,T;H^1)\cap L^4(-T,T;L^4)
\end{multline*}
for every $T\in\mathbb N$ to some function $u\in L^\infty(\mathbb R;\mathcal H)\cap L^2_{loc}(\mathbb R,H^1)$ which also satisfies estimate \eqref{1.uniform}. Moreover, analogously to \eqref{1.neg}, we conclude that $\Dt u_n$ is uniformly bounded in $L^2(t,t+1;\mathcal H^{-1})$. Thus, without loss of generality
$$
\Dt u^n\rightharpoondown\Dt u
$$
in the space $L^2(t,t+1;\mathcal H^{-1})$ for all $t\in\mathbb R$. Using
the embedding
$$
L^2(t,t+1;\mathcal H^{-1})\cap L^2(t,t+1;\mathcal H^1)\subset C(t,t+1;\mathcal H),
$$
see \cite[Lemma III.1.2]{temam},
we see that $u_n(0)\rightharpoondown u(0)$. Thus, to verify
\eqref{1.weak-n}, we only need to check that $u(t)$, $t\in\mathbb R$ is
a weak solution of \eqref{1.ns-main}. In other words, we need to
pass to the limit as $n\to\infty$ in the analogue of \eqref{1.weak}:
\begin{multline}\label{1.weak1}
-\int_{\mathbb R}(u_n,\Dt\phi)\,dt+\int_{\mathbb R}((u_n,\Nx)u_n,\varphi)\,dt+
\alpha\int_{\mathbb R}(u_n,\varphi)\,dt+\\+\nu\int_{\mathbb R}
(\Nx u_n,\Nx\varphi)\,dt=\int_{\mathbb R}(g,\varphi)\,dt,
\end{multline}
where $\varphi(t,x)\in C_0^\infty(\mathbb R^3)^2$ is an arbitrary fixed
divergence free function. Since passing to the limit in the linear
terms is evident, we only need to pass to the limit $n\to\infty$ in
the non-linear term. Integrating by parts and rewriting the nonlinear term in the form
$$
((u_n(t),\Nx)u_n(t),\varphi(t))=-\sum_{i,j=1}^2
(u_n^i(t)u_n^j(t)\partial_{x_i}\varphi^j(t))
$$
and using the fact that the support of $\varphi$ is finite, we conclude that, for passing to the limit in the nonlinear term, it is enough to verify that, for every fixed $R,T>0$,
\begin{equation}\label{1.strong}
u_n\to u \ \text{ strongly in } L^2(-T,T;L^2(B^R_0)),
\end{equation}
where $B^R_0$ stands for the ball of radius $R$ in $\mathbb R^2$ centered
at the origin. To verify the convergence \eqref{1.strong}, we
recall that the sequence $u_n$ is bounded in the space
$L^2(-T,T;\mathcal H^1(B^R_0))$ due to the dissipative estimate
\eqref{1.dis}. Moreover, analogously to \eqref{1.neg} but using
the test functions $\varphi\in L^2(-T,T;\mathcal H^1_0(B^R_0))$,
we see that $\Dt u_n$ is bounded in the space
$L^2(-T,T;\mathcal H^{-1}(B^R_0))$. Thus, since
$$
\mathcal H^{1}(B^R_0))\subset\mathcal H(B^R_0)\subset\mathcal H^{-1}(B^R_0)
$$
and the first embedding is compact, the
compactness theorem (see, for instance~\cite[Theorem~III.2.1]{temam})
implies that the embedding
$$
H^1(-T,T;\mathcal H^{-1}(B^R_0))\cap L^2(-T,T;\mathcal H^1(B^R_0))\subset L^2(-T,T;\mathcal H)
$$
is compact. Therefore, $u_n$ is precompact in
$L^2(-T,T;L^2(B^R_0))$ for every $R>0$, $T>0$ and passing to
a subsequence if necessary, we conclude that the convergence
\eqref{1.strong} indeed holds. Thus, passing to the limit in the
nonlinear term of \eqref{1.weak1} is verified and $u$ is a weak
solution of \eqref{1.ns-main} which is defined for all $t\in\mathbb{R}$ and bounded, so
$u\in\mathcal K$. This means that the
convergence \eqref{1.weak-n} is verified.
\par
We are now ready to verify that
\begin{equation}\label{1.strong-n}
u_n(0)\to u(0) \ \text{strongly in } \ \mathcal H
\end{equation}
and finish the proof of the theorem. We multiply the energy equality~\eqref{1.energy}
for the solutions $u_n$ by $ e^{2\alpha t}$ and integrate from $-t_n$ to $0$:
\begin{multline}\label{1.energy-n}
\|u_n(0)\|^2=-2\int_{-t_n}^0e^{2\alpha s}\|\Nx u_n(s)\|^2\,ds+\\
+\|u_n(-t_n)\|^2e^{-2\alpha t_n}+2\int_{-t_n}^0e^{2\alpha s}(g,u_n(s))\,ds.
\end{multline}
We want to pass to the limit $n\to\infty$ in this equality.
Indeed, using the weak convergence $u_n\to u$ in $L^2_{loc}(\mathbb R,H^1)$ implying that
\begin{multline*}
\limsup_{n\to\infty}-2\int_{-t_n}^0e^{2\alpha s}\|\Nx u_n(s)\|^2\,ds=\\=
-2\liminf_{n\to\infty}\int_{-t_n}^0e^{2\alpha s}\|\Nx u_n(s)\|^2\,ds\le
-2\int_{-\infty}^0e^{2\alpha s}\|\Nx u(s)\|^2\,ds
\end{multline*}
and the uniform bounds \eqref{1.uniform}, we see from~\eqref{1.energy-n} that
$$
\limsup_{n\to\infty}\|u_n(0)\|^2\le-2\int_{-\infty}^0e^{2\alpha s}\|\Nx u(s)\|^2\,ds
+ 2\int_{-\infty}^0e^{2\alpha s}(g,u(s))\,ds.
$$
On the other hand, thanks to the energy equality, for the whole-line $L^2$-bounded solution
$u\in\mathcal K$ we have
$$
\|u(0)\|^2=-2\int_{-\infty}^0e^{2\alpha s}\|\Nx u(s)\|^2\,ds+
2\int_{-\infty}^0e^{2\alpha s}(g,u(s))\,ds
$$
and, therefore, taking into the account the weak convergence
\eqref{1.weak-n}, we finally arrive at
$$
\limsup_{n\to\infty}\|u_n(0)\|^2\le\|u(0)\|^2\le\liminf_{n\to\infty}\|u_n(0)\|^2.
$$
Thus, $\lim_{n\to\infty}\|u_n(0)\|=\|u(0)\|$, and the strong
convergence \eqref{1.strong-n} is proved:
$$
\lim_{n\to\infty}\|u_n(0)-u(0)\|^2=
\lim_{n\to\infty}\left(\|u_n(0)\|^2+\|u(0)\|^2-2(u_n(0),u(0))\right)
=0,
$$
and the theorem is also proved.
\end{proof}
To conclude the section, we also discuss the extra regularity of the attractor $\mathcal A$. We say that $u:\mathbb R\to\mathcal H^1$ is a {\it strong} solution of the damped Navier-Stokes equations \eqref{1.ns-main} if
$$
u\in C(0,T;\mathcal H^1)\cap L^2(0,T;H^2(\mathbb R^2))
$$
and equation \eqref{1.ns-main} is satisfied in the sense of distributions. The following analogue of Theorem \ref{Th1.main} gives the global well-posedness of strong solutions of the Navier-Stokes problem.
\begin{theorem}\label{Th1.strong}
Let $u_0\in\mathcal H^1$. Then, the weak solution $u$ constructed
in Theorem \ref{Th1.main} is actually a strong solution and
satisfies for almost all $t\ge0$ the following analogue of the
energy equality:
\begin{equation}\label{1.strong-en}
\frac12\frac d{dt}\|\Nx u(t)\|^2+\nu\|\Dx u(t)\|^2
+\alpha\|\Nx u(t)\|^2=(g,\Dx u(t)).
\end{equation}
\end{theorem}
The proof of this theorem is analogous to Theorem \ref{Th1.main} and even
simpler since the solution $u$ is a priori more regular now, see, for instance,
\cite{temam} for more details. We only mention here that the identity
\eqref{1.strong-en} follows by multiplication of equation \eqref{1.ns-main}
by $\Dx u$, integration over $x\in\mathbb R^2$ and using the well-known
orthogonality relation
$
((u,\Nx)u,\Dx u)=0
$
which holds for any $u\in [H^2(\mathbb R^2)]^2$, $\divv u=0$. In fact, integrating
by parts and setting $\omega:=\rot u=\partial_{x_1}u^2-\partial_{x_2}u^1$, the divergence theorem gives
$$
((u,\Nx)u,\Dx u)=\frac12\int_{\mathbb{R}^2}\divv (u \omega^2)\,dx=0
$$
for a smooth $u\in C^\infty_0(\mathbb{R}^2)^2$, $\divv u=0$. The general case follows
by a standard approximation procedure.
The next corollary gives the dissipative estimate for the solutions
of \eqref{1.ns-main} in $\mathcal H^1$, and its proof is similar to
that of~\eqref{1.dis}.
\begin{corollary}\label{Cor1.strong-dis}
The strong solution
satisfies the following estimate:
\begin{equation}\label{1.strong-dis}
\aligned
\|\Nx u(t)\|^2&\le\|\Nx u_0\|^2e^{-\alpha t}+
(2\alpha\nu)^{-1}\|g\|^2,\\
\nu\int_t^{t+T}\|\Dx u(s)\|^2ds&\le\nu^{-1}T\|g\|^2+\|\Nx u(t)\|^2.
\endaligned
\end{equation}
\end{corollary}
In the following corollary we establish the smoothing
property for the weak solutions of equation \eqref{1.ns-main}.
\begin{corollary}\label{Cor1.smo}
For $t>0$, $u(t)\in\mathcal H^1$ and the following estimate holds:
\begin{equation}\label{1.smo}
\|u(t)\|^2_{H^1}\le Ct^{-1}\(\|u(0)\|^2+\|g\|^2\),\ 0<t\le1,
\end{equation}
where $C=C(\alpha,\nu)$.
In particular, the attractor $\mathcal A$ is bounded in $\mathcal H^1$.
\end{corollary}
\begin{proof}
Since $\|\Nx u(t)\|^2$ is integrable,
for every $0<T\le1$, there exists $\tau\le T$ such
that $u(\tau)\in\mathcal H^1$ and
$$
\|\Nx u(\tau)\|^2\le \frac1T\int_0^T\|\Nx u(s)\|^2ds\le
\frac1{2\nu T}\(\|u(0)\|^2+\alpha^{-1}T\|g\|^2\),
$$
where the second inequality follows from~\eqref{1.dis}.
By \eqref{1.strong-dis} with the initial time $t=\tau$, we have
$$
\|\Nx u(T)\|^2\le \|\Nx u(\tau)\|^2+
(2\alpha\nu)^{-1}\|g\|^2\le CT^{-1}\(\|u(0)\|^2+\|g\|^2\).
$$
This and \eqref{1.dis} prove \eqref{1.smo}
and guarantee that the $\mathcal H^1$-ball
$$
\mathcal B_1:=\{u_0\in\mathcal H^1,\ \ \|u_0\|^2_{\mathcal H^1}\le R\}
$$
is a bounded absorbing set for the
solution semigroup $S(t)$
if $R$ is large enough.
Thus, $\mathcal A\subset\mathcal B_1$ is bounded
in $\mathcal H^1$ and the proof is complete.
\end{proof}
\begin{remark} Arguing analogously, it is not difficult to show that the smoothness of the global attractor $\mathcal A$ is restricted by the smoothness of the external forces $g$. In particular, the attractor will be $C^\infty$-smooth (analytic) if the external forces are $C^\infty$-smooth (analytic).
\end{remark}
\section{Fractal dimension of the attractor: upper bounds}\label{s3}
In this section, we prove that the global attractor constructed
above has finite fractal (box-counting) dimension and give
explicit upper bounds for this dimension in terms of the physical
parameters $\nu$ and $\alpha$ and the norms of the right-hand
side $g$ in homogeneous Sobolev spaces. To this end, we first need to remind some
definitions, see \cite{temam,BV} for more detailed exposition.
\begin{definition}\label{Def2.dim} Let $\mathcal A\subset\mathcal H$ be a compact set in a metric space $\mathcal H$. Then, by the Hausdorff criterion, for every $\varepsilon>0$ it can be covered by the finite number of $\varepsilon$-balls in $\mathcal H$. Let $N_\varepsilon(\mathcal A,\mathcal H)$ be the minimal number of such balls. Then, the fractal (box-counting) dimension of $\mathcal A$ in $\mathcal H$ is defined via the following expression:
$$
\dim_f(\mathcal A,\mathcal H):=\limsup_{\varepsilon\to0}\frac{\ln N_{\varepsilon}(\mathcal A,\mathcal H)}{\ln\frac1\varepsilon}.
$$
The fractal dimension coincides with the usual dimension if the set
$\mathcal A$ is regular enough (for instance, for the case
where $\mathcal A$ is a Lipschitz manifold in $\mathcal H$),
but may be non-integer for irregular sets (for instance,
for the standard ternary Cantor set $K\subset[0,1]$ this
dimension is $\frac{\ln2}{\ln3}$),
see, for instance, \cite{robinson} for more details.
\end{definition}
We now estimate the dimension of the global attractor $\mathcal{A}$
by means of the so-called volume contraction method~\cite{CF85},
\cite{temam}. The solution semigroup $S_t$ is uniformly quasi-differentiable
on the attractor in the sense that there exists (for a fixed $t\ge0$) a linear bounded operator
$D S(t,u_0)$ such that
\begin{equation}\label{Differ}
\|S(t)u_1-S(t)u_0-D S(t,u_0)\cdot(u_1-u_0)\|\le h(\|u_1-u_0\|),
\end{equation}
where $h(r)/r\to0$ as $r\to0$, and $u_0,u_1\in\mathcal{A}$.
The quasi-differential $D S(t,u_0)$ is the solution operator
$\xi\to v(t)$ of the following equation of variations:
\begin{equation}\label{2.eq-var2}
\partial_t v=L(t,u_0)v:=-\Pi\bigl((v,\Nx)u(t)+(u(t),\Nx)v\bigr)-\alpha v+\nu\Dx v,\ v(0)=\xi,
\end{equation}
where $\Pi:[L^2(\mathbb R^2)]^2\to\mathcal H$ is
the Leray ortho-projection onto the divergence free vector fields, and $u(t)=S(t)u_0$
is the solution lying on the attractor and parameterized by $u_0\in\mathcal{A}$.
For the proof see~\cite{BV1}, where it is also shown that the
solution semigroup is even differentiable for all $u_0\in\mathcal{H}$ and
the differential $D S(t,u_0)$ depends continuously on the point $u_0$.
We define for $m=1,2\dots$ the numbers $q(m)$ (the sums of the first
$m$ global Lyapunov exponents)
$$
q(m)=\limsup_{t\to\infty}\ \sup_{u_0\in {\mathcal A}}\ \
\sup_{\{v_j\}_{j=1}^m\in\mathcal H^1}
\frac{1}t
\int_0^t \sum_{j=1}^m\bigl({ L}(\tau,u_0)v_j,v_j\bigr)d\tau,
$$
where the supremum closest to the integral is taken with respect
to all $L^2$-orthonormal families $\{v_j\}_{j=1}^m\in\mathcal H^1$.
To define $q(m)$ for all real $m\ge1$ we just linearly
interpolate between $q(m)$ and $q(m+1)$ so that $q(m)$
is now a piece-wise linear continuous function of $m$.
The following theorem is the key technical tool for estimating the dimension of the attractor via the so-called volume contraction method.
\begin{theorem}\label{Th2.vol-contr} Let $S(t):\mathcal H\to\mathcal H$ be the solution semigroup associated with problem \eqref{1.ns-main}
in a Hilbert space $\mathcal H$ and let $\mathcal A$ be a compact invariant
set of $S_t$ in $\mathcal H$: $S(T)\mathcal{A}=\mathcal{A}$.
Suppose that
the semigroup $S(t)$ is uniformly quasi-differentiable for every fixed $t$ on $\mathcal A$
in the sense of~\eqref{Differ}.
Suppose further that the quasi-differential $D S(t,u_0)$
depends continuously on the initial point $u_0\in\mathcal{A}$ as
a map $D S(t,\cdot):u_0\to\mathcal{L}(\mathcal{H},\mathcal{H})$.
Suppose that there exists number $m>0$ such that $q(m)<0$. Then
$$
\dim_f\mathcal{A}<m.
$$
\end{theorem}
For the proof of this theorem see \cite{CF85}, \cite{temam} in the case of Hausdorff dimension and \cite{CI} for the fractal dimension.
\begin{remark}
The condition on the continuity of the quasi-differentials with respect to the initial point
is redundant in the case of the Hausdorff dimension.
It is also redundant in the case of the fractal dimension if the graph of
$q(m)$ lies below the straight line joining the
points $(m-1,q(m-1))$ and $(m,q(m))$, where $q(m)<0$ and $q(m-1)\ge0$, see
\cite{Ch-V-book,CI01}.
Also, in applications to infinite dimensional dissipative dynamical
systems an upper bound for $q(m)$ is usually found in the
form
$$
q(m)\le -c_1m^{\gamma}+c_2,\quad\gamma\ge1.
$$
For example, as we shall shortly see, $\gamma=1$ in our case. In this case,
also without the continuity condition, we have
$$
\dim_f\mathcal{A}\le (c_2/c_1)^{1/\gamma}.
$$
\end{remark}
To apply this theorem for obtaining the fractal dimension of the attractor $\mathcal A$ of
the semigroup $S(t)$ generated by damped Navier-Stokes equation
\eqref{1.ns-main} we need to state the Lieb-Thirring inequality which plays a fundamental role in estimating the quantities $q(m)$.
\begin{lemma}[{\bf Lieb--Thirring inequality}]\label{T:L-T}
Let
$\{v_j\}_{j=1}^m\in {H}^1(\mathbb{R}^2)^2$
be a family of orthonormal vector-functions and let
$\divv v_j=0$.
Then the following inequality holds for
$\rho(x)=\sum_{k=1}^m|v_k(x)|^2$:
\begin{equation}\label{L-T}
\|\rho\|^2=
\int_{\mathbb{R}^2}\biggl(\sum_{j=1}^m|v_j(x)|^2\biggr)^2dx \le
c_\mathrm{LT}
\sum_{j=1}^m \| \Nx v_j\|^2,\quad
c_\mathrm{LT}
\le\frac1{2\sqrt{3}}\,.
\end{equation}
\end{lemma}
\begin{proof}
The proof (see~\cite{I-Stokes}) is a reduction to the scalar case
(which works in two dimensions) \cite{CI} and the
use of the main result in ~\cite{D-L-L}.
\end{proof}
We first set $\nu=1$ and $\alpha=1$ and begin
with estimating the $m$-trace of the operator $L(t,u_0)$
in the equation of variations~\eqref{2.eq-var2} for this particular case.
The general case will be reduced later to the case $\nu=\alpha=1$ by the proper scaling.
\begin{proposition}\label{Prop1.trace} Let $\nu=1$ and $\alpha=1$. Then, the following estimate holds:
\begin{multline}\label{2.tr-est}
\limsup_{T\to\infty}\sup_{u_0\in\mathcal A}\sup_{\{v_j\}_{j=1}^m\in\mathcal H^1}
\frac1T\int_0^T
\sum_{j=1}^m\bigl({ L}(t,u_0)v_j,v_j\bigr)\,dt\le
\\\le-m +\frac{1}{16\sqrt{3}} \,\limsup_{T\to\infty}\sup_{u_0\in\mathcal A}\frac1T
\int_0^T\|\Nx u(t)\|^2\,dt,
\end{multline}
where $u(t)=S(t)u_0$.
\end{proposition}
\begin{proof} Let $v_1,\dots,v_m\in\mathcal H^1$ be an orthonormal family in $\mathcal H$.
Then, integrating by parts and using that the vector fields $v_i$ are divergence free, we get
$$
\allowdisplaybreaks
\aligned
&\sum_{j=1}^m(L(u(t))v_j,v_j)=\\&=-\sum_{j=1}^m\|\Nx v_j\|^2-
\sum_{j=1}^m\|v_j\|^2
-\int_{\mathbb{R}^2}\sum_{j=1}^m\sum_{i,k=1}^2v_j^k\partial_{x_k}u^iv_j^i\,dx=\\&=
-m-\sum_{i=1}^m\|\Nx v_i\|^2-\int_{\mathbb{R}^2}\sum_{j=1}^m\sum_{i,k=1}^2v_j^k\partial_{x_k}u^iv_j^i\,dx.
\endaligned
$$
To estimate the right-hand side above, we use the following point-wise inequality
\begin{equation}\label{strange}
\big|\sum_{k,i=1}^2
v^k\partial_{x_k}u^iv^i\big|\le
2^{-1/2}|\Nx u||v|^2
\end{equation}
which holds for any $v=(v^1,v^2)\in\mathbb R^2$ and any Jacobi matrix $\Nx u=\(\partial_{x_i}u^j\)_{i,j=1}^2\in\mathbb R^4$ such that $\partial_{x_1}u^1+\partial_{x_2}u^2=0$, see \cite[Lemma~4.1]{CI}. Indeed,
setting
$$
v=(\xi,\eta),\qquad \Nx u=A:=\left(
\begin{array}{cr}
a & b \\
c & -a \\
\end{array}
\right),
$$
we have by the Cauchy-Schwartz inequality
$$
\aligned
&(Av,v)=a(\xi^2-\eta^2)+(b+c)\xi\eta=a(\xi^2-\eta^2)+((b+c)/2)\,2\xi\eta\le\\\le
&\sqrt{a^2+(b+c)^2/4}\,\sqrt{(\xi^2-\eta^2)^2+4\xi^2\eta^2}=
\sqrt{a^2+(b+c)^2/4}\cdot |v|^2\le\\\le
&\sqrt{a^2+(b^2+c^2)/2}\cdot|v|^2=\frac1{\sqrt{2}}\sqrt{a^2+b^2+c^2+a^2}\cdot|v|^2,
\endaligned
$$
and estimate \eqref{strange} is proved. Using this pointwise estimate and
the Lieb--Thirring inequality for
divergence free vector fields~\eqref{L-T}, we finally have
$$
\aligned
&\sum_{j=1}^m(L(u(t))v_j,v_j)\le-m-\sum_{i=1}^m\|\Nx v_i\|^2+\frac1{\sqrt2}\int_{\mathbb R^2}\rho(x)|\Nx u(x)|\,dx\le\\
\le& -m-\sum_{i=1}^m\|\Nx v_i\|^2+\frac1{\sqrt{2}}\|\rho\|\|\Nx u(t)\|\le\\\le
&-m-\sum_{i=1}^m\|\Nx v_i\|^2+\frac1{\sqrt{2}}\biggl(c_\mathrm{LT}
\sum_{j=1}^m \| \Nx v_j\|^2\biggr)^{1/2}\|\Nx u(t)\|\le\\\le
&-m+\frac{c_\mathrm{LT}}8\|\Nx u(t)\|^2=
-m+\frac1{16\sqrt{3}}\|\Nx u(t)\|^2.
\endaligned
$$
Integrating this inequality for $t\in[0,T]$ and taking the supremum over all $u_0\in\mathcal A$,
we obtain \eqref{2.tr-est} and finish the proof of the proposition.
\end{proof}
\begin{theorem}\label{Cor2.dim} Let the assumptions of Theorem \ref{Th1.main} hold
and let $\nu=\alpha=1$. Then the fractal dimension of the global attractor
$\mathcal A$ of problem \eqref{1.ns-main} satisfies the following estimate:
\begin{equation}\label{2.dim-nu}
\dim_f\mathcal A\le\frac{1}{16\sqrt{3}}
\limsup_{T\to\infty}\frac1T\sup_{u_0\in\mathcal A}\int_0^T\|\Nx u(t)\|^2\,dt,
\end{equation}
where $u(t)=S(t)u_0$.
\end{theorem}
\begin{proof}
This estimate is a corollary of Theorem \ref{Th2.vol-contr} and estimate \eqref{2.tr-est}.
\end{proof}
Thus, we only need to estimate the integral in the RHS of \eqref{2.dim-nu}.
We assume below that the right-hand side $g$ belongs to
the homogeneous Sobolev space $\dot H^s(\mathbb R^2)=(-\Dx)^{-s/2}L^2(\mathbb R^2)$, $s\in\mathbb R$
with norm:
\begin{equation}\label{2.hom}
\|u\|^2_{\dot H^s}:=\int_{\mathbb R^2}|\xi|^{2s}|\hat u(\xi)|^2\,d\xi,
\end{equation}
where $$\hat u(\xi):=\frac1{2\pi}\int_{\mathbb R^2}u(x)e^{- i\xi x}\,dx$$ is the
Fourier transform of $u$, see \cite{triebel} for more details. Then, obviously,
\begin{equation}\label{2.hom-sp}
\|u\|_{\dot H^0}=\|u\|,\ \ \|u\|_{\dot H^1}=\|\Nx u\|,
\ \ \|u\|_{\dot H^2}=\|\Dx u\|,
\end{equation}
and the following interpolation inequalities immediately follow from
definition~\eqref{2.hom} and the Cauchy--Schwartz inequality:
\begin{equation}\label{2.hom-int}
\aligned
&\|u\|_{\dot H^s}\le \|u\|^{1-s}\|\Nx u\|^s,\\
&\|\Dx u\|_{\dot H^{-s}}=\|\Nx u\|_{\dot H^{1-s}}\le
\|\Nx u\|^{s}\|\Dx u\|^{1-s},\ \ s\in[0,1].
\endaligned
\end{equation}
\begin{corollary}\label{Cor2.dim-est} Let the assumptions of
Theorem \ref{Th1.main} hold and let, in addition,
$g\in\dot H^{-s}$ for some $s\in[0,1]$ and $\nu=\alpha=1$.
Then the fractal dimension of the attractor $\mathcal A$
of problem \eqref{1.ns-main} satisfies the following estimate:
\begin{equation}\label{2.dim-attr-s}
\dim_f \mathcal A\le \frac{1-s^2}{64\sqrt{3}}\(\frac{1+s}{1-s}\)^s\|g\|^2_{\dot H^{-s}}.
\end{equation}
\end{corollary}
\begin{proof} From the energy equality \eqref{1.energy} with $\nu=\alpha=1$, interpolation inequalities \eqref{2.hom-int} and the Young inequality, we get
\begin{multline*}
\frac12\frac d{dt}\|u\|^2_{L^2}+\|u\|^2_{L^2}+\|\Nx u\|^2_{L^2}=
(g,u)\le\|g\|_{\dot H^{-s}}\|u\|_{\dot H^s}\le\\\le \|g\|_{\dot H^{-s}}\|\Nx u\|^s_{L^2}\|u\|^{1-s}_{L^2}\le
\frac{\varepsilon^2}2\|g\|^2_{\dot H^{-s}}+\frac{\delta^p}p\|\Nx u\|^2_{L^2}+\frac{\gamma^q}q\|u\|^2_{L^2},
\end{multline*}
where $p=\frac2{s}$, $q=\frac2{1-s}$ and positive numbers $\varepsilon,\delta,\gamma$ are such that $\varepsilon\delta\gamma=1$. Fixing now the parameter $\gamma$ in such way that
$$
\frac{\gamma^q}q=1\ \ \Rightarrow \ \ \gamma=\(\frac{1-s}2\)^{-\frac{1-s}2},
$$
and integrating over $t\in[0,T]$, we arrive at
$$
\frac1{2T}\(\|u(t)\|^2-\|u(0)\|^2\)+\(1-\frac{\delta^p}p\)\frac1T\int_0^T\|\Nx u(t)\|^2
\,dt\le \frac{\varepsilon^2}2\|g\|_{\dot H^{-s}}^2.
$$
From the dissipative estimate \eqref{1.dis} we conclude that
$$
\lim_{T\to\infty}\sup_{u_0\in\mathcal A}\frac1{2T}\(\|u(t)\|^2-\|u(0)\|^2\)=0
$$
and, therefore,
\begin{equation}\label{2.huge}
\limsup_{T\to\infty}\frac1T\sup_{u_0\in\mathcal A}\int_0^T\|\Nx u(t)\|^2
\,dt\le \frac{\varepsilon^2}2\(1-\frac{\delta^p}p\)^{-1}\|g\|_{\dot H^{-s}}^2
\end{equation}
and we only need to optimize the coefficient in the RHS with respect
to $\varepsilon$ and $\delta$. Indeed, since $\varepsilon\delta\gamma=1$, we conclude that
$$
\varepsilon\delta=\(\frac{1-s}2\)^{\frac{1-s}2}\ \Rightarrow\ \varepsilon=\delta^{-1}\(\frac{1-s}2\)^{\frac{1-s}2}
$$
and
$$
\frac{\varepsilon^2}2\(1-\frac{\delta^p}p\)^{-1}=
\frac12\(\frac{1-s}2\)^{1-s}\frac1{x\left(1-\frac s2 x^{\frac1s}\right)},
$$
where $x:=\delta^2$. Thus, it only remains to maximize the function
$$
f(x):=x\left(1-\frac s2 x^{\frac1s}\right)
$$
on the interval $x\ge0$:
$$
f'(x)=1-\frac{s+1}2x^{\frac1s}=0 \Rightarrow
x=\(\frac2{s+1}\)^s \Rightarrow f(x)=\frac1{s+1}\(\frac2{s+1}\)^s.
$$
Inserting the obtained estimates into the right-hand side of \eqref{2.huge}, we finally get
$$
\limsup_{T\to\infty}\frac1T\sup_{u_0\in\mathcal A}\int_0^T\|\Nx u(t)\|^2\,dt\le \frac{1-s^2}4\(\frac{1+s}{1-s}\)^s\|g\|_{\dot H^{-s}}^2.
$$
This estimate together with \eqref{2.dim-nu} completes the proof of the corollary.
\end{proof}
The next corollary gives the analogous result for $g\in\dot H^s$ with $s\ge0$.
\begin{corollary}\label{Cor2.dim-est1} Let the assumptions of
Theorem~\ref{Th1.main} hold and let, in addition, $g\in\dot H^{s}$
for some $s\in[0,1]$ and $\nu=\alpha=1$. Then the fractal dimension
of the attractor $\mathcal A$ of problem \eqref{1.ns-main} satisfies
the following estimate:
\begin{equation}\label{2.dim-attr-s1}
\dim_f\mathcal A\le \frac{1-s^2}{64\sqrt{3}}\(\frac{1+s}{1-s}\)^s\|g\|^2_{\dot H^{s}}.
\end{equation}
\end{corollary}
\begin{proof} From the second energy equality \eqref{1.strong-en} with $\nu=\alpha=1$, interpolation inequalities \eqref{2.hom-int} and the Young inequality, we get
\begin{multline*}
\frac12\frac d{dt}\|\Nx u\|^2+\|\Nx u\|^2+\|\Dx u\|^2=
(g,\Dx u)\le\|g\|_{\dot H^{s}}\|\Dx u\|_{\dot H^{-s}}\le\\\le \|g\|_{\dot H^{s}}
\|\Nx u\|^s\|\Dx u\|^{1-s}\le\\\le
\frac{\varepsilon^2}2\|g\|^2_{\dot H^{s}}+\frac{\delta^p}p\|\Nx u\|^2+\frac{\gamma^q}q\|\Dx u\|^2,
\end{multline*}
where the exponents $p,q$ and the constants $\varepsilon,\delta,\gamma$ are {\it exactly} the same as in the proof of the previous corollary.
Thus, using the strong dissipative estimate \eqref{1.strong-dis} and arguing exactly as in the proof of the previous corollary, we end up with
$$
\limsup_{T\to\infty}\frac1T\sup_{u_0\in\mathcal A}\int_0^T\|\Nx u(t)\|^2
\,dt\le \frac{1-s^2}4\(\frac{1+s}{1-s}\)^s\|g\|_{\dot H^{s}}^2
$$
which together with \eqref{2.dim-nu} finishes the proof of the corollary.
\end{proof}
Thus, combining estimates \eqref{2.dim-attr-s} and \eqref{2.dim-attr-s1}, we end up with
\begin{equation}\label{2.dim-attr-s2}
\dim_f\mathcal A\le\frac{1-s^2}{64\sqrt{3}}\(\frac{1+|s|}{1-|s|}\)^{|s|}\|g\|^2_{\dot H^{s}}
\end{equation}
which holds for the case $\nu=\alpha=1$ if $g\in\dot H^{-1}\cap \dot H^1$ and $s\in[-1,1]$.
\par
Finally, we need the analogue of estimate \eqref{2.dim-attr-s2} for general $\nu,\alpha>0$.
We reduce this general case to the particular case
$\nu=\alpha=1$ by the proper scaling of $t$, $x$ and $u$.
Indeed, let $u=u(t,x)$ be a solution of \eqref{1.ns-main}
with arbitrary $\nu,\alpha>0$. Then, taking
$$
t':=\alpha t,\ \ x'=\(\frac\alpha\nu\)^{1/2}x,\ \ \tilde u=\frac1{(\alpha\nu)^{1/2}}u
$$
we see that the function $\tilde u(t',x'):=\frac1{(\alpha\nu)^{1/2}}u(t,x)$
solves equation \eqref{1.ns-main} with $\nu=\alpha=1$ and the
external forces $\tilde g(x')=\alpha^{-1}(\alpha\nu)^{-1/2}g(x)$.
Since the fractal dimension of the attractor does not change under
this scaling, using the obvious scaling properties of the $\dot H^s$-norm:
$$
\|g\(\gamma\cdot\)\|^2_{\dot H^s}=\gamma^{2(s-1)}\|g\|^2_{\dot H^s},
$$
we have proved the following result.
\begin{theorem} Let the assumptions of Theorem \ref{Th1.main}
hold (now with arbitrary positive $\nu$ and $\alpha$) and let,
in addition $g\in \dot H^s$ for some $s\in[-1,1]$. Then, the
fractal dimension of the attractor $\mathcal A$ in $\mathcal H$ satisfies the following estimate:
\begin{equation}\label{2.main-est}
\dim_f\mathcal A\le \frac{1-s^2}{64\sqrt{3}}
\(\frac{1+|s|}{1-|s|}\)^{|s|}\frac1{\alpha^2\nu^2}\(\frac\nu\alpha\)^s\|g\|^2_{\dot H^{s}}.
\end{equation}
\end{theorem}
\begin{proof} Indeed, the above estimate follows from \eqref{2.dim-attr-s2} and the identity
$$
\|\tilde g\|^2_{\dot H^s}=\frac1{\alpha^2\nu^2}\(\frac\nu\alpha\)^s\|g\|^2_{\dot H^s}.
$$
\end{proof}
If $g\in \dot H^{-1}\cap\dot H^1$, then the rate of growth of the
estimate~\eqref{2.main-est} with respect to $\nu$ as $\nu\to0$ is
the smallest when $s=1$. In this case we have
\begin{corollary}\label{Cor:s=1} Suppose that
$g\in \dot H^1$. Then the
fractal dimension of the attractor $\mathcal A$ satisfies
\begin{equation}\label{2.main-est-s=1}
\dim_f\mathcal A\le \frac{1}{16\sqrt{3}}
\frac{\|\rot g\|^2}{\alpha^3\nu}.
\end{equation}
\end{corollary}
\begin{proof}
Since $\divv g=0$, it follows that
$$
\|g\|^2_{\dot H^1}=\|\Nx g\|^2=\|\divv g\|^2+\|\rot g\|^2=
\|\rot g\|^2.
$$
\end{proof}
\subsubsection*{\bf{Acknowledgements}}\label{SS:Acknow}
This work was done with the financial support
from the Russian Science Foundation (grant no. 14-21-00025).
|
\section{Introduction}
\label{sec:introduction}
Spreadsheet systems are the software system of choice for many
non-professional programmers, often called end-user programmers~\cite{enduser}, like
for example, teachers, accountants, secretaries, engineers, managers,
etc. In the last century such end users would develop their
spreadsheets individually in their own desktops, and sharing and reuse
was not the usual procedure. The recent advent of powerful mobile
devices, and, as a consequence, the availability of powerful cloud-based
spreadsheet systems (like, for example, Google Drive), has
dramatically changed this situation. Nowadays, spreadsheets are
complex software systems, developed and maintained by many (end) users.
Very much like in the development of other software systems, different
developers (end users in this case) are concerned with
different aspects of the functionality of the
system. However, while modern programming languages offer modularity
mechanisms providing powerful abstractions to develop software
collaboratively, spreadsheet systems offer no such support to their
users. As a result, a spreadsheet tends to evolve into a single
software artifact where all business logic from all different users is
defined! In such a collaborative setting if a new user needs to
express a new business rule on the spreadsheet data, he/she has to do it
by intrusively adding formulas to the existing spreadsheet.
The programming language community developed advanced modularity
mechanisms to avoid this problem in regular programming languages.
In that sense, aspect-oriented programming (AOP) is a popular and advanced
technique that enables the modular implementation of the so-called
crosscutting concerns. The crosscutting structure tends to appear
tangled and scatted across several artifacts of a software system.
While implementing such crosscutting concerns, the crosscutting
structure can appear in common software development concerns,
such as distribution and persistence~\cite{Soares-Laureano-Borba02},
error handling~\cite{Lippert-Lopes00,Castor-etal09}, certain
design patterns~\cite{Hannemann-Kiczales02}, tracing~\cite{Kiczales-etal01},
or design by contract~\cite{Kiczales-etal01,Rebelo-etal14,Rebelo-Lima-Leavens11}.
In this paper we introduce the concept of AOP to spreadsheets\xspace.
We start by introducing a running example in Section~\ref{sec:motivation}.
Our first contribution is presented in Section~\ref{sec:architecture}
where we adapt the AOP concept to spreadsheets. For instance,
a key feature of a common aspect-oriented language is
the possibility to add advice before, after or around a join point.
Since spreadsheets\xspace are two-dimensional, AOP features like advising need to
be adapted in order to be applied to spreadsheets\xspace.
The second contribution we do in this work is the design of a new
language to implement AOP for spreadsheets.
This language is presented in Section~\ref{sec:language}.
An overview of the architecture of the system is presented
in Section~\ref{sec:weaver}. Finally, Section~\ref{sec:related} discusses related work,
and Section~\ref{sec:conclusion} presents our conclusions.
\section{Motivation}
\label{sec:motivation}
In this section we present an example of a popular setting to use spreadsheets:
manage the students marks in a course. Such a spreadsheet is shown in
Fig.~\ref{fig:example}.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.5]{images/example}
\caption{Student grading spreadsheet.}
\label{fig:example}
\end{figure}
In this simple spreadsheet, the final mark of a student is the average
of three evaluation items: two exams and one essay. Thus, this mark is
obtained by the formula \texttt{=AVERAGE(B2:D2)} (for the student in
line~2).
Let us consider now a real scenario where three different users
share, access, and update this spreadsheet: a \textit{teaching
assistant}, that mainly structures the spreadsheet and compiles the
different marks, the \textit{teaching coordinator} who has to validate
the spreadsheet, decide on borderline cases, and send the final
marks to the academic services, and finally, the \textit{Erasmus
coordinator} who has to adapt the marks of Erasmus students to the
ECTS system\footnote{For more details on the ECTS system please refer to
\url{http://ec.europa.eu/education/tools/docs/ects-guide_en.pdf}.}.
Because spreadsheets offer no modularity mechanisms, a typical spreadsheet
development environment allows a spreadsheet to quickly evolve into a more
complex one. That is, each of the users adds/updates data and
formulas in order to express the (crosscutting) logic they need. In the following,
we show the spreadsheet after the teaching and Erasmus coordinator update it.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.5]{images/example-round-ects}
\caption{Example of a student grading spreadsheet with mark adjustments (cell
E3) and ECTS marks (column F).}
\label{fig:example-round-ects}
\end{figure}
The data edited by one user can be seen as intrusive code for the
others. For example, the teaching coordinator and its assistant are
not concerned with the ECTS system, although this concern is
now part of their spreadsheet.
Let us analyze the teaching coordinator role in more detail. Because
he decides on borderline cases, he decided to approve only one of such
cases (student in line~3). Instead of changing that particular final
mark formula, and as often occurs in reality, he just replaces this
formula by the constant \texttt{5.0}. Although the result of such an
operation is a correct spreadsheet, this action has several problems.
First, there is no documentation on how borderline cases have been
decided (the threshold is not known). Second, the original spreadsheet
data is lost, and it may be difficult to recover it (for
example, if other borderline students ask for explanations).
This is the natural setting to apply AOP. In this style of programming
users do not intrusively modify the original program (a spreadsheet in
our case), but instead, a new software fragment is defined in order to
specify the program transformations needed to express such new
concern. These software fragments are called \textit{aspects}. Then, a specific
AOP mechanism, called \textit{weaver}, given the original program and the
aspect(s), weaves them into a coherent executable software program.
We present in Listing~\ref{lst:borderline} our first spreadsheet aspect, the one that
specifies the coordinator concerns on borderline cases.
\begin{lstlisting}[caption={An aspect to handle borderline marks.},label={lst:borderline}]
aspect BorderlineCase
finalmark : select sheet{*}.column{*}.cell{*}
around finalmark {
#{cell.result >= 4.8 && cell.result < 5 ? 5 : cell.value}
} when {
cell.column[0].value = "Final Mark"
}
end
\end{lstlisting}
An aspect-oriented language has three main parts. First it is necessary to select
the \textit{join points} of interest by means of pointcut declarations.
In the example, the pointcut \texttt{finalmark} selects any cell within
any worksheet. This is done by the \textit{select} command in our language.
Then, the \textit{around} advice declaration defines the actions
(transformations in our case) to be applied.
Hence, it specifies when the \textit{result} of evaluating a formula is greater
than of equal to \texttt{4.8}, and if it is, the cell is replaced by the constant \texttt{5.0}.
To access the result of the computation of a cell, that is, the result
of the formula in the cell, we use \lstinline|cell.result|. This accesses
a cell, and after that, its computed \textit{result}.
The \textit{when} statement specifies when the action is applied (in
this case to all columns labeled \texttt{Final Mark}).
This aspect clearly and non-intrusively defines the crosscutting rule used
to decide on borderline cases: all students with marks greater
than or equal to \texttt{4.8} are approved in the course.
Similarly, the Erasmus coordinator can define an aspect where the
rules to define the ECTS marks are expressed.
This aspect is presented in Listing~\ref{lst:ectsmark}.
\begin{lstlisting}[caption={Aspect to add ECTS marks.},label={lst:ectsmark}]
aspect AddECTSMark
finalmarks : select sheet{*}.column{*}.cell{*}
right finalmarks {
=IF(#{cell.name}<=10 && #{cell.name}>=9.5
, "A"
, IF(#{cell.name}<9.5 && #{cell.name}>=8.5
, "B"
, IF(#{cell.name}<8.5 && #{cell.name}>=6.5
, "C"
, IF(#{cell.name}<6.5 && #{cell.name}>=5.5
, "D"
, IF(#{cell.name}<5.5 && #{cell.name}>=5
, "E"
, "F")))))
} when {
column[0].value == "Final Mark"
&& cell.row <> 0
}
right finalmarks {
ECTS Mark
} when {
column[0].value = "Final Mark"
&& cell.row = 0
}
end
\end{lstlisting}
In this aspect, we use \lstinline|cell.name| to access the cell
address (for instance ``E4''). Thus, when the \lstinline|IF| formula
is placed in the corresponding cells, the correct references
will be calculated and inserted in the formula replacing the
use of \lstinline|cell.name|.
An aspect weaver can then weave this aspects into the original
spreadsheet individually, or by composing them. Actually,
Fig.~\ref{fig:example-round-ects} shows the result of weaving
the aspect \textit{AddECTSMark} after the
aspect \textit{BorderlineCase}. Thus, aspects can be combined. It
should also be noticed that such a spreadsheet based weaver has to
dynamically weave aspects since for some aspects
(\textit{BorderlineCase}) only after computing formulas, aspects can
be weaved. Indeed, it is necessary to execute the spreadsheet
to get the results of formulas, for instance, when the \lstinline|cell.result|
operator is used.
We have just presented two aspects of our AOP spreadsheet language and
briefly explained how aspects are weaved into a spreadsheet system. Next
sections present in detail the design and implementation of this
language and weaver.
\section{Applying Aspect-Oriented Programming\\to Spreadsheets}
\label{sec:architecture}
In this section, we discuss how AOP concepts can be applied to spreadsheets.
Specific examples are described in Section~\ref{sec:language}.
\subsection{Spreadsheet Join Point Model}
In order to apply AOP to a language, it is necessary to identify the
\textit{join point} model that the new aspect language supports. A join point is
a well-defined point in the program that is specified by a \textit{pointcut},
that is, an expression to match specific elements within the language. In our
case, these elements are set to be the main elements of the spreadsheet, where
users will want to perform operations on, for instance, use an alternate
worksheet for testing or add more cells for debugging. For spreadsheets, we
define the following join points of interest:
\begin{itemize}
\item worksheets;
\item ranges; and,
\item cells.
\end{itemize}
With these join points the spreadsheet is accessible from aspects, hence users
can separate concerns at different levels:
\begin{center}
\includegraphics[width=0.35\columnwidth]{images/hierarchy}
\end{center}
When a join point is instantiated, it is possible to perform an action using
advice. Advice are additional behavior that one wants for the underlying program, that can
be new worksheets, ranges of cells, or single cells. This allows to separate the
business logic of the spreadsheet into several concerns when developing the
spreadsheet and then join everything in order to achieve the wanted application.
\subsection{Worksheet}
A spreadsheet file is composed of a set of worksheets, each of which containing
the cells with the data and formulas. Worksheets are the top-level artifacts in
a spreadsheet\xspace.
As a join point, a worksheet (see Fig.~\ref{fig:sheets}) can be modified by the
standard AOP advice in the following manner:
\begin{itemize}
\item {\bf before} -- insert a worksheet before the current join point;
\item {\bf after} -- insert a worksheet after the current join point;
\item {\bf around} -- insert a worksheet before and/or after the
current joint point and/or define an alternative
worksheet for the current join point.
\end{itemize}
\begin{figure}[ht]
\centering
\includegraphics[width=0.7\columnwidth]{images/sheets}
\caption{List of three worksheets within a spreadsheet file.}\label{fig:sheets}
\end{figure}
As an example, consider Fig.~\ref{fig:sheets} with \texttt{Sheet2} as the join
point. A \textit{before} advice for this join point results in a new worksheet
between the worksheets \texttt{Sheet1} and \texttt{Sheet2}. On the other hand,
if we consider an \textit{after} advice, the resulting worksheet is between
worksheets \texttt{Sheet2} and \texttt{Sheet3}.
To specify a worksheet, two options are available: either defining the cells for
the new worksheet, or referencing a worksheet to be duplicated.
\subsection{Cell}
Cells are the finest grained join points possible in the spreadsheets world.
Unlike common programming languages, they are inserted in a two-dimensional
plan. Thus, the usual \textit{before} and \textit{after} advice declarations are
not completely adequate since ambiguities may arise about which of the two
dimensions (vertical or horizontal) should be used. To overcome this issue,
each kind of advice declaration is separated in two for each dimension in the
plane, resulting in the following:
\begin{center}
\begin{tabular}{|l|l|l|}
\hline
\bf \backslashbox{direction}{advice} & \it \textbf{before} & \it \textbf{after} \\\hline
\textbf{vertical} & above & below \\\hline
\textbf{horizontal} & left & right \\\hline
\end{tabular}
\end{center}
The advice that can be defined for cell join points is:
\begin{itemize}
\item {\bf left} -- add a cell, or range of cells, to the left of the current
join point;
\item {\bf above} -- add a cell, or range of cells, above the current join
point;
\item {\bf right} -- add a cell, or range of cells, to the right of the
current join point;
\item {\bf below} -- add a cell, or range of cells, below the current join
point;
\item {\bf around} -- define an alternative cell for the current join point.
\end{itemize}
\begin{figure}[ht]
\centering
\includegraphics[width=0.7\columnwidth]{images/advices2d}
\caption{Two-dimensional advice.}\label{fig:advices2d}
\end{figure}
A cell is specified just by stating its contents. A range of cells is specified
by defining the cells that it contains.
\subsection{Range}
A range is a set of cells contained in a worksheet. As it regularly happens in
spreadsheets\xspace, our setting only handles rectangular ranges, that is, the range must be a
rectangle and all the cells in that rectangle must part of the range.
Advice declaration for ranges is similar to the ones for cells, but special
care must be taken to match range sizes when adding new cells, or ranges before
or after it. Thus, we have the following possible advice declarations:
\begin{itemize}
\item {\bf left} -- add a range of cells, to the left of the current join
point;
\item {\bf above} -- add a range of cells, above the current join point;
\item {\bf right} -- add a range of cells, to the right of the current join
point;
\item {\bf below} -- add a range of cells, below the current join point;
\item {\bf around} -- define an alternative range for the current join point.
\end{itemize}
\section{Language to Specify Aspects for Spreadsheets}
\label{sec:language}
In this section we present our language for aspects for spreadsheets. This
language is based on existing ones for common purposes programming languages,
and implements the vision we described in Section~\ref{sec:architecture}. The
proposed language allows to specify aspects defining its pointcuts and advice.
For each component of the aspects to be written by the user we present next the
corresponding grammar. This grammar is the artifact used to validate, though a
compiler, the correctness of the aspects written by the users. We start by
introducing the grammar for pointcuts.
\subsection{Pointcuts}
Pointcuts are defined by an expression pattern, which is
specified by the following grammar:
\medskip
\begin{grammar}
<join\_point> ::= <jp\_name> `:' `select' <pExpr>
\end{grammar}
\medskip
\noindent
where \prodiden{pExpr} is a pattern expression. The expression is used to
define the kind of join point to be selected: \emph{sheet}, \emph{range}, or
\emph{cell}.
The allowed expressions in join points are instances of the
following productions:
\medskip
\begin{grammar}
<pExpr> ::= <pSheet>
\alt <pSheet> `.' <pRange>
\alt <pSheet> (`.' <pRange>)? `.' <pCell>
<pSheet> ::= `sheet' `{' <pSheetExpr> `}'
<pSheetExpr> ::= `name' <bComp> <string>
\alt `number' <bComp> <integer>
\alt `*'
<pRange> ::= `range' `{' <pRangeExpr> `}'
\alt `column' `{' <pRangeExpr> `}'
\alt `row' `{' <pRangeExpr> `}'
<pRangeExpr> ::= `name' <bComp> <string>
\alt `*'
<pCell> ::= `cell' `{' <pCellExpr> `}'
<pCellExpr> ::= `name' <bComp> <string>
\alt `match' <bComp> <string>
\alt `*'
<bComp> ::= `==' | `<' | `<=' | `>' | `>=' | `<>'
\end{grammar}
\medskip
The pattern expression \prodiden{pExpr} allows to select worksheets
\prodiden{sheet}, ranges \prodiden{range}, or cells \prodiden{cell}.
When
specifying a worksheet as a join point, it is possible to select a worksheet relative
to a given name, to a worksheet index, or to select all worksheets.
For ranges,
there are three kinds that can be selected: a column range (\texttt{\small
column}) with a width of one cell, a row range (\texttt{\small row}) with a
height of one cell, or a range with any rectangular shape (\texttt{\small range}).
For cells, it is possible to specify its address (\texttt{\small name}), or a pattern-match expression
(\texttt{\small match}). When
no specific worksheet, range or cell is necessary, the wildcard symbol \lit{*} can be used.
With this, we can select, for instance, the second worksheet
\medskip
\lstinline|worksheet_jp: select worksheet{number=2}|
\medskip
\noindent
or a rectangular range of 3 columns by 2 rows starting at cell \texttt{A2} in any
worksheet:
\medskip
\lstinline|range_jp: select worksheet{*}.range{name="A2:C3"}|
\medskip
\noindent
or the cells in the first row of any worksheet:
\medskip
\lstinline|cell_jp: select worksheet{*}.range{row=1}.cell{*}|
\medskip
Depending on the kind of join point selected, different artifacts are made
available to work with within the advice. For worksheets, we can use the variable
\lstinline|worksheet|, and then one of the attributes: \texttt{\small name} or
\texttt{\small number}. For ranges, depending on its kind, the available
variable can be: \texttt{\small range}, \texttt{\small column}, or
\texttt{\small row}; they have the attribute \texttt{\small name}.
Moreover, since ranges are sets of cells, indexation can be used, for instance, to
select the first row of a column, one can write \texttt{\small column[0]}. For
cells, we have the variable \texttt{\small cell} which has the attribute
\texttt{\small name} (its cell reference).
\subsection{Advice}
Advice are the actions to apply to the join points. They are defined by the
following grammar:
\medskip
\begin{grammar}
<advice> ::= $\left(<advice\_name> `:'\right)?$ <advice\_position> <jp\_name> `{' <code> `}' <advice\_condition>
<advice\_position> ::= `left' |~`above' |~`right' |~`below' |~`around'
<advice\_condition> ::= $\left(`when {' \; <bExpr> \; `}'\right)?$
\end{grammar}
\medskip
\noindent
where \prodiden{code} is a cell or a list of cells and respective contents:
\medskip
\begin{grammar}
<code> ::= <string> | <cellList>
<cellList> ::= <cellRef> `=' <string> $\left(`;' <cellList>\right)?$
\end{grammar}
\medskip
The contents of the cell \prodiden{string} can be defined using interpolation of values made
available in the context of the advice (for instance, join point contents) using
interpolation. For example, to add a row which evaluates the total of a
column (for instance, the join point is a column range) where the join point is named
\textit{myColumn}:
\medskip
\lstinline|below myColumn { =SUM(#{range.name}) }|
\medskip
In the above example, interpolation is used to introduce a value that is
available in the join point, but is not accessible from common spreadsheet
formulas. If the column range is \texttt{\small C1:C20}, then the formula would
be \texttt{\small =SUM(C1:C20)}.
Note that the order by which the advice are applied is important.
They are applied according to their precedence which
is defined by the order they are defined. In the case of \lstinline|before|
the ones defined first are the ones with more precedence. Thus,
the ones defined earlier are applied first.
For the \lstinline|after| it is the other way around, that is,
although the ones declared first are the ones with more priority,
they are executed last. This is the common behavior of aspects
for other programming languages.
Advice can be applied conditionally, that is, when a specified criterion is met.
This is specified with a boolean expression as defined by the following grammar:
\medskip
\begin{grammar}
<bExpr> ::= `!' <bExpr>
\alt <bExpr> (`&&' | `||') <bExpr>
\alt <expr> <bComp> <expr>
<expression> ::= <bExpr> |~<var> |~<string> |~<number> |~<cellRef>
|~<rangeRef>
\end{grammar}
\subsection{Aspect}
An aspect is composed of the definition of the pointcuts and the advice to apply
to them, as defined by the following grammar:
\medskip
\begin{grammar}
<aspect> ::= `aspect' <string>\\
<join\_point>+ <advice>+\\
`end'
\end{grammar}
\medskip
\section{A Weaver For Spreadsheets}
\label{sec:weaver}
In this section we present our architectural model for aspect spreadsheets\xspace. The aspects
are defined using the language presented in Section~\ref{sec:language}, and
together with the spreadsheet\xspace, they are interpreted by the \textit{weaver}. In this
context, the spreadsheet\xspace is only complete when considered together with the aspects.
Some aspects can be handled statically, that is, without executing the spreadsheet\xspace. For
instance, the example shown is Listing~\ref{lst:ectsmark} does not need to
execute the spreadsheet\xspace to know what to do.
However, some aspects require the execution of the program, in our case, of the
spreadsheet\xspace (formulas). This is the case illustrate in Listing~\ref{lst:borderline}
where it is necessary to \textit{compute} the student grade to decide if the
final grade is changed or not. In that case we used the operator
\lstinline|cell.result|, which requires to evaluate the formula of the
underlying cell to obtain its result. Thus, the weaver must be integrated with
a spreadsheet\xspace execution engine. Indeed we intend to build the weaver inside Excel
itself, so it can reuse its recalculation mechanism.
In these cases, the spreadsheet\xspace is kept untouched, but since now the spreadsheet is only
complete when considered with the aspects defined, the values it shows may
change.
If the user wants to see the original spreadsheet\xspace then, it is only necessary to
deactivate the aspects.
Fig.~\ref{fig:arc} illustrates the integration of our weaver with a spreadsheet
system to create an aspect-oriented spreadsheet\xspace system.
\begin{figure}[!htb]
\includegraphics[width=\columnwidth]{images/arc.pdf}
\caption{A weaver for aspect-oriented spreadsheets.}
\label{fig:arc}
\end{figure}
\section{Related Work}
\label{sec:related}
Aspect-oriented programming has been applied to several programming languages,
for instance, Java~\cite{kiczales2001overview}, C++~\cite{aspectc++2002}, or
\textsc{Matlab}\xspace~\cite{cardoso2010domain,aspectmatlab2010}, targeting specific
characteristics of each language in order to improve them. Some of these
applications contributed to shape the use of aspects, and introduced new
concepts to this paradigm.
Java was one of the first languages to be exposed to aspects through a language
dubbed AspectJ~\cite{kiczales2001overview}. This language is similar to Java so
it feels familiar to Java users. It provides a dynamic join point model, where
some advice are applied at compile time, but others are applied only during run
time when the complete information about the execution is available. The nature
of spreadsheets, where both data and computation are at the same level, imposes
a dynamic join point model in order to have access to run-time values.
AOP was also used to support development of embedded systems. The LARA
language~\cite{cardoso2012lara} was purposely designed with this goal, but has a
wider range of applications. Since it can target several languages, we inspired
ourselves on it as the basis for the specification of aspects for spreadsheets.
In the context of spreasheets, there are several works presenting techniques to
transform spreadsheets by using spreadsheet specific transformation languages
\cite{felienne2014bumblebee,badame2012refbook}, and querying languages
\cite{vlhcc-td13,vlhcc13}. Moreover, such transformations can be refactorings
to remove spreadsheet
smells~\cite{felienne2012smells,Cunha:2012:TCS:2346340.2346357,
CunhaFMMS12,quatic14} and thus improve their usage and reduce possible error
entry points.
BumbleBee~\cite{felienne2014bumblebee} is a Microsoft Excel add-in mainly for
performing refactorings to remove smells, but can also perform other kinds of
transformations. It finds cells where it can apply a previously defined set of
transformations, lets the user select the transformation to apply, and then
applies the transformation to a selected range, to the entire worksheet, or to
the entire file. A limitation is that BumbleBee only supports intra-formula
transformations. Our language uses the BumbleBee's transformation language to
support cell value transformations.
RefBook~\cite{badame2012refbook} is another tool to perform refactorings to
remove spreadsheet smells. It implements a set of seven refactorings in a
Microsoft Excel add-in allowing users to perform refactorings when working with
their spreadsheets. Our approach can also be used to perform these refactorings.
More generic transformations for spreadsheets, introduced by end-users' needs,
can be performed using program synthesis. This technique added the ability to
transform strings of text~\cite{Gulwani2011}, or tables of
data~\cite{Harris2011} from user-supplied examples, providing a familiar way to
solve common tasks when dealing with spreadsheets.
Another kind of transformation, targeting spreadsheet testing, is
mutation~\cite{spreadsheetmutation}. The goal is to perform mutations in the
spreadsheet, that is, small modifications, in order to analyze a test set for
the spreadsheet being tested. Using our aspect system, mutation is also
possible, by using the BumbleBee transformation language to specify cell
mutations.
In \cite{vlhcc10,ase14} we presented techniques to infer the business logic of
spreadsheet data. Such techniques restructure the spreadsheet data into
different (relational) tables. Such tables can be viewed as aspects of the
business logic/spreadsheet data. Thus, this approach may be used to infer
aspects in spreadsheets and to evolve a legacy spreadsheet into an
aspect-oriented one.
\section{Conclusion}
\label{sec:conclusion}
This paper proposes the use of aspect-oriented programming for spreadsheets. We
have designed an aspect language that considers spreadsheet peculiarities, and a
dynamic weaver that is embedded in the evaluation mechanism of a spreadsheet
system.
Although AOP provides a powerful modular approach that is particularly suitable
to be used in software that is shared and being collaboratively developed, our
work opens some important questions that we intend to answer in future work by
conducting empirical studies, namely:
\begin{itemize}
\item Are end users able to understand the abstractions provided by AOP
and to use it in practice?\\
This is not only related to our proposed
language as it is a more general question. Nevertheless, we need
to answer it so we can understand how to better make our AOP language
available for user. This is specially important when dealing with end users.
\item Does AOP improves end-users' productivity?\\
We have shown that some of the model-driven approaches we introduced
in the past can do that. We will conduct similar studies to evaluate
this new proposal.
\item Is the textual definition of the AOP language adequate or should
we use a more spreadsheet-like one? \\
When dealing with more advanced users, it is not always the case they
prefer visual languages~\cite{2015arXiv150207948C}. However, for end users
this is probably the case. We will extend our work with a visual
language to allow end users to define aspects in a more friendly way.
\end{itemize}
\section*{Acknowledgment}
This work is financed by the FCT – Fundação para a Ciência e a Tecnologia
(Portuguese Foundation for Science and Technology) within project UID/EEA/50014/2013.
This work has also been partially funded by FLAD/NSF
through a project grant (ref. 233/2014).
The last author is supported
by CAPES through a \textit{Programa Professor Visitante do
Exterior (PVE)} grant (ref. 15075133).
\IEEEtriggeratref{19}
|
\section{Introduction}
Superfluid $^3$He is a condensate of spin-triplet $p$-wave Cooper pairs, which undergoes complex symmetry breaking, in addition to $U(1)_N$ gauge symmetry breaking~\cite{vollhardt:book}.
Without magnetic fields, the superfluid phases of $^3$He consist of the nodal gapped A-phase in a high-temperature and high-pressure region and the B-phase with an isotropic gap in another wide parameter region.
In the full gapped B-phase, the residual symmetry group is
\begin{align}
H_{\rm B}=SO(3)_{{\bm L}+{\bm S}}\times T\times C,
\end{align}
where $SO(3)_{{\bm L}+{\bm S}}$ is simultaneous rotation symmetry of orbital and spin spaces, $T$ is time-reversal symmetry, and $C$ is particle-hole symmetry.
Three-dimensional spin-triplet superfluids with $T$ and $C$, such as the B-phase, are topological phases belonging to class DIII in the Altland-Zrinbauer symmetry classes~\cite{schnyder:2008}.
In consequence of the topological order, the Majorana bound states and helical spin current exist on surfaces of the superfluid $^3$He B-phase~\cite{chung:2009,volovik:2009a,nagato:2009,tsutsumi:2011b,tsutsumi:2012c,wu:2013,okuda:2012}.
The topological phases are classified by whether there are discrete symmetries $T$ and $C$~\cite{schnyder:2008}.
It has been known that also spatial discrete symmetries in point group can give rise to a nontrivial topology of excitations in the topological phases~\cite{fu:2011}.
The topological excitations without a gap are robust if the symmetries are preserved.
Host matters of the symmetry protected excitations are called topological crystalline insulators~\cite{fu:2011} or topological crystalline superconductors~\cite{ueno:2013,tsutsumi:2013}.
The symmetry protected excitations are mainly discussed on gapless boundary states with reflection symmetries~\cite{chiu:2013,morimoto:2013}.
Recently, the topological classification is generalized for excitations in a defect with a two-fold symmetry including magnetic point group symmetry~\cite{shiozaki:2014}.
An example of the symmetry protected excitation in the superfluid $^3$He B-phase is the surface Majorana bound state under magnetic fields~\cite{mizushima:arXiv}.
The gapless Majorana bound state is preserved under weak magnetic fields parallel to the surface although time-reversal symmetry is broken~\cite{mizushima:2012b,mizushima:2012}.
This system has magnetic $\pi$-rotation symmetry instead of time-reversal symmetry because the flipped magnetic field by time-reversal operation is recovered by the $\pi$-rotation around the axis perpendicular to the surface.
The gapless Majorana bound state under magnetic fields is due to the protection by the magnetic $\pi$-rotation symmetry.
The symmetry protected Majorana excitations are also expected in the vortex bound state, where time-reversal symmetry is broken by the phase winding of the vortex.
For an axisymmetric vortex in the B-phase, the residual symmetry group is
\begin{align}
H_{\rm vortex}=U(1)_Q\times P\times C,
\label{eq:sym_vortex}
\end{align}
where the generator of the $U(1)_Q$ symmetry is $Q=L_z+S_z-N/2$ for a singly quantized vortex along the $z$-axis~\cite{salomaa:1987,volovik:book}.
The residual spatial symmetry $P$ is $P_2\times P_3$ for the most symmetric vortex called $o$ vortex~\cite{ohmi:1983}, where $P_2$ is magnetic reflection symmetry on a plane including the vortex line and $P_3$ is magnetic $\pi$-rotation symmetry around a axis perpendicular to the vortex line.
The combination of $P_2$ and $P_3$ gives inversion symmetry $P_1=P_2P_3$.
A vortex with a symmetry $P=P_1$, $P_2$, or $P_3$ is called $u$, $v$, or $w$ vortex, respectively, and the lowest symmetric vortex without the symmetry $P$ is called $uvw$ vortex.
In experiment, two types of vortices, V1 vortex in a high-temperature and high-pressure region and V2 vortex in a low-temperature and low-pressure region, were observed within the B-phase~\cite{volovik:book}.
The transition line between the vortex states was determined by the NMR measurement~\cite{hakonen:1983,hakonen:1983b,pekola:1984} and the measurement of critical velocity~\cite{pekola:1984b,pekola:1985}.
Since the transition pressure on the superfluid critical temperature is reproduced by the Ginzburg-Landau theory, V1 vortex and V2 vortex are believed to be axisymmetric $v$ vortex and non-axisymmetric $v$ vortex, respectively~\cite{thuneberg:1986b,thuneberg:1987}.
Breaking of the axial symmetry over the transition from V1 vortex to V2 vortex was confirmed by the observation of the Goldstone mode related to twisting of the anisotropic vortex core~\cite{kondo:1991}.
The existence of $P_2$ symmetry which is a direct evidence to form $v$ vortex, however, has not been verified.
For the identification of the kind of vortex in the superfluid $^3$He B-phase, the symmetry protected vortex bound state can be utilized.
Indeed, we will demonstrate that $P_3$ symmetry guarantees the existence of Majorana zero energy modes in the vortex bound state.
On the other hand, $P_2$ symmetry does not provide symmetry protected excitations in a vortex.
Therefore, many zero energy modes in axisymmetric $v$ vortex~\cite{silaev:2009} are gapped out by the axial symmetry breaking.
The formation of the gap in low energy excitations through the vortex transition implies the realization of $v$ vortex with only $P_2$ symmetry.
This paper is arranged as follows: In Sec.~II, we formulate the Bogoliubov-de Gennes (BdG) theory and the quasiclassical theory.
Vortex bound states are calculated by the BdG theory with order parameters (OPs) which are self-consistently obtained by the quasiclassical theory.
The possible spatial symmetries $P_1$, $P_2$, and $P_3$ for an axisymmetric vortex in the B-phase are summarized in Sec.~III.
The spatial symmetries play an important role in the topological classification of the vortex bound state.
In Sec.~IV, we discuss the bound state in the most symmetric $o$ vortex with $P_1$, $P_2$, and $P_3$ symmetries.
The vortex bound state has symmetry protected Majorana zero modes; however, $o$ vortex is not realized in the actual B-phase.
The $v$ vortex with $P_2$ symmetry is believed to be realized in the B-phase, whose bound state is discussed in Sec.~V.
There is a difference between low energy excitations in axisymmetric $v$ vortex and non-axisymmetric $v$ vortex in consequence of that $P_2$ symmetry does not protect zero energy modes.
We also discuss the bound state in $w$ vortex with $P_3$ symmetry in Sec.~VI.
Concerning the OP, the difference between $w$ vortex and $v$ vortex is only a phase of induced components of the OP around the vortex.
However, the bound state in $w$ vortex dramatically changes from that in $v$ vortex and has symmetry protected Majorana zero modes owing to $P_3$ symmetry.
We devote the final section to the summary in which we also mention topologically trivial bound states in $uvw$ vortex and $u$ vortex.
\section{Formulation}
\subsection{Bogoliubov-de Gennes theory}
We have numerically calculated the vortex bound states for some kinds of vortices in the B-phase by the BdG theory.
The BdG theory gives discretized modes in the vortex bound state at $\Delta^2/E_{\rm F}$ intervals, where $\Delta$ is a superfluid gap and $E_{\rm F}$ is the Fermi energy~\cite{caroli:1964}.
Note that the discreteness can be negligible when we consider physical quantities in the superfluid $^3$He owing to $\Delta/E_{\rm F}\sim 10^{-3}$~\cite{vollhardt:book};
however, we use the BdG theory in order to discuss the symmetry protection for the discretized modes in the vortex bound state.
The BdG equation in spin-triplet superfluid states is given as~\cite{kawakami:2011}
\begin{multline}
\int d{\bm r}_2
\begin{pmatrix}
\hat{h}({\bm r}_1,{\bm r}_2) & \hat{\Delta }({\bm r}_1,{\bm r}_2) \\
-\hat{\Delta }^{\dagger }({\bm r}_1,{\bm r}_2) & -\hat{h}^{\rm T}({\bm r}_1,{\bm r}_2)
\end{pmatrix}
\vec{u}_{\nu }({\bm r}_2)\\
=E_{\nu }\vec{u}_{\nu }({\bm r}_1),
\end{multline}
where the quasiparticle wave function with the $\nu$-th excited state with the eigenvalue $E_{\nu }$ is $\vec{u}_{\nu }({\bm r})=[u_{\nu }^{\uparrow }({\bm r}),u_{\nu }^{\downarrow }({\bm r}),v_{\nu }^{\uparrow }({\bm r}),v_{\nu }^{\downarrow }({\bm r})]^{\rm T}$.
The single particle Hamiltonian and the OP are described as
\begin{align}
\hat{h}({\bm r}_1,{\bm r}_2)=&\left[-\frac{\hbar^2\nabla_1^2}{2m}-E_{\rm F}\right]\delta({\bm r}_1-{\bm r}_2)\hat{1},\\
\hat{\Delta}({\bm r}_1,{\bm r}_2)
=&
\begin{pmatrix}
\Delta_{\uparrow\uparrow }({\bm r}_1,{\bm r}_2) & \Delta_{\uparrow\downarrow }({\bm r}_1,{\bm r}_2)\\
\Delta_{\downarrow\uparrow }({\bm r}_1,{\bm r}_2) & \Delta_{\downarrow\downarrow }({\bm r}_1,{\bm r}_2)
\end{pmatrix},
\end{align}
where $m$ is mass of the condensed particle and $\Delta_{\sigma\sigma'}({\bm r}_1,{\bm r}_2)=V(r')\sum_{\nu }u_{\nu }^{\sigma }({\bm r}_1)[v_{\nu }^{\sigma'}({\bm r}_2)]^*f(E_{\nu })$ with the interparticle interaction $V(r')$ at $r'=|{\bm r}_1-{\bm r}_2|$ and the Fermi distribution function $f(E_{\nu })$.
Here, we consider the bound state in a vortex along the $z$-axis; then the wave number $k_z$ becomes a well-defined quantum number.
The BdG equation for the $k_z$-resolved two-dimensional (2D) form is given as~\cite{kaneko:2012}
\begin{multline}
\int d{\bm \rho }_2
\begin{pmatrix}
\hat{h}_{k_z}(\bi{\rho }_1,\bi{\rho }_2) & \hat{\Delta }_{k_z}(\bi{\rho }_1,\bi{\rho }_2) \\
-\hat{\Delta }_{-k_z}^{\dagger }(\bi{\rho }_1,\bi{\rho }_2) & -\hat{h}_{-k_z}^{\rm T}(\bi{\rho }_1,\bi{\rho }_2)
\end{pmatrix}
\vec{u}_{\nu,k_z}(\bi{\rho }_2)\\
=E_{\nu,k_z}\vec{u}_{\nu,k_z}(\bi{\rho }_1),
\label{2DBdG}
\end{multline}
where
\begin{align}
\hat{h}_{k_z}(\bi{\rho }_1,\bi{\rho }_2)=\left[-\frac{\hbar^2\nabla_{\rm 2D}^2}{2m}-E_{\rm F}^{2{\rm D}}(k_z)\right]\delta(\bi{\rho }_1\!-\!\bi{\rho }_2)\hat{1},
\end{align}
with $\nabla_{\rm 2D}^2=\partial_{x_1}^2+\partial_{y_1}^2$.
The 2D form of the Fermi energy $E_{\rm F}^{2{\rm D}}(k_z)=(\hbar^2/2m)(k_{\rm F}^2-k_z^2)$ reflects the $k_z$-cross section of the Fermi surface, where $k_{\rm F}$ is the Fermi wave number.
Although the OP $\hat{\Delta }_{k_z}(\bi{\rho }_1,\bi{\rho }_2)$ should be calculated self-consistently with the quasiparticle wave function $\vec{u}_{\nu,k_z}({\bm \rho_i})$ and the eigenvalue $E_{\nu,k_z}$, an approximate solution can be derived from the quasiclassical theory for superfluid states with a small gap $\Delta/E_{\rm F}\ll 1$.
We expand the OP to the Fourier integral with the relative coordinate ${\bm \rho }'={\bm \rho }_1-{\bm \rho }_2$ as
\begin{align}\label{fourier}
\hat{\Delta }_{k_z}(\bi{\rho }_1,\bi{\rho }_2)=\int\frac{d\bi{k}^{\rm 2D}}{(2\pi)^2}e^{i\bi{k}^{\rm 2D}\cdot\bi{\rho }'}\hat{\Delta }(\bi{k},\bi{\rho })\Gamma(k),
\end{align}
where ${\bm \rho }=({\bm \rho }_1+{\bm \rho }_2)/2$ is the center-of-mass coordinate and ${\bm k}^{\rm 2D}=(k_x,k_y)$ is the 2D component of the relative momentum ${\bm k}$.
In this model, the Gaussian factor $\Gamma(k)\!=\!e^{-(k^2-k_{\rm F}^2)\xi_{\rm p}^2}$ indicates that the pairing interaction is non-zero near the Fermi surface $k=k_{\rm F}$ with the range of the interaction $\xi_{\rm p}$~\cite{mizushima:2010b}.
The OP $\hat{\Delta }({\bm k},{\bm \rho })$ is obtained by the quasiclassical theory in Sec.~II B.
If the $k_z$-resolved quasiparticle wave function $\vec{u}_{\nu,k_z}=\left( u_{\nu,k_z}^{\uparrow }, u_{\nu,k_z}^{\downarrow }, v_{\nu,-k_z}^{\uparrow }, v_{\nu,-k_z}^{\downarrow }\right)^{\rm T}$ satisfies the condition $u_{\nu,k_z}^{\uparrow }=\left(v_{\nu,-k_z}^{\uparrow }\right)^*$ and $u_{\nu,k_z}^{\downarrow }=\left(v_{\nu,-k_z}^{\downarrow }\right)^*$, the creation and annihilation operators are equivalent for the zero energy quasiparticle with $k_z=0$~\cite{kawakami:2011,sato:2014}.
Then, the quasiparticle is a Majorana zero mode.
\subsection{Quasiclassical theory}
The spatial structure of OP with a vortex has been calculated by the quasiclassical theory, which is valid for superfluids and superconductors with $\Delta\ll E_F$, such as the superfluid $^3$He.
We have found self-consistent solutions of the OP $\hat{\Delta }(\bar{\bm k},{\bm \rho })$ with the quasiclassical Green's function $\widehat{g}(\bar{\bi{k}},\bi{\rho },\omega_n)$ by the Eilenberger equation~\cite{eilenberger:1968,serene:1983,fogelstrom:1995}
\begin{multline}
-i\hbar\bi{v}(\bar{\bi{k}})\cdot\bi{\nabla }\widehat{g}(\bar{\bi{k}},\bi{\rho },\omega_n)\\
= \left[
\begin{pmatrix}
i\omega_n\hat{1} & -\hat{\Delta }(\bar{\bi{k}},\bi{\rho }) \\
\hat{\Delta }^{\dagger }(\bar{\bi{k}},\bi{\rho }) & -i\omega_n\hat{1}
\end{pmatrix}
,\widehat{g}(\bar{\bi{k}},\bi{\rho },\omega_n) \right],
\label{Eilenberger eq}
\end{multline}
where the wide hat indicates the 4 $\times$ 4 matrix in particle-hole and spin spaces.
The quasiclassical Green's function is described in particle-hole space by
\begin{align}
\widehat{g}(\bar{\bi{k}},\bi{\rho },\omega_n) = -i\pi
\begin{pmatrix}
\hat{g}(\bar{\bi{k}},\bi{\rho },\omega_n) & i\hat{f}(\bar{\bi{k}},\bi{\rho },\omega_n) \\
-i\underline{\hat{f}}(\bar{\bi{k}},\bi{\rho },\omega_n) & -\underline{\hat{g}}(\bar{\bi{k}},\bi{\rho },\omega_n)
\end{pmatrix},
\end{align}
where $\omega_n=(2n+1)\pi k_B T$ is the Matsubara frequency and $\bar{\bm k}$ is the normalized relative momentum on the Fermi surface.
The quasiclassical Green's function satisfies the normalization condition $\widehat{g}^2=-\pi^2\widehat{1}$.
The Fermi velocity is given as $\bi{v}(\bar{\bi{k}})=v_F\bar{\bi{k}}$ on a three-dimensional Fermi sphere.
The spin-triplet OP is defined by
\begin{align}
\hat{\Delta }(\bar{\bi{k}},\bi{\rho })=i\bi{d}(\bar{\bi{k}},\bi{\rho })\cdot\hat{\bi{\sigma }}\hat{\sigma_y},
\label{OP}
\end{align}
where $\hat{\bi{\sigma }}$ is the Pauli matrix.
The $d$-vector is perpendicular to the spin $\bi{S}$ of a Cooper pair, namely, $\bi{d}\cdot\bi{S}=0$.
The description using projections of spin angular momentum is more convenient than the description by the $d$-vector for the OP with an axisymmetric vortex, namely,
\begin{align}
\hat{\Delta }(\bar{\bi{k}},\bi{\rho })=
\begin{pmatrix}
-\sqrt{2}C_+(\bar{\bi{k}},\bi{\rho }) & C_0(\bar{\bi{k}},\bi{\rho }) \\
C_0(\bar{\bi{k}},\bi{\rho }) & \sqrt{2}C_-(\bar{\bi{k}},\bi{\rho })
\end{pmatrix},
\end{align}
where $C_{\pm }=(d_x\mp id_y)/\sqrt{2}$ and $C_0=d_z$.
Each coefficient can be expanded in projections of orbital angular momentum,
\begin{multline}
C_a(\bar{\bi{k}},\bi{\rho })=C_{a+}(\bi{\rho })\bar{k}_+ + C_{a0}(\bi{\rho })\bar{k}_0 + C_{a-}(\bi{\rho })\bar{k}_-,
\label{eq:C}
\end{multline}
with $a=0$ or $\pm$, where $\bar{k}_{\pm }=(\bar{k}_x\pm i\bar{k}_y)/\sqrt{2}$ and $\bar{k}_0=\bar{k}_z$~\cite{salomaa:1987}.
The self-consistent condition for $\hat{\Delta }$ is given as
\begin{align}
\hat{\Delta }(\bar{\bi{k}},\bi{\rho }) = N_0\pi k_BT\sum_{|\omega_n| \le \omega_{\rm c}}\left\langle V(\bar{\bi{k}}, \bar{\bi{k}}') \hat{f}(\bar{\bi{k}}',\bi{\rho },\omega_n)\right\rangle_{\bar{\bi{k}}'},
\label{order parameter}
\end{align}
where $N_0$ is the density of states in the normal state,
$\omega_{\rm c}$ is a cutoff energy setting $\omega_{\rm c}=20k_B T_{\rm c}$ with the critical temperature $T_{\rm c}$,
and $\langle\cdots\rangle_{\bar{\bi{k}}}$ indicates the Fermi surface average.
The pairing interaction $V(\bar{\bi{k}}, \bar{\bi{k}}')=3g\bar{\bi{k}}\cdot\bar{\bi{k}}'$, where $g$ is a coupling constant with the relation $(gN_0)^{-1}=\ln(T/T_c)+\pi k_BT\sum_{|\omega_n| \le \omega_c}|\omega_n|^{-1}$.
We solve Eq.~\eqref{Eilenberger eq} and Eq.~\eqref{order parameter} alternately at $T=0.2T_{\rm c}$, and obtain a self-consistent solution.
Then, we use the self-consistent OP $\hat{\Delta }(\bar{\bm k},{\bm \rho })$ after the replacement of $\bar{\bm k}$ by ${\bm k}/k_{\rm F}$ as $\hat{\Delta }({\bm k},{\bm \rho })$ in Eq.~\eqref{fourier}.
\section{Spatial symmetries for an axisymmetric vortex}
For the OP with an axisymmetric vortex, the coefficients in Eq.~\eqref{eq:C} are described by $C_{ab}(\bm{\rho })=C_{ab}(\rho)e^{in_{ab}\phi}$, where $\rho$ is a radial distance from a vortex core and $\phi$ is an azimuthal angle in the $xy$-plane.
Since the OP has $U(1)_Q$ symmetry with $Q=L_z+S_z-N/2$ as Eq.~\eqref{eq:sym_vortex}, the phase winding number $n_{ab}$ in each coefficient satisfies
\begin{equation}
\begin{split}
n_{a+1,b}=&n_{a,b}-1,\\
n_{a,b+1}=&n_{a,b}-1.
\end{split}
\end{equation}
The following coefficients are finite in the bulk B-phase:
\begin{align}
C_{+-}=C_{00}=C_{-+}=\Delta_{\rm B},
\end{align}
where $\Delta_{\rm B}$ is the amplitude of a gap in the bulk, because the total angular momentum ${\bm J}={\bm L}+{\bm S}={\bm 0}$ in the B-phase~\cite{vollhardt:book}.
With a singly quantized vortex, the OP has $2\pi$-phase winding around the vortex, namely, phase winding numbers
\begin{align}
n_{+-}=n_{00}=n_{-+}=1.
\end{align}
Then, the general description of the coefficients with a singly quantized axisymmetric vortex in the superfluid $^3$He B-phase is
\begin{align}
C(\bi{\rho })=
\begin{pmatrix}
C_{++}(\rho)e^{-i\phi } & C_{+0}(\rho) & C_{+-}(\rho)e^{i\phi } \\
C_{0+}(\rho) & C_{00}(\rho)e^{i\phi } & C_{0-}(\rho)e^{2i\phi } \\
C_{-+}(\rho)e^{i\phi } & C_{-0}(\rho)e^{2i\phi } & C_{--}(\rho)e^{3i\phi }
\end{pmatrix},
\label{eq:axivortex}
\end{align}
where $C_{ab}$ approaches $\Delta_{\rm B}$ for $a+b=0$ and vanishes for $a+b\neq 0$ when $\rho\to\infty$.
The B-phase with a vortex can have the additional symmetry $P$ as Eq.~\eqref{eq:sym_vortex}.
Under the symmetry operation, the semi-classical BdG Hamiltonian
\begin{align}
\widehat{H}_{\rm BdG}({\bm k},{\bm \rho})=
\begin{pmatrix}
\hat{h}(\bm k) & \hat{\Delta }({\bm k},{\bm \rho}) \\
\hat{\Delta }^{\dagger }({\bm k},{\bm \rho}) & -\hat{h}^{\rm T}(-{\bm k})
\end{pmatrix},
\end{align}
satisfies
\begin{align}
\widehat{\mathcal{P}}\widehat{H}_{\rm BdG}({\bm k},{\bm \rho})\widehat{\mathcal{P}}^{-1}=\widehat{H}_{\rm BdG}({\bm k}',{\bm \rho}'),
\end{align}
where the symmetry operator is $\widehat{\mathcal{P}}={\rm diag}(\hat{\mathcal{P}},\hat{\mathcal{P}}^*)$ and ${\bm k}$ and ${\bm \rho}$ are transformed into ${\bm k}'$ and ${\bm \rho}'$, respectively, by the operation.
The normal state Hamiltonian $\hat{h}({\bm k})=(\hbar^2/2m)({\bm k}^2-k_{\rm F}^2)\hat{1}$ implies $\hat{\mathcal{P}}\hat{h}({\bm k})\hat{\mathcal{P}}^{-1}=\hat{h}({\bm k})$.
Since the operation $\hat{\mathcal{P}}\hat{\Delta }({\bm k},{\bm \rho})\left(\hat{\mathcal{P}}^*\right)^{-1}$ can be regarded as $\mathcal{P}C({\bm k},{\bm \rho})\left(\mathcal{P}^*\right)^{-1}$, the inversion symmetry operator $\mathcal{P}_1$ acts on $C_a({\bm k},\rho,\phi)$ as
\begin{multline}
\mathcal{P}_1C({\bm k},\rho,\phi)\left(\mathcal{P}_1^*\right)^{-1}=C(-{\bm k},\rho,\phi+\pi)\\
=\begin{pmatrix}
C_{++}(\rho)e^{-i\phi } & -C_{+0}(\rho) & C_{+-}(\rho)e^{i\phi } \\
-C_{0+}(\rho) & C_{00}(\rho)e^{i\phi } & -C_{0-}(\rho)e^{2i\phi } \\
C_{-+}(\rho)e^{i\phi } & -C_{-0}(\rho)e^{2i\phi } & C_{--}(\rho)e^{3i\phi }
\end{pmatrix}
\begin{pmatrix}
k_+ \\ k_0 \\ k_-
\end{pmatrix}.
\label{eq:p1sym}
\end{multline}
The symmetry $P_2$ is the combined symmetry with the time-reversal and mirror reflection on a plane including the vortex line, namely the magnetic reflection symmetry.
The time-reversal operator $\mathcal{T}$ acts as $\mathcal{T}C_{a,b}(\rho,\phi)\left(\mathcal{T}^*\right)^{-1}=-C_{-a,-b}^*(\rho,\phi)$ and the mirror reflection operator $\mathcal{M}$ acts as $\mathcal{M}C_{a,b}(\rho,\phi)\left(\mathcal{M}^*\right)^{-1}=-C_{-a,-b}(\rho,-\phi)$; therefore,
\begin{multline}
\mathcal{P}_2C({\bm k},\rho,\phi)\left(\mathcal{P}_2^*\right)^{-1}\\
=\begin{pmatrix}
C_{++}^*(\rho)e^{-i\phi } & C_{+0}^*(\rho) & C_{+-}^*(\rho)e^{i\phi } \\
C_{0+}^*(\rho) & C_{00}^*(\rho)e^{i\phi } & C_{0-}^*(\rho)e^{2i\phi } \\
C_{-+}^*(\rho)e^{i\phi } & C_{-0}^*(\rho)e^{2i\phi } & C_{--}^*(\rho)e^{3i\phi }
\end{pmatrix}
\begin{pmatrix}
k_+ \\ k_0 \\ k_-
\end{pmatrix}.
\label{eq:p2sym}
\end{multline}
The symmetry $P_3$ is the magnetic $\pi$-rotation symmetry around an axis perpendicular to the vortex line.
The $\pi$-rotation operator $\mathcal{R}$ acts as $\mathcal{R}C_{a,b}(\rho,\phi)\left(\mathcal{R}^*\right)^{-1}=C_{-a,-b}(\rho,\pi -\phi)$.
Then, the combination of $\mathcal{T}$ and $\mathcal{R}$ gives
\begin{multline}
\mathcal{P}_3C({\bm k},\rho,\phi)\left(\mathcal{P}_3^*\right)^{-1}\\
=\begin{pmatrix}
C_{++}^*(\rho)e^{-i\phi } & -C_{+0}^*(\rho) & C_{+-}^*(\rho)e^{i\phi } \\
-C_{0+}^*(\rho) & C_{00}^*(\rho)e^{i\phi } & -C_{0-}^*(\rho)e^{2i\phi } \\
C_{-+}^*(\rho)e^{i\phi } & -C_{-0}^*(\rho)e^{2i\phi } & C_{--}^*(\rho)e^{3i\phi }
\end{pmatrix}
\begin{pmatrix}
k_+ \\ k_0 \\ k_-
\end{pmatrix}.
\label{eq:p3sym}
\end{multline}
\section{$o$ vortex}
\label{sec:o_vortex}
The $o$ vortex is the most symmetric vortex in the superfluid $^3$He B-phase.
Since it has the all possible discrete symmetries $P_1$, $P_2$, and $P_3$, several coefficients should vanish as
\begin{align}
C(\bi{\rho })=
\begin{pmatrix}
C_{++}(\rho)e^{-i\phi } & 0 & C_{+-}(\rho)e^{i\phi } \\
0 & C_{00}(\rho)e^{i\phi } & 0 \\
C_{-+}(\rho)e^{i\phi } & 0 & C_{--}(\rho)e^{3i\phi }
\end{pmatrix},
\end{align}
where all remaining coefficients $C_{ab}(\rho)$ are real and they approach $C_{+-}=C_{00}=C_{-+}=\Delta_{\rm B}$ and $C_{++}=C_{--}=0$ when $\rho\rightarrow\infty$.
The component $C_{++}$ ($C_{--}$) is induced around the vortex by the spatial variation of the bulk component $C_{+-}$ ($C_{-+}$)~\cite{ohmi:1983}.
The self-consistently obtained OP by the quasiclassical theory is shown in Fig.~\ref{ovor}(a).
The bulk components $C_{+-}$, $C_{00}$, and $C_{-+}$ rise up with $\rho$-linear from a vortex core, where $C_{00}$ recovers the bulk gap $\Delta_{\rm B}$ with slightly shorter length than $C_{+-}$ and $C_{-+}$.
Rises of $C_{++}$ and $C_{--}$ are $\rho$-linear and $\rho$-cubic, respectively, which reflect phase winding numbers $n_{++}=-1$ and $n_{--}=3$.
This difference makes a little variance between $C_{+-}$ and $C_{-+}$ via a coupling between the same orbital state.
The discretized eigenvalues of the vortex bound state, which is derived from the BdG theory with the self-consistent OP, are shown in Figs.~\ref{ovor}(d) and \ref{ovor}(e).
In Fig.~\ref{ovor}(d), the eigenvalues for $k_z=0$ are classified into quantized orbital angular momentum $l$.
The low energy eigenvalues are discretized at the order of $\Delta_{\rm B}^2/E_{\rm F}$ intervals, where we take $\Delta_{\rm B}/E_{\rm F}=0.1$.
The vortex bound state has spin degenerate exact two zero energy modes at $l=0$ indicated by an arrow.
The wave functions of the degenerate zero energy modes for $\uparrow$-spin and $\downarrow$-spin states are shown in Figs.~\ref{ovor}(b) and \ref{ovor}(c), respectively.
Since the orbital chirality of the $\uparrow$-spin state is antiparallel to the vorticity but the orbital chirality of the $\downarrow$-spin state is parallel to the vorticity, $u_0^{\uparrow }({\bm\rho})=\left[v_0^{\uparrow }({\bm\rho})\right]^*\propto J_0(k_{\rm F}\rho)\exp(-\rho/\xi)$ and $u_0^{\downarrow }({\bm\rho})=\left[v_0^{\downarrow }({\bm\rho})\right]^*\propto J_1(k_{\rm F}\rho)\exp(-\rho/\xi)e^{i\phi }$~\cite{gurarie:2007,mizushima:2010}, where $J_n$ is the $n$-th order Bessel function and $(k_{\rm F}\xi)^{-1}=\Delta_{\rm B}/E_{\rm F}=0.1$.
These zero energy modes are Majorana zero modes because the quasiparticle wave functions for both spin states satisfy the condition $u_0^{\uparrow }=\left(v_0^{\uparrow }\right)^*$ and $u_0^{\downarrow }=\left(v_0^{\downarrow }\right)^*$, as demonstrated in Fig.~\ref{ovor}(b) and \ref{ovor}(c).
The $k_z$-dispersion of the vortex bound state for various $l$'s is shown in Fig.~\ref{ovor}(e), where two different spin states exist for each $l$.
As demonstrated in Appendix~\ref{app:ovortex}, the $l=0$ branches linearly cross the zero energy at $k_z=0$ as $E_{\pm}\propto \pm k_z$.
Their wave functions are distinguished into $\vec{u}_{\pm}=\left(\vec{u}_0^{\uparrow}\pm\vec{u}_0^{\downarrow}\right)/\sqrt{2} + \vec{\mathcal{O}}(k_z/k_{\rm F})$, where $\vec{u}_0^{\uparrow}=\left(u_0^{\uparrow},0,\left[u_0^{\uparrow}\right]^*,0\right)^{\rm T}$ and $\vec{u}_0^{\downarrow}=\left(0,u_0^{\downarrow},0,\left[u_0^{\downarrow}\right]^*\right)^{\rm T}$ are the wave functions of the Majorana zero modes in the $\uparrow$-spin and $\downarrow$-spin states, respectively.
Note that the inversion symmetry $P_1$ maps each eigenstate to another one, namely, $\vec{u}_{\pm}=\widehat{\mathcal{P}}_1\vec{u}_{\mp}$.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{fig1.eps}
\end{center}
\caption{\label{ovor}(Color online) Numerical results for $o$ vortex.
(a) OP configuration $C_{ab}(\rho)$ for finite components $C_{+-}$, $C_{00}$, $C_{-+}$, $C_{++}$, and $C_{--}$.
Real part of wave functions $u^{\sigma}_0(\rho)$ and $v^{\sigma}_0(\rho)$ with $E_{\nu,k_z}=0$ for the $\uparrow$-spin state (b) and $\downarrow$-spin state (c).
(d) Eigenvalues for quasiparticles with $k_z=0$. The degenerate eigenstates indicated by an arrow correspond to the two wave functions in (b) and (c).
(e) $k_z$-dispersion of the vortex bound state whose color denotes angular momentum $l$ of quasiparticles.
}
\end{figure}
The spin degenerate Majorana zero modes are protected by spatial symmetry.
The vortex bound state with $P_2$ and $P_3$ symmetries is classified by a topological invariant $\mathbb{Z}$ which is demonstrated in Appendix~\ref{app:clifford} by using Clifford algebras.
Note that $P_1$ symmetry does not provide symmetry protected defect zero modes in general.
The appropriate topological invariant for $o$ vortex is calculated by
the one-dimensional winding number, which is evaluated from the BdG
Hamiltonian at an infinite point from a vortex
\begin{align}
\widehat{H}_{\rm BdG}({\bm k},\phi)=
\begin{pmatrix}
\hat{h}({\bm k}) & \hat{\Delta }({\bm k},\phi) \\
\hat{\Delta }^{\dagger }({\bm k},\phi) & -\hat{h}^{\rm T}(-{\bm k})
\end{pmatrix},
\end{align}
where $\hat{h}({\bm k})=(\hbar^2/2m)({\bm k}^2-k_{\rm F}^2)\hat{1}$ and
\begin{align}
\hat{\Delta }({\bm k},\phi)=\frac{\Delta_{\rm B}}{k_{\rm F}}
\begin{pmatrix}
-k_x+ik_y & k_z \\
k_z & k_x+ik_y
\end{pmatrix}e^{i\phi }.\nn
\end{align}
From combination of $P_3$ and particle-hole symmetry,
the BdG Hamiltonian for $o$ vortex obeys
$(\widehat{\mathcal{C}}\widehat{\mathcal{P}}_3)\widehat{H}_{\rm
BdG}(k_x,-k_y,-k_z,-\phi)(\widehat{\mathcal{C}}\widehat{\mathcal{P}}_3)^{-1}=-\widehat{H}_{\rm
BdG}({\bm k},\phi)$ where $\widehat{\mathcal{P}}_3=i\hat{\sigma}_z\widehat{\tau}_z\mathcal{K}$ and $\widehat{\mathcal{C}}=\widehat{\tau
}_x\mathcal{K}$ are operators for the magnetic $\pi$-rotation and
particle-hole symmetry with the complex conjugation operator
$\mathcal{K}$.
The combined symmetry defines the chiral symmetry,
$\widehat{\Gamma }\widehat{H}_{\rm
BdG}(k_x,k_y=0,k_z=0,\phi=0,\pi)\widehat{\Gamma }^{-1}=-\widehat{H}_{\rm
BdG}(k_x,k_y=0,k_z=0,\phi=0,\pi)$ with $\widehat{\Gamma }=\widehat{\tau
}_y\hat{\sigma }_z$, in the symmetric space $k_y=k_z=0$
and $\phi=0$ or $\pi$.
The chiral symmetry enables us to introduce the one-dimensional winding number as~\cite{sato:2009c,sato:2011}
\begin{widetext}
\begin{align}
w^{\phi=0,\pi }=-\frac{1}{4\pi i}\int dk_x{\rm tr}\left[\widehat{\Gamma }\widehat{H}^{-1}_{\rm BdG}(k_x,k_y=0,k_z=0,\phi=0,\pi)\partial_{k_x}\widehat{H}_{\rm BdG}(k_x,k_y=0,k_z=0,\phi=0,\pi)\right].
\label{eq:1dwinding}
\end{align}
\end{widetext}
For $o$ vortex, the one-dimensional winding number is evaluated as
$w^0=2$ and $w^{\pi }=-2$.
The difference of the winding $(w^0-w^{\pi })/2=2$ provides
the $\mathbb{Z}$ topological invariant for $o$ vortex.
This topological number corresponds to the number of the zero energy states at
$k_z=0$. (More precisely, the $\mathbb{Z}$ topological invariant is
equal to the index ${\rm tr}\widehat{\Gamma}$ of the quasiparticle states
at $k_z=0$. Since the chiral symmetry requires that the zero energy states at
$k_z=0$ are eigenstates of $\widehat{\Gamma}$, the index ${\rm
tr}\widehat{\Gamma}$ reduces to the difference between the number of
the zero energy states with eigenvalue $\widehat{\Gamma}=+1$ and that
with $\widehat{\Gamma}=-1$~\cite{sato:2011}. Hence, if
$(w^0-w^{\pi})/2=N$, there exist at least $|N|$ zero energy states at $k_z=0$.)
The obtained topological number, i.e. $(w^0-w^\pi)/2=2$, guarantees the
existence of two zero energy states at $k_z=0$, which are indeed realized
as the two $l=0$ zero modes in Fig.~\ref{ovor}.
Note that the second Chern number, which characterizes bound states in a line defect for the symmetry class D, vanishes in the presence of $P_1$ symmetry.
The second Chern number ${\rm Ch}_2$ is obtained by~\cite{qi:2008, teo:2010}
\begin{align}
{\rm Ch}_2=\frac{1}{8\pi^2}\int d{\bm k}d\phi\epsilon^{ijk}{\rm tr}[f_{\phi i}f_{jk}],
\label{eq:secondCh}
\end{align}
where $f_{\alpha\beta}^{mn}=\partial_{\alpha}a_{\beta}^{mn}
-\partial_{\beta}a_{\alpha}^{mn}+i[a_{\alpha},a_{\beta}]^{mn}$ is the
curvature of the non-Abelian Berry connection $a_\alpha$ with
$\alpha=(\phi, i)=(\phi,k_x,k_y,k_z)$, and
the non-Abelian Berry connection is given by
\begin{align}
a_{\alpha}^{mn}=-i\langle m,\bi{k},\phi|\partial_{\alpha}|n,\bi{k},\phi\rangle,
\end{align}
with eigenstates $|m, \bi{k}, \phi\rangle$ and $|n, \bi{k}, \phi\rangle$
of the BdG Hamiltonian $\widehat{H}_{\rm BdG}({\bm k},\phi)$.
The indices $m$ and $n$ label quasiparticle states with negative energies.
For $o$ vortex, the eigenstates have $P_1$ symmetry which gives the periodicity of Berry curvature, $f_{\phi i}({\bm k},\phi+\pi)=-f_{\phi i}(-{\bm k},\phi)$ and $f_{jk}({\bm k},\phi+\pi)=f_{jk}(-{\bm k},\phi)$.
Therefore, the integral in Eq.~\eqref{eq:secondCh} yields ${\rm Ch}_2=0$.
Since the second Chern number generally vanishes when there is $P_1$ symmetry, it is not appropriate for the topological number of the bound states in $o$ vortex.
\section{$v$ vortex}
In the actual B-phase, it is believed that two kinds of $v$ vortices are realized.
One $v$ vortex has axial symmetry and the other $v$ vortex breaks the axial symmetry.
In this section, we show the difference of the vortex bound states between the two kinds of $v$ vortices.
\subsection{Axisymmetric $v$ vortex}
The axisymmetric $v$ vortex has the magnetic reflection symmetry $P_2$.
From Eq.~\eqref{eq:p2sym}, all coefficients $C_{ab}(\rho)$ in Eq.~\eqref{eq:axivortex} are real, where $C_{+-}=C_{00}=C_{-+}=\Delta_{\rm B}$ and the other coefficients vanish when $\rho\rightarrow\infty$.
Since $C_{0+}$ and $C_{+0}$ corresponding to the A-phase and $\beta$-phase components, respectively, do not have any phase winding, they can compensate a vortex core.
The self-consistently obtained OP by the quasiclassical theory is shown in Fig.~\ref{vvor}(a).
The A-phase component $C_{0+}$ compensates the vortex core with larger amplitude than that of the bulk components in the B-phase.
The $\beta$-phase component $C_{+0}$ also compensates the vortex core; however, the amplitude is smaller than that of $C_{0+}$ and the sign is opposite.
Away from the vortex core, $C_{0+}$ and $C_{+0}$ are identical to $C_{0-}$ and $C_{-0}$, respectively, in order that total angular momentum becomes zero.
A core radius $\xi_{\rm c}$ of the $v$ vortex, which is characterized by the healing length of the bulk components, is longer than that of $o$ vortex because a loss of the condensation energy at the $v$ vortex core is small.
Other $C_{++}$ and $C_{--}$ components are also induced slightly, which change the sign away from the vortex core.
The discretized eigenvalues derived from the BdG equation by using the self-consistent OP are shown in Figs.~\ref{vvor}(d) and \ref{vvor}(e).
In Fig.~\ref{vvor}(d), the eigenvalues for $k_z=0$ are classified into quantized orbital angular momentum $l$ along the vortex line.
The quasiparticles with $l=0$ and $k_z=0$ have the finite energy by the $\pi/2$-phase shift owing to the A-phase component compensating the vortex core [see Appendix~\ref{app:shift}].
Discretized eigenvalues for $k_z=0$ may be situated on the zero energy with finite $\pm l$, accidentally, according to a core radius.
The wave function of an accidental zero energy mode, indicated by an arrow in Fig.~\ref{vvor}(d), is shown in Figs.~\ref{vvor}(b) and \ref{vvor}(c) for $\uparrow$-spin and $\downarrow$-spin parts, respectively.
The quasiparticle wave function satisfies $u_{\nu,k_z=0}^{\uparrow }=\left(v_{\nu,k_z=0}^{\uparrow }\right)^*$ while $u_{\nu,k_z=0}^{\downarrow }\ne\left(v_{\nu,k_z=0}^{\downarrow }\right)^*$;
therefore, the quasiparticle is not the Majorana zero mode~\cite{kawakami:2011,sato:2014}.
It is a natural consequence because the particle-hole operation changes the sign of orbital angular momentum.
The $k_z$-dispersion of the vortex bound state for $l\ge 0$ is shown in Fig.~\ref{vvor}(e).
Eigenvalues for $l'=-l\le 0$ are given by $E'=-E$ with $k_z'=-k_z$ owing to the particle-hole symmetry.
The eigenvalues with $l=0$ approach the zero energy at $k_z=\pm k_F$ owing to an induced $C_{-0}$ component around the vortex, which is different from the result by Silaev~\cite{silaev:2009} [see Appendix~\ref{app:shift}].
Since momentum $k_z$ is a continuous quantity, the eigenvalues with $|l|\lesssim\xi_{\rm c}k_{\rm F}$ cross the zero energy at finite $k_z$.
The zero energy modes with finite $k_z$ are, however, also not Majorana zero modes because signs of $l$ and $k_z$ for the zero energy modes are changed by the particle-hole operation, that is, the zero energy modes are mapped to different zero energy modes by the particle-hole operation.
The trivial vortex bound state without Majorana zero modes can be understood by the topological arguments in Appendix~\ref{app:clifford}.
The vortex bound states without any symmetry protections are clearly shown as gaps of low energy excitations in non-axisymmetric $v$ vortex which breaks axial symmetry but keeps $P_2$ symmetry.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{fig2.eps}
\end{center}
\caption{\label{vvor}(Color online) Numerical results for axisymmetric $v$ vortex.
(a) OP configuration $C_{ab}(\rho)$ whose legends are identical to that in Fig.~\ref{ovor}(a).
(b) and (c) Real part of wave function $u^{\sigma}_{\nu,k_z=0}(\rho)$ and $v^{\sigma}_{\nu,k_z=0}(\rho)$ for a state with $E_{\nu,k_z=0}\approx0$.
(d) Eigenvalues for quasiparticles with $k_z=0$. The eigenstate indicated by an arrow corresponds to the wave function in (b) and (c).
(e) $k_z$-dispersion of quasiparticles with $l\ge0$ whose color denotes angular momentum $l$ of quasiparticles.
}
\end{figure}
\subsection{Non-axisymmetric $v$ vortex}
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{fig3.eps}
\end{center}
\caption{\label{dvor}(Color online) Numerical results for non-axisymmetric $v$ vortex.
2D profile for each OP component $C_{ab}(\bm{\rho})$ (a) and the root-mean-square value of the gap, $|\Delta|=\sqrt{\langle{\rm Tr}[\hat{\Delta }\hat{\Delta }^{\dagger }]\rangle_{\bar{\bi{k}}}/2}$, (b).
(c) Eigenvalues for quasiparticles with $k_z=0$.
(d) $k_z$-dispersion of the vortex bound state whose color denotes angular momentum $l$ of quasiparticles.
}
\end{figure}
The self-consistently obtained OP for non-axisymmetric $v$ vortex is shown in Figs.~\ref{dvor}(a) and \ref{dvor}(b).
Far away from the vortex core, non-axisymmetric $v$ vortex is also described by the coefficients in Eq.~\eqref{eq:axivortex}; however, the original $4\pi$-phase singularity of $C_{-0}$ and $C_{0-}$ is split into two singularities with the $2\pi$-phase winding as shown in Fig.~\ref{dvor}(a).
This deformation breaks $U(1)_Q$ symmetry but keeps $P_2$ symmetry.
In Fig.~\ref{dvor}(b), we show the root-mean-square value of the gap, $|\Delta|=\sqrt{\langle{\rm Tr}[\hat{\Delta }\hat{\Delta }^{\dagger }]\rangle_{\bar{\bi{k}}}/2}$.
Since four components $C_{0+}$, $C_{+0}$, $C_{-0}$, and $C_{0-}$ are finite at the vortex center, substantial gap opens on the vortex center.
In return for the gap on the vortex center, finite minima of $|\Delta|$ appear on the phase singularity of $C_{-0}$ and $C_{0-}$.
The discretized eigenvalues derived from the BdG equation by using the self-consistent OP are shown in Figs.~\ref{dvor}(c) and \ref{dvor}(d).
In Fig.~\ref{dvor}(c), the eigenvalues for $k_z=0$ are classified into orbital angular momentum $l$ along the vortex line.
The $k_z$-dispersion of the vortex bound state with values of $l$ is shown in Fig.~\ref{dvor}(d).
Note that $l$ is not a quantum number but it is calculated by
\begin{align}
l=-i\hbar\int d\bi{\rho }\vec{u}_{\nu,k_z}^{\dagger }({\bm \rho})\partial_{\phi }\vec{u}_{\nu,k_z}({\bm \rho}),
\label{eq:ang_moment}
\end{align}
for each eigenstate labeled $\nu$ and $k_z$.
The quasiparticle excitations in the vortex bound state have a gap by the hybridization of different $l$ eigenstates as discussed below.
This is a consequence of that $P_2$ symmetry does not protect zero energy modes in a line defect.
The amplitude of the excitation gap is of the order of $\Delta/E_{\rm F}$.
Therefore, the gap becomes larger when $k_z$ approach $k_{\rm F}$ because the effective Fermi energy $E_{\rm F}^{\rm 2D}=(\hbar^2/2m)(k_{\rm F}^2-k_z^2)$ decreases.
Here, let us consider the hybridization of different $l$ eigenstates for non-axisymmetric $v$ vortex.
For axisymmetric $v$ vortex, since angular momentum $l$ is a well-defined quantum number, different $l$ eigenstates do not hybridize each other.
The reduction of axial symmetry to $n$-fold rotational symmetry, however, gives hybridization between the states with angular momentum $l$ and $l+nm$, where $m\in\mathbb{Z}$.
Then, $l$ and $l+2m$ eigenstates for non-axisymmetric $v$ vortex, which has two-fold rotational symmetry, are hybridized.
For the $k_z$-dispersion of the original $v$ vortex bound state, branches of $l$ and $-l$ eigenstates are crossed on the zero energy, where $l\ge 0$ eigenstates are shown in Fig.~\ref{vvor}(e) and $l'=-l$ eigenstates have eigenvalues $E'=-E$.
Since the $l$ and $-l$ eigenstates are hybridized for non-axisymmetric $v$ vortex, the quasiparticle excitations have a gap.
\section{$w$ vortex}
\label{sec:w_vortex}
The $w$ vortex with the magnetic $\pi$-rotation symmetry $P_3$ is described by the coefficients $C_{ab}(\rho)$ in Eq.~\eqref{eq:axivortex} as well as $v$ vortex.
However, the coefficients $C_{+0}$, $C_{0+}$, $C_{0-}$, and $C_{-0}$ are pure imaginary numbers.
Thus, the relative phase between the compensate A-phase and $\beta$-phase components and the bulk B-phase component is $\pi/2$, which is the difference of the OP from $v$ vortex.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm]{fig4.eps}
\end{center}
\caption{\label{wvor}(Color online) Numerical results for $w$ vortex.
(a) and (b) Degenerate wave functions $u^{\sigma}_0(\rho)$ and $v^{\sigma}_0(\rho)$ with $E_{\nu,k_z}=0$.
(c) Eigenvalues for quasiparticles with $k_z=0$. The degenerate eigenstates indicated by an arrow correspond to the two wave functions in (a) and (b).
(d) $k_z$-dispersion of the vortex bound state whose color denotes angular momentum $l$ of quasiparticles.
(e) Schematic $k_z$-dispersion for $l=0$ quasiparticles, which is precisely degenerate near $k_z=0$.
}
\end{figure}
We calculate the vortex bound state for $w$ vortex by the BdG theory with the $v$ vortex OP shown in Fig.~\ref{vvor}(a) after changing the phases of $C_{+0}$, $C_{0+}$, $C_{0-}$, and $C_{-0}$ by $\pi/2$.
The obtained eigenvalues are shown in Figs.~\ref{wvor}(c) and \ref{wvor}(d).
In Fig.~\ref{wvor}(c), the eigenvalues for $k_z=0$ are classified into quantized orbital angular momentum $l$ along the vortex line.
The vortex bound state has degenerate two exact zero energy modes at $l=0$ indicated by an arrow.
The $k_z$-dispersion of the vortex bound state for various $l$'s is shown in Fig.~\ref{wvor}(d).
The quasiparticles with $l=0$ are degenerate in small $k_z$ and linearly cross the zero energy at $k_z=0$.
The $l=0$ branches cross the zero energy twice at $k_z=0$ and $|k_z|\sim k_{\rm F}$ schematically shown in Fig.~\ref{wvor}(e), which is consistent with the value of second Chern number discussed below.
The wave functions of the degenerate zero energy modes indicated by the arrow in Fig.~\ref{wvor}(c) are shown in Figs.~\ref{wvor}(a) and \ref{wvor}(b).
The quasiparticle wave functions for each mode satisfy ${\rm Re}\!\left(u_0^{\uparrow }\right)={\rm Re}\!\left(v_0^{\uparrow }\right)$ and ${\rm Im}\!\left(u_0^{\downarrow }\right)=-{\rm Im}\!\left(v_0^{\downarrow }\right)$, namely, $u_0^{\uparrow }=\left(v_0^{\uparrow }\right)^*$ and $u_0^{\downarrow }=\left(v_0^{\downarrow }\right)^*$.
Therefore, the degenerate zero energy modes are the Majorana zero modes.
The existence of the Majorana zero modes can be understood by the topological arguments in Appendix~\ref{app:clifford}.
The additional magnetic $\pi$-rotation symmetry $P_3$ in the
Altland-Zrinbauer symmetry class D gives a topological classification
$\mathbb{Z}\oplus\mathbb{Z}$ for the vortex bound state in three-dimensional space~\cite{shiozaki:2014}.
One of the topological numbers $\mathbb{Z}$ is the second Chern number.
For $w$ vortex, the second Chern number is found to be zero as well as $o$ vortex
because the BdG Hamiltonian for $w$ vortex
is identical to that for $o$ vortex at infinity from the vortex.
The vanishing second Chern number is consistent with the obtained
$k_z$-dispersion of the vortex bound states in Fig.~\ref{wvor}(d):
Some of vortex bound states cross the zero
energy in the $k_z$-direction, but the crossing always occurs twice, so
they can smoothly
merge into the bulk state without closing bulk gap.
Therefore, the zero energy states are topologically unstable, except
for the $l=0$ bound state.
The topological stability of the $l=0$ bound state is ensured by the other
$\mathbb{Z}$ topological invariant, namely the index ${\rm tr}\widehat{\Gamma }$.
Since the chiral symmetry $\widehat{\Gamma }=\widehat{\mathcal{C}}\widehat{\mathcal{P}}_3$ is also defined in the symmetric space $k_y=k_z=0$ and $\phi=0$ or $\pi$ for $w$ vortex, the one-dimensional winding numbers $w^{\phi=0}$ and $w^{\phi=\pi }$ are evaluated as the same manner in Sec.~\ref{sec:o_vortex}.
The difference of the winding numbers $(w^0-w^{\pi })/2=2$ provides the $\mathbb{Z}$ topological invariant and guarantees the existence of two zero energy states at $k_z=0$, which are indeed realized as the two $l=0$ zero modes in Fig.~\ref{wvor}.
\section{Summary}
We have calculated the bound state in $o$ vortex, $v$ vortex, and $w$ vortex by the full quantum BdG theory with the self-consistent OP obtained by using the quasiclassical theory.
Moreover, we have discussed symmetry protection of zero energy excitations in the vortex bound states with additional symmetry.
Our results are summarized in the following and in Table I.
Characteristic features of the bound states in $u$ vortex and $uvw$ vortex are also listed in Table I.
The most symmetric $o$ vortex has $P_1$, $P_2$, and $P_3$ symmetries in which induced components around the vortex core are fixed for $C_{++}$ and $C_{--}$ in real numbers, and $C_{+0}$, $C_{0+}$, $C_{-0}$, and $C_{0-}$ as zero.
The vortex bound state for the quasiparticles with angular momentum $l=0$ has spin degenerate Majorana zero modes at $k_z=0$.
The Majorana zero modes are protected by $P_3$ symmetry and characterized by a topological invariant $\mathbb{Z}$ which is the chiral index $(w^0-w^{\pi})/2=2$.
\begin{table*}
\begin{center}
\begin{minipage}{15cm}
\caption{Classified vortices by $P_1$, $P_2$, and $P_3$ symmetries which fix the element of induced components $C_{ab}(\rho)$.
For each vortex, we show the presence or absence of zero energy modes (ZEM) and Majorana zero modes, as well as topological invariant classifying the vortex bound state if it is present.
The candidates for the topological invariant are the chiral index $N\equiv(w^0-w^{\pi})/2=2$ and the second Chern number ${\rm Ch}_2=0$.}
\label{tab}
\end{minipage}
\begin{tabular*}{15cm}{@{\extracolsep{\fill}}ccccccc}
\hline \hline
\rule{0pt}{2.5ex} \multirow{2}{*}{vortex} & \multirow{2}{*}{symmetry} & \multirow{2}{*}{$C_{++}$, $C_{--}$} & $C_{+0}$, $C_{0+}$, & \multirow{2}{*}{ZEM} & \multirow{2}{*}{Majorana} & \multirow{2}{*}{top.~inv.} \\
\rule{0pt}{2.5ex} & & & $C_{-0}$, $C_{0-}$ \\ \hline
\rule{0pt}{2.5ex} $o$ vortex & $P_1$, $P_2$, $P_3$ & Real & -- & $\checkmark$ & $\checkmark$ & $N$ \\
\rule{0pt}{2.5ex} axisym.~$v$ vortex & $P_2$ & Real & Real & $\checkmark$ & -- & -- \\
\rule{0pt}{2.5ex} non-axisym.~$v$ vortex & $P_2$ & Real & Real & -- & -- & -- \\
\rule{0pt}{2.5ex} $w$ vortex & $P_3$ & Real & Imaginary & $\checkmark$ & $\checkmark$ & $N\oplus{\rm Ch}_2$ \\
\rule{0pt}{2.5ex} $uvw$ vortex & -- & Complex & Complex & $\checkmark$ & -- & ${\rm Ch}_2$ \\
\rule{0pt}{2.5ex} $u$ vortex & $P_1$ & Complex & -- & $\checkmark$ & -- & -- \\
\hline \hline
\end{tabular*}
\end{center}
\end{table*}
The $v$ vortex has $P_2$ symmetry in which all induced components around the vortex core are real numbers.
In the vortex bound state, the quasiparticles with $|l|\lesssim\xi_{\rm c}k_{\rm F}$ cross the zero energy twice at finite $\pm k_z$.
However, the zero energy modes are topologically trivial, that is, they are not protected any symmetry.
Then, low energy excitations in non-axisymmetric $v$ vortex has a gap because the deformation breaking axial symmetry lifts the zero energy modes in spite of keeping $P_2$ symmetry.
If we observe the excitation gap accompanying the vortex transition in the B-phase, that will be a strong evidence of that the realized vortex has only $P_2$ symmetry, namely, $v$ vortex.
Although the order of the gap is $\Delta/E_{\rm F}$, that becomes large for quasiparticles with $k_z\sim k_{\rm F}$ owing to the small effective Fermi energy $E_{\rm F}^{\rm 2D}=(\hbar^2/2m)(k_{\rm F}^2-k_z^2)$.
The excitation gap may have been observed as the difference of critical velocity between V1 and V2 vortices~\cite{pekola:1984b,pekola:1985}.
The $w$ vortex has $P_3$ symmetry.
The difference between $w$ vortex and $v$ vortex is only phases of induced components $C_{+0}$, $C_{0+}$, $C_{-0}$, and $C_{0-}$ concerning the OP.
However, the vortex bound state in superfluids belonging to the symmetry class D with additional symmetry $P_3$ is characterized by the topological invariants $\mathbb{Z}\oplus\mathbb{Z}$, the second Chern number and the chiral index.
For $w$ vortex, the second Chern number ${\rm Ch}_2=0$, but the chiral index $(w^0-w^{\pi})/2=2$ indicating difference between the numbers of the zero energy mode with opposite chirality.
Thus, the two-fold degenerate Majorana zero modes at the symmetric point $k_z=0$ and $l=0$ are protected by $P_3$ symmetry.
The $uvw$ vortex without any $P_1$, $P_2$, and $P_3$ symmetries has many zero energy modes, as calculated by Silaev~\cite{silaev:2009}.
However, the zero energy modes are topologically unstable because the vortex bound state in superfluids belonging to the class D without additional symmetry is characterized by the second Chern number, which is zero for the vortex bound state in the superfluid $^3$He B-phase.
The $u$ vortex with $P_1$ symmetry is achieved by adding imaginary parts to real $C_{++}$ and $C_{--}$ components in $o$ vortex.
The imaginary parts lift the Majorana zero modes at the symmetric point $k_z=0$ and $l=0$ by breaking $P_3$ symmetry.
Other accidental zero modes will be left but they are not characterized by the second Chern number which generally vanishes under $P_1$ symmetry.
In conclusion, Majorana zero modes protected by $P_3$ symmetry are bound in $o$ vortex and $w$ vortex which are not stable in the bulk B-phase.
The $v$ vortex is more stable than the $o$ vortex owing to the condensation energy by A-phase and $\beta$-phase components which compensates the vortex core.
Then, when we confine the B-phase in a thin slab with hight along the vortex line to suppress the $\beta$-phase component and simultaneously apply a magnetic field along the vortex line to suppress the A-phase component, $o$ vortex with Majorana zero modes will be achieved.
Note that the confinement and the magnetic field do not break any $P_1$, $P_2$, and $P_3$ symmetries.
\section*{Acknowledgments}
We thank M.~Ichioka, T.~Mizushima, T.~Morimoto, A.~Furusaki, and S.~Kobayashi for helpful discussion.
A part of the numerical calculations was performed by using the RIKEN Integrated Cluster of Clusters (RICC) and ICE8200EX in NIMS.
This work was supported by KAKENHI (Nos. 24840048, 21340103, 2200247703, 25287085, and 22103005) and WPI Initiative on Material nanoarchtectonics, MEXT, Japan.
K.S.~is supported by a JSPS Fellowship for Young Scientists.
|
\section{\label{sec:intro}Introduction}
Advanced LIGO~\citep{aLIGO} and Advanced Virgo~\citep{aVirgo} will be the most sensitive observatories in the gravitational-wave (GW) spectrum between 10 Hz to a few kHz.
Binary systems comprised of compact stellar remnants, such as stellar mass black holes (BH) and neutron stars (NS) merge at frequencies between $\sim\,100$ and $\sim\,1000$ Hz and are the primary science target for the LIGO/Virgo (LV) network.
GW observations will investigate these systems in ways not accessible to electromagnetic observations.
Detailed observations of individual sources will provide unparalleled insight into strong field gravity and the NS equation of state~\citep{Flanagan:2007ix, Read:2009yp, Lackey:2013axa, Wade:2014vqa, Li:2011cg, Sampson:2013lpa}.
Inferred characteristics of the compact binary population will feed back into complicated, ill-constrained physics of compact object formation and binary evolution~(\citet{Abadie:2010cf} and references therein).
We investigate the capabilities of an Advanced LV network of detectors to constrain the individual component masses of a compact binary and thereby distinguish between NSs and BHs. We identify the mass intervals for which neutron star black-hole binaries (NSBH), and binary black hole (BBH) systems can be securely distinguished. This is the first large-scale study to characterize compact binary posterior distribution functions (PDFs) including spin-precession effects over a broad range of masses and spins.
We use our results to assess LV's role in resolving the debate over the compact object ``mass gap.''
Observations of X-ray binaries suggest a depletion in the mass distribution of compact remnants between the highest mass NSs ($\sim 2 {\rm M}_\odot$) and the lowest mass BHs ($\gtrsim 5 {\rm M}_\odot$) ~\citep{Ozel:2010su, Farr:2010tu}.
Inferring masses electromagnetically is challenging and systematic errors may dominate. Including variable emission from the accretion flow in the analysis of the same X-ray binaries systematically finds lower masses and disfavors the mass gap~\citep{Kreidberg:2012ud}.
The mass distribution of NSs and BHs has implications on plausible explanations for the core-collapse supernova mechanism.
\citet{Belczynski:2011bn} suggests the mass gap as observational evidence that supernovae develop rapidly (within 100 to 200 ms), while a ``filled gap'' favors longer timescale explosion mechanisms.
\citet{Kochanek:2013yca} and \citet{Clausen:2014wia} find that a bimodal mass distribution for compact remnants is a natural consequence of high mass red supergiants ending their lives as failed supernovae.
GW observations have been suggested as a means of resolving this controversy because the component masses are directly encoded in the signal.
We use our study to forecast whether LV observations can confirm or falsify a mass gap between NSs and BHs.
Our results are to be contrasted with the contemporaneous paper from \citet{Mandel:2015spa}, in which they assume a specific mass distribution (generated from population synthesis) and found it is possible to distinguish BHNS and BBHs and infer a gap after tens of detections.
\section{Inferring physical parameters from gravitational-wave data}\label{sec:pe}
Inspiral waveforms are well understood from post-Newtonian (PN) expansions of the binary dynamics \citep[e.g.][]{Blanchet:2006zz}.
PN waveforms enable template-base data analysis methods where many trial waveforms are compared to the data.
The model waveforms, or templates, are parameterized by the masses, spins, location, and orientation of the binary.
GW signals from compact binaries will be in the LV sensitivity band for tens of seconds to minutes, evolving through $N_{\rm cycle}\sim10^2$--$10^4$ cycles before exiting the band or merging.
Long duration signals place strict demands on the acceptable phase difference between the template waveforms and the signal.
To leading order in the PN expansion, the phase evolution depends only on the ``chirp mass''~\citep[see][]{Peters:1963ux} $\mathcal{M}_c \equiv \left(m_1 m_2\right)^{3/5}\left(m_1+m_2\right)^{-1/5}$ where $m_1>m_2$ are the component masses of the binary.
The uncertainty in $\mathcal{M}_c$ scales as $( N_{\rm cycle} )^{-1/2}$~\citep{Sathyaprakash:2009xs} and is thereby constrained to high precision while the component masses are completely degenerate.
Higher order corrections introduce the mass ratio $q=m_2/m_1$ and couplings between the intrinsic angular momentum of the component bodies $\vec{S}_{1,2}$ and the orbital angular momenta of the system $\vec{L}$ breaking the degeneracy between $m_1$ and $m_2$, though large correlations remain \citep{Cutler:1994ys, Poisson:1995ef}.
\citet{Hannam:2013uu} investigated how well LV can distinguish between NSs and BHs by approximating parameter confidence intervals using the match between a proposed signal and a grid of templates.
Their study used a simplified version of the full PN waveforms parameterized by a single ``effective spin''
and concluded that most LV detections will not provide unambiguous separation between NSs and BHs.
The effective spin approximation overstates the degeneracy between mass and spin because it ignores precession of the orbital plane and spin alignments.
\citet{Chatziioannou:2014coa} revisited the topic of component-mass determination with a Markov chain Monte Carlo (MCMC) analysis using waveforms that included spin and orbital precession .
Their study found that spin precession reduces mass-spin correlations, concluding that BNS systems with components consistent with known NSs in binaries would not be misidentified as either low-mass black holes or ``exotic'' neutron stars which they define as having masses either below $1\ {\rm M}_\odot$ or above $2.5\ {\rm M}_\odot$, and/or with dimensionless spins $\chi>0.05.$
In this work we use the same methodology of \citet{Chatziioannou:2014coa} -- employing an MCMC with templates that include full two-spin precession effects to infer the PDF -- but use a large ensemble of NSBH and BBH signals, paying particular attention to LV's ability to identify sources which have one, or both, components in the mass gap.
\section{\label{sec:method} Method}
Our study uses the \emph{LALInference}\ software library for recovering the parameters of compact binary systems.
In our work we elect to use the MCMC implementation \texttt{lalinference\_mcmc} though results do not depend on the chosen sampler.
A complete description of the software is found in~\citet{LALInference}.
\emph{LALInference}\ is part the LSC Analysis Library (LAL) which has a wide variety of template waveforms available.
{\bf For this work we use the SpinTaylorT2 waveform implemented in LAL. A detailed description of the waveform can be found in Appendix B of \citet{Nitz:2013mxa}}.
To simulate a population of plausible LV detections we draw 1000 binary parameter combinations from a uniform distribution in component masses with $m_1\geq m_2$, $m_i\geq 1\ {\rm M}_\odot$ and $m_1+m_2 \leq 30\ {\rm M}_\odot$.
The maximum total mass of $30\ {\rm M}_\odot$ is chosen so that the merger and ring-down portion of the waveform (where the PN approximation is invalid) does not dominate the signal as observed by LV.
Dimensionless spin magnitudes $\chi$ are drawn uniformly from $[0,1]$ and the spin vector orientation is randomly distributed over a sphere with respect to $\vec{L}$.
Notice that our choice of simulated signals includes the possibility of neutron stars with anomalously high spins $\chi>0.7$~\citep{Lo:2011}.
The sky location is distributed uniformly over the celestial sphere, orientations are randomly distributed, and the distance to the binary is uniform in volume.
We reject sources which do not have signal-to-noise ratio $\rm{SNR}>5$ in two or more detectors.
Figure~\ref{fig:masses} shows the mass distributions of our population.
The bottom panel is a scatter plot of $m_1$ and $m_2$ colored by the source SNR over the Advanced LV network.
{\bf It is important to note that our simulated population includes binaries at sufficiently high mass ($M\gtrsim12 {\rm M}_\odot$) that the merger and ringdown portion of the signal is detectable~\citep{Buonanno:2009zt}.
Our study is focused on quantifying how well masses can be determined purely from the inspiral part of the waveform, for which we have reliable waveforms appropriate for parameter estimation (i.e., valid across all parameter space).
Using inspiral waveforms for both our signal simulations and parameter recovery allows us to assess the inspiral effects separate from merger-ringdown effects.
However we are in urgent need of precessing waveforms that include the merger/ringdown and are valid over the full prior range to avoid systematic errors in analyses of real data.
Currently no such waveforms are available. Available inspiral models that include precession and merger ringdown must be calibrated to numerical relativity simulations and to date are only valid at mass ratios $q>1/4$~\citep{Hannam:2013oca}.
Such limited-validity waveforms lead to non-quantified, systematic biases and cannot be used for parameter estimation at present.
In the absence of generic waveforms it may be necessary to apply a low-pass filter on real data to mask any merger and ringdown signals. }
{\bf Because GW emission is strongest along the orbital angular momentum direction, the detected binaries from a population uniform in orientation and volume is biased against systems whose orbital plane is edge-on to the observer (defined as having an inclination angle near $90^\circ$). Fig.~\ref{fig:inclination} shows the cumulative distribution function of our observed population (red) compared to an underlying population distributed uniformly in orientation. This selection effect has an important role in mass measurement for LV observations. Precession-induced effects on the waveforms, which are relied upon to break the degeneracy between mass ratio and spin, are less detectable for face-on systems \citep{Vitale:2014mka}. Previous studies which have used hand-selected populations of signals to study the effects of precession have not accounted for this selection bias, overemphasizing the role precession can play for generic signals. For example, the inclination used by \citet{Chatziioannou:2014coa} (63$^\circ$) is larger than $\sim80\%$ of our sources so the improvement they found will be fully realized for a small fraction of our simulated population.}
For each binary we compute the response of Advanced LV at design sensitivity \citet{ObservingScenarios} with a low frequency cutoff at 20 Hz.
Each simulated ``detection'' is then analyzed with {\texttt lalinference\_mcmc} which returns independent samples from the PDF for the model parameters.
We do not simulate instrument noise in this study because our focus is on the degree to which LV's frequency-dependent sensitivity, and the flexibility of the PN waveforms, limit mass measurements.
Adding simulated noise to our signals introduces uncontrollable contributions to the PDF without adding any value to our assessment of each simulated signal.
Posterior distributions in true detections will be altered by the particular noise realization in which the signal is embedded.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig1.eps}
\caption{\small{The mass distribution of our simulated population of compact binaries in units of ${\rm M}_\odot$.
[Top panel] Histogram for $m_1$ (red, solid) and $m_2$ (blue dotted).
The bottom panel is a scatter plot of $m_1$ and $m_2$ colored by the source's SNR over the advanced LV network.}}
\label{fig:masses}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig2.eps}
\caption{\small{Cumulative distribution function for detected inclination angles (red) compared to the isotropicly oriented underlying population.
There is a significant selection effect in favor of face-on systems (inclination $\sim0^\circ$) because GW emission is strongest along the direction of the orbital angular momentum vector of the binary. Modulations in the waveform from spin-induced precession are less detectable for face-on systems. This selection effect will suppress the ability for precession to break the degeneracy between spin magnitude and mass ratio.}}
\label{fig:inclination}
\end{figure}
\section{\label{sec:results} Results}
\subsection{\label{sec:m1m2}Distinguishing black holes from neutron stars}
We take $m=3\ {\rm M}_\odot$ to be the dividing line between the masses of NSs and BHs, with all NSs located below this threshold and all BHs located above.
We will use $5\ {\rm M}_\odot$ as the minimum mass of a BH when assuming the existence of a mass gap.
The distinction between NSBHs and BBHs in our sample is determined by the measurement of the smaller mass, $m_2$.
Figure~\ref{fig:m2pe} shows the $90\%$ credible intervals of the PDF for $m_2$ as a function of the true value from the simulated population.
Each entry is colored by the SNR of the signal in the 3-detector network.
Horizontal gray dashed lines denote the mass gap.
We find all of our simulations with $m_2\lesssim1.5\ {\rm M}_\odot$ to be clearly identified as containing at least one neutron star, however our population yields only ten such systems so these are small-number statistics. Most binaries with a NS below 2 ${\rm M}_\odot$ constrain the NS to have a mass below 3 ${\rm M}_\odot$ although occasional the posterior supports the smaller object being a low-mass BH .
Detections of larger mass neutron stars ($2\leq m_2 \leq 3\ {\rm M}_\odot)$ have $90\%$ credible intervals which consistently extend into the BH regime.
The tendency for recovery of high-mass neutron stars in NSBH systems to be consistent with low-mass BHs poses a challenge for determining the maximum NS mass from LV observations alone.
At what point can we \emph{rule out} the possibility that the system contains a neutron star, and definitively declare that we have detected a binary black hole?
In Figure~\ref{fig:m2pe} we see that the true mass of the smaller object must exceed $\sim 6\ {\rm M}_\odot$ before the $90\%$ credible intervals rule out a NS. {\bf Depending on details of spin alignment and orientation, systems with $m_2$ as low as $4\ {\rm M}_\odot$ can be unambiguously identified as BBHs.} {\bf Within this range} the $m_2$ posteriors for BBH sources seldom reach below $2\ {\rm M}_\odot$, so if a maximum NS mass were independently confirmed to be consistent with current observations LV's classification of NSs and BHs would improve.
{\bf It may seem surprising in Fig~\ref{fig:m2pe} that the width of the credible intervals do not exhibit the $1/{\rm SNR}$ scaling predicted by Fisher matrix approximations~\citep{Cutler:1994ys}. The Fisher approximation is only suitable at sufficiently high SNR that the posterior distribution function is well approximated by a multivariate Gaussian, in which case the inverse Fisher matrix is the covariance of the posterior. Implicit in this condition is the assumption that the model waveform is a linear function of the source parameters. See~\citet{Vallisneri:2007ev} for a thorough deconstruction of Fisher matrix-based intuition being applied to GW signals. For typical LIGO/Virgo binaries these conditions are not satisfied for the mass parameters. The width of the $m_2$ credible intervals is driven by uncertainty in $q$ which, due to the degeneracy with spin, is extremely non-Gaussian and can span the entire prior range. The chirp mass, on the other hand, \emph{is} a sufficiently well constrained parameter at the SNRs in our simulated population. In Fig.~\ref{fig:mcpe} we show the fractional $1-\sigma$ uncertainty in $\mathcal{M}_c$ inferred from the Markov chains as a function of the source value, with each event colored by the network SNR. The expected $1/\rm{SNR}$ dependence is apparent for this parameter. Notice also that the chirp mass errors grow with increasing total mass. Higher mass binaries are in band for fewer GW cycles which directly impacts measurability as discussed in Sec.~\ref{sec:pe}. }
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig3.eps}
\caption{\small{90\% credible intervals for recovered $m_2$ as a function of the true mass. Each entry is colored by the network SNR of the source. Horizontal lines denote the mass gap~\citep{Ozel:2010su, Farr:2010tu}. Credible intervals for high mass systems are limited from above by our prior on the total mass $M<30$ and the $m_2\leq m_1$ convention. Axes are in units of ${\rm M}_\odot$.}}
\label{fig:m2pe}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig4.eps}
\caption{\small{Fractional $1-\sigma$ error in chirp mass as a function of the true value. Each entry is colored by the network SNR of the source. The expected SNR scaling of the errors is evidence in the chirp mass recovery, while it is absent in the $m_2$ measurement due to the degeneracy between mass ratio and spin. The x-axis is in units of ${\rm M}_\odot$.}}
\label{fig:mcpe}
\end{figure}
\subsection{\label{sec:gap}Identifying systems in the mass gap}
Following \cite{Mandel:2010} we use our simulated LV detections to infer the relative fraction of NSs, BHs in the mass gap, and BHs above the mass gap.
Because of the large mass-measurement uncertainties the number of detections needed to conclude BHs inhabit the mass gap is highly dependent on the underlying mass distribution. Depending on whether the observed population is heavily dominated by low- or high-mass systems, we find
between ten and many hundreds of detections are necessary to conclude (at three-sigma confidence) the gap is populated.
Our initial set of simulated signals features many high-mass ratio binaries. We simulated an additional 100 sources with both objects having $3<m<5\ {\rm M}_\odot$ to check whether comparable-mass systems are easier to identify in the gap. The mass-gap population still suffers from large mass errors with $>95\%$ of the sources having posterior support for a NSBH system.
Figure~\ref{fig:m2_gap} shows the distribution of the 90\% credible interval widths for $m_1$ (red, solid) and $m_2$ (blue, dotted) of the gap sources.
The majority of plausible mass-gap sources yield credible intervals that are similar to or exceed the width of the mass gap.
However, $\sim 25\%$ of the mass gap sources' $m_2$ posteriors do not reach below $2\ {\rm M}_\odot$.
While we will not be able to say with any certainty that an individual source occupies the mass gap, we can often conclude that the binary contains
either an unusually high-mass neutron star, or a pair of unusually low-mass black holes.
Three of the mass-gap sources were constrained to be $3<m<5\ {\rm M}_\odot$. A careful investigation of these systems revealed they were in low-probability alignments -- all with the spin of the larger mass close to the orbital plane, and the best constrained having $\vec L$ nearly perpendicular to the line of sight.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig5.eps}
\caption{\small{Distribution of $90\%$ credible interval widths in ${\rm M}_\odot$ from 100 mass gap sources. The width of the mass gap is 2 ${\rm M}_\odot$, and therefore most of the BHs cannot be constrained to fall within the gap.}}
\label{fig:m2_gap}
\end{figure}
\subsection{\label{sec:spin}Restricting allowed spins for neutron stars}
All of the results thus far presented used a uniform prior on spin magnitude between $[0,1]$.
The maximum upper limit on the spin of a NS is $\chi\lesssim0.7$ while observed NS spins are lower~\citep{Lo:2011}.
We investigate if using a physical prior on NS spins can improve $m_2$ measurement.
We resample the posterior by imposing a maximum spin for component masses below $3\ {\rm M}_\odot$, rejecting samples from the Markov chain with $\chi_2>0.7$.
We found no indication that restricting the range of neutron star spins significantly improves uncertainty in the determination of $m_2$.
Figure~\ref{fig:m1m2a2} demonstrates the limited role of $m_2$'s spin on mass determination.
Open circles mark the true values of component masses for nine representative examples.
Going through each circle is the scatter plot of the posterior samples colored by $\chi_2$.
The arcs traced out by the samples are lines of constant $\mathcal{M}_c$.
The vertical and horizontal dashed lines denote the mass gap.
Notice that there is no obvious correlation between position along the arc and $\chi_2$ -- the spin of the smaller body is generally not constrained and therefore does not help with mass determination.
Restricting the spin of the smaller mass does not help because at high mass ratios (and therefore NS-like $m_2$), the contribution to the PN phase from $\chi_2$ is suppressed.
The leading order spin corrections enter the PN phase with magnitude $\chi_i m_i^2$, so $\chi_2$'s influence to the phase evolution is down-weighted relative to $\chi_1$ by $\mathcal{O}\left(({\rm mass\ ratio})^2\right)$.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig6.eps}
\caption{\small{Component mass recovery colored by the spin of the lower mass. Open circles mark the true values for nine representative binary systems from the simulated population. The spin of the lower mass is not well constrained, and does not strongly correlate with the mass parameters. Therefore, restricting the spin of the smaller body to be consistent with our prior expectations does not impact our ability to distinguish between black holes and neutron stars. Both x- and y-axes are in units of ${\rm M}_\odot$.}}
\label{fig:m1m2a2}
\end{figure}
\section{\label{sec:discuss} Discussion}
In this paper we investigated the capability of the Advanced LV network to distinguish between NSs and BHs {\bf from the inspiral-only waveforms} using a large population of plausible detections.
This is the first large-scale study to characterize compact binary PDFs including spin-precession effects over a broad range of masses, mass ratios, and spins. Our study does not factor in systematic effects from real detector noise or differences between template waveforms and the true gravitational wave signal.
{\bf Our study is limited to inspiral-only waveforms for simulation of signals and template waveforms for recovery because available precessing merger/ringdown models are not valid over the full prior volume. For many of our simulated signals the merger will be detectable and may help improve component mass estimates. Further improvements may come from including PN amplitude corrections.}
We arrive at four main conclusions from our analysis:
\begin{enumerate}[leftmargin=*]
\item \emph{When are we certain of at least one NS?} For most systems with $m_2 \leq 2\ {\rm M}_\odot$, and all systems with $m_2 \leq1 .5\ {\rm M}_\odot$.
\end{enumerate}
For larger-mass neutron stars ($2 \leq m_2 \leq 3$) the 90\% credible intervals frequently extend into the low-mass BH regime. This tendency for high-mass neutron stars in NSBH systems to be consistent with low-mass black holes poses a challenge for determining the maximum NS mass from LV observations alone.
\begin{enumerate}[resume,leftmargin=*]
\item \emph{When are we certain of a BBH with both masses above 3 ${\rm M}_\odot$?} When the mass of the smaller object exceeds $\sim 6\ {\rm M}_\odot$.
\end{enumerate}
The $m_2$ posteriors for BBH signals seldom reach below 2 ${\rm M}_\odot$, so if a maximum neutron star mass were independently confirmed to be $m_{\rm max}\sim2{\rm M}_\odot$, then LV's ability to discriminate between BBH and BNS/NSBH populations would be significantly improved.
\begin{enumerate}[resume,leftmargin=*]
\item \emph{When is a black hole definitely not in the mass gap?} When component mass $m_2 \gtrsim 10\ {\rm M}_\odot$ the 90\% credible intervals do not reach into the mass gap.
\end{enumerate}
It may prove challenging to confirm its existence because NSs with masses above $\sim 2\ {\rm M}_\odot$ have error bars reaching into the $m_2\in[3,5]\ {\rm M}_\odot$ interval from below while BBH systems with $m_2 \lesssim 10\ {\rm M}_\odot$ can have credible intervals that extend below $5\ {\rm M}_\odot$. Our findings suggest that inferring a bimodal mass distribution based solely on the observed sample of compact binaries especially without reliable predictions for the underlying compact object mass distribution will be challenging.
\begin{enumerate}[resume,leftmargin=*]
\item \emph{When are we certain of a mass-gap object?} Only in rare circumstances when the binary is nearly edge-on and spins are oriented in the orbital plane. For typical binaries component mass errors are larger than the mass gap.
\end{enumerate}
We simulated an additional 100 injections with both masses in the mass gap between $[3,5]$ ${\rm M}_\odot$. We found $> 95\%$ of the systems had some posterior support for a NSBH system. The majority of plausible mass gap sources yield credible intervals that are similar to, or exceed the width of the mass gap itself (2 ${\rm M}_\odot$).
{\bf Only sources in low-probability alignments (spins in the orbital plane and/or edge-on orientations) were constrained to be $3<m<5\ {\rm M}_\odot$.}
However, $m_2$ posteriors from mass gap BBHs do not reach below 2 ${\rm M}_\odot$. While we will not be able to say with any certainty that a given source occupies the mass gap, we will at least be alerted to the fact that the binary contains either an unusually high-mass neutron star, or a pair of unusually low mass black holes, with either possibility providing plenty of intrigue for the astrophysics community.
Assuming flat mass distributions, we used simulated LV detections to infer the relative fraction of NSs and BHs in and above the mass gap, and estimate how many detections would be needed to confidently conclude that gap BHs exist. Depending on whether the underlying mass distribution is heavily dominated by low- or high-mass systems, ten to hundreds of detections are needed to confirm objects in the gap. Assuming a plausible population-synthesis model where BNSs, NSBHs, and BBHs are well separated in chirp-mass space, \citet{Mandel:2015spa} found that a mass gap was statistically distinguishable with a only few tens of detection
It is clear that forecasts for what we may learn from LV observations of compact binaries are subject to large variance while uncertainty about the true mass distribution persists.
Continued effort in theoretically understanding that distribution, and how LIGO/Virgo observations can be used to test such theories, will be of great value as we begin assembling a catalog of compact binaries coalescences.
\section{Acknowledgments}
We thank Atul Adhikari, Claudeson Azurin, Brian Klein, Brandon Miller, Leah Perri, Ben Sandeen, Jeremy Vollen, Michael Zevin for help running the MCMC analysis. Will Farr, Carl-Johan Haster, and Ilya Mandel provided important comments about our calculations and results.
TBL and VK acknowledge NSF award PHY-1307020. BF was supported by the Enrico Fermi Institute at the University of Chicago as a McCormick Fellow. SC thanks the US-UK Fulbright Commission for personal financial support during this research period. DEH acknowledges NSF CAREER grant PHY-1151836, the Kavli Institute for Cosmological Physics at the University of Chicago through NSF grant PHY-1125897, and an endowment from the Kavli Foundation. Computational resources were provided by the Northwestern University Grail cluster through NSF MRI award PHY-1126812.
|
\section{Introduction}
Examinee ability is measured by the scores obtained by an examinee in
a test designed to assess such ability. As with all other
measurements, this ability measurement too is fundamentally
uncertain or inclusive of errors. In Classical Test Theory,
quantification of the complementary certainty, or reliability of a
test, is defined as the ratio of the true score variance and the
observed score variance i.e. reliability is defined as the proportion
of observed test score variance that is attributable to true
score. Here, the observed score is treated as inclusive of the
measurement error. This theoretical definition notwithstanding, there
are different methods of obtaining reliability in practice, and
problems arise from the implementation of these different techniques
even for the same test. Importantly, it is to be noted that the
different methods of estimating reliability coefficients differ
differently from the aforementioned classical or theoretical
definition of reliability. This can potentially result in different
estimates of reliability of a particular test even for the same
examinee sample. \ctn{berkowitz_2000} defined reliability as the
degree to which test scores for a group of test takers are consistent
over repeated applications of the test, and are
therefore inferred to be dependable and repeatable for an individual
test taker. In this framework, high reliability means that the test is
trustworthy. \ctn{jacobs_1991} and \ctn{satterly_1994} have opined
that another way to express reliability is in terms of the standard
error of measurement. \ctn{rudner_2002} mentioned that it is
impossible to calculate a reliability coefficient that conforms to the
theoretical definition of the ratio of true score variance and
observed score variance. \ctn{webb_2006} have suggested that the
theoretical reliability coefficient is not practical since true scores
of individuals taking the test are not known. In fact, none of the
existing methods of finding reliability is found to be isomorphic to
the classical definition.
In this paper we present a new methodology to compute the classically
defined reliability coefficient of a multiple-choice binary test,
i.e. a test, responses to the questions of which is either correct or
incorrect, fetching a score of 1 or 0 respectively. Thus, our method
gives the reliability as per the classical definition--thereby
avoiding confusion about different reliability values of a given
test. The method that we advance does not resort to multiple testing,
i.e. administering the same test multiple times to a given examinee
cohort. While avoiding multiple testing, this method has the important
additional benefit of identifying the way of splitting of the test, in
order to ensure that the split halves are equivalent (or parallel), so
that they approach maximal correlation and therefore maximal
split-half reliability. We offer an interval estimate of the true
score of any individual examinee whose obtained score is known.
This paper is organised as follows. In the following section, we
discuss the methods of finding reliability as found in the current
literature. In Section~\ref{sec:batt_rel}, we present a similar
discussion, but this time, in regard to a set or battery of tests. Our method
of finding the reliability as per the classical definition, is
expounded upon in Section~\ref{sec:method}; this is based upon a novel
method of splitting a test into two parallel halves using a new
iterative algorithm that is discussed in Section~\ref{sec:dichot}
along with the benefits of this iterative scheme. Subsequently, we
proceed to discuss the computation of the reliability of a battery of
tests using summative scores and weighted scores
(Section~\ref{sec:weights}). Empirical verification of the presented
methodology is undertaken using four sets of simulated data (in
Section~\ref{sec:simulated}). Implementation of the method to real data
is discussed in Section~\ref{sec:real}. We round up the paper in
Section~\ref{sec:conclu}, by recalling the salient outcomes of the
work.
\section{Currently available methods}
\label{sec:lit}
In the multiple testing (or the test-retest) approach, the main
concern is that the correlation between test scores coming from
distinct administrations, depends on the time gap between the
administrations, since the sample may learn or forget something in
between and also get acquainted to the test. Thus, different values of
reliability can be obtained depending on the time gap
\cite{meadows_2005}. Also, test-retest reliability primarily reflects
stability of scores and may vary depending on the homogeneity of
groups of examinees \cite{gualtieri_2006}.
In the split half method,
the test is divided into two sub-tests, i.e. the whole set of questions is divided into two sets of questions and the entire instrument is administered to one
examinee sample. The split-half reliability estimate is the
correlation between scores of these two parallel halves, though
researchers like \ctn{kaplan_2001} recommend finding
reliability of the entire test using the Spearman-Brown formula. Value
of the split-half reliability depends on how the test is
dichotomised. We acknowledge that a test consisting of 2$\eta$-number
of items can be split in half in $\displaystyle{2\eta\choose\eta}$
number of ways and that each method of splitting will
imply a distinct value of reliability in general. In this context, it
merits mention that it is of crucial importance to identify that way
of splitting that ensures that the split halves are parallel and
the split-half reliability is maximum.
In attempts to compute reliability based on ``internal consistency'',
the aim is to quantify how well the questions in the test, or test
``items'', reflect the same construct. These include the usage of
Cronbach's Alpha ($\alpha$): the average of all possible split-half
estimates of reliability, computed as a function of the ratio of the
total of the variance of the items and variance of the examinee
scores. Here, the ``item score'' is the total of the scores obtained
in that item by all the examinees who are taking the test, while an
examinee score is the total score of such an examinee, across all the
items in the test.
$\alpha$ can be shown to
provide a lower bound for the reliability under certain
assumptions. Limitations of this method have been reported by
\ctn{eisinga_2012} and \ctn{ritter_2010}. \ctn{panayides_2013}
observed that
high values of $\alpha$ do not necessarily mean higher reliability
and better quality of scales or tests. In fact very high values of
$\alpha$ could be indicating lengthy scales, a narrow coverage of the
construct under consideration and/or parallel items
\cite{boyle_1991,kline_1979} or redundancy of items
\cite{streiner_2003}.
In another approach, referred to as the ``parallel forms'' approach,
the test is divided into two forms or sub-tests by randomly selecting
the test items to comprise either sub-test, under the assumption that
the randomly divided halves are parallel or equivalent. Such a demand
of parallelity requires generating lots of items that reflect the same
construct, which is difficult to achieve in practice. This approach is
very similar to the split-half reliability described above, except
that the parallel forms constructed herein, can be used independently
of each other and are considered equivalent. In practice, this
assumption of strict parallelity is too restrictive.
Each of the above method of estimation of reliability has certain
advantages and disadvantages. However, the estimation of the
reliability coefficient under each such method deviates differently
from the theoretical definition and consequently, gives different
values for reliability of a single test. In general, test-retest
reliability with significant time gap is lower in value than the
parallel forms reliability and reliability computed under the model of
internal consistency amongst the test items.
To summarise, there is a potency for confusion over the
trustworthiness of a test, emanating out of the inconsistencies
amongst the different available methods that are implemented to
compute reliability coefficients. This state of affairs motivates a
need to estimate reliability in an easily reproducible way, from a
single administration of the test, using the theoretical definition of
reliability i.e. as a ratio of true score variance and observed score
variance.
\subsection{Battery reliability}
\label{sec:batt_rel}
\noindent
\ctn{bock_1966} derived reliability of a battery using Multivariate
Analysis of Variance (MANOVA), where $r$ number of parallel forms of a
test battery were given to each individual in the examinee sample. It
can be proved that average canonical reliabilities from MANOVA
coincides with the canonical reliability for the Mahalanobis distant
function.
\ctn{wood_1987} compared three multivariate models
(canonical reliability model, maximum generalisability model,
canonical correlation model) for estimating reliability of test
battery and observed that the maximum generalisability model showed
the least degree of bias, smallest errors in estimation, and the
greatest relative efficiency across all experimental
conditions. \ctn{ogasawara_2009} considered four estimators of the
reliability of a composite score based on a factor analysis approach,
and five estimators of the maximal reliability for the composite
scores and found that the same estimators are obtained when Wishart
Maximum Likelihood estimators are used. \ctn{conger_1973}
considered canonical reliability as the ratio of the average squared
distance among the true scores to the average squared distance among
the observed scores. They thereby showed that the canonical
reliability is consistent with multivariate analogues of parallel
forms of correlation, squared correlation between true and observed
scores and an analysis of variance formulation. They opined that the
factors corresponding to the small eigenvalues might show significant
differences and should not be discarded. Equivalences amongst
canonical factor analysis, canonical reliability and principal
component analysis were studied by \ctn{conger_1976}; they
found that if variables are scaled so that they have equal measurement
errors, then canonical factor analysis on all non-error variance,
principal component analysis and canonical reliability analysis, give
rise to equivalent results.
In this paper, the objective is to present a methodology for obtaining
reliability as per its theoretical definition from a single
administration of a test, and to extend this methodology to find ways
of obtaining reliability of a battery of tests.
\section{Methodology}
\label{sec:method}
\noindent
Consider a test consisting of $n$ items, administered to $N$
persons. The test score of the $i$-th examinee is $X_i$,
$i=1,\ldots,N$ and $\boldsymbol{X} =(X_1, X_2, \ldots, X_N)^T$ is referred to as
the test score vector. The vector $\boldsymbol{I} = (I_1, I_2, \ldots,I_N)^T$
depicts the maximum possible scores for the test where $I_i = n$
$\forall i =1,2,\ldots,N$. Here $\boldsymbol{X},\boldsymbol{I}\in{\cal X}\subset {\mathbb
R}^N$ where we refer to ${\cal X}$ as the $N$-dimensional person
space. Sorting the components of $\boldsymbol{X}$ in increasing order will
produce a merit list for thesample of examinees, in terms of the
observed scores and thereby help infer the ability distribution of the
sample of examinees. However, parameters of the test can be obtained
primarily in terms of the 2-norm of the vectors $\boldsymbol{X}$ and $bI$,
i.e. $\parallel\boldsymbol{X}\parallel$, $\parallel\boldsymbol{I}\parallel$, and the angle
$\theta_X$ between these two vectors.
Mean of the test scores ${\bar{X}}$ is given as follows. Since
$\cos\theta_X = \displaystyle{\frac{\sum_i^N X_i I_i}{\parallel\boldsymbol{X}\parallel \parallel\boldsymbol{I}\parallel}}\Longrightarrow \displaystyle{ {\sum_i^N X_i}}/N =
\displaystyle{\frac{\parallel\boldsymbol{X}\parallel \cos\theta_X}{\sqrt{N}}}={\bar{X}}$.
Similarly, mean of the true score vector $\boldsymbol{T}$ is
${\bar{T}} = \displaystyle{\frac{\parallel\boldsymbol{T}\parallel \cos\theta_T}{\sqrt{N}}}$ where $\theta_T$ is the angle
between the true score vector and the ideal score vector $\boldsymbol{I}$.
Test variance $S_X^2$ can also be
obtained in this framework. For $N$ persons who take the test,
$S_X^2 = \displaystyle{\frac{1}{N}\parallel\boldsymbol{x}\parallel^2}
\Longrightarrow S_X = \displaystyle{\frac{1}{\sqrt{N}}\parallel\boldsymbol{x}\parallel}$
where $x_i=\displaystyle{X_i-{\bar{X}}}$ is the $i$-th deviation
score, $i=1,2,\ldots,N$. The
relationship between the norm of deviation score and norm of test score can be easily derived as
$\parallel\boldsymbol{x}\parallel^2 = \parallel\boldsymbol{X}\parallel^2 \sin^2\theta_X$
Similarly,
$\parallel\boldsymbol{t}\parallel^2 = \parallel\boldsymbol{T}\parallel^2 \sin^2\theta_T$
where $t_i=\displaystyle{T_i-{\bar{T}}}$.
Using this relationship between deviation and test scores, test variance can be
rephrased as
\begin{equation}
S_X^2 = \displaystyle{\frac{\parallel \boldsymbol{X}\parallel^2\sin^2\theta_X}{N}}.\quad{\textrm{Also}}\quad
S_T^2 = \displaystyle{\frac{\parallel \boldsymbol{T}\parallel^2\sin^2\theta_T}{N}}.
\label{eqn:variance}
\end{equation}
The aforementioned relationship between test and deviation scores gives
$\sin^2\theta_X = \displaystyle{\frac{\parallel\boldsymbol{x}\parallel^2}{\parallel\boldsymbol{X}\parallel^2}}$ and
$\sin^2\theta_T = \displaystyle{\frac{\parallel\boldsymbol{t}\parallel^2}{\parallel\boldsymbol{T}\parallel^2}}$.
At the same time, in Classical Test Theory, the means of the observed
and true scores coincide, i.e. ${\bar{X}}={\bar{T}}$, as in this
framework, the observed score is represented as the sum of true score
(obtained under the paradigm of no errors) and the error, where the
error is assumed to be distributed as a normal with zero mean and
variance referred to as the error variance. Then recalling the
definition of these sample means, we get
\begin{equation}
\cos\theta_T = \displaystyle{\frac{\parallel
X\parallel}{\parallel T\parallel} \cos\theta_X}.
\label{eqn:theta_T_X}
\end{equation}
This gives the relationship amongst $\parallel\boldsymbol{X}\parallel$,
$\parallel\boldsymbol{T}\parallel$, $\theta_T$, $\theta_X$. The stage is
now set for us to develop the methodology for computing reliability
using the classical definition.
\section{Our method}
\subsection{Background}
\label{sec:reliability}
\noindent
Reliability, $r_{tt}$, of a test is defined clasically as $r_{tt}=\displaystyle{\frac{S_T^2}{S_X^2}}$.
Thus, to know $r_{tt}$ one needs to know the value of the true score variance or the error variance.
For this, let us concentrate on parallel tests. As per the classical definition, two tests ``$g$'' and ``$h$'' are parallel if
\begin{equation}
T_{i}^{(g)} = T_{i}^{(h)}\quad{\textrm{and}}\quad S_{e}^{(g)} = S_{e}^{(h)}\quad i=1,2,\ldots,N
\label{eqn:ssaresame}
\end{equation}
where the superscript ``$g$'' refers to sub-test $g$ and superscript ``$h$'' to sub-test $h$, and $S_{e}^{(p)}$ is the standard deviation of the
error scores in the $p$-th sub-test, $p=g,h$. Thus, $T_i^{(g)}$ is the true score of the $i$-th examinee in the $g$-th test and $T_i^{(h)}$ that in the $h$-th test. Similarly, the observed scores of the $i$-th examinee, in the two tests, are $X_i^{(g)}$ and $X_i^{(g)}$.
From here on, we interpret ``g'' and ``h'' as the two parallel
sub-tests that a given test is dichotomised into. This implies that
the observed score vectors of these two parallel tests can be
represented by two points $\boldsymbol{X}_g$ and $\boldsymbol{X}_h$, in ${\cal X}$ such that
$\boldsymbol{X}_g = \boldsymbol{T}_g + \boldsymbol{E}_g$ and $\boldsymbol{X}_h = \boldsymbol{T}_h + \boldsymbol{E}_h$ in the paradigm of
Classical Test Theory, where $\boldsymbol{E}_p$ is the error score in the $p$-th
test, $p=g, h$. (Here $\boldsymbol{X}_p=(X_1^{(p)},\ldots,X_N^{(p)})$,
$p=g,h$). Now, recalling that for parallel tests $g$ and $h$,
$T_i^{(g)}=T_i^{(h)}\forall i=1,\ldots,N\Longrightarrow
\boldsymbol{T}_g=\boldsymbol{T}_h\Longrightarrow \boldsymbol{X}_g-\boldsymbol{X}_h = \boldsymbol{E}_g-\boldsymbol{E}_h$, so that
\begin{equation}
\parallel \boldsymbol{X}_g\parallel^2 + \parallel \boldsymbol{X}_h\parallel^2 - 2\parallel \boldsymbol{X}_g\parallel \parallel \boldsymbol{X}_h\parallel\cos\theta_{gh} = \parallel \boldsymbol{E}_g\parallel^2 + \parallel \boldsymbol{E}_h\parallel^2 - 2\parallel \boldsymbol{E}_g\parallel \parallel \boldsymbol{E}_h\parallel\cos\theta^{(E)}_{gh}
\label{eqn:1sstep}
\end{equation}
where $\theta_{gh}$ is the angle between $\boldsymbol{X}_g$ and $\boldsymbol{X}_h$ while
$\theta^{(E)}_{gh}$ is the angle between $\boldsymbol{E}_g$ and $\boldsymbol{E}_h$. But, the correlation of the errors in two parallel tests is zero (follows from equality of standard deviation of error sores in the parallel sub-tests and equality of means of the error scores--see Equation~\ref{eqn:theta_T_X}). The geometrical interpretation of this is that the error vectors of the two parallel tests are orthogonal, i.e. $\cos\theta^{(E)}_{gh}=0$. Then Equation~\ref{eqn:1sstep} can be written as\\
\vspace{-.5cm}
\begin{center}
$\parallel \boldsymbol{X}_g\parallel^2 + \parallel \boldsymbol{X}_h\parallel^2 - 2\parallel \boldsymbol{X}_g\parallel \parallel \boldsymbol{X}_h\parallel\cos\theta_{gh}$
\end{center}
\begin{equation}
=\parallel \boldsymbol{E}_g\parallel^2 + \parallel \boldsymbol{E}_h\parallel^2 = N(S_e^{(g)})^2 + N(S_e^{(h)})^2 = 2N(S_e^{(g)})^2
\label{eqn:2ndstep}
\end{equation}
where we have used equality of error variances of parallel tests in
the last step. Equations~\ref{eqn:2ndstep} can be employed to find the
value of $S_e^{(g)}$ from the data (on observed scores). In other words, we can use the
available test score data in this equation to achieve the error
variance of either parallel sub-test that a test can be dichotomised
into. Alternatively, if two parallel tests exist, then
Equations~\ref{eqn:2ndstep} can be invoked to compute the error
variance in either of the two parallel tests.
Now, Equations~\ref{eqn:2ndstep} suggest that the error variance $(S_e^{(test)})^2$ of the entire test is
\begin{equation}
(S_e^{(test)})^2 = 2(S_e^{(g)})^2 = \displaystyle{\frac{\parallel \boldsymbol{X}_g\parallel^2 + \parallel \boldsymbol{X}_h\parallel^2 - 2\parallel \boldsymbol{X}_g\parallel \parallel \boldsymbol{X}_h\parallel\cos\theta_{gh}}{N}}
\label{eqn:finally}
\end{equation}
Then recalling that in the paradigm of Classical Test Theory, the true score is by definition independent of the error, it follows that the observed score variance $S_X^2$ is sum of true score variance $S_T^2$ and error variance $(S_e^{(test)})^2$ the classical definition of reliability gives
\vspace{-.2cm}
\begin{center}
$r_{tt} = \displaystyle{\frac{S_T^2}{S_X^2}} = \displaystyle{1 - \frac{(S_e^{(test)})^2}{S_X^2}}= $
\end{center}
\begin{equation}
1 - \displaystyle{\frac{\parallel \boldsymbol{X}_g\parallel^2 + \parallel \boldsymbol{X}_h\parallel^2 - 2\parallel \boldsymbol{X}_g\parallel \parallel \boldsymbol{X}_h\parallel\cos\theta_{gh}}{N S_X^2}}
= 1 - \displaystyle{\frac{\parallel \boldsymbol{X}_g\parallel^2 + \parallel \boldsymbol{X}_h\parallel^2 - 2\sum_{i=1}^N X_i^{(g)}X_i^{(h)}}{N S_X^2}}
\label{eqn:3rdstep}
\end{equation}
As $\parallel\boldsymbol{X}_g\parallel \parallel\boldsymbol{X}_h\parallel\cos\theta_{gh} = \displaystyle{\sum_{i=1}^N X_i^{(g)}X_i^{(h)}}$, Equation~\ref{eqn:3rdstep} can be simplified to give
\begin{equation}
r_{tt} = 1 - \displaystyle{\frac{2\parallel X_g\parallel^2 - 2\sum_{i=1}^N X_i^{(g)}X_i^{(h)}}{N S_X^2} }
\label{eqn:4thstep}
\end{equation}
since $\parallel\boldsymbol{X}_g\parallel= \parallel\boldsymbol{X}_h\parallel$ for parallel
tests, given that ${\bar{X}}_g={\bar{X}}_h$ and $S_e^{(g)}=
S_e^{(h)}$. Equation~\ref{eqn:3rdstep} and Equation~\ref{eqn:4thstep}
give a unique way of finding reliability of a test from a single
administration, using the classical definition of reliability as long
as dichotomisation of the test is performed into two parallel
sub-tests. Importantly, in this framework, the hypothesis of equality
of error variances of the two halves of a given test can be tested
using an F-test.
In addition, the method also provides a way to
estimate true scores from the data. We discuss this
later in this section.
\subsection{Proposed method of split-half}
\label{sec:dichot}
\noindent
\ctn{chakrabartty_2011} gave a method for splitting a test into 2
parallel halves. Here we give a novel method of splitting a test into
2 parallel halves--$g$ and $h$--that have nearly equal means and
variances of the observed scores. The splitting is initiated by the
determination of the item-wise total score for each item. So let the
$j$-th item in the test have the item-wise score
$\tau_j:=\displaystyle{\sum\limits_{i=1}^N X_i^{(j)}}$, where
$X_i^{(j)}$ is the $i$-th examinee's score in the $j$-th item,
$j=1,\ldots,n$. Our method of splitting is as follows.
\begin{enumerate}
\item[Step-I] The item-wise scores are sorted in an ascending order
resulting in the ordered sequence
$\tau_1,\tau_2,\ldots,\tau_n$. Following this, the item with the
highest total score is identified and allocated to the $g$-th
sub-test. The item with second highest total score is then allocated
to the $h$-th test, while the item with the third highest score is
assigned to $h$-th test and the fourth highest to the $g$-th test,
and so on. In other words, allocation of items is performed to
ensure realisation of the following structure.
\begin{center}{
\begin{tabular}{l l l }
{\underline{sub-test $g$}} &
{\underline{sub-test $h$}} & {\underline{difference in item-wise scores of 2 sub-test}} \\
$\tau_1$ & $\tau_2$ & $\tau_1 - \tau_2 \geq 0$ \\
$\tau_4$ & $\tau_3$ & $\tau_4 - \tau_3 \leq 0$ \\
$\vdots$ & $\vdots$ & $\vdots$ \\
\end{tabular}}
\end{center}
where we assume $n$ to be even; for tests with an odd number of items,
we ignore the last item for the purposes of dichotomisation. The
sub-tests obtained after the very first slotting of the sequence
$\{\tau_k\}_{k=1}^n$ into the sub-tests, following this suggested
method of distribution, is referred to as the ``seed sub-tests''.
\item[Step-II] Next, the difference of item-wise scores in every item
of the $g$-th and $h$-th sub-tests is recorded and the sum ${\cal
S}$ of these differences is computed (total of column 3
in the above table). If the value of ${\cal S}$ is zero,
we terminate the process of distribution of items across the 2
sub-tests, otherwise we proceed to the next step.
\item[Step-III] We identify rows in the above table, the swapping of
the entries of columns 1 and 2 of which, results in the reduction of
$\vert{\cal S}\vert$, where $\vert\cdot\vert$ denotes absolute
value. Let the row numbers of such rows be $\rho^{(\star_\ell)}$,
$\ell=1,2,\ldots,n^{(\star)}$ where $n^{(\star)}\leq n/2$. We swap the
$\rho^{(\star_\ell)}$-th item of the $g$-th sub-test with the
$\rho^{(\star_\ell)}$-th item of the $h$-th sub-test and
re-calculate sum of the entries of the revised $g$-th sub-test and
$h$-th sub-test. If the revised value of $\vert{\cal S}\vert$ is
zero or a number close to zero that does not reduce upon further
iterations, we stop the iteration; otherwise we return to the
identification of the row numbers $\rho^{(\star_\ell)}$ and proceed
therefrom again.
\end{enumerate}
We considered other methods of splitting of the test as well,
including methods in which swapping of items of the two sub-tests is
allowed not just for a given row, but also across rows, and
minimisation of the sum of the differences between the items in the
$g$-th and $h$-th sub-tests is not the only criterion. To be precise,
we conduct methods of dichotomisation in which in Step-I we compute
the sum ${\cal S}$ of the differences between the item scores in the 2
sub-tests, as well as consider the sum ${\cal S}_{sq}$ of differences
of squares of the item scores in the $g$-th and $h$-th sub-tests. We
then identify row numbers $\rho^{(\star_1)}$ and $\rho^{(\star_2)}$ in
the above table, such that swapping the item in row $\rho^{(\star_1)}$
of the $g$-th sub-test with the entry in row $\rho^{(\star_2)}$ of the
$h$-th sub-test results in a reduction of $\vert {\cal S}\vert\times
\vert{\cal S}_{sq}\vert$; $\rho^{(\star_1)}, \rho^{(\star_1)}
\leq n/2$. When this product can no longer be reduced over a chosen
number of iterations, the scheme is stopped, otherwise the search for
the parallel halves that result in the minimum value of this product,
is maintained. However, parallelisation obtained with this method of
splitting that seeks to minimise $\vert {\cal S}\vert\times\vert{\cal
S}_{sq}\vert$ was empirically found to yield similar results as with
the method that seeks to minimise $\vert {\cal S}\vert$. Here by
``similar results'' is implied dichotomisation of the given test into
two sub-tests, the means and variances of which are equally close to
each other in both methods, i.e. sub-tests are approximately equally
parallel. The reason for such empirically noted similarity is
discussed in the following section. Given this, we advance the method
enumerated above in Steps-I to III, as our chosen method of
dichotomisation.
By the mean (or variance) of a sub-test, is implied mean (or variance)
of the vector of examinee scores in the $n/2$ items that comprise that
sub-test, i.e. the mean (or variance) of the score vector of the
sub-test. Thus, if the item numbers $g_1,\ldots,g_{n/2}$ comprise the
$g$-th sub-test, $g_j\in\{1,2,\ldots,n\},\:j=1,\ldots,n/2$, then the
$N$-dimensional score vector of the $g$-th sub-test is
$\boldsymbol{X}^{(g)}:=(X_1^{(g)}, \ldots, X_N^{(g)})^T$, where
$X_i^{(g)}=\displaystyle{\sum\limits_{j=1}^{{n/2}} X_i^{(g_j)}}$, i.e. the score acieved by the $i$-th examinee across all the items that are included in the $g$-th sub-test; this is to be distinguished from $\tau_j$--the score in the $j$-th item, summed over all examinees. Given that $X_i^{(j)}$ is either 0 or 1, we get that $X_i^{(g)}\leq {n/2}$ and $\tau_j\leq N,\:\forall\:i, j$. We
similarly define the score-vector $\boldsymbol{X}^{(h)}$ of the $h$-th sub-test.
Also, in he following section we identify score in the $j$-th item
that is constituent of the $g$-th sub-test as $\tau_j^{(g)},
j=1\ldots,n/2$, Similarly we define $\tau_j^{(h)}$.
\begin{remark}
\label{remark:remark1}
The order of our algorithm is independent of the examinee number and driven by the number of items in each sub-test that the test of $n$ items is dichotomised into; the order is ${\cal O}((n/2)^2)$.
\end{remark}
\subsection{Benefits of our method of parallelisation}
\label{sec:benefits_parallel}
\noindent
\begin{theorem}
\label{theorem:means}
Difference between means of $g$-th and $h$-th sub-tests is minimised.
\end{theorem}
\begin{proof}
Let the score vectors of the $g$-th and $h$-th sub-tests be
$\boldsymbol{X}^{(g)}:=(X_1^{(g)}, \ldots, X_N^{(g)})^T$ and
$\boldsymbol{X}^{(h)}:=(X_1^{(h)}, \ldots, X_N^{(h)})^T$,
where
$X_i^{(\cdot)}=\displaystyle{\sum\limits_{j=1}^{{n/2}} X_i^{(\cdot_j)}}.$
Now, mean of the $g$-th sub-test is
${\bar{X}}_g=\displaystyle{\frac{\sum\limits_{i=1}^{N} X_i^{(g)}}{N}}$
and mean of the $h$-th sub-test is
${\bar{X}}_h=\displaystyle{\frac{\sum\limits_{i=1}^{N} X_i^{(h)}}{N}}$.
Let item score of the $j$-th item in the $g$-th sub-test be $\tau_j^{(g)}$, $j=1,\ldots,n/2$. Similarly, let score of $j$-th item in the $h$-th sub-test be $\tau_j^{(h)}$.
But, sum of item scores in all $n/2$ items in a sub-test, is equal to sum of examinee scores achieved in these $n/2$ items, i.e.
\begin{eqnarray}
\displaystyle{\sum\limits_{i=1}^{N} X_i^{(g)}} &= & \displaystyle{\sum\limits_{j=1}^{n/2} \tau_j^{(g)}}\nonumber\\
\displaystyle{\sum\limits_{i=1}^{N} X_i^{(h)}} &= & \displaystyle{\sum\limits_{j=1}^{n/2} \tau_j^{(h)}}
\label{eqn:item-ex}
\end{eqnarray}
Now, by definition,
\begin{eqnarray}
\vert{\cal S}\vert &=& \displaystyle{{\Big\vert} \left[\tau_1^{(g)}-\tau_1^{(h)}\right] + \ldots + \left[\tau_{n/2}^{(g)}-\tau_{n/2}^{(h)}\right]{\Big\vert}} \\ \nonumber
&=& \displaystyle{{\Big\vert}\left[\tau_1^{(g)}+\ldots+\tau_{n/2}^{(g)}\right] -
\left[\tau_1^{(h)}+\ldots+\tau_{n/2}^{(h)}\right]{\Big\vert}}. \nonumber
\end{eqnarray}
At the end of the splitting of the test, let $\vert{\cal S}\vert=\epsilon$. Then $\epsilon$ is the minimum value of $\vert{\cal S}\vert$ by our method.
Then using
$$\displaystyle{{\Big\vert}\left[\tau_1^{(g)}+\ldots+\tau_{n/2}^{(g)}\right] -
\left[\tau_1^{(h)}+\ldots+\tau_{n/2}^{(h)}\right]{\Big\vert}} =\epsilon$$
in Equations~\ref{eqn:item-ex}, we get
\begin{equation}
\displaystyle{{\Big\vert}\sum\limits_{i=1}^{N} X_i^{(g)} - \sum\limits_{i=1}^{N} X_i^{(h)}{\Big\vert}}
= \displaystyle{{\Big\vert}\sum\limits_{j=1}^{n/2} \tau_j^{(g)} - \sum\limits_{j=1}^{n/2} \tau_j^{(h)}{\Big\vert}} = \epsilon
\label{eqn:inside}
\end{equation}
Then ${\Big\vert}{\bar{X}}_g-{\bar{X}}_h{\Big\vert} = \displaystyle{{\Bigg\vert}\frac{\sum\limits_{i=1}^{N} X_i^{(g)}}{N} -\frac{\sum\limits_{i=1}^{N} X_i^{(h)}}{N}{\Bigg\vert}}= \epsilon/N$ (using Equation~\ref{eqn:inside}), i.e. difference between means is minimised by our method.
\end{proof}
\vspace{1cm}
\begin{remark}
In our work, by ``near-equal means'', we imply means of the sub-tests, the difference between which is minimised using our method; typically, this minimum value is close to 0. Thus, the means of the two sub-tests, are near-equal.
\end{remark}
\vspace{1cm}
\begin{theorem}
\label{theorem:vars}
Absolute difference between sums of squares of examinee scores in the $g$-th and $h$-th sub-tests is of the order of $\epsilon^2$, if absolute difference between sums of examinee scores is $\epsilon$.
\end{theorem}
\begin{proof}
Absolute difference between sums of scores $g$-th and $h$-th sub-tests
is minimised in our method (Equation~\ref{eqn:inside}), with
\begin{equation}
\displaystyle{\sum\limits_{i=1}^{N} X_i^{(g)}}= \displaystyle{\sum\limits_{i=1}^{N} X_i^{(h)}}\pm\epsilon. \nonumber
\end{equation}
For our method, $\epsilon$ is small. Thus we state:
\begin{equation}
\displaystyle{\sum\limits_{i=1}^{N} X_i^{(g)}} \approx \displaystyle{\sum\limits_{i=1}^{N} X_i^{(h)}},
\label{eqn:basic}
\end{equation}
so that $T:=\displaystyle{\sum\limits_{i=1}^{N} X_i^{(g)}}\Longrightarrow \displaystyle{\sum\limits_{i=1}^{N} X_i^{(h)}}=T \mp \epsilon.$
Here we define
$$X_i^{(g)}=\displaystyle{\sum\limits_{j=1}^{n/2} X_i^{(g_j)}}$$
and similarly, we define $X_i^{(h)}$, $\forall i=1,\ldots,N$.
Now,
\begin{eqnarray}
\displaystyle{\left(\sum\limits_{i=1}^{N} X_i^{(g)}\right)^2 }
&=& \displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(g)}\right)^2}+\nonumber \\
&& \displaystyle{X_1^{(g)}X_2^{(g)} + \ldots + X_1^{(g)}X_N^{(g)} +} \nonumber \\
&& \ldots + \nonumber \\
&& \displaystyle{X_N^{(g)}X_1^{(g)} + \ldots + X_N^{(g)}X_{N-1}^{(g)}} \nonumber \\
&=& \displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(g)}\right)^2+
\sum\limits_{i=1}^{N}\left[\sum\limits_{k=1; k\neq i}^{N} X_i^{(g)}X_k^{(g)}\right]} \nonumber \\
\label{eqn:1}
\end{eqnarray}
Now, for $k\neq i$,
\begin{eqnarray}
X_i^{(g)}X_k^{(g)}&=& \displaystyle{\left(X_i^{(g_1)}+\ldots+X_i^{(g_{n/2})}\right)\left(X_k^{(g_1)}+\ldots+X_k^{(g_{n/2})}\right)}\nonumber \\
&=& \displaystyle{X_i^{(g_1)}X_k^{(g_1)} + X_i^{(g_1)}X_k^{(g_2)} +\ldots + X_i^{(g_1)}X_{k}^{(g_{n/2})}+} \nonumber \\
&&\ldots +\nonumber \\
&& \displaystyle{X_i^{(g_{n/2})}X_k^{(g_1)} + X_i^{(g_{n/2})}X_k^{(g_2)} +\ldots + X_i^{(g_{n/2})}X_{k}^{(g_{n/2})}+} \nonumber \\
&=& \displaystyle{\sum\limits_{j=1}^{n/2}\left[\sum\limits_{j'=1}^{n/2}\left(\mbox{number of times each of $X_i^{(j)}$ and $X_k^{(j')}$ is 1}\right)\right]}\nonumber \\
&=& \displaystyle{\left[\sum\limits_{j=1}^{n/2}\left(\mbox{number of times $X_i^{(g_j)}=1$}\right)\right] \left[\sum\limits_{j=1}^{n/2}\left(\mbox{number of times $X_k^{(g_j)}=1$}\right)\right]}\nonumber \\
&\approx& (n/2)^2 \displaystyle{\left[\sum\limits_{j=1}^{n/2} \Pr(X_i^{(g_j)}=1)\right]\left[\sum\limits_{j=1}^{n/2} \Pr(X_k^{(g_j)}=1)\right]}
\label{eqn:2}
\end{eqnarray}
Here $X_i^{(g_j)}$ is the score of the $i$-th examinee in the $j$-th item of the $g$-th sub-test and can attain values of either 1 or 0, with probability $p_i^{(g_j)}$ or 1-$p_i^{(g_j)}$ respectively. Thus,
$X_i^{(g_j)}\sim{\textrm{Bernoulli}}(p_i^{(g_j)})$, i.e. $\Pr(X_i^{(g_j)}=1)=p_i^{(g_j)}$. The approximation in Equation~\ref{eqn:2} stems from approximating the probability for an event, with its relative frequency.
Then following Equation~\ref{eqn:2}, we get
\begin{eqnarray}
\displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N} X_i^{(g)}X_k^{(g)}}
&\approx & (n/2)^2 \displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N}
\left[\sum\limits_{j=1}^{n/2} p_i^{(g_j)}\sum\limits_{j=1}^{n/2} p_k^{(g_j)}\right]}\nonumber \\
&=& (n/2)^2 \displaystyle{\sum\limits_{i=1}^{N}\left[
\left(\sum\limits_{j=1}^{n/2} p_i^{(g_j)}\right)\left(\sum\limits_{k=1; k\neq i}^{N}\sum\limits_{j=1}^{n/2} p_k^{(g_j)}\right)\right]}
\label{eqn:3}
\end{eqnarray}
Now, Equation~\ref{eqn:basic} implies that
\begin{eqnarray}
\displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} X_i^{(g_j)}\right]} &\approx& \displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} X_i^{(h_j)}\right]}\quad{\mbox{i.e.}}\nonumber \\
\displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} \left(\mbox{number of times $X_i^{(g_j)}=1$}\right)\right]} &\approx& \displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} \left(\mbox{number of times $X_i^{(h_j)}=1$}\right)\right]}\quad{\mbox{i.e.}}\nonumber \\
\displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} p_i^{(g_j)}\right]} &\approx& \displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} p_i^{(h_j)}\right]}
\label{eqn:4}
\end{eqnarray}
Then if we delete any 1 out of the $N$ examinees, over which the outer summation on the RHS and LHS of the last approximate equality is carried out, it is expected that the approximation expressed in statement~\ref{eqn:4} would still be valid. This is especially the case if $N$ is large. In other words, bigger the $N$, smaller is the distortion affected on the structure of the sub-tests generaetd by splitting the test data obtained after deleting the score of any 1 of the $N$ examinees from the original full test data. Then using statement~\ref{eqn:4} for a large $N$, we can write
\begin{equation}
\displaystyle{\sum\limits_{k=1, k\neq i}^{N}\left[\sum\limits_{j=1}^{n/2} p_k^{(g_j)}\right]} \approx \displaystyle{\sum\limits_{k=1;k\neq i}^{N}\left[\sum\limits_{j=1}^{n/2} p_k^{(h_j)}\right]}
\label{eqn:5}
\end{equation}
where
$$\displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} p_i^{(g_j)}\right]}= \displaystyle{\sum\limits_{i=1}^{N}\left[\sum\limits_{j=1}^{n/2} p_i^{(h_j)}\right] \pm \epsilon'},$$
follows from Equation~\ref{eqn:basic} that suggests that $\displaystyle{\sum\limits_{i=1}^{N} X_i^{(g)}} = \displaystyle{\sum\limits_{i=1}^{N} X_i^{(h)}}\pm \epsilon$. Thus $\epsilon\in{\mathbb Z}_{\geq 0}$ and $\epsilon'\in{\mathbb R}_{\geq 0}$ such that $\epsilon \geq \epsilon'$, given that $\epsilon'$ is the absolute difference between sums of probability of correct response in the $g$-th sub-test and that in the $h$-th sub-test while $\epsilon$ is the absolute difference between sums of scores in the two sub-tests.
In other words, if we define the sum of probabilities of correct
response in the $g$-th sub-test, $T'$, as
$$T' := \displaystyle{\sum\limits_{k=1, k\neq i}^{N}\left[\sum\limits_{j=1}^{n/2} p_k^{(g_j)}\right]},$$ then
\begin{equation}
\displaystyle{\sum\limits_{k=1, k\neq i}^{N}\left[\sum\limits_{j=1}^{n/2} p_k^{(h_j)}\right]}\approx T'\mp \epsilon'.
\label{eqn:7}
\end{equation}
Using this, for sub-test $g$, in the last line of Equations~\ref{eqn:3} we get
\begin{equation}
\displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N} X_i^{(g)}X_k^{(g)}} = (n/2)^2 T'\displaystyle{\sum\limits_{i=1}^{N}\left[
\left(\sum\limits_{j=1}^{n/2} p_i^{(g_j)}\right)\right]}.
\label{eqn:6}
\end{equation}
For sub-test $h$,
$$
\displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N} X_i^{(h)}X_k^{(h)}} \approx (n/2)^2 T'\displaystyle{\sum\limits_{i=1}^{N}\left[
\left(\sum\limits_{j=1}^{n/2} p_i^{(h_j)}\right)\right]},
$$
where the approximation in the above statement is of the order as in statement~\ref{eqn:6}, enhanced by $\mp(n/2)^2 T'\epsilon'$, following statement~\ref{eqn:7}. Then
\begin{equation}
\displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N} X_i^{(g)}X_k^{(g)}} \approx
\displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N} X_i^{(h)}X_k^{(h)}},
\label{eqn:8}
\end{equation}
where the approximation is of the order of $(n/2)^2 T'\epsilon'$.
Since $\displaystyle{\left(\sum\limits_{i=1}^{N} X_i^{(g)}\right)^2 }- \displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(g)}\right)^2}=\displaystyle{\sum\limits_{i=1}^{N}\sum\limits_{k=1; k\neq i}^{N} X_i^{(g)}X_k^{(g)}}$, as in the last line of Equations~\ref{eqn:1}--statement~\ref{eqn:8} tells us that for the $g$-th and the $h$-th sub-tests,
\begin{equation}
\displaystyle{\left(\sum\limits_{i=1}^{N} X_i^{(g)}\right)^2 }- \displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(g)}\right)^2} \approx
\displaystyle{\left(\sum\limits_{i=1}^{N} X_i^{(h)}\right)^2 }- \displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(h)}\right)^2},
\label{eqn:lastbut1}
\end{equation}
where, as for statement~\ref{eqn:8}, the approximation is of the order of $(n/2)^2 T'\epsilon'$.
But by squaring both sides of Equation~\ref{eqn:basic} we get $\displaystyle{\left(\sum\limits_{i=1}^{N} X_i^{(g)}\right)^2}\approx \displaystyle{\left(\sum\limits_{i=1}^{N} X_i^{(g)}\right)^2}$, where the approximation is of the order of $\epsilon^2\mp 2 T\epsilon$. Then in statement~\ref{eqn:lastbut1} we get
\begin{equation}
\displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(g)}\right)^2} \approx
\displaystyle{\sum\limits_{i=1}^{N}\left(X_i^{(h)}\right)^2},
\end{equation}
the approximation in which is of the order of $\epsilon^2\mp 2T\epsilon \pm (n/2)^2 T'\epsilon'$, i.e. of the order of $\epsilon^2$.
\end{proof}
\begin{remark}
{It is to be noted in the proof of Theorem~\ref{theorem:vars} that even if the sum of scores of the two sub-tests are equal, as per our method of splitting, i.e. even if $\epsilon$ is achieved to be 0, absolute difference between sums of squares of scores in the two sub-tests is not necessarily 0, since $T$ and $(n/2)^2 T'$ are not necessarily equal.}
\end{remark}
Now, variance of sub-tests $g$ and $h$ is respectively,
$\displaystyle{\sum\limits_{i=1}^{N}
\frac{\left(X_i^{(g)}\right)^2}{N} - {\bar{X}}_g^2}$ and
$\displaystyle{\sum\limits_{i=1}^{N}
\frac{\left(X_i^{(h)}\right)^2}{N} - {\bar{X}}_h^2}$. Then near-equality of sums of squares of sub-test scores, (Theorem~\ref{theorem:vars}) and near-equality of means (Theorem~\ref{theorem:means}) imply that the difference between sub-test variances is small.
Empirical confirmation of
near-equality of sub-test variances is presented in later sections.
Now item scores manifest item difficulty. Therefore, splitting using
item scores is equivalent to splitting using item difficulty values.
From the near-equality of variances it follows that, if instead of
allocating items into the two sub-tests, on the basis of the
difficulty value of the item, we split the test according to item
variance, we would get nearly the same sub-tests. Thus our method of
dichotomisation with respect to item scores (or difficulty values), is
nearly equivalent to dichotomisation with respect to item variance.
The near-equality of means and variances of the sub-tests are
indicative of the sub-tests being nearly parallel, since parallel
tests have equal means and variances. Then the item scores of the two
sub-tests can be taken as coming from nearly the same populations with
nearly same density functions having two parameters, namely mean and
variance.
For parallel sub-tests, variances are equal, and the regression of the
item score vector of the $g$-th sub-test on the $h$-th, coincides with
the regression of the scores of the $h$-th sub-test on that of the
$g$-th; such coinciding regression lines imply that the Pearson
product moment correlation $r_{gh}$ between $\{X_i^{(g)}\}_{i=1}^N$
and $\{X_i^{(h)}\}_{i=1}^N$, or the split-half regression
coefficient $r_{gh}$, is maximal, i.e. unity \cite{stepniak_2009}. Our
method of splitting the test results in nearly parallel sub-tests, so
that the split-half regression is close to, but less than unity. The
closer $\vert{\cal S}\vert$ is to 0, the more parallel are the
resulting sub-tests $g$ and $h$, and higher is the value of the
attained $r_{gh}$. Thus, the method gives a simple way of splitting a
test in halves, while ensuring that the split-half reliability
$r_{tt}$ is maximum.
The iterative method also ensures that the two sub-tests are almost
equi-correlated with a third variable. Thus, these (near-parallel)
sub-tests will have almost equal validity.
The problem of splitting an even number of non-negative item scores
into two sub-tests with the same number of items, such that absolute
difference between sums of sub-test item scores is minimised, is a
simpler example of the partition problem that has been addressed in
the literature \cite[among others]{hayes,borgs_2001}. Detailed
discussion of these methods is outside the scope of this paper, but
within the context of test dichotomisation aimed at computing
reliability, we can see that the method of assigning even numbered
items to one sub-test and odd-numbered ones to another as a method of
splitting a test into two halves, cannot yield sub-tests that are as
parallel as the sub-tests achieved via our method of splitting,
because an odd-even assignment does not guarantee that the sum of item
scores in one sub-test approaches that of the other closely, unless
the difficulty value of all items are nearly equal. Thus, a test in
which consecutive items are of widely different difficulty values,
will yeld sub-tests with means that are far from each other, if the
test is dichotomised using an odd-even method of assignment of items
to the sub-tests; in comparison, our method of splitting is designed
to yield sub-tests with close values of variances and even closer
means. Following on from the low order of our algorithm (see
Remark~\ref{remark:remark1}), a salient feature of our method is that
it is a fast method, the order of which depends on the number of items
in the test and not the examinee number, thus allowing for fast
splitting of the test score data and consequently, fast computation of
reliability.
\subsection{Estimation of True Score}
\label{sec:true_score}
\noindent
When the reliability is computed according to the
method laid out above, it is possible to perform interval estimation
of true score of an individual whose obtained score is known, where
the interval itself represents the ±1 standard deviation of the
distribution of the error score $E$. In other words, the
distribution of the error score $E$ is assumed--normal with zero mean
and variance $S_E^2 = S_X^2 (1-r_{tt})$ (as per the classical
definition of reliability); using the error variance, errors on the
estimated true scores are given as ±$S_E$. Below we discuss this
choice of model for the interval on the estimate of the true scores.
At a given observed score $X$, the true score can be estimated using the linear regression:
\begin{equation}
\hat{T}=\alpha_1 + \beta_1 X
\label{eqn:te}
\end{equation}
where the regression coefficients are
$\beta_1=r_{XT}\displaystyle{\frac{S_T}{S_X}}=r_{tt}$ and
$\alpha_1={\bar{X}}(1-\beta_1)$.
That the error on the estimated true score is given by $\pm S_E$, is
indeed a model choice \cite{klinebook} and is in fact, an
over-estimation given that the error $\varepsilon_i$ of estimation is
$$\varepsilon_i^2=(\alpha_1 + \beta_1 X_i - X_i)^2 =
(\bar{X}(1-r_{tt}) + r_{tt} X_i - X_i)^2=
[(X_i-\bar{X})(1-r_{tt})]^2$$ (recalling that
$\alpha_1=\bar{X}(1-r_{tt})$ and $\beta_1=r_{tt}$). Then
$$\displaystyle{\frac{\sum\limits_{i=1}^{N} \varepsilon_i^2}{N}}=(1 -
r_{tt})^2 S_X^2 \leq (1 - r_{tt}) S_X^2 = S_E^2,$$ i.e. $S_E^2 -
\displaystyle{\frac{\sum\limits_{i=1}^{N} \varepsilon_i^2}{N}}= (1 -
r_{tt}) S_X^2 - (1 - r_{tt})^2 S_X^2=r_{tt} (1 - r_{tt}) S_X^2$. In
other words, the standard error of prediction of the true score from a
linear regression model, falls short of the error variance $S_E^2$ by
the amount $r_{tt}(1- r_{tt})S_X^2$. Thus, the higher the reliability
$r_{tt}$ of the test, the less is the difference between the standard
error of prediction and our model of the error (on ${\hat{T}}$). In
general, our model choice of the error on ${\hat{T}}$ is higher than the standard error of prediction, for a given $X$.
Using ${\hat{T}}$ of each individual taking the test, one may
undertake computation of the probability that the percentile true
score of the $i$-th examinee is $t$ given the observed percentile
score of the examinee is $x$ and reliability is $r$, i.e. $\Pr(T \leq
t\vert r_{tt}=r, X\leq x)$. We illustrate this in our analysis of
real and simulated test score data.
In the context of estimating true scores using a computed reliability, we realise that using the split-half reliability $r_{gh}$ to estimate true scores ${\widehat{T}}_{split-half}$ will be in excess of the estimate ${\widehat{T}}_{classical}$ obtained by using the classically defined reliability $r_{tt}$, for high values of the observed scores, and under-estimated compared to ${\widehat{T}}_{classical}$ for low values of $X$. This can be easily realised.
\begin{theorem}
\label{theorem:over}
${\widehat{T}}_{split-half} \geq {\widehat{T}}_{classical}$ for $X > {\bar{X}}$ and ${\widehat{T}}_{split-half} \leq {\widehat{T}}_{classical}$ for $X < {\bar{X}}$.
\end{theorem}
\begin{proof}
\begin{eqnarray}
{\widehat{T}}_{split-half} - {\widehat{T}}_{classical} &=& [{\bar{X}} + r_{gh}(X-{\bar{X}})] - [{\bar{X}} + r_{tt}(X-{\bar{X}})]\\ \nonumber
&=& (X - {\bar{X}})(r_{gh} - r_{tt}) \\ \nonumber
\geq 0 \quad {\mbox{for}}\quad X > {\bar{X}} &{\mbox{and}}& \leq 0 \quad {\mbox{for}} \quad X < {\bar{X}}
\end{eqnarray}
given that $\quad r_{gh} \geq r_{tt}$.\\
Therefore ${\widehat{T}}_{split-half} \geq {\widehat{T}}_{classical}$ for $X \geq {\bar{X}}$ and ${\widehat{T}}_{split-half} \leq {\widehat{T}}_{classical}$ for $X \leq {\bar{X}}$.
\end{proof}
\section{Reliability of a battery}
\noindent
The above method of finding reliability, as per the classical
definition, can be extended to compute the reliability of a battery of
tests. Battery scores of a set of examinees are obtained as a function
of the scores of the constituent tests. After administration of the
battery to $N$ individuals, $S_X^2$, $S_T^2$, $S_E^2$ and $r_{tt}$ of
each constituent test are known as per the method described above. The
method of obtaining the reliability of the battery will depend on
these parameters as well as on the definition of the battery
scores. Usually, battery scores are computed as the sum of the scores
in the individual constituent tests (summative scores) or as a
weighted sum of these test scores. However, the possibility of a
non-linear combination of test scores to depict battery scores cannot
be ruled out. Here we discuss the weighted sum of component test scores
after motivating the concept of summative scores; we illustrate our
method of finding the weights on real data
(Section~\ref{sec:real}).
\subsection{Weighted sum of component tests}
\label{sec:weights}
\noindent
Suppose a battery consists of two tests: Test-1 and Test-2. Let the
$i$-th examinee's scores in the 2 constituent tests be $X_{1i}$ and
$X_{2i}$. Let this examinee's true scores and error score of the
$j$-th constituent test be $T_{ji}$ and $E_{ji}$ respectively,
$j=1,2$. Then the summative score of the $i$-th examinee is $X_i =
X_{1i} + X_{2i}$ assuming additivity and equal weights. Let
$r_{tt(1)}$ and $r_{tt(2)}$ denote reliability of the respective
constituent tests.
Now,
\begin{equation}
{\textrm var}(X) = {\textrm var}(X_1) + {\textrm var}(X_2) + 2{\textrm cov}(X_1, X_2)
\label{eqn:varcov}
\end{equation}
suggests
\begin{eqnarray}
\displaystyle{\sum X_{1i}X_{2i}} &=& \displaystyle{\sum(T_{1i} + E_{1i}) (T_{2i} + E_{2i})} \nonumber \\
& =& \displaystyle{\sum T_{1i}T_{2i}} \nonumber \\
\end{eqnarray}
since $\displaystyle{\sum T_{1i}E_{2i}}=\displaystyle{\sum T_{2i}E_{1i}}=\displaystyle{\sum E_{1i}E_{2i}}=0$.
Again
\begin{eqnarray}
{\textrm cov}(X_1, X_2) &=& \displaystyle{\frac{\sum X_{1i}X_{2i}}{N} - {\bar X}_1 {\bar X}_2 }\nonumber \\
&=& \displaystyle{\frac{\sum T_{1i}T_{2i}}{N} - {\bar T}_1 {\bar T}_2 }\nonumber \\
&=&{\textrm cov}(T_1, T_2)
\label{eqn:t1t2}
\end{eqnarray}
Now variance of true score of the battery is
\begin{eqnarray}
S_{T(Battery)} &=& {\textrm var}(T_1) + {\textrm var}(T_2) + 2{\textrm cov}(T_1, T_2)\nonumber \\
&=& {\textrm var}(T_1) + {\textrm var}(T_2) + 2{\textrm cov}(X_1, X_2)\nonumber \\
&=& r_{tt(1)}S_{X_1} + r_{tt(2)}S_{X_2} + 2{\textrm cov}(X_1, X_2)
\label{eqn:rel_bat}
\end{eqnarray}
where we have used Equation~\ref{eqn:t1t2}. Thus, reliability of a battery can be found using Equation~\ref{eqn:rel_bat}. Using these equations,
reliability of the battery $r_{tt(battery)}$ is given by
\begin{equation}
r_{tt(battery)} = \displaystyle{\frac{r_{tt(1)}S_{X_1} + r_{tt(2)}S_{X_2} + 2{\textrm cov}(X_1, X_2)}{S_{X_1} + S_{X_2} + 2{\textrm cov}(X_1, X_2)}}
\end{equation}
In general, if a battery has $K$-tests and all tests are equally important then the reliability of the battery, where the battery score is a summative score of the $K$ constituent tests, will be given by
\begin{equation}
r_{tt(battery)} = \displaystyle{\frac{\sum_{i=1}^K r_{tt(i)}S_{X_i} +
\sum_{i=1, i\neq j}^K \sum_{j=1}^K 2{\textrm cov}(X_i, X_j)}{\sum_{i=1}^K S_{X_i} + \sum_{i=1, i\neq j}^K \sum_{j=1}^K 2{\textrm cov}(X_i, X_j)}}
\label{eqn:rel_bat_gen}
\end{equation}
We can extend this idea to weighted summative scores.
Consider a battery of $K$-tests with weights
$W_1, W_2, \ldots, W_k$ where $W_i > 0\forall\:i= 1,2,\ldots,k$ and
$\displaystyle{\sum_{i=1}^K W_i} = 1$. To find reliability of such a
battery, let us consider the composite score or the battery score $Y=
\displaystyle{\sum_{i=1}^K W_i X_i}$ where $X_i$ is the score of the
$i$-th constituent test. Clearly, ${\textrm{var}}(Y) =
\displaystyle{\sum_{i=1}^K W_i^2{\textrm{var}}(X_i)}$.
Then
\begin{equation}
r_{tt(battery)} = \displaystyle{\frac{\sum_{i=1}^K r_{tt(i)}W_i^2 S_{X_i} +
\sum_{i=1, i\neq j}^K \sum_{j=1}^K 2 W_i W_j {\textrm cov}(X_i, X_j)}{\sum_{i=1}^K W_i^2 S_{X_i} + \sum_{i=1, i\neq j}^K \sum_{j=1}^K 2W_i W_j {\textrm cov}(X_i, X_j)}}
\label{eqn:rel_bat_wt}
\end{equation}
However, a major problem in the computation of the reliability
coefficient in this fashion could be experienced in determination of the
weights.
\subsection{Our method for determination of weights}
\label{sec:wtdetermination}
\noindent
We propose a method towards the determination of the vector of
weights $\boldsymbol{W} = (W_1,W_2,\ldots,W_K)^T$ with
$\displaystyle{\sum_{i=1}^K W_i} = 1$, such that the variance of the
battery score is a minimum where the battery score vector is $\boldsymbol{Y}$
with ${\textrm{var}}(\boldsymbol{Y}) = \boldsymbol{W}^T \boldsymbol{D} \boldsymbol{W}$ and $\boldsymbol{D}$ is the
variance-covariance matrix of the component tests. From a single
administration of a battery to $N$ examinees, the variance-covariance
matrix $\boldsymbol{D}$ is such that the $i$-th diagonal element of this matrix
is the variance $S_{X_i}$ of the $i$-th constituent test, and the
$ij$-th off diagonal element is ${\textrm{cov}}(X_i,X_j)$. This is
subject to the condition $\boldsymbol{W}^T\boldsymbol{e}=1$ where $\boldsymbol{e}$ is the
$K$-dimensional vector, the $i$-th component of which is 1,
$\forall\:i=1,\ldots,K$. It is possible to determine the weights using
the method of Lagrangian multipliers $\lambda$. To this effect, we
define $A := \boldsymbol{W}^T \boldsymbol{D} \boldsymbol{W} + \lambda(1 - \boldsymbol{W}^T\boldsymbol{e})$ and set the
derivative of $A$ taken with respect to $\boldsymbol{W}$ and $\lambda$ to zero
each, to respectively attain:
\begin{equation}
2\boldsymbol{D}\boldsymbol{W} - \lambda\boldsymbol{e}= 0 \quad\mbox{and}\quad 1 - \boldsymbol{W}^T\boldsymbol{e} = 0.\nonumber
\end{equation}
Thus,
\begin{eqnarray}
\boldsymbol{W} = \displaystyle{\frac{\lambda}{2}{\boldsymbol{D}}^{-1}\boldsymbol{e}}\quad&\mbox{and}&\quad
\boldsymbol{W}^T\boldsymbol{e} = 1,\nonumber \\
\mbox{or}\quad \boldsymbol{W}= \displaystyle{
\frac{\boldsymbol{D}^{-1}\boldsymbol{e}}{\boldsymbol{e}^T\boldsymbol{D}^{-1}\boldsymbol{e}}
} \quad&\mbox{and}&\quad \lambda= \displaystyle{\frac{2}{\boldsymbol{e}^T\boldsymbol{D}^{-1}\boldsymbol{e}}}
\label{eqn:de}
\end{eqnarray}
\subsection{Benefits of this method of finding weights}
\noindent
This method of finding reliability of a battery of a number of tests
can be used in the context of summation score, difference score or
weighted sum of scores. Weights found as above have the advantage that
the battery score or composite score ($\boldsymbol{Y}$) has minimum variance.
Also, covariance between the battery score and the test score of an
individual test is a constant, i.e. ${\textrm{cov}}(Y_i,X_i)=
\displaystyle{\frac{1}{\boldsymbol{e}^T\boldsymbol{D}^{-1}\boldsymbol{e}}}$ $\forall\:i$. If the
available test scores are standardised and independent such that the
$i$-th score is $Z_i$, then weights are equal, and correlation between
$Y$ and $Z_i$ is the same as correlation between $Y$ and $Z_j$ =
$\displaystyle{\frac{1}{\sqrt{\boldsymbol{e}^T\boldsymbol{R}^{-1}\boldsymbol{e}}}}$ $\forall\:i=j$,
$i,j=1,2,\ldots,k$, where $\boldsymbol{R}$ is the correlation matrix. In other
words, the battery score is equi-correlated with the standardised
score of each constituent test. The method alone does not guarantee
that all weights be non-negative. This is especially true if the tests
are highly independent so that $\boldsymbol{R}$ is too sparse. Non-negative weights
can be ensured by adding the physically motivated constraint that $W_i
\geq 0$ $\forall\: i$, in which case, the problem boils down to a
Quadratic Programming formulation. This is similar to the data-driven
weight determination method presented by Chakrabartty (2013).
If it is found desirable to get weights so that $X_i$ is
proportional to ${\textrm{cov}}(Y, X_i)$ then it can be proved that
$\boldsymbol{W}$ is the eigenvector corresponding to the maximum eigenvalue of
the variance-covariance matrix of the test scores.
If on the other hand, we want $W_i$ as proportional to
$\displaystyle{\frac{{\textrm{cov}}(Y,X_i)}{{\textrm{var}}(X_i)}}$,
then $\boldsymbol{W} = \displaystyle{\frac{\boldsymbol{S}^{-1} U}{\boldsymbol{e}^T \boldsymbol{S}^{-1} U}}$
where $\boldsymbol{S}$ is the diagonal matrix of the standard deviations of the
constituent tests and $U$ is the maximum eigen-value of the
correlation matrix.
In order to find the battery scores, weights of test scores can be
found by principal component analysis, factor analysis, canonical
reliability, etc. However, it is suggested that reliability of the
composite scores be found as weighted scores
(Equation~\ref{eqn:rel_bat_wt}) where the weights are determined as per
Equation~\ref{eqn:de}.
So far, we have discussed additive models. We could also have
multiplicative models like $Y= \displaystyle{\prod X_i^{W_i}}$ so that
on a log-scale, we retrieve the additive model.
We illustrate our method of finding weights using real data in
Section~\ref{sec:real}.
\section{Application to simulated data}
\label{sec:simulated}
\noindent
In order to validate our method of computing the classically defined
reliability following dichotomisation of a test into parallel groups,
we use our method to find the value of $r_{tt}$ of 4 toy tests, the
scores of which are simulated from chosen models, as described
below. We simulate the 4 test data sets $D_1, D_2, D_3, D_4$ under 4
distinct model choices; the underlying standard model is that the score
$X_{i}^{(j)}$ is obtained by the $i$-th examinee to the $j$-th item in a test,
is a Bernoulli variate, i.e.
$$X_{i}^{(j)}\sim{\mbox{Bernoulli}}(p)\quad {\mbox{implying}}\quad\Pr(X_{i}^{(j)}=1)=p,\:\:\Pr(X_{i}^{(j)}=0)=1-p$$
where the probability $Pr(X_{i}^{(j)}=1)$, of answering the $j$-th item correctly by the $i$-th examinee
\begin{itemize}
\item is held a constant $p_i$ for the $i$-th examinee $\forall\:j=1,\ldots,n$, with $p_i$ sampled from a uniform distribution in [0,1], $\forall\:i=1,\ldots,N$, in data $D_1$.
\item is held a constant $p_i$ for the $i$-th examinee $\forall\:j=1,\ldots,n$, with $p_i$ sampled from a Normal distribution ${\cal N}(0.5,0.2)$, $\forall\:i=1,\ldots,N$, in data $D_3$.
\item is held a constant $p_j$ for the $j$-th item for all examinees, with $p_j$ sampled from a uniform distribution in [0,1], $\forall\:j=1,\ldots,n$, in data $D_2$.
\item is held a constant $p_j$ for the $j$-th item for all examinees, with $p_j$ sampled from a Normal distribution ${\cal N}(0.5,0.2)$, $\forall\:j=1,\ldots,n$, in data $D_4$.
\end{itemize}
We use $n$=50 and $N$=999 in our simulations.
Thus we realise that data sets $D_1$ and $D_3$ resemble test data in
reality, with the ability of the $i$-th examinee represented by
$p_i$. Our simulation models are restrictive though in the sense that
variation with items is ignored. Such a test, if administered, will
not be expected to have a low reliability. On the other hand, the data
sets $D_2$ and $D_4$ are toy data sets that are utterly unlike real
test data, in which the probability of correct response to a given
item is a constant, irrespective of examinee ability, and examinee
ability varies with item--equally for all examinees. A toy test data
generated under such an unrealistic model would manifest low test
variance and therefore low reliability. Given this background, we
proceed to analyse these simulated tests with our method of computing
$r_{tt}$.
We implement our method to compute reliabilities of all 4 data sets (using Equation~\ref{eqn:4thstep}). The results are given in Table~1.
\begin{table}
\renewcommand\baselinestretch{1.5}{
\caption{Table showing results of using our method of dichotomisation
of 4 simulated test data sets, $D_1,\ldots,D_4$, into 2 parallel
sub-tests, $g$ and $h$ , resulting in the computation of
classically-defined reliability $r_{tt}$ of the test (see
Section~\ref{sec:dichot} for details of our method).}}
\label{table:table1}
\hspace*{-1cm}
{\footnotesize
\begin{tabular}{|c||c||c|c||c|c||c||c|}\hline
\hline
\multicolumn{1}{|c||}{Data set}& \multicolumn{1}{c||}{Test}& \multicolumn{2}{|c||}{Sum of item scores $X_g \& X_h$ in sub-tests} & \multicolumn{2}{|c||}{Sum of squares of item scores in sub-test} & \multicolumn{1}{c||}{$\sum_{i=1}^{N}X^{(g)}_iX^{(h)}_i$}& \multicolumn{1}{|c|}{Reliability}\\
& variance & $g$ & $h$ & $g$ & $h$ & & $r_{tt}$\\
\hline
${D_1}$ & 222.89 & 12014$\quad\quad\quad\quad$ & 12013 & 202506$\quad\quad\quad\quad$ & 201523 & 198257 & 0.9662 \\
${D_2}$ & 7.96 & 10440$\quad\quad\quad\quad$ & 10439 & 113212$\quad\quad\quad\quad$ & 112843 & 109133 & 0.02050 \\
${D_3}$ & 110.06 & 12597$\quad\quad\quad\quad$ & 12597 & 188727$\quad\quad\quad\quad$ & 189465 & 183564 & 0.8994 \\
${D_4}$ & 10.86 & 12683$\quad\quad\quad\quad$ & 12684 & 166199$\quad\quad\quad\quad$ & 166520 & 161130 & 0.0361\\
\hline
\end{tabular}
}
\vspace{-0.1in}
\end{table}
\begin{figure}[!h]
\centering
\vspace*{1cm}
\includegraphics[width=18cm]{simulated.ps}
\vspace*{-1cm}
\caption{\small{Figure showing
histograms of the scores of 999
examinees in the 2 sub-tests (in solid and broken lines
respectively), obtained by splitting the simulated test data set
$D_3$ which has been generated under the choice that examinee
ability is normally distributed (right panel). The left panel
includes histograms of the 2 sub-tests that result from splitting
the test data $D_4$ that was simulated using examinee
ability as item-dependent, with probability for correct response
to the $j$-th item given as a normal variate, unrealistically
fixed for all examinees, $\forall\:j=1,\ldots, 50$.}}
\label{fig:new}
\end{figure}
The reliabilities of tests with data $D_1$ and $D_3$ are as expected
high, while the same for the infeasible tests $D_2$ and $D_4$ are low,
as per prediction. We take these results as a form of validation our method
of computing $r_{tt}$ based on the splitting of a test.
In Figure~\ref{fig:new}, we present the histograms of the 2 sub-tests
that result from splitting the realistic data $D_3$ as well histograms
of the 2 sub-tests that result from the splitting of the unrealistic
test data $D_4$.
\subsection{Results from Large Simulated Tests}
\noindent
We also conducted a number of experiments with finding reliabilities of larger test data sets that were simulated. The simulations were performed such that the test score variable has a Bernoulli distribution with parameter $p$. in these simulations, we chose $p_i$ as fixed for the $i$-th examinee, with $p_i$ randomly sampled from a chosen Gaussian $pdf$, i.e. $p_i\sim{\cal N}(0.5, 0.2)$ by choice, $i=1,\ldots, N$. We simulated different test score data sets in this way, including
\begin{enumerate}
\item[--] a test data set for 5$\times$10$^5$ examinees taking a 50-item test.
\item[--] a test data set for 5$\times$10$^4$ examinees taking a 50-item test.
\item[--] a test data set for 1000 examinees taking a 100-item test.
\item[--] a test data set for 1000 examinees taking a 1000-item test.
\end{enumerate}
In each case, the test data was split using our method and reliability of the test was computed as per the classical definition. The 4 simulated test data sets mentioned above, yielded reliabilities of 0.96, 0.98, 0.93, 0.85, in order of the above enumeration. Histograms of the sub-tests obtained by splitting each test data were over-plotted to confirm their concurrence.
Importantly, the run-time of reliability computation of these large cohorts of examinees ($N$=500,000 and 50,000), who take the 50-item long test, is very short--from about 0.8 seconds for the 50,000 cohort to about 6.2 seconds for the 500,000 cohort. On the other hand the order of our splitting algorithm being ${\cal O}((n/2)^2)$, the run-times increase rapidly for the 1000-item test, from the 100-item one, with the fixed examinee number. These experiments indicate that the computation of reliabilities for very large cohorts of examinees, in a test with a realistic number of items, is rendered very fast indeed, using our method.
\section{Application to real data}
\label{sec:real}
\noindent
We apply our method of computing reliability using the classical
definition, following the splitting of a real test into 2 parallel
halves by the method discussed in Section~\ref{sec:dichot}. We also
estimate the true scores and offer the error in this estimation, using
the computed reliability and a real test score data set. The used real
test data was obtained by examining 912 examinees in a multiple choice
examination administered with the aim of achieving selection to a
position. The test had 50 items and maximum time given was 90
minutes. To distinguish between this real test data and other real
test data sets that we will employ to determine reliability of a
battery, we refer to the data set used in this section as DATA-I. For
the real test data DATA-I, the histograms of the item scores and that
of the examinee scores are shown in Figure~\ref{figure:histos}. The
histogram of the examinee scores shows that $X_i
\leq 26\:\forall i=1,\ldots,912$, with the biggest mode of the score
distribution around 11. In other words, the test was a low-scoring
one. DATA-I indicates 2 other smaller modes at about 15
and 22. Some other parameters of the test data DATA-I are as listed below.
\begin{enumerate}
\item[-] number of items $n$=50,
\item[-] number of examinees is $N$=912,
\item[-] magnitude of the vector of maximum possible scores is
$\parallel \boldsymbol{I} \parallel$ = $\sqrt{912\times 50^2}\approx 1509.97$,
\item[-] magnitude of the observed score vector is
$\parallel \boldsymbol{X} \parallel\approx$ 357.82,
\item[-] $\parallel \boldsymbol{X} \parallel^2$ = 128032,
\item[-] $\cos\theta_X\approx$0.9275
\item[-] test mean ${\bar X}\approx$10.99,
\item[-] test variance $S_X^2 = \approx 19.63$.
\end{enumerate}
\begin{figure}[!t]
\centering
\vspace*{1cm}
\includegraphics[width=18cm]{histos_new.ps}
\vspace*{-1cm}
\caption{\small{Figure showing histograms of the 912 examinee scores (left panel) and of 50 item scores (right panel) of the real test data DATA-I that we use to implement our method of splitting the test, aimed at computation of reliability as per the classical definition. The left panel also includes histograms of the $g$ and $h$ sub-tests that result from the splitting of this test; the histograms of the sub-tests are shown in dotted and dashed-dotted lines, superimposed on the histogram of examinee scores for the full test (in solid black line).}}
\label{figure:histos}
\end{figure}
\begin{figure}[!t]
\centering
\vspace*{-4cm}
\includegraphics[width=18cm]{regress.ps}
\vspace*{-14cm}
\caption{\small{Figure showing regression of the examinee scores in the sub-test $g$ on the scores in sub-test $h$ (right panel) and vice versa (left panel), where the 2 sub-tests are obtained by splitting real data DATA-I with our method of dichotomisation (Section~\ref{sec:dichot}). The score data are shown in filled circles while the regression lines are also drawn. The regression coefficients are marked on the top left hand corner of the respective panel.} }
\label{figure:regress}
\end{figure}
The test was dichotomised into parallel halves by the iterative
process discussed in Section~\ref{sec:dichot}. Let the resulting
sub-tests be referred to as the $g$ and $h$ sub-tests. Then each of
these sub-tests had 25 items in it and sum of examinee scores is
$\displaystyle{\sum\limits_{i=1}^{912} X_i^{(g)}}=
\displaystyle{\sum\limits_{i=1}^{912}
X_i^{(h)}}=5011$, so that the mean of each of the two sub-tests is equal to about 5.49. Also, variances of the 2 sub-tests are approximately equal at 6.81 and 6.49.
The histograms of the examinee scores in these 2 sub-tests are drawn
in blue and red in the left panel of Figure~\ref{figure:histos}. That the
histograms for the $g$ and $h$ sub-tests overlap very well, is
indicative of these sub-tests being strongly parallel. We regress the
score vector $(X_1^{(g)},\ldots,X_{912}^{(g)})^T$ on the score vector
$(X_1^{(h)},\ldots,X_{912}^{(h)})^T$; the regression lines are shown
in the right panel of Figure~\ref{figure:regress}. Similarly, regressing
$(X_1^{(h)},\ldots,X_{912}^{(h)})^T$ on the score vector
$(X_1^{(g)},\ldots,X_{912}^{(g)})^T$ results in a similar linear
regression line (shown in the left panel of
Figure~\ref{figure:regress}). Table~2 gives the details of splitting of
the 50 items of the full test into the 2 sub-tests.
\begin{table}
\caption{Splitting of real data DATA-I of a selection test with 50 items, administered to 912 examinees. The dichotomisation has been carried out according to the algorithm discussed in Section~\ref{sec:dichot}.}
\label{table:table1}
\begin{center}
{\footnotesize
\begin{tabular}{|c|c||c|c||c|}\hline
\multicolumn{2}{|c||}{$g$-th sub-test} & \multicolumn{2}{c||}{$h$-th sub-test} & \multicolumn{1}{c|}{Difference between scores of 2 tests}\\
\hline
Score & Item No. & Score & Item no. & \\
\hline
75 & 25 & 75 & 1 & 0\\
84 & 39 & 80 & 46 & 4\\
85 & 43 & 90 & 24 & -5\\
100 & 20 & 96 & 9 & 4\\
102 & 44 & 103 & 34 & -1\\
111 & 50 & 106 & 31 & 5\\
112 & 5 & 113 & 41 & -1\\
124 & 32 & 115 & 45 & 9\\
127 & 36 & 128 & 8 & -1\\
131 & 33 & 129 & 26 & 2\\
134 & 18 & 135 & 3 & -1\\
151 & 29 & 144 & 4 & 7\\
172 & 7 & 178 & 48 & -6\\
193 & 19 & 190 & 21 & 3\\
195 & 37 & 196 & 14 & -1\\
205 & 28 & 199 & 47 & 6\\
221 & 30 & 230 & 6 & -9\\
239 & 27 & 236 & 49 & 3\\
256 & 11 & 263 & 12 & -7\\
284 & 2 & 285 & 17 & -1\\
292 & 22 & 294 & 10 & -2\\
337 & 15 & 309 & 23 & 28\\
393 & 13 & 376 & 42 & 17\\
405 & 16 & 453 & 38 & -48\\
483 & 35 & 488 & 40 & -5\\
\hline
Sum of item scores & 5011 & & 5011 & \\
Sum of squares of item scores & 33453 & & 33747 & \\
\hline
\end{tabular}
}
\vspace{-0.1in}
\end{center}
\end{table}
\subsection{Computation of reliability}
\noindent
We implement the splitting of the real test score data into the scores
in the $g$ and $h$ sub-tests to compute the reliability as per the
theoretical definition, i.e. as given in
Equation~\ref{eqn:4thstep}. Then using the observed sub-test score
vectors $\boldsymbol{X}^{(g)}$ and $\boldsymbol{X}^{(h)}$ we get the classically defined reliability to be $r_{tt} \approx 0.66$. Then
$r_{XT}:=\sqrt{r_{tt}}\approx 0.8128$. On the other hand, the Pearson
product moment correlation between $\boldsymbol{X}^{(g)}$ and $\boldsymbol{X}^{(h)}$ is
$r_{gh}\approx$0.9941. In other words, the split-half reliability is
$r_{gh}\approx 0.9941$.
We can now proceed to estimate the true scores following the
discussion presented in Section~\ref{sec:true_score}. For the
observed score $X_i$, the estimated true score is
${\widehat{T}_i}=\beta_1 X_i + \alpha_1$ (as per Equation~\ref{eqn:te}
and discussions thereafter), where $\beta_1=r_{tt}$ and
$\alpha_1={\bar{X}}(1 - \beta_1)$. The true scores can also be
estimated using the split-half reliability instead of the reliability
from the classical definition. Then the above regression coefficients
$\beta_q$ and $\alpha_q$ are computed as above, except this time,
$r_{tt}$ is replaced by $r_{gh}$, $q=1,2$. We present the true scores
estimated using both the reliability from the classical definition
(${\widehat{T}}_{classical}$ in black) as well as the split-half
reliability (${\widehat{T}}_{split-half}$ in red) in
Figure~\ref{figure:t_est}. In this figure, the errors in the estimated
true scores are superimposed as error bars on this estimate, in
respective colours. This error is considered to be $\pm S_E$, where
the error variance is $S_E^2 = (1- r_{tt})S_X^2$ when the classically defined reliability is implemented and $S_E^2 = (1- r_{gh})S_X^2$ when the split-half reliability is used.
In fact, the method of using $r_{gh}$ in the estimation of the true
scores will result in the true scores ${\widehat{T}}_{split-half}$
being over-estimated for $X>{\bar{X}}$ and under-estimated for $X <
{\bar{X}}$, as shown in Theorem~\ref{theorem:over}. This can be easily
corroborated in our results in Figure~\ref{figure:t_est}; the higher
value of $r_{gh}$ (than of $r_{tt}$) results in
${\widehat{T}}_{split-half} \geq {\widehat{T}}_{classical}$ for
$X>{\bar{X}}$ and in ${\widehat{T}}_{split-half} \leq
{\widehat{T}}_{classical}$ for $X<{\bar{X}}$.
Figure~\ref{figure:t_est} also includes the sample probability
distribution of the point estimate of the true score obtained from the
linear regression model suggested in Equation~\ref{eqn:te}, given the
observed scores and using $r_{tt}$ and $r_{gh}$, i.e. $\Pr(T \leq t\vert
r, \boldsymbol{X})$, where $r$ is the reliability.
We invoke the lookup table that gives the examinee indices
corresponding to a point estimate of the true score. Then the
percentile rank of all those examinees whose true score is (point)
estimated as $T$, is given as $100\times\Pr(T \leq t\vert r,
\boldsymbol{X})$. (This lookup table is not included in the text for brevity's
sake). For example, the true score
${\widehat{T}}_{split-half}\in[20,21)$ is estimated using $r=r_{gh}$,
for examinees with indices 893, 867, 210, 837, 834, 408, 706, 690,
655, 653, 161, 638, 312, 308, 149, 290, 539, 260. Then the
percentile rank of all these examinees is about 93.
\begin{figure}[!t]
\centering
\includegraphics[width=14cm]{t_es.ps}
\vspace*{-.3cm}
\caption{\small{{\it Left}: figure showing the true scores estimated at observed examinee scores of the real test data DATA-I. The true scores estimated using the reliability computed from the classical definition is shown in black filled circles; the corresponding estimates of the error scores are superimposed as error bars. The true scores estimated using the split-half reliability are plotted in red crosses, as are the corresponding error scores. {\it Right}: the sample cumulative probability distributions of point estimate of true scores obtained from the linear regression model given in Equation~\ref{eqn:te} where implementation of the classically defined reliability results in the plot in black circles while that obtained using split-half reliability is in red crosses. This cumulative distribution can be implemented to identify the percentile rank of an examinee, in conjuction with the lookup table containing the ranking of examinee indices by the point estimate of the true score.}}
\label{figure:t_est}
\end{figure}
Thus, our method of splitting the test into 2 parallel halves helps to
find a unique measure of reliability as per the classical definition,
for a given real data set. Using this we can then estimate true scores
for each observed score in the data.
\subsection{Weighted battery score using real data}
\label{sec:weightedresults}
\noindent
In order to illustrate our method of finding weights relevant to the
computation of the classically defined weighted reliability of a test
battery (Equation~\ref{eqn:rel_bat_wt}), we employ real life test data sets
DATA-II(a) and DATA-II(b). These data sets comprise the examinee scores of
784 examinees who took 2 tests (aimed at selection for a position). The
test that resulted in DATA-II(a) aimed at measuring examinee
verbal ability, while the test for which DATA-II(b) was the result, measured
ability to interpret data. The former test contained 18 items and the latter 22
items. Histograms of examinee scores and item scores of these 2 tests
are shown in Figure~\ref{figure:histos_batt}. The mean and variance of
the 2 tests corresponding to DATA-II(a) and DATA-II(b) are
approximately 4.30, 6.25 and 4.20, 4.06 respectively. The battery
comprising these two tests is considered for its reliability.
\begin{figure}[!h]
\centering
\includegraphics[width=18cm]{histo_batt_new.ps}
\vspace*{-1cm}
\caption{\small{{\it Left:} figure showing histograms of the examinee scores in the tests, the scores of which constitute data sets DATA-II(a) (in black) and DATA-II(b) (in red dotted line). {\it Right:} figure showing histograms of item scores in real data DATA-II(a) (in black solid lines) and in DATA-II(b) (in red broken lines).} }
\label{figure:histos_batt}
\end{figure}
We split each test into 2 parallel halves using our method of
dichotomisation delineated in Section~\ref{sec:dichot}. Splitting
DATA-II(b) results in 2 halves, the means of the examinee scores of
which are about 2.10 and the variances of the examinee scores are
about 1.80 and 1.65. Reliability defined classically
(Equation~\ref{eqn:4thstep}) computed using these split halves is
about 0.35. The correlation between the split halves is about 0.90. On
the other hand, splitting DATA-II(a) results in 2 halves, the means of
which are equal to about 2.15 and the variances of which are about
1.87 and 2.19. Classically defined reliability of this test using
these split halves, turns out be 0.66 approximately. The correlation
coefficient between the two halves is about 0.95. The true scores
estimated for both observed data sets are shown in
Figure~\ref{figure:batt}, with errors of estimation superimposed as
$\pm S_E$ where $S_E^2$ is the variance of the error scores that are
modelled as distributed as ${\cal N}(0, S_E^2)$.
\begin{figure}[!ht]
\centering
\includegraphics[width=10cm]{batt_new.ps}
\vspace*{-1cm}
\caption{\small{Figure showing true scores estimated using the observed scores in real data set DATA-II(a) (in black filled circles) and in DATA-II(b) (in red crosses). Errors of estimate are depicted as error bars that correspond to $\pm (1-r_{tt})^2S_X^2$ where the $r_{tt}$ is the classically defined reliability computed using the respective real data set.}}
\label{figure:batt}
\end{figure}
Using the scores in the 2 tests in the considered battery, namely sets DATA-II(a) and DATA-II(b), we compute the variance-covariance matrix $\boldsymbol{D}$ of the observed scores (see Equation~\ref{eqn:de}). Then in this example, $\boldsymbol{D}$ is a 2$\times$2 matrix, the diagonal elements of which are the variances of the 2 tests and the off-diagonal elements are the covariances of the examinee test score vectors $\boldsymbol{X}_1$ and $\boldsymbol{X}_2$ of these 2 tests. Then in this example real-life test battery, recalling that $\boldsymbol{e} = \displaystyle{\left(1, 1\right)^T}$ we get\\
$$ \boldsymbol{D}\approx\left( \begin{array}{cc}
6.24 & 2.46 \\
2.46 & 4.06 \end{array} \right),\quad\mbox{so that}\quad \boldsymbol{D}^{-1}\boldsymbol{e} \approx \displaystyle{\left(0.08, 0.20\right)^T},\quad \boldsymbol{e}^T\boldsymbol{D}^{-1}\boldsymbol{e}\approx 0.28.$$
Then the weights are $\approx$ 0.7028 and 0.2972 (to 4 significant figures).
Then recalling that (upto 4 decimal places) the reliability and test variance of the 2 tests the score data of which are DATA-II(a) and DATA-II(b) are 0.6571, 6.2405 and 0.3488, 4.0571 respectively, the covariance between the 2 tests is 2.4571, we use Equation~\ref{eqn:rel_bat_wt} to compute the reliability of this real battery to be approximately \\
$$\displaystyle{\frac{0.6571\times 6.2405\times 0.7028^2 +
0.3488\times 4.0571\times 0.2972^2 +
2\times 0.7028\times 0.2972\times 2.4571}
{0.7028^2\times 6.2405 + 0.2972^2\times 4.0571 + 2\times 0.7028\times 0.2972\times 2.4571}}\approx 0.7112.$$
Thus, using our method of splitting a test in two parallel halves, we
have been able to compute reliability of the test as per the classical
definition and extended this to compute the reliability of a real
battery comprising two tests. For the used real data sets DATA-II(a)
and DATA-II(b), the battery reliability turns out to be about 0.71.
\section{Summary $\&$ Discussions}
\label{sec:conclu}
\noindent
The paper presents a new, easily calculable split-half method of
achieving reliability of tests, using the classical definition where
the basic idea implemented is that the square of the magnitude of the
difference between the score vectors $\boldsymbol{X}_g$ and $\boldsymbol{X}_h$ of $N$
examinees in the $g$ and $h$ sub-tests obtained by splitting the full
test, is proportional to the variance $S_E^2$ of the errors in the
scores obtained by the examinees who take the test, i.e. $\parallel
\boldsymbol{X}_g - \boldsymbol{X}_h\parallel^2 = N S_E^2$. Here, working within the paradigm
of Classical Test Theory, the error in an examinee's score is the
difference between the observed and true scores of the examinee. Our
method of splitting the test is iterative in nature and has the
desirable properties that the sample distribution of the split halves
are nearly coincident, indicating the approximately equal means and
variances of the split halves. Importantly, the splitting method that
we use, ensures maximum split-half correlation between the split
halves and the splitting is performed on the basis of difficulty of
the items (or questions) of the test, rather than examinee
attributes. A crucial feature of this method of splitting is that the
splitting being in terms of item difficulty, the method requires very
low computational resources to split a very large test data set into
two nearly coincident halves. In other words, our method can easily be
implemented to find the classically defined reliability using test
data that is obtained by collating responses from a very large sample
of examinees, on whom a test of as large or as small a number of items
has been administered. In our demonstration of this method, a
moderately large real test on 912 examinees and 50 items, convergence
to optimum splitting (splitting into halves that share equal means and
nearly equal variances) was achieved in about 0.024
seconds. Implementation of sets of toy data, generated under different
model choices for examinee ability was undertaken: 999 examinees
responding to a 50-item test, as well as much larger cohorts of
examinees--500,000 to 50,000--taking a 50-item test, and a cohort of
1000 examinees taking 100 to 1000-item tests. The order of our
splitting algorithm is corroborated to be quadratic in half of the
number of items in the test, while computational time for reliability
computation (input+output times, in addition to splitting of the test)
varies linearly with examinee sample size, so that even for the
500,000 examinees taking the 50-item test, reliability is computed to
be less than 10 seconds. Once the reliability of the test is computed,
it is exploited to perform interval estimation of the true score of
each examinee, where the error of this estimation is modelled as the
test error variance.
Subsequent to the dichotomisation of the test, we invoke a simple
linear regression model for the true score of an examinee, given the
observed score $X$, to achieve an interval estimate of the true score,
where the interval is modelled as $\pm 1S_E$. We recognise this
interval to be in excess of $\pm 1$ standard deviation of the error of
estimation of the true score for a given $X$, as provided by the
regression model; in other words, our estimation of uncertainty on the
estimated true score is pessimistic.
This method of splitting a test into 2 parallel halves, forms the
basis of our computation of the reliability of a battery, i.e. a set
of tests, as per the classical definition. A weighted battery score is
used in this computation where we implement a new way for the
determination of the weights, by invoking a Lagrange multiplier based
solution. We illustrate the implementation of this method of
determining weights--and thereby of computing the reliability of a
test battery following the computation of the reliabilities of the
component test as per the classical definition--on a real test battery
that comprises 2 component tests.
We have presented a new method of computing reliability as per
the classical definition, and demonstrated its proficiency and
simplicity. Thus, it is possible to uniquely find test characteristics like
reliability or error variance from the data, and such can be adopted
while reporting the results of the administered test. In this
paradigm, testing of hypothesis of equality of error variance from two
tests will help to compare the tests.
\renewcommand\baselinestretch{1.5}
\small
\bibliographystyle{ECA_jasa}
|
\section{Introduction}
In the recent decades, $\gamma$-ray astrophysics has made great advances enabled by improvements in instrumentation in the high energy technologies. The most recent and most accurate $\gamma$-ray source catalog is the \textit{Fermi} Large Area Telescope (LAT) Second Source Catalog \citep[2FGL,][]{2FGL}, based on the first 24 months of data from LAT. Thanks to its silicon strip pair production and modern analysis processes, LAT has drastically reduced the positional error of the sources with respect to previous studies, like those performed by the Energetic Gamma-Ray Experiment Telescope (EGRET) on board the Compton Gamma-Ray Observatory \citep{CGRO}.
As the {\it Fermi}-LAT $\gamma$-ray observations of the sky continue, the task of finding clear counterparts to all the detected sources becomes increasingly challenging, mostly due to the large positional uncertainty of the faintest sources. Since a strong connection between $\gamma$-ray and radio emission has been clearly demonstrated \citep{Ghirlanda2010,Mahony2010,Ackermann2011a}, it is natural to exploit radio surveys for finding new associations.
Several association methods have been proposed to match the $\gamma$-ray sources detected with source catalogs at lower frequencies \citep {U4,Masetti2013}, to give a proper counterpart to unidentified sources, where possible. Of the 1873 sources in the 2FGL catalog, 575 have unknown physical nature. Regarding the \textit{Fermi} sources of known types, blazars are the largest known population representing more than 80\%. It is, therefore, fair to presume that a significant fraction of the unidentified $\gamma$-ray sources (UGSs) can be unrecognized blazar-like sources. Finding them is the aim of our work. \par
Blazars are compact radio sources with flat (i.e., spectral index $\alpha< 0.5$, where $\rm{S}_\nu \sim \nu^{-\alpha}$) radio spectra that steepen towards the infrared-optical bands; their overall spectral energy distribution shows two broadband components: the low energy one peaking in the infrared (IR) to X-ray frequency range, while the high energy one extends up to the MeV-to-TeV range. Their emission features high and variable polarization, apparent superluminal motions of radio jet components and high apparent luminosity, coupled with rapid flux variability over the whole electromagnetic spectrum. Blazars come into two classes: flat-spectrum radio quasars and BL Lac objects, which we label here as BZQs and BZBs respectively, following Roma-BZCAT \citep{RomaBZCAT} nomenclature. Blazar emission interpreted as originating in radio loud active galaxies, with one of the relativistic jets emerging from the super-massive black hole in the galaxy nucleus pointing almost directly to the observer \citep{Giommi2012}. \par
This paper is the latest of a series that focus on the nature of UGSs: \citet{U1} (hereafter Paper I), which investigates sources as UGSs counterparts as described in \citet[][Paper II]{U2}; \citet[][Paper III]{U3} adds another feature for the search of blazar-like sources focusing on the low-frequency radio feature of blazars; \citet[][Paper IV]{U4} involves the X-ray emission as a distinctive feature and \citet[][Paper V]{U5} propose a renewed IR approach, based on a 2-dimensional kernel density estimation (KDE) technique, all for the same purpose.
In this work, we apply the method proposed in {\citetalias{U3}}\ to search for blazar-like candidates within the $\gamma$-ray positional uncertainty regions of the UGSs listed in the 2FGL and to select counterpart among them. This method is based on a multifrequency procedure, relying primarily on the flatness of the radio spectrum. The innovative point in this method is that it is based on low-frequency (i.e., $<1$ GHz) measurements, which is unprecedented in the investigation the emission of blazars and their association with UGSs. In {\citetalias{U3}}, we combined the observations at 325\,MHz from the Westerbork Northern Sky Survey (WENSS) at 325\,MHz with those of the NRAO Very Large Array Sky Survey (NVSS) at 1.4\,GHz.
In this work, we apply the same approach to the region covered by the counterpart of WENSS in the southern sky: the Westerbork in the Southern Hemisphere \citep[WISH,][]{De Breuck2002}.
Thanks to the combined results of this new search and of {\citetalias{U3}}, we build a reasonably sized population of new low-frequency selected blazar candidates. We study the spectral properties of individual candidates and explore the flux density distribution of the whole sample in comparison to the $\gamma$-ray blazars already listed in the 2nd \textit{Fermi} LAT Catalog of Active Galactic Nuclei \citep[2LAC,][]{2LAC}. We further refine our search by looking for IR, optical, UV and X-ray counterparts within the catalogs available.
The paper is organized as follows: in Sect.\ 2 we describe our method; in Sect.\ 3 we report and characterize our list of candidates; in Sect.\ 4 we discuss the spectral properties of all the sources found with the low-frequency method and make some projections regarding expectations from new low-frequency radio telescopes and surveys.
\section{Method}
\subsection{Blazar characteristics in the low-frequency radio band}
The characteristics of blazars (in particular $\gamma$-ray emitting blazars) in this range of the electromagnetic spectrum are discussed in {\citetalias{U3}}; we briefly summarize those results here for the sake of clarity.
Based on the cross correlation between the 2FGL catalog, the Roma-BZCAT v4.1 (the most comprehensive blazar catalog up to now), and the WENSS, we defined one sample labeled as Low radio frequency Blazar (LB) and a subsample labeled Low radio frequency $\gamma$-ray Blazar (LGB).
Since the Roma-BZCAT catalog is based on the NVSS survey, we computed the low-frequency radio spectral index values for these samples as:
\begin{equation}
\centering
\label{eq:index}
\alpha_{\nu}^{1400}=\log{ \left( \frac{S_{1400}}{S_{\nu}} \right )}/\log{\left( \frac{\nu}{1400} \right)}
\end{equation}
{where $\nu$ is the low radio frequency measured in MHz (i.e., 325 MHz for WENSS and 352 MHz for WISH), $S_{1400}$ is the flux density measured at 1400 MHz in the NVSS survey and $S_\nu$ is the flux density measured at $\nu$ frequency. Both flux densities are measured in mJy.} \par
The indices of about 80\% of the sources from both the LB and LGB samples are smaller than 0.5 and 99\% have indices below 1.0 \citep{U3}.
The flatness of the radio spectrum can be seen as a consequence of the dominance, also at low radio frequencies, of the inner jet departing from the core relative to the emission radiated by the larger structures of the blazar. \par
In particular for the $\gamma$-ray blazar sample, we consider as \textit{A class} candidates the radio sources characterized by $-1.0 \leq \alpha_{352}^{1400}\leq 0.55$ and \textit{B class} as those with $0.55\leq \alpha_{352}^{1400}\leq 0.65$. These two classes have been defined to represent respectively the 80\% (\textit{A class}) and 90\% (\textit{B class}) of $\gamma$-ray blazar radio spectral indices. In fact, the flatness of radio spectrum above 1\,GHz is indeed a well known feature and was already used to discern $\gamma$-ray blazar associations in the past \citep{CGRABS}. The low-frequency radio observations allow us to confirm this flatness down to 325\,MHz.\par
\subsection{The WISH survey}
The WISH survey is the natural extension of the WENSS survey to 1.60 sr of the southern sky. at 352\,MHz. Both surveys were performed in the 1990s by the Westerbork Synthesis Radio Telescope (WSRT), using the bandwidth synthesis mosaicing technique to combine 8 different bands of 5 MHz. WISH covers the area between $-9^\circ < {\rm Dec} < -26^{\circ}$ to a limiting flux density of $\sim 18$ mJy ($5 \sigma$), the same limiting flux density as the WENSS. Due to the low elevation of the observations, the survey has a lower resolution in declination ($54^{\prime\prime} \times \csc(\delta)$) than in right ascension ($54^{\prime\prime}$), resulting poorer source localization than for WENSS by a factor $\sim 2\times$ on average. Aside from this consideration, the WISH shares with the WENSS the same features upon which this method was calibrated, except for a negligible difference in frequency ($\Delta \nu = 27$ MHz). For this reason, we are confident in applying the same association procedure that we used for WENSS. \par
\subsection{Association procedure}
For each UGS in the WISH footprint, we search for low-frequency sources in a circular region of radius equal to the semi-major axis of the 95\% confidence level positional uncertainty ellipse of the UGS itself.
For each WISH source found, we look for a corresponding NVSS source in a circular region of radius equal to 8.5\arcsec\ and we then calculate the radio spectral index $\alpha_{352}^{1400}$.
For this study, we only consider WISH sources classified as single component sources and featuring a NVSS association.
In addition, we made local maps of the search regions for each UGS, overlaying on the WISH background the brightness contours of the NVSS map, the \textit{Fermi} positional uncertainty ellipse, and any possible blazar-like \textit{WISE} detection up to 3.3\arcsec\ from the position of the matching NVSS source (see Fig.~\ref{fig:Map2}). This gave us, in addition to a qualitative comparison of the relative positions of various possible candidates and the UGS, a clue about their potential non-blazar nature if complex structures are visible. Also, the angular separation between a candidate and the center of the {\it Fermi}\ ellipse was taken in account in cases of multiple matches, with the nearest source preferred. \par
For associations between WISH and NVSS and between NVSS and \textit{WISE}, the search radii (of 8.5\arcsec, and 3.3\arcsec, respectively) were selected based on a statistical study, as described in {\citetalias{U3}}.
\subsection{Multi-frequency data}
Finally, we looked for additional multifrequency data for our new blazar candidates and those found in {\citetalias{U3}}. In particular, we searched for matches with sources in the Australia Telescope 20 GHz survey\citep[AT20G,][]{AT20G} for the WISH candidates, and in the Green Bank 4.85 GHz northern sky survey \citep[GB6,][]{GB6} for the WENSS candidates. The AT20G survey, performed between 2005 and 2006 with the Australia Telescope Compact Array (ATCA), contains sources with $S_{20.0}\geq40$ mJy; the GB6, performed between 1986 and 1987 with the NRAO seven-beam receiver, takes into account sources with $S_{4.85}\geq18$ mJy. For these surveys, source association is automatically provided by NASA/IPAC Extragalactic Database (NED). We also refer to \citet{U4} for the analysis of X-ray emission observed by the {\it Swift} X-ray Telescope. \par
\section{Results}
\subsection{UGSs in the WISH footprint}
Of the 575 UGSs in the 2FGL catalog, we discard all that feature a `$c$' analysis flag to rule out potentially confused or spurious detections due to the strong interstellar emission \citep{2FGL}. The resulting list of UGSs on the WISH footprint has 27 $\gamma$-ray sources. Of these, 17 have at least one WISH and NVSS match (Table~\ref{tab:table}).\par
The total number of candidates is 31. Seven of these candidates have a radio spectral index in the range $\alpha_{325}^{1400}<0.55$ and are considered {\it class A} candidates while six are in the range $0.55 < \alpha_{325}^{1400} < 0.65$ and are considered {\it class B} candidates.
Sixteen out of the 31 candidates in the WISH footprint feature a {\it WISE}\ detection in at least two IR bands; three of them are {\it A class} candidates and two are {\it B class}. However, we do not consider {\it WISE}\ detection to be necessary for selection. IR data (see Table ~\ref{tab:WISE}) are taken from the \textit{AllWise} November 2013 release \citet{ALLWISE}. Flat spectrum radio quasars at high redshift or high synchrotron peaked BZBs could be too faint in the IR to be detected in one or more {\it WISE}\ bands; moreover, the relative astrometry of the WISH and NVSS {might be poorer} in this region \citep{U5} in respect to WENSS and NVSS.
All considered, we propose eight UGSs blazar-like associations (i.e., five \textit{A} and two \textit{B} class candidates and one special case for WBN 2255.7$-$1833). In particular, we increase the number of {\it class A} and {\it class B} associations combined with {\citetalias{U3}} to 19 and 8, respectively.
\par
\subsection{Multiwavelength observations and radio flatness of low-frequency selected blazar candidates}
In this work and in \citetalias{U3}, we evaluate the flatness of the radio spectrum according to Eqn.~\ref{eq:index} on the basis of just two non-simultaneous flux density measurements, at 1.4\,GHz (NVSS) and at low-frequency (352\,MHz for the WISH or 325\,MHz for the WENSS). For this reason, it is important to find other radio flux density data to better investigate the spectral and variability properties of the candidates. We searched the literature for other radio surveys performed in the WENSS and WISH footprints; we consider here all the candidates found both in \citetalias{U3} and in the present work.\par
In the WENSS footprint, we found that six \textit{A class} and two \textit{B class} candidates are listed in the GB6 survey at 5\,GHz. We report them in Table \ref{tab:regr1} along with the radio spectral index, obtained with a weighted linear regression on the measurements at the three frequencies. For all these sources, the 5 GHz flux density is in good agreement with the extrapolation of the low-frequency spectrum, confirming that the radio spectrum is flat and flux density variability is not dramatic. A sample spectrum is shown in Fig.~\ref{fig:regr1} (left panel).\par
In the WISH footprint, we find that one \textit{B class} (WNB 1251.8$-$2148, alias NVSS 125429$-$220419) is also listed in the AT20G. Data and radio spectral index regression are reported in Table \ref{tab:regr2} (see also Fig.~\ref{fig:regr1}, right panel). In this case, the extrapolation of the low-frequency power law clearly fails to match the high frequency data, suggesting prominent variability or a rather complex spectral shape, with a strongly inverted component above a few GHz; either way, the source behavior is consistent with the BZB class.\par
\section{Discussion and conclusions}
\subsection{Comparison with Paper III}
As we already noted, the WISH and WENSS surveys are very similar but, a posteriori, we can draw some useful conclusions regarding the impacts their different features had on our study.
The poorer resolution in declination of WISH ($\sim2\times$ on average) mainly affects the spatial association and not the parameters on which the counterpart is selected. Since spatial association is calibrated with a higher resolution survey like WENSS, applying it to the WISH region and catalog results in a more conservative approach. The fact that we found almost the same ratio (i.e. $\sim1.8$) of candidates per UGS in both WENSS and WISH (58 WENSS candidates for 32 UGSs and 31 WISH candidates for 17 UGSs) indicates that, on average, spatial association is not compromised. \par
Similarly, the resolution discrepancy cannot be blamed for any impact on the quality of the associations proposed among the candidates, given the similar rate of $\gamma$-ray blazar-like associations in the two surveys. In fact, in \citetalias{U3} we propose 21 (including one special association that is neither {\it class A} nor {\it class B}; see {\citetalias{U3}}) new $\gamma$-ray blazars out of 65 UGSs in the WENSS region (i.e., \ $32\%$) and 8 out of 27 in WISH region (i.e., \ $30\%$)\footnote{The small discrepancy { [with respect to what?]} can be explained, in addition to simple statistics, on the grounds of the slightly different starting list, since in {\citetalias{U3}}\ we considered only $\gamma$-ray sources without {\it any} analysis flag, while here we excluded only the sources with the {\it c} $\gamma$-ray analysis flag in the 2FGL \citep{2FGL}.}. \par
\subsection{Comparison with other methods}
Other methods have been suggested to recognize low-energy counterparts to UGSs, or at least to provide a statistically significant classification of these sources. For instance, \citet{Ackermann2012} have developed a statistical approach to classify UGSs in the first catalog of {\it Fermi}\ sources \citep[1FGL,][]{Abdo2010}. Six of the 2FGL UGSs in the WISH footprint are associated with 1FGL sources analyzed by \citet{Ackermann2012}, five of which are AGN-like and one of which is pulsar-like. Within the former sample, there are two sources for which we propose an association with a blazar counterpart on the basis of the low-frequency spectrum: 2FGL J2017.5$-$1618 and 2FGL J2358.4$-$1811. Interestingly, for the single UGS for which \citet{Ackermann2012} propose a pulsar classification (2FGL J1544.5$-$1126), our method finds a WENSS-NVSS match with quite steep spectral index ($\alpha=0.74\pm0.03$), rejecting a blazar scenario.
We further compared our results with the proposed associations found in the other papers of this series. In Figure~\ref{f.wgs} we show the comparison between the distribution in the IR $[3.4]-[4.6]-[12]\, \mu$m color-color plane provided by {\it WISE}\ for the $\gamma$-ray emitting blazars and the sources selected in this work. The overall distribution of the whole set of the simultaneous WISH-NVSS matches (black, red and green dots) is quite more scattered than the $\gamma$-ray blazar population (orange dots). However, when we only consider the low-frequency selected blazar candidates, all of the most prominent outliers are excluded and the {remaining five sources} (black and red dots) are in much better agreement with the IR colors of $\gamma$-ray blazars. In two cases, the $A$-class candidates NVSS 120900$-$231335 and NVSS 222830$-$163643, the agreement is very good. For the latter source, associated with 2FGL\, J2228.6$-$1633, our method also provides the same counterpart candidate selected on the basis of the kernel density estimator technique to IR colors of WISE counterparts applied to X-ray \citep{U4} and radio \citep{U5} data.
\subsection{Radio flux density analysis}
The $\gamma$-ray blazar candidates selected in this work and in {\citetalias{U3}}\ have by selection the same radio spectral properties of confirmed Roma-BZCAT blazars and of 2FGL blazar associations. It is natural to wonder why these sources have not been detected in the 2FGL catalog and eventually associated, e.g.\ in the second catalog of AGNs detected by {\it Fermi}\ \citep{2LAC}. The sources could have been excluded from the 2LAC because they do not formally pass the threshold for being considered high confidence associations (a test that does not take into account the spectral index); typically, this could happen for low flux density radio sources, which have a larger spatial density. It is thus likely that, together with a general similarity to the already known blazars, our candidates also have some peculiar characteristics.
For this reason, we show in Fig.~\ref{fig:hist} the distribution of radio flux density at 1.4\,GHz for all the blazars in the 2LAC and for the candidates selected here and in {\citetalias{U3}}. The two distributions are clearly different, as confirmed by a Kolmogorov–Smirnov test which yields a probability of $2.7\times10^{-15}$ of being obtained from the same population. In particular, the 2LAC blazar flux density distribution is shifted to much larger values. This strongly suggests that our method is very efficient at selecting faint blazars. These sources are potentially of great interest: if they are of the BZB type, it could mean that they could be of the extreme and elusive class of ultra-high synchrotron peaked sources; one prominent example could be WNB 2225.8$-$1652, which has an inverted low-frequency radio spectral index of $\alpha=-0.19\pm0.10$ and is our proposed association for the UGS 2FGL\, J2228.6$-$1633, characterized by a $\gamma$-ray spectrum as hard as $\Gamma=2.07\pm0.16$. On the other hand, some of our candidates could also be faint BZQs, and in this case their low flux densities could stem from their high redshifts \citep[like WN 1500.0+4815 with an estimated redshift of 2.78, see][]{U3}, which would also make them of scientific value. No WISH sources have a measured redshift.
\subsection{Summary and outlook}
We have searched for counterparts of the the 27 UGSs in the 1.6 sr footprint of the WISH survey, following the methods described in {\citetalias{U3}}\ and based on the flat spectrum at low-frequency characteristic of blazars. We have found blazar-like associations for 8 UGSs, that together with the 23 sources selected in the WENSS footprint with the same method provides a blazar association for a sample of 30 new $\gamma$-ray blazar associations. This sample extends the distribution of the radio flux density of $\gamma$-ray blazars to lower values, allowing us to study otherwise elusive AGNs.
The application of our method thus shows promising results in terms of numbers of counterparts proposed and their physical features. In particular, the possibility of using low-frequency radio data to find UGSs counterparts is even more important in the light of the imminent start of numerous studies in this frequency range, like the LOw Frequency ARray \citep[LOFAR,][]{van Haarlem2013}, the Murchison Widefield Array \citep[MWA,][]{Tingay2013}, the Long Wavelength Array \citep[LWA,][]{Ellingson2009}, and eventually the Square Kilometer Array \citep[SKA, e.g.][]{Dewdney2010}. These facilities will allow our method to be extended using an even deeper and simultaneous dataset, while dedicated targeting of the candidates with optical spectroscopy and Very Long Baseline Interferometry observations will confirm the natures of the proposed associations. \par
\acknowledgements
We are grateful to S. Digel for providing us useful suggestions that improved the presentation of our results.
M.\ Giroletti acknowledges financial contribution from grant PRIN-INAF-2011. The work is supported by the NASA grants NNX12AO97G and NNX13AP20G.
R.\ D'Abrusco gratefully acknowledges the financial support of the US Virtual Astronomical Observatory, which is sponsored by the National Science Foundation and the National Aeronautics and Space Administration.
The work by G.\ Tosti is supported by the ASI/INAF contract I/005/12/0.
This research has made use of data obtained from the High Energy Astrophysics Science Archive Research Center (HEASARC) provided by NASA's Goddard Space Flight Center; the SIMBAD database operated at CDS, Strasbourg, France; the NASA/IPAC Extragalactic Database (NED) operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Part of this work is based on the NVSS (NRAO VLA Sky Survey); The National Radio Astronomy Observatory is operated by Associated Universities, Inc., under contract with the National Science Foundation. This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration.
|
\section{Introduction}
\label{introduction}
Spreadsheets are interactive software applications for manipulation and storage of data. In a spreadsheet data are organized in worksheets, any of which is represented by a matrix of cells each containing either data or formulas that are automatically calculated at any variation of data.
Modern spreadsheets systems (e.g. Microsoft Excel) are integrated with scripting languages (e.g., Microsoft Visual Basic for Applications).
An end-user can extend the presentation layer of the application with Visual Basic for Application (VBA in the following) by defining new \textit{User Forms} and can develop new business functions by means of \textit{Procedures}. The execution of these procedures is event-driven: the procedures can be attached to events related to any element of the Excel Object Model, including Workbooks, Worksheets, Shapes and the same User Forms. Whenever one of these events occurs, the attached event handler code is executed. VBA can also be used to programmatically access and modify the underlying Excel Object Model, i.e. the hierarchy of objects contained in Excel representing all its accessible resources.
Very often developers create complex spreadsheet applications without any documentation at design level, making hard any task related to their maintenance. In order to reconstruct design models of an existing spreadsheet application reverse engineering techniques and tools are needed.
The research community has devoted great attention to the analysis and comprehension of spreadsheet applications. However, most papers focus on analyzing and testing the formulas in a spreadsheet, while they leave out the analysis of the spreadsheet’s embedded code and of the static and dynamic relationships between the code and the spreadsheets cells.
In this paper we will present in a nutshell two reverse engineering techniques and a tool supporting the second technique allowing the abstraction of design models of an existing spreadsheet application. In details, in section \ref{dmre} we will present a technique for abstracting conceptual data models by analyzing spreadsheets data while in section \ref{smre} we will present a tool supporting the abstraction of views representing the relationships between VBA procedures, user forms and spreadsheet cells. Finally, in section \ref{conclusions} some conclusions and future works will be presented.
\section{Data Model Reverse Engineering}
\label{dmre}
The first technique presented in this paper regards the abstraction of conceptual data models from the analysis of the structure and of the information included in an Excel spreadsheet by means of heuristic rules.
This technique is based on heuristic rules automatically applicable on a spreadsheet. By means of these rules, set of candidate classes with attributes, relationships between them and the corresponding cardinalities are abstracted on the basis of the structure and of the properties of spreadsheets and of their components, such as sheets, cells, cell headers, etc.. In particular, the heuristics consider cells labels and data by looking for repeated data, synonyms and group of cells containing well-defined data structures such as array strings, integer matrixes, etc.
The considered rules has been presented in \cite{Amalfitano2014a} and \cite{Amalfitano2014} where they have been used with success to abstract the conceptual data model underlying some complex spreadsheet applications used in the automotive context. Some of the proposed rules were derived from the literature \cite{Abraham2006},\cite{Hermans2010}.
Figure \ref{fig:dmre} shows an example of a possible application of some of the proposed rules. First of all, the sheet of the spreadsheet shown in Figure \ref{fig:dmre} may be abstracted as a class. Moreover, the two distinct rectangular areas separated by a blank column and composed of labels and data may be abstracted as two classes. Two composition relationships between these two classes and the class representing the sheet may be abstracted, too. The first rows of the two areas have cells with bold texts and different background colors: they may be abstracted as attributes of the classes representing the areas. The texts of these cells may be abstracted as names of the attributes of the corresponding classes.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{Fig02.png}
\caption{An example of application of heuristic rules to abstract a conceptual data model of a spreadsheet}
\label{fig:dmre}
\end{figure}
\section{Business Logic Model Reverse Engineering}
\label{smre}
The second contribution presented in this position paper is related to a reverse engineering technique that we have realized to abstract models of the business logic of a spreadsheet application by statically analyzing sheets, user forms, VBA procedures and their inter-relationships. The technique is completely supported by an interactive Excel add-in called EXACT (EXcel Application Comprehension Tool) that we have realized and that is available at \url{https://github.com/reverse-unina/EXACT}. It provides features for the extraction of information and the abstraction of several views of an existing spreadsheet application.
The tool extracts information from a spreadsheet by exploiting the features offered by the Office Primary Interop Assemblies for the Excel Application that exposed the Microsoft Excel Object Library and by statically analyzing the source VBA code. The tool is able to reconstruct the set of elements composing the spreadsheet including sheets, user forms, event handlers, classes and procedures. Furthermore the tool recovers information related to the relationships between these elements such as procedure calls, relationships between events and the corresponding handlers, and dependencies between procedures and cells.
The EXACT tool provides several features of software visualization. It offers multiple views at different levels of detail and provides cross-referencing functions for switching between these views. In example, the left part of Figure \ref{fig:smre} exemplifies the structural view proposed by EXACT for an Excel spreadsheet. The figure shows that the spreadsheet is composed of a single workbook with 4 worksheets, a VBProject including 6 code modules (that implement a total of 10 procedures) and a user form with 12 controls. In the right part of the figure there are some detailed information provided by the EXACT tool about a procedure selected by the user and a graph showing the dependencies between this procedure and the other application components. In particular, the considered procedure writes in 9 different groups of cells, reads the values of a group of cells and calls two procedures.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{structuralView.png}
\caption{Structural View of an Excel Spreadsheet}
\label{fig:smre}
\end{figure}
\section{Conclusions and Future Works}
\label{conclusions}
In this position paper we have presented in a very concise way two reverse engineering techniques and a tool abstracting conceptual data models and business logic models of Excel spreadsheets, taking into account both the data structure, the VBA code procedures, the User Forms and their inter-relationships.
In future, we plan to extend the EXACT tool by implementing the data model reverse engineering technique proposed in section \ref{dmre} with the aim to make possible comprehension processes of Excel spreadsheets applications.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{Introduction}
Cataclysmic variables (CVs) are well known to be short-period interacting
binary systems where a Roche-lobe filling star, the secondary, transfers
matter via an accretion disk to a white dwarf primary. One of the most
striking photometric characteristics common to all CVs are short-term
variations of the optical light occurring on time scales of seconds to a
few dozen minutes with amplitudes of some hundredths of a magnitude up to
more than an entire magnitude. This phenomenon is called flickering and appears
to the eye as a continuous sequence of overlapping flares in the light
curves of virtually all CVs (but note that this does not necessarily mean
that flickering consists of a linear superposition of independent events).
Indeed, it can safely be said that if it does not flicker, is is no CV.
Flickering has been observed for more then one and a half centuries: In 1856
Pogson
(\cite{Pogson57}) visually observed variations of the dwarf nova U~Gem that
might be interpreted as flickering. So did Baxendell in 1858 (Turner,
\cite{Turner}). Modern observations of flickering started in the middle
of the twentieth$^{\rm }$ century (Henize \cite{Henize}, Linnell \cite{Linnell},
Lenouvel \& Daguillon \cite{Lenouvel}, Johnson et al.\ \cite{Johnson},
Walker \& Herbig \cite{Walker54}, Walker \cite{Walker57}).
The amplitude of the flickering immediately shows that the
underlying light source is a major contributor to the optical light of the
respective systems. On the mean over all CVs, it emits
at least 25\% of the radiation of
the primary but can easily contribute as much as 50\% and in strong flares
can outshine the non-flickering light sources even by a factor of almost 2
(Bruch \cite{Bruch89}, \cite{Bruch92}). Even though flickering is thus due to
one of the optically most luminous light sources, it
remains among the least understood phenomena observed in CVs.
Flickering is not restricted to CVs or to optical wavelengths. The phenomenon
appears rather to be intimately related to accretion of matter from a disk
onto a central body. Herbst \& Shevchenko (\cite{Herbst}), for instance,
interpreted flickering-like variations in pre-main-sequence variables as being
due to unsteady accretion (see also Kenyon et al.\ \cite{Kenyon}). A
comprehensive review of aperiodic X-ray variability in X-ray binaries with
neutron stars and black holes was presented by van der Klis (\cite{vanderKlis}).
Flickering-like variability also occurs in active galactic nuclei [see e.g.\
Garcia et al.\ (\cite{Garcia}) and references therein].
The time scale of the flickering-type variations depends on the dimensions of
the underlying system, and the frequency range where it is best observed
depends on the compactness of the accretor.
The first encompassing systematic study of the flickering
properties in cataclysmic variables has been performed more than two decades ago by Bruch
(\cite{Bruch89}, \cite{Bruch92}), who defined several
statistical parameters to describe the phenomenon. In addition
to quantifying
the already mentioned strong contribution of the flickering light
source to the total optical light, he showed that the increase of its spectrum
to shorter wavelength is steeper than can be explained by simple models such
as a black body or a power law. Investigating various physical scenarios,
Bruch (\cite{Bruch89}, \cite{Bruch92}) concluded that instabilities in the
inner accretion disk and/or the boundary layer between the disk and the white
dwarf are most likely to give rise to the observed phenomena.
Later work on flickering includes
attempts to identify the exact location where it
occurs in CV systems. Already in 1985, Horne \& Stiening (\cite{Horne})
had analysed the variations of RW~Tri through eclipse and concluded that
the flickering light source has the same centre as the accretion disk, but
is less extended. This was later confirmed by Bennie et al.\ (\cite{Bennie}).
Similarly, Welsh et al.\ (\cite{Welsh}) found that flickering in HT~Cas is
concentrated to the centre of the disk, but is not confined to the immediate
vicinity of the white dwarf. Modifying the technique introduced by Horne \&
Stiening (\cite{Horne}) to avoid systematic effects, Bruch
(\cite{Bruch96}, \cite{Bruch00}) found similar results for several more
eclipsing CVs, but that to a lesser degree the hot spot also contributes.
However, in IP~Peg the hot-spot flickering dominates. In V2051~Oph,
Baptista \& Bortoletto (\cite{Bap04}) identified two flickering components,
namely a slowly varying one that they associated with the disk overflowing gas
stream from the secondary star, and a more rapidly varying one with a radial
distribution equal to the steady disk light.
The behaviour of flickering in the frequency domain has been
the subject of many studies. Fritz \& Bruch (\cite{Fritz}) studied wavelet
transforms of numerous light curves of many CVs. They parametrized the
resulting scalegrams (Scargle et al. \cite{Scargle}) in terms of its
inclination $\alpha$ (frequency behaviour) and its value $\Sigma$ (flickering
strength) at a reference time scale. For a given system, $\alpha$ and $\Sigma$
are stable over many years. On average,
flickering is somewhat bluer on short time-scales than on longer ones. CVs of
different types and photometric states occupy distinct regions in the
$\alpha$ -- $\Sigma$ -- plane. A similar study limited to the intermediate
polar V709~Cas was published by Tamburini et al.\ (\cite{Tamburini}).
A wavelet study in X-rays was performed
by Anzolin et al.\ (\cite{Anzolin}). The X-ray data are distributed in a much
smaller area of the $\alpha - \Sigma$ parameter space than the optical data.
The authors explain the similarity of the X-ray flickering in objects of
different classes together with the predominance of a persistent stochastic
behaviour in terms of magnetically driven accretion processes acting in a
considerable fraction of the analysed objects.
Many studies of the frequency characteristics focus on power spectra where
flickering light curves in general exhibit a behaviour similar to red noise,
that is, on a double logarithmic scale, a linear decrease of power
towards high frequencies (but with a flattening towards low frequencies;
Bruch \cite{Bruch89}, \cite{Bruch92}; Schimpke \cite{Schimpke}).
Dobrotka et al.\ (\cite{Dobrotka14}) detected two red-noise and two white-noise
components in the X-ray power spectrum of RU~Peg, indicating the presence
of two turbulent regions. The long and low-noise high-cadence light curves
of some CVs observed by the Kepler satellite represent superb data from which to study
flickering. Scaringi et al.\ (\cite{Scaringi12a}, \cite{Scaringi12b})
used these data to construct
the flickering power spectrum of MV~Lyr, which they decomposed into a broken
power law and a series of Lorentzian components, following Belloni et al.\
(\cite{Belloni}) and Novak et al.\ (\cite{Novak}). They detected a linear
rms -- flux relation such that the light source becomes more variable when
it becomes brighter, as has earlier been observed in X-ray binaries and AGNs
(Uttley \& ${\rm M^c}$Hardy \cite{Uttley01}, Uttley et al. \cite{Uttley05},
Heil \& Vaughan \cite{Heil}). This is interpreted as an indication that
flickering variations are due to a multiplicative as opposed to an additive
process, such as shot noise, since in the latter case the rms -- flux relation
would be destroyed. In addition to MV~Lyr, rapid variations were also analysed
in Kepler data of other CVs by Scaringi et al.\ (\cite{Scaringietal14}).
Interpretations of the frequency behaviour in terms of the propagation of
disturbances in the accretion disk leading to flickering have repeatedly
been discussed in the literature. For instance, Lyubarskii (\cite{Lyubarskii}),
Yonehara et al.\ (\cite{Yonehara}), Pavlidou et al.\ (\cite{Pavlidou}), and
Dobrotka et al.\ (\cite{Dobrotka10}) developed conceptually similar models
that differ in details concerning the origin and the propagation of the
disturbances. Of these, the model of Lyubarskii (\cite{Lyubarskii}),
formulated originally to explain X-ray variations in galactic and
extragalactic X-ray source, and which invokes viscosity fluctuations
at different radii, has found most widespread attention. In the
optical range it has sucessfully been applied to MV~Lyr by Scaringi
(\cite{Scaringi14}) [but note that Dobrotka et al.\ (\cite{Dobrotka15})
was able to interpret the same observational data also in the framework of the
model of Dobrotka et al.\ (\cite{Dobrotka10})].
This scanario is quite attractive, in particular because it seeks
to explain optical flickering in CVs and X-ray fluctuations in neutron star
and black hole binaries as well as AGNs in a unified model.
However, it has not yet been
convincingly shown that it can explain all flickering characteristics observed
in CVs (or, alternatively, that certain properties cited in the literature
are based on incorrect assumptions or faulty techniques). Therefore, alternative
models should not yet be discarded.
None of these mentioned models is concerned with a detailed description
of the emission mechanism that leads to the observed flickering.
Of the few attempts to construct a realistic physical model for the temporal
and spectral evolution of a flickering flare, the work of Pearson et al.\
(\cite{Pearson}) deserves to be mentioned, which I discuss
in more detail in Sect.~\ref{Fireballs and the case of SS Cyg}.
As a broad-band phenomenon, in the optical range flickering is apparent from
the blue to the red atmospheric cutoff and, not surprisingly, in different
photometric bands it is not independent from each other but strongly
correlated (see, e.g.\ Fig.~2 of Bruch 1992). However, depending on the
temporal evolution of the properties of the emitting light source (e.g.\
temperature and opacity), it is conceivable that individual flares do not
reach their highest intensity simultaneously at all wavelengths, thus
leading to time lags between the flickering in different
photometric bands. Detection and quantification of such lags will
provide constraints on physical mechanisms explaining flickering.
A temporal shift of the flickering at different wavelengths can be traced by
measuring the location of the peak of the cross-correlation function (CCF) of
light curves in various bands. This technique has been applied by
Jensen et al.\ (\cite{Jensen}), who
observed in this way a lag of about 1\hoch{m} between the flickering
in the optical and X-ray bands in TT~Ari. Similarly, Balman \&
Revnivtsev (\cite{Balman}) detected lags between 96\hoch{s} and 181\hoch{s}
between UV and X-ray observations in five more CVs.
Using optical observations of AM~Her, Szkody \& Margon (\cite{Szkody})
deduced from the asymmetry of the maximum of the CCF a
lag of the flickering between the $U$ and $B$ bands of 8\hoch{s} in the sense
that variations appear earlier in $U$ than in $B$. On the other hand,
Bachev et al.\ (\cite{Bachev}) found no conclusive evidence for delays of the
flickering in different bands of the nova-like variable KR~Aur. This may
be due to the comparatively coarse time resolution of 30\hoch{s} of their
data. Bruch (\cite{Bruch89}, \cite{Bruch92}) also obtained null results,
but this can probably be attributed to insufficient quality and quantity
of the observational data.
A more general approach makes use of the coherence function, that is,\ a
Fourier frequency-dependent measure of the linear correlation between time
series measured simultaneously in two spectral ranges (Vaughan \& Novak
\cite{Vaughan}). This permits detecting time lags as a function of the
(temporal) frequency of the underlying signals. This technique has found
widespread application in the study of X-ray binaries and AGNs, but has only
recently successfully been applied to optical data of CVs. Scaringi et
al.\ (\cite{Scaringi13}) studied the Fourier-frequency-dependent
coherence of light curves of MV~Lyr and LU~Cam observed at high time
resolution (0\hochpunkt{s}844 -- 2\hochpunkt{s}276) in the SDSS
bands $u'$, $g'$
and $r'$. They also described in detail the applied data reduction
method. Scaringi et al.\ (\cite{Scaringi13}) found that low-frequency
variations (below $\approx 10^{-3}$~Hz) occur about 10 seconds later in the
in $r'$ than in $u'$ in LU~Cam. The corresponding time lag is $\approx
3$ seconds in MV~Lyr. At higher frequencies, the errors become too large
to permit reliable measurement.
Here, I investigate numerous light curves of various CVs observed in
several photometric systems to measure time shifts of the
flickering as a function of wavelength. As mentioned, the time lags
observed by Scaringi et al.\ (\cite{Scaringi13}) are most significant for low-frequency variations. Moreover, the (approximate) power-law behaviour of
the flickering power as a function of frequency (Bruch \cite{Bruch89},
\cite{Bruch92})
means that low-frequency variations strongly dominate high-frequency
flickering. Considering furthermore that the CCF will be most sensitive to the
strongest variations as well as the limitations of the available data
(time resolution, noise, length of individual data sets; see also
Sect.~\ref{Data}), I prefer here to use the cross-correlation function, which is
probably more robust than the much more sophisticated method applied by
Scaringi et al.\ (\cite{Scaringi13}). Not using the coherence function
implies the loss of some information. But in view of the data limitation,
this disadvantage is compensated by more reliable results than would be
possible using more sophisticated techniques. To measure
the time lag of the flickering as a function of wavelength, I
therefore examine
the CCFs of the light curves, that is,\ the normalized cross
covariant as a function of the shift of the data in a reference band
with respect to the other (comparison) bands of the photometric system. In Sect.~\ref{Reduction method} the reduction method is
explained. The observational data are briefly introduced in
Sect.~\ref{Data} before the results of this study are presented in
Sect.~\ref{Results}. Scenarios to explain the results are discussed in
Sect.~\ref{Discussion}, and finally, conclusions are drawn in
Sect.~\ref{Conclusions}.
\section{Reduction method}
\label{Reduction method}
The CCF is calculated in the following way: A light curve is expected to
have been sampled simultaneously in the reference and the comparison
band in discrete time intervals separated by $\Delta t$. In the data sets
used in this study this is often the case, but sometimes only approximately so.
Then, the light curves are re-sampled by linear interpolation between
data points at the instances $t_1 + j\, \Delta t$, where $t_1$ is the instance
of the first data point in the reference band, $j$ is an integer number, and
$\Delta t$ is chosen to be the average time interval between data points in
the original light curve. Comparison and reference data (designated $x$ and
$y$, respectively) are then normalized
to their average value before the mean is subtracted, yielding $n$ pairs of
data points ($x , y$).
The CCF is evaluated at discrete time shifts $\tau_k$, such that
$\tau_k = k\, \Delta t$ and $k$ is an integral number,
\begin{displaymath}
CCF(\tau_k) = \frac{\sum_{i=1}^m \left( x_i y_{i+k} \right) \, - \,
\frac{1}{m}{\sum_{i=1}^m x_i \sum_{i=1}^m y_{i+k}}}
{\sqrt{\left[ \sum_{i=1}^m x_i^2 -
\frac{\left(\sum_{i=1}^m x_i \right)^2}{m} \right]
\left[ \sum_{i=1}^m y_{i+k}^2 -
\frac{\left( \sum_{i=1}^m y_{i+k} \right)^2}{m} \right]}}
\end{displaymath}
Here, $i = 1 \ldots m$ indicate those data pairs for which reference and
comparison
data overlap after the comparison data have been shifted in time by $\tau_k$.
The CCF is thus discretely sampled at a time resolution of $\Delta t$.
If reference and comparison data exhibit similar variations but the pattern
of variability is shifted by a time interval ${t_{\rm s}}$ between them,
the CCF will assume a maximum at $\tau = {t_{\rm s}} \equiv \Delta T_{\rm max}$.
In principal, it is straightforward to determine the maximum of the CCF
of the data in a reference and a comparison band of light curves
observed simultaneously or quasi-simultaneously at different wavelengths.
However, there are a few details arising from the structure of the individual
data sets available for this study that deserve to be mentioned. As an
example, the upper panel of Fig.~\ref{ccexlres} shows the $B$ -band light
curve of the old nova V603~Aql observed in the Walraven photometric system
on 1983, July 18 (see Walraven \& Walraven, \cite{Walraven}, and Lub \& Pel,
\cite{Lub}, for the peculiarities of the Walraven system). In
addition to
flickering, it shows another feature frequently observed in CVs, that is,\
variations on longer time scales that may not be immediately associated
with flickering (at least not in the light curve discussed here;
see Bruch \cite{Bruch91}).
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{ccexlres.eps}}
\caption{Example of a light curve and the CCF of the brightness variations
in two photometric bands. {\em Top:} $B$-band light curve
of V603~Aql observed on 1983 July 18. The solid line is a
fit of a high-order polynomial to the data.
{\em Centre:} Residuals between the original light curve and
the polynomial fit.
{\em Bottom left:} CCF of the light
curve shown in the top panel with the simultaneously
observed light curve in the $U$-band (broken line) and of
the residuals between the original data and the polynomial
fit (solid line). {\em Bottom right:}
Enlarged section of the maximum of the CCF shown as a
solid line in the lower left panel
(dots) together with a fit of a high-order polynomial
to the data (broken curve). The broken
vertical line indicates the location of the maximum of the
fitted polynomial.}
\label{ccexlres}
\end{figure}
The CCF of the light curves in the $B$ band and the simultaneously observed
$U$ band is shown as a broken line in the lower left frame of
Fig.~\ref{ccexlres}. The central peak is caused by the correlated
flickering activity in the two bands, but on the whole, the
profile is rather shallow. This is caused by the longer time-scale modulations.
Even if these (in a more general case than for the present example) were a
manifestation of flickering, the limited length of the light curves -- being
of the same order of magnitude -- would inhibit a reliable measurement of
an associated time lag of the order of a few seconds between different bands.
Therefore, to better isolate the variability due to flickering
on time scales where a lag can be measured more reliably, it
is appropriate to subtract the slow modulations. This
is done by fitting a polynomial of suitable order to the original data in both
bands, as shown by the solid line in the upper frame of Fig.~\ref{ccexlres}.
In all cases when a polynomial was subtracted, the order was chosen such
that the polynomial followed the long-term variations satisfactorily, and the
same order was adopted for the reference and the comparison bands.
The light curve in the central fram of Fig.~\ref{ccexlres}
constitutes the residuals between the
original data and the fitted polynomial in the $B$ band. The solid graph
in the lower left frame represents the CCF of these residual light curves in
$B$ and $U$. Here, the central peak due to the correlated flickering stands
out much more clearly.
The lower right frame of Fig.~\ref{ccexlres} contains a magnified version
of the upper part of the CCF maximum (solid dots).
The coarse resolution reflects
the rather low time resolution of 21\hoch{s} of the example light curve.
To estimate the location of the maximum, a polynomial was fit
to the data of the CCF. Its order depends on the number of available points
in the peak of the CCF (range: $-120\hoch{s} \le \tau \le 120\hoch{s}$ in
this case). This is the dashed line in the figure. In light curves with a
very high time resolution, the CCF often contains a sharp spike centred on
zero time-shift. It is caused by correlated atmospheric noise in the
reference and comparison bands and is masked in the fit process. The quality
of the fit indicates that the maximum of the CCF can be
determined with high precision by calculating the maximum of the fit
polynomial. Its location is marked by the broken vertical line.
The corresponding $\Delta T_{\rm max}$ is slightly negative in this example.
To obtain a notion of the accuracy to which $\Delta T_{\rm max}$ can
be measured, a simple shot-noise model was used to generate artificial
flickering light curves. While this may not reproduce the flickering properties
observed at least in some CVs and in many X-ray binaries such as the linear
rms - flux relation, the dependence of the accuray of $\Delta T_{\rm max}$ on
properties such as statistical noise in the data or their time resolution will
not depend strongly on such details. Thus,
100 artificial light curves were generated, each of
which subsequently served as reference data. One of them
is shown in the upper frame of Fig.~\ref{artlc}.
The light curves were constructed in the following way:
1\,500 simulated flares were distributed randomly on a time base of three
hours, which is the typical duration of the real light curves studied here.
Their amplitudes are distributed between $0\hoch{m} < A \le 0\hochpunkt{m}1$
such that the probability of a flare with a particular amplitude to occur
decays linearly from the minimum to the maximum value of $A$ and the
probability for a flare with the maximum amplitude goes to zero.
The flares are symmetric and rise and decay at a mean rate of
0\hochpunkt{m}5 per hour, allowing for a random
scatter equally distributed between $0 < {\rm d}m/{\rm d}t \le 1$
magnitudes per hour around this value. The superposition of these
flares, sampled in 1\hoch{s} intervals, plus an arbitrary constant,
constitutes the light curves. To make the light curve displayed in
Fig.~\ref{artlc} look more natural, random Gaussian noise with
$\sigma = 0\hochpunkt{m}01$ was added. Second versions of the same data sets
were created by applying a constant time shift of 5\hoch{s} to the originals.
These served subsequently as comparison light curves. Using a constant
is acceptable here because a possible
dependence of the time lag on flickering frequency as observed in MV~Lyr
and LU~Cam (Scaringi et al.\ \cite{Scaringi13}) will not have a strong
bearing on the accuracy of the $\Delta T_{\rm max}$ determination, the more
so because the CCF is -- as mentioned -- most sensitive to the stronger
low-frequency variations.
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{artlc.eps}}
\caption[]{{\em Top:} Artificial flickering light curve used to estimate
the accuracy of the determination of flickering time lags.
{\em Bottom left:} Dependence of $\Delta T_{\rm max}$ on the
noise level. {\em Bottom right:} Expected error of
$\Delta T_{\rm max}$ as a function of the time resolution of
the light curve for low (0\hochpunkt{m}02; open symbols) and
moderately high (0\hochpunkt{m}05; filled symbols) noise levels.}
\label{artlc}
\end{figure}
To study the accuracy of $\Delta T_{\rm max}$ on the noise level,
Gaussian noise was added to the reference and the
comparison light curves before $\Delta T_{\rm max}$ was calculated.
This was repeated for different noise amplitudes.
The results are shown in the lower left frame of Fig.~\ref{artlc},
where for each noise level the mean value of $\Delta T_{\rm max}$ derived
from the 100 trial light curves is plotted. The error bars are the
standard deviation $\sigma$ around the mean and thus indicate the
accuracy with which $\Delta T_{\rm max}$ can be measured in an individual
light curve.
The dependence of the accuracy on the time resolution of the data was
determined in a similar way: The artificial light curves were binned to
simulate a lower resolution before noise was added. Then, the standard
deviation of $\Delta T_{\rm max}$ was calculated. It is
shown in the bottom right frame of Fig.~\ref{artlc} as a function of the
time resolution for low (0\hochpunkt{m}02; open circles) and intermediate
(0\hochpunkt{m}05; filled circles) noise levels.
The accuracy of the results also depends on the length of the light curves
(or, equivalently, on the number of data points).
Tests with 100 data sets with half (twice) the
length of the original ones but with otherwise identical characteristics,
subjected to the same noise levels as shown in the lower left frame of
Fig.~\ref{artlc}, resulted in errors of $\Delta T_{\rm max}$ that were
1.44$\pm$0.21 (0.70$\pm$0.06) larger. Thus, not unexpectedly,
the errors scale to a high degree of accuracy
with the inverse of the square root of the
length of the light curves.
This exercise shows that $\Delta T_{\rm max}$ can be measured to an
accuracy of the order of a tenth of a second in high-quality (long, low noise,
high time resolution) light curves, but is limited to several seconds if they
are shorter, noisier and/or of low time resolution.
\begin{table*}
\caption{Mean flickering time lags (in seconds) in cataclysmic variables}
\label{Mean flickering time lags}
\begin{tabular}{l*{12}{@{\hspace{1ex}}c}}
\hline\hline
\multicolumn{1}{l}{Name} &
\multicolumn{1}{c}{B/W} &
\multicolumn{1}{c}{B/U} &
\multicolumn{1}{c}{B/L} &
\multicolumn{1}{c}{B/V} &
\multicolumn{1}{c}{B/R} &
\multicolumn{1}{c}{B/I} \\
\hline
\object{RX And} &
&
\phantom{0}-3.12\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
\phantom{0}-5.19\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
&
\\
\object{AE Aqr} &
\phantom{0}-2.16\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
\phantom{-0}2.59$\pm$\phantom{0}7.97 (55) &
\phantom{0}-5.93$\pm$17.29 (\phantom{0}2) &
\phantom{0}-2.82$\pm$11.14 (54) &
\phantom{-0}4.38$\pm$\phantom{0}7.25 (47) &
\phantom{-0}5.02$\pm$\phantom{0}9.77 (31) \\
\object{V603 Aql} &
\phantom{0}-2.40$\pm$\phantom{0}3.53 (\phantom{0}4) &
\phantom{0}-3.85$\pm$\phantom{0}2.46 (15) &
\phantom{0}-3.75$\pm$\phantom{0}4.18 (\phantom{0}4) &
\phantom{-0}2.44$\pm$\phantom{0}4.04 (15) &
\phantom{-}10.58$\pm$\phantom{0}2.92 (\phantom{0}7) &
\phantom{-}17.15$\pm$\phantom{0}7.21 (\phantom{0}2) \\
\object{TT Ari} &
\phantom{-0}0.02$\pm$\phantom{0}1.29 (\phantom{0}3) &
\phantom{0}-1.45$\pm$\phantom{0}1.16 (24) &
\phantom{-0}0.55$\pm$\phantom{0}0.86 (\phantom{0}3) &
\phantom{-0}1.20$\pm$\phantom{0}1.30 (24) &
\phantom{-0}2.25$\pm$\phantom{0}0.53 (13) &
\\
\object{QU Car} &
&
\phantom{0}-1.65$\pm$\phantom{0}0.82 (\phantom{0}3) &
&
\phantom{-0}4.90$\pm$\phantom{0}2.56 (\phantom{0}3) &
&
\\ [1ex]
\object{BV Cen} &
\phantom{-0}1.88$\pm$\phantom{0}3.04 (\phantom{0}5) &
\phantom{0}-1.53$\pm$\phantom{0}1.72 (\phantom{0}5) &
\phantom{0}-3.18$\pm$\phantom{0}4.12 (\phantom{0}5) &
\phantom{0}-0.05$\pm$\phantom{0}4.41 (\phantom{0}5) &
&
\\
\object{WW Cet} &
\phantom{0}-1.26$\pm$\phantom{0}2.13 (\phantom{0}6) &
\phantom{-0}1.04$\pm$\phantom{0}3.53 (\phantom{0}7) &
\phantom{-0}2.07$\pm$\phantom{0}4.60 (\phantom{0}6) &
\phantom{-0}0.84$\pm$\phantom{0}3.47 (\phantom{0}7) &
\phantom{0}-1.78\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
\phantom{0}-1.77\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) \\
\object{T CrB} &
&
\phantom{-0}0.68$\pm$\phantom{0}1.71 (\phantom{0}2) &
&
\phantom{-0}0.43$\pm$\phantom{0}3.11 (\phantom{0}2) &
&
\\
\object{SS Cyg} &
&
\phantom{-0}1.06$\pm$\phantom{0}3.12 (\phantom{0}7) &
&
\phantom{-0}4.22$\pm$\phantom{0}1.41 (\phantom{0}6) &
&
\\
\object{HR Del} &
&
\phantom{0}-0.75$\pm$17.79 (19) &
&
\phantom{-0}4.23$\pm$10.70 (19) &
\phantom{-0}5.39$\pm$\phantom{0}4.14 (\phantom{0}4) &
\\ [1ex]
\object{DO Dra} &
&
\phantom{-0}1.09$\pm$\phantom{0}0.06 (\phantom{0}2) &
&
\phantom{-0}2.13$\pm$\phantom{0}0.42 (\phantom{0}2) &
\phantom{-0}1.65$\pm$\phantom{0}0.22 (\phantom{0}2) &
\\
\object{V795 Her} &
&
\phantom{-0}1.63$\pm$\phantom{0}0.01 (\phantom{0}2) &
&
\phantom{-0}0.45$\pm$\phantom{0}0.58 (\phantom{0}2) &
\phantom{0}-0.07$\pm$\phantom{0}0.00 (\phantom{0}2) &
\\
\object{RS Oph} &
\phantom{-}10.93\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
\phantom{-0}5.10$\pm$\phantom{0}5.28 (\phantom{0}8) &
\phantom{0}-3.28$\pm$\phantom{0}0.19 (\phantom{0}2) &
\phantom{-0}3.96$\pm$\phantom{0}3.29 (\phantom{0}8) &
\phantom{-0}8.29$\pm$\phantom{0}0.97 (\phantom{0}2) &
\phantom{-}10.20$\pm$\phantom{0}1.37 (\phantom{0}2) \\
\object{V426 Oph} &
\phantom{-0}2.87$\pm$\phantom{0}1.96 (\phantom{0}3) &
\phantom{-0}0.67$\pm$\phantom{0}2.37 (\phantom{0}3) &
\phantom{-0}1.63$\pm$\phantom{0}0.62 (\phantom{0}3) &
\phantom{-0}0.67$\pm$\phantom{0}1.31 (\phantom{0}3) &
&
\\
\object{CN Ori} &
&
\phantom{0}-5.64\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
\phantom{-0}0.94\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
\phantom{-0}4.93\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
\\ [1ex]
\object{RU Peg} &
&
\phantom{-0}2.61\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
\phantom{-0}5.28\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
\\
\object{IP Peg} &
&
\phantom{0}-1.15\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
\phantom{0}-1.17$\pm$\phantom{0}0.57 (\phantom{0}2) &
\phantom{0}-0.08$\pm$\phantom{0}0.49 (\phantom{0}2) &
\\
\object{GK Per} &
&
\phantom{-0}0.19$\pm$\phantom{0}2.66 (15) &
&
\phantom{-0}1.97$\pm$\phantom{0}2.93 (15) &
\phantom{-0}2.46$\pm$\phantom{0}2.86 (12) &
\\
\object{AO Psc} &
&
\phantom{0}-0.39$\pm$\phantom{0}0.98 (11) &
&
\phantom{-0}0.44$\pm$\phantom{0}1.11 (11) &
\phantom{-0}0.19$\pm$\phantom{0}0.94 (11) &
\\
\object{VY Scl} &
\phantom{0}-0.15$\pm$\phantom{0}0.83 (\phantom{0}4) &
\phantom{-0}0.65$\pm$\phantom{0}1.21 (\phantom{0}4) &
\phantom{-0}0.50$\pm$\phantom{0}1.90 (\phantom{0}4) &
\phantom{-0}2.66$\pm$\phantom{0}0.39 (\phantom{0}4) &
&
\\ [1ex]
\object{RW Tri} &
&
\phantom{-0}6.16$\pm$\phantom{0}3.80 (\phantom{0}4) &
&
\phantom{-0}4.35$\pm$\phantom{0}3.84 (\phantom{0}4) &
\phantom{-0}8.81$\pm$\phantom{0}6.55 (\phantom{0}4) &
\\
\object{UX UMa} &
&
\phantom{-}15.74\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
&
-13.89\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
-30.96\phantom{$\pm$\phantom{0}0.00} (\phantom{0}1) &
\\
\object{IX Vel} &
&
\phantom{-0}0.21$\pm$\phantom{0}2.48 (\phantom{0}5) &
&
\phantom{0}-5.60$\pm$\phantom{0}8.23 (\phantom{0}5) &
&
\\
\hline
\end{tabular}
\end{table*}
\section{Data}
\label{Data}
A total of 197 light curves of 23 different CVs were used in this study.
They were observed in several photometric systems over an interval of about
two decades at different telescopes by many observers with a variety of
scientific purposes in mind. They are thus quite
heterogeneous. The data can roughly be grouped and characterized as follows:
$VBLUW$ (Walraven) light curves: They were obtained by various observers at
the Dutch telescope at ESO/La Silla, equipped with the Walraven photometer
(Walraven \& Walraven, \cite{Walraven}).
The instrument permits only a limited choice
of configurations, resulting in a time resolution of about 21\hoch{s} of the
present data, which may be at the limit of usefulness for the current purpose.
The five passbands with decreasing wavelengths from V in the visual to W in the
ultraviolet are observed simultaneously.
$UBV$ light curves: Data in this photometric system were generally observed
sequentially in the three bands. This also inhibited a high time resolution.
Most of the $UBV$ light curves regarded here are limited to a resolution of
$\approx$15\hoch{s} or lower.
$UBVRI$ light curves: Many suffer from the same disadvantages as their $UBV$
counterparts, resulting thus in a relatively coarse time resolution. However,
some have been obtained using a photometer equipped with a rapidly rotating
filter wheel (Jablonski et al.\ \cite{Jablonski}),
which permits obtaining quasi-simultaneous measurements in all
passbands. These data sets were observed at a time resolution of 5\hoch{s},
and are thus more suitable for the present purpose.
$UBVR*$ (Stiening) light curves:
These are data sets were obtained with the Stiening
photometer (Horne \& Stiening \cite{Horne}), which is equipped with four filters
similar but not identical to the standard $UBVR$ filters. All passbands
are observed strictly simultaneously. The disadvantage that this photometric
system is not well calibrated is by far outweighed in the present context by
the high time resolution used to observe the light curves employed here:
Most have a sampling interval of 0\hochpunkt{s}5 or 1\hoch{s}, while
some even reach 0\hochpunkt{s}2.
Thus, the most suitable data sets are those observed at high time resolution
in the $UBVRI$ system and, above all, the Stiening data.
\section{Results}
\label{Results}
All available light curves were subjected to the reduction method outlined in
Sect.~\ref{Reduction method}. The $B$
band was used as reference band\footnote{One of the light curves of AE~Aqr
was observed only in $U$, $V$ and $R$.
In this case, the shortest wavelength band served as reference.}.
The light curves of all other bands were cross-correlated with the $B$
-band light curve, and the time lags $\Delta T_{\rm max}$ of the maxima
of the respective CCFs were determined. In some cases, only parts
of the light curves were used. This applies in particular to AE~Aqr, which
sometimes exhibits periods of no or very weak flaring activity that were
then excluded. Eclipses in some light curves of eclipsing CVs were
also removed. Strong long-term variations (time scale:
hours), if present, were removed by subtracting a fitted polynomial of
suitable degree (see Fig.~\ref{ccexlres}).
Table~\ref{Mean flickering time lags}
contains average values of
$\Delta T_{\rm max}$ of the maximum of the CCF of each comparison band
with respect to the reference band as indicated in the table header.
This is interpreted as the time lags
of the flickering in these bands. Since short data sets should contribute
to the average to a lesser degree than long ones, the results for the
individual light curves were weighted with their total duration. The number
$n$ of light curves contributing to the mean are given in brackets. The
errors are standard deviations derived from all light curves of a given
star. No
selection was made. All light curves, regardless of the photometric system
or the time resolution, enter the average.
\subsection{The ensemble}
\label{The ensemble}
No clear picture emerges for
the individual systems. In most cases,
the standard deviation is considerable, often significantly larger than the
average value of $\Delta T_{\rm max}$. This is not surprising in view of
the expected errors determined from the simulations performed
in Sect.~\ref{Reduction method} and
the limited quality of most light curves. Moreover, more often than not
(disregarding the errors),
there is no monotonical dependency of $\Delta T_{\rm max}$ as a function of
wavelength difference between reference and comparison bands, as might be
naively expected. However,
when the investigated stars are considered as an ensemble, at least a trend
appears.
\begin{table}
\caption{Isophotal wavelengths}
\label{Isophotal wavelengths}
\vspace{1em}
\begin{tabular}{ll|ll}
\hline
\multicolumn{2}{l|}{UBVRI system} & \multicolumn{2}{l}{Walraven system} \\
Band & $\lambda_{\rm iso}$ (\AA) & Band & $\lambda_{\rm iso}$ (\AA) \\
\hline
U & 3573.2 & W & 3231.9 \\
B & 4347.6 & U & 3618.2 \\
V & 5458.6 & L & 3832.6 \\
R & 6454.9 & B & 4283.1 \\
I & 7992.0 & V & 5417.9 \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{d-deltat.eps}}
\caption[]{{\em Left:} Average $\Delta T_{\rm max}$ for all investigated
CVs as
a function of isophotal wavelength of the comparison band.
{\em Right:} Average $\Delta T_{\rm max}$ of all individual light
curves, weighted in proportion to their duration. The broken
line is a linear least-squares fit to the data points
(excluding the points referring to the $W$ and $U$ bands),
weighted by the inverse of their standard deviations. The
error bars represent the mean error of the mean.
In both frames the location of the reference band ($B$) is
indicated by a cross.}
\label{d-deltat}
\end{figure}
In Fig.~\ref{d-deltat} (left)
the average $\Delta T_{\rm max}$ for each system is shown as a
function of the isophotal wavelength $\lambda_{\rm iso}$ of the comparison band.
$\lambda_{\rm iso}$ was calculated using the transmission curves as listed
by Bessell (\cite{Bessell}) for the $UBVRI$ system and by Lub \& Pel
(\cite{Lub}) for the
Walraven system [applying a correction based on the revision of the
mean wavelengths by de Ruiter \& Lub (\cite{deRuiter})], together with a first-order
approximation of the spectral density $S_\lambda$ of a CVs according to the
law $S_\lambda \propto \lambda^{-7/3}$, valid over a wide wavelength range for
an accretion disk radiating locally as a black body (Lydnen-Bell
\cite{Lynden-Bell}). The results are listed in
Table~\ref{Isophotal wavelengths}. Since the transmission curves of the Stiening
system are unknown, $\lambda_{\rm iso}$ was taken to be the same as for the
respective bands of the $UBVRI$ system. Moreover, for simplicity in the case
of the Walraven system the isophotal wavelength of the $U$, $B$ and $V$ bands
were substituted by those of the $UBVRI$ system in Fig.~\ref{d-deltat}.
Finally, $\Delta T_{\rm max}$ between
the $B$ and $R$ band of the only light curve of UX~UMa was discarded because
of a very high negative value, deviating strongly from all other results. This
is due to a peculiar and highly asymmetric CCF. The location of the (reference)
$B$ band (where $\Delta T_{\rm max}$ should evidently be zero)
is indicated by a cross in Fig.~\ref{d-deltat}.
There is a certain trend for $\Delta T_{\rm max}$ to increase
with wavelength for the photometric bands redward of $U$, meaning that on
average there is a slight offset between the flickering in the blue and in the
visual or red or infrared in the sense that flickering occurs later at longer
wavelength. But this trend is not continued to the ultraviolet $W$ and $U$
bands. It may not be a coincidence that the points deviating from the linear
relationship refer to those bands that have an isophotal wavelength shorter
than the Balmer limit (see Sect.~\ref{Fireballs and the case of SS Cyg}).
The right frame of Fig.~\ref{d-deltat} is similar to the left one. However,
instead of showing the average $\Delta T_{\rm max}$ values of the respective
systems, the average of all individual light curves, weighted in proportion to
their lengths, is shown as a function of the wavelength of the comparison band.
The error bars represent the mean error of the mean (i.e.\ the
$\sigma/\sqrt{n}$, where $\sigma$ is the standard deviation and $n$ is the
number of data points contributing to the average). The broken line is a
linear least-squares fit (excluding the $W$ and $U$ bands)
where the data points were weighted according to the
inverse of the standard deviation. It has an inclination of
$(1.9 \pm 0.3) \, 10^{-3}$ sec/\AA\footnote{Weighting alternatively
the data points
according to the number of contributing light curves changes this value only
slightly to $(2.1 \pm 0.4) \, 10^{-3}$ sec/\AA.}. Again, the cross indicates
the location of the reference band that did not enter the fit.
The trends observed in Fig.~\ref{d-deltat} suggests the consistent
presence of flickering time lags in CVs, but may not yet be considered
conclusive evidence. This is the more so because the data points shortward of
the Balmer jump deviate from the linear relationship. It may be that well-defined time lags in some system with high-quality data are ``diluted'' by
large errors from lower quality light-curves when regarding the entire
ensemble. This is suggested by more convincing results on some individual
systems that I discuss subesequently.
\subsection{Individual systems}
\label{Individual systems}
While there is thus a trend for a time lag of the flickering at different
wavelengths when regarding the whole ensemble of investigated light curves,
this is, as mentioned, in general not evident when regarding individual stars.
There are exceptions, however. Concentrating on high-quality, low-noise light
curves with a time resolution of $\le$5\hoch{s} and using as criteria
(i) that a reasonable number of light curves are available to
render the results statistically reliable, (ii) that the standard deviation of
the $\Delta T_{\rm max}$ values of the individual curves is significantly smaller
than the difference of the average $\Delta T_{\rm max}$ at short and long
wavelengths, and (iii) that the average $\Delta T_{\rm max}$ is a monotonic
function of wavelength and has a different sign for comparison bands shorter
and longer than the reference band, two systems were identified where
$\Delta T_{\rm max}$ has a convincing dependence on wavelength. These are
V603~Aql and TT~Ari.
\begin{table}
\caption{Restricted results for individual systems}
\label{Restricted results for individual systems}
\vspace{1em}
\begin{tabular}{l@{\hspace{1ex}}lll}
\hline
& & \phantom{-}V603 Aql & TT Ari \\
\hline
$\Delta T_{\rm max}\, (B/U)$ & (sec) &
\phantom{0}-4.5$\pm$2.3 (7) &
-2.0$\pm$0.6 (16) \\
$\Delta T_{\rm max}\, (B/V)$ & (sec) &
\phantom{0-}4.1$\pm$1.8 (7) &
\phantom{-}1.1$\pm$0.8 (16) \\
$\Delta T_{\rm max}\, (B/R)$ & (sec) &
\phantom{-}10.6$\pm$2.9 (7) &
\phantom{-}2.3$\pm$0.5 (13) \\
$\Delta T_{\rm max}\, (B/I)$ & (sec) &
\phantom{-}17.2$\pm$7.2 (2) &
\phantom{-}-- \\ [1ex]
d$\Delta T_{\rm max}$/d$\lambda$ & (sec/\AA) &
\phantom{-}(5.0$\pm$0.3)\,$10^{-3}$ &
\phantom{-}(1.5$\pm$0.1)\,$10^{-3}$ \\
\hline
\end{tabular}
\end{table}
For both stars, these results are based almost exclusively on the high
time resolution $UBVR*$ light curves. Only for V603~Aql, additionally,
two $UBVRI$ data sets were used.
Table~\ref{Restricted results for individual systems}
lists the average $\Delta T_{\rm max}$ values
together with their statistical errors. The number of contributing light
curves is given in brackets. Figure~\ref{individu} (upper frames), which is
organized in
the same way as the right frame of Fig.~\ref{d-deltat}, show the same
results graphically. Clearly, $\Delta T_{\rm max}$ strictly obeys a linear
relationship with wavelength $\lambda,$ which, in these cases, continues to
the $U$ band. The derivative of $\Delta T_{\rm max}$
with respect to $\lambda$, that is,\ the inclination of a linear fit to
the data points weighted by the inverse of their standard deviation, is
listed in the last line of
Table~\ref{Restricted results for individual systems}.
If I relax the criteria formulated in the first paragraph of this
subsection, permitting that in the bands blue-ward of the Balmer jump
$\Delta T_{\rm max}$ deviates from a monotonical relationship with wavelength,
and including also light curves observed with lower time resolution
(but requiring instead that at least four remaining data points are available),
Table~\ref{Mean flickering time lags}
shows that the relationship
between $\Delta T_{\rm max}$ and $\lambda$ is also quite significant for
RS~Oph (lower left frame of Fig.~\ref{individu}) for wavelengths longer
than the Balmer limit. For this system, the derivative
${\rm d}\Delta T_{\rm max}/{\rm d}\lambda = (3.6 \pm 0.5)\, 10^{-3}$ sec/\AA.
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{individu.eps}}
\caption[]{Average $\Delta T_{\rm max}$ derived from high time
resolution light curves of V603~Aql, TT~Ari and RS~Oph as
a function of isophotal wavelength of the comparison band.
The broken
lines are linear least-squares fits to the data points (in
the case of RS~Oph without considering the points referring
to the $W$ and $U$ bands),
weighted by the inverse of their standard deviations.
The location of the reference band ($B$) is
indicated by a cross. The error bars represent the mean error
of the mean.}
\label{individu}
\end{figure}
${\rm d}\Delta T_{\rm max}/{\rm d}\lambda$
differs for the three stars, being more than three times higher in V603~Aql
than in TT~Ari, with the value for RS~Oph lying in between. This holds true
for the average, but also for inclinations
derived from the individual light curves. The magnitude of the flickering
time lags thus depends on the particular star, but within a given system
it appears to be more or less stable\footnote{In RS~Oph the
scatter of ${\rm d}\Delta T_{\rm max}/{\rm d}\lambda$ of individual light
curves is larger than in V603~Aql and TT~Ari, which may be due to the
inferior quality of the available data.}.
However, this statement being based
here on only three stars still requires confirmation from high-quality data of
other CVs.
Scaringi et al.\ (\cite{Scaringi13}) found that the time lags in the
stars studied by them depend on the frequency of the fluctuations, which are
larger for variations occuring on longer time scales (and even may change
sign on short time-scales). The CCF technique applied here is not able to
distinguish between different temporal frequencies. Therefore, the observed
time lags should represent some kind of weighted mean over all time scales.
Since the CCF is most sensitive to the strongest variations, the mean is
expected to be biased to the time lags of the dominating fluctuations. Bruch
(\cite{Bruch89}, \cite{Bruch92}) has shown that the many CVs these occur on
time scales of the order of 2\hoch{m} -- 4\hoch{m} as measured
from the width of the central peak of the auto-correlation function of their
light curves.
The fact that a time lags between the flickering in different photometric
bands is clearly measured in the two systems with the highest quality data
available for this study [plus at least one additional system (RS~Oph)
and also the stars observed by Scaringi et al. (\cite{Scaringi13})]
gives confidence that the trend observed in the
entire ensemble is not accidental, but a strong indication that a similar
time lag is a universal feature of the flickering in CVs.
\section{Discussion}
\label{Discussion}
As mentioned in the introduction, reports about a time lag of the
flickering in different spectral regions have been published earlier, but
the scope and quantity of these studies remains quite limited. In two older
papers and in two more recent studies a positive detection is claimed.
The finding of Szkody \& Margon (\cite{Szkody}) was that in AM~Her the flickering
in the $V$ band lags that in the $U$ band by 8\hoch{s} is not based on the
measurement of the peak of the CCF, but on its asymmetry. They measured the
lag at the half-intensity point. But it is by no means obvious how
an asymmetry in the peak of a CCF is related to a time lag between the
correlated functions. Indeed, tests with artificial light
curves showed that it is not easy to introduce asymmetries in the CCFs
by shifting the flares in one of the bands or by changing their
shape\footnote{Even the extreme case where the flares are triangular in
the reference band and saw-tooth-shaped in the comparison band (setting the
rising branch of the original triangle to zero) only introduced a very
slight asymmetry (but changed the location of the maximum significantly).}.
However, such asymmetries can be caused by longer time-scale variations if
these are significantly different in the reference and comparison bands.
Therefore, it is not clear how reliable the results of Szkody \& Margon
(\cite{Szkody}) are. They interpreted their findings in the framework of the
AM~Her star model, which does not apply to the CVs studied here.
More relevant in the present connection is the study of Jensen et al.\
(\cite{Jensen}),
who found a correlations between the flickering in X-rays and the optical
in TT~Ari, the same system where I have detected a highly significant
dependence of $\Delta T_{\rm max}$ on the wavelength in
Sect.~\ref{Individual systems}. The X-rays lag the optical
variations by about a minute. Jensen et al.\ (\cite{Jensen}) interpreted their
results within a model where X-rays are emitted by a
corona above the accretion disk. The correlation of variations in the
different bands can then be understood as due to the efficiency of acoustic
and magneto-hydrodynamic transport processes from the disk to the corona.
However, Jensen et al.\ (\cite{Jensen}) did not investigate this scenario
quantitatively. Balman \& Revnivtsev (\cite{Balman}) obtained similar results
for additional CVs, but interpreted them
differently: In the framework of a model of propagating
fluctuations (e.g., Lyubarskii \cite{Lyubarskii}), the lag would be due to
the travel time of matter from the innermost part of a truncated accretion
disk to the surface of the white dwarf.
While Szkody \& Margon (\cite{Szkody}) focused on the special
case of a strongly magnetic CV, and Jensen et al.\ (\cite{Jensen}) as
well as Balman \& Revnivtsev (\cite{Balman}) compared
flickering in widely differing wavelength ranges, the study of flickering
time delays in MV~Lyr and LU~Cam by Scaringi et al.\ (\cite{Scaringi13}) can
be compared much more directly to the present one. As mentioned, at low
flickering frequencies (several times $10^{-2}$~Hz as read from their
Fig.~3), they measured a delay of $\approx 10$ sec in
LU~Cam and $\approx 3$ sec in MV~Lyr of the $r'$ band variations with respect
to those in
the $u'$ band. With the average SDSS filter wavelengths as informed on the
SDSS web page (and neglecting the small difference between average and
isophotal wavelength), this translates into a time lag of $3.8\, 10^{-3}$
sec/\AA\, and $1.1\, 10^{-3}$ sec/\AA, respectively. These numbers are
quite similar to the time lags observed here, strengthening the case for
such lags to be ubiquitous in CVs.
\subsection{A heuristic model}
\label{A heuristic model}
The time lags observed here in the optical bands are much smaller than
those seen by Jensen et al.\ (\cite{Jensen}) and Balman \& Revnivtsev
(\cite{Balman}) and have the opposite wavelength
dependence. This is not surprising because the reasons for the lags in the
optical bands probably are quite different from those for the lag
between the optical and X-ray range. Here, I find that
variations in the blue occur slightly earlier than in the red.
In principle this can be understood if during the development of a flickering
flare the radiation characteristics of the underlying light source change
such that in the early phases of the flare more short wavelength radiation is
emitted, and later on, the peak of the emission shifts to the red.
To verify that this leads to results that are compatible
with the observations, a very simple scenario was investigated.
I stress that this is not meant as a physically realistic model but
just to show that based on simple but sensible assumptions, it is
possible to reproduce the results found in Sect.~\ref{Results}.
For this purpose the light source underlying a flickering flare
was approximated by a black body with a temperature
$\Theta$\footnote{I chose the symbol $\Theta$ for the temperature to avoid confusion with the symbol $T$ that is used in the quantity
$\Delta T_{\rm max}$ introduced earlier
for the time lag of the flickering in different bands.}
that over time $t$ varies according to some function $\Theta(t)$. The emitting
area $\cal A$ of the black body was also considered a function of time:
${\cal A}(t)$. The radiation
emitted by the black body as a function of wavelength $\lambda$ is then
given by the Planck function $B_\lambda [\Theta(t)]$ and ${\cal A}(t)$. The
radiation flux $F(t)$ detected in a passband of a photometric system is thus
\begin{displaymath}
F(t) = {\rm const} \int_{\lambda_1}^{\lambda_2} B_\lambda[\Theta(t)] \,
{\cal A}(t) \, \cal{T}(\lambda) \, {\rm d}\lambda,
\end{displaymath}
where $\cal{T}(\lambda)$ is the transmission function of the photometric
band, and the integration extends between its upper and lower wavelength
cutoff. The constant depends on the distance of the radiation source and is
of no relevance in the present context.
In the absence of a specific physical model for the flickering flares,
the functions $\Theta(t)$ and ${\cal A}(t)$ are free parameters. They are chosen
here by trial and error such that the resulting flare profiles and their time
lag $\Delta T_{\rm max}$ as a function of wavelength are compatible with the
observations. Of course, the choices made here can by no means be
considered as unique. The combination of different functions may well lead
to similar results.
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{flaresim.eps}}
\caption[]{{\em Top:} Assumed temperature (solid line, left-hand scale)
and emitting area (broken line, right-hand scale) evolution
of a black body used in simulations of the brightness
development of a flickering flare as a function of time.
{\em Bottom:}
Calculated profiles of a simulated flickering flare in the
bands of the $UBVRI$ system. The inset shows the times of
maxima (dots) as a function of the isophotal wavelength of the
respective bands, together with a linear least-square fit
(broken line). See text for details.}
\label{flaresim}
\end{figure}
The time development of the emitting area is taken to be
${\cal A}(t) \propto t^2$.
This is based on the simple notion of a light source that expands linearly
with time (which may not necessarily be realistic!). With respect to the
temperature evolution, good results are obtained with a function that
rises rapidly from low temperatures to a maximum of just above 30\,000~K
and then gradually drops to about 6\,000~K at the end of the flare. Both
${\cal A}(t)$ (right-hand scale; normalized to a maximum value of 1; broken
line) and $\Theta(t)$ (left-hand scale; solid line) are shown as a
function of time (the flare duration being
normalized to $t_{\rm fl}=1$) in the upper frame of Fig.~\ref{flaresim}.
With this choice of temperature and area evolution, and using the transmission
functions of Bessell (\cite{Bessell}), $F(t)$ (i.e.\ the
flare light curve) was calculated for the
passbands of the $UBVRI$ system. The constant was chosen such that $F(t)$
obtains a maximum of $F(t)=1$ in the $U$ band.
To mimic a constant background
light, a quantity $F_{\rm bg}=1$ was added to all light curves before
transforming them into magnitudes with an arbitrary zero point equal to
the magnitude of the first point in the light curve corresponding to each
passband. These light curves, depicting the magnitude development of the
flare, are shown in the lower frame of Fig.~\ref{flaresim}.
The simulated flare has several characteristics in common with real flickering
flares: (i) Bruch (\cite{Bruch92}) found the mean ratio of the gradients of
the rise and decline of flickering flares in the $B$ band of many light
curves (measured between the base and the peak of the flare) to be
$1.18 \pm 0.07$, meaning that on average the rise is slightly more rapid
than the decline. In the simulated flare a very similar gradient ratio of
1.21 is measured. But note that this value depends quite strongly on the
choice of $\Theta(t)$ and ${\cal A}(t)$.
(ii) The amplitude of the flare is highest in the $U$
band and declines monotonically in longer wavelength bands\footnote{The
similarity of the amplitudes in $V$ and $R$ is a consequence of the
significantly larger width of the $R$ passband.}. This agrees with
observations. However, it is more difficult to make a quantitative assessment
here because the amplitude depends on the amount of the constant background
light, which was (unrealistically) assumed to be the same in all passbands in
the current simulations. (iii) There is a clear dependence of the time of
maximum on the passband. This is shown in the inset in the lower frame of
Fig.~\ref{flaresim}, where the maximum times (dots) are plotted as a function
of the isophotal wavelength of the photometric band. The broken
line is a linear least-squares fit to the data.
Thus, there is a time lag of the flickering observed at different
wavelengths. The least-squares fit
yields ${\rm d}\Delta T_{\rm max}/{\rm d}\lambda = 1.4 \times 10^{-5}$ time
units/\AA. To estimate $t_{\rm fl}$ , the average width
$\sigma$ of a Gauss fit to the central peak of the auto correlation function
of the $B$ band
of the light curves of TT~Ari used to derive the results quoted in
Sect.~\ref{Individual systems} was measured:
$\sigma = 145\hoch{s} \pm 23\hoch{s}$.
Taking this as the typcial flare duration (i.e.\ $t_{\rm fl} \equiv \sigma$),
${\rm d}\Delta T_{\rm max}/{\rm d}\lambda = 2.1 \times 10^{-3}$
sec/\AA. This is of the same order of magnitude as the observed time lag
for TT~Ari (Table~\ref{Restricted results for individual systems}).
Thus, the simple scenario explored here can explain important characteristics
of the flickering, in particular the main result of the present study, namely
the observed time lag of the flickering at different optical wavelengths.
The numerical value of ${\rm d}\Delta T_{\rm max}/{\rm d}\lambda$ is of the
same order of magnitude
as that measured in real light curves and depends on the choice of the
temperature and emission area evolution. Although not investigated here, it
may be expected that deviations from black-body radiation have similar
effects.
The results of this exercise hold for a single flare. But what about an
entire light curve composed of many flares? If flickering were
to consist of
the additive superposition of many independent flares, it would be easy to
construct such a light curve and investigate its properties. However, there
are strong indications in CVs (e.g.\ Scaringi et al. \cite{Scaringi12b}) and
even more so in X-ray binaries (e.g.\ Uttley et al. \cite{Uttley05}) that
flickering is the result not of an additive, but of a multiplicative process.
The origins of disturbances in CV system that gives rise to flickering are
largely speculative, and the physical mechanisms of how they lead to the
emission of optical radiation are unknown. Therefore, the properties of an
ensemble of (nonlinearly) superposed flares are not straightforward to predict,
and I refrain from investigating this issue further.
\subsection{Fireballs and the case of SS Cyg}
\label{Fireballs and the case of SS Cyg}
The heuristic model discussed in the previous subsection can obviously not
substitute a physical model for flickering flares. It was meant to show
that parameter combinations for a light source exist that can plausibly
explain the observed time lag of the flickering as a function of wavelength.
The evolution of the (black body) temperature and the emitting area during
the flare event was adjusted such that the flicker shape resembles
that observed in real light curves.
A more sophisticated model must take into account the physical conditions in
the flare light source. An interesting attempt in this direction has been
published by Pearson et al.\ (\cite{Pearson}).
They modelled the light source as a fireball,
that is,\ a hot, spherically symmetric expanding ball of gas, and calculated
the emitting spectrum as a function of the input parameters and of time,
permitting them to construct light curves at various wavelengths. They
compared the results with light curves (time resolution: 2\hoch{s}) in six
bands between 3590~\AA\, and 7550~\AA\, of a particular flare extracted from
high-speed spectro-photometric
observations of the dwarf nova SS~Cyg. Adjusting the model parameters such
that the observed light curves are well fitted yields encouraging results if
the fireball is assumed to be in isothermal expansion.
To investigate how the results of Pearson et al.\ (\cite{Pearson})
fit in with the present study, I extracted the light curves of SS~Cyg
(consisting just of a single flare) from their Fig.~14. For convenience,
the one corresponding to the
band centred on 4545~\AA\, is shown in the left panel of Fig.~\ref{lc-sscyg}.
In the same way as has been done for the other systems investigated in this
study, the light curves of SS~Cyg were cross-correlated, using the one centred
on 4225~\AA\, as reference. The time lag $\Delta T_{\rm max}$ with respect
to the reference band is shown as a function of wavelength in the right panel
of Fig.~\ref{lc-sscyg} (filled circles). The wavelength of the reference band
is marked by a cross.
\begin{figure}
\centering
\resizebox{\hsize}{!}{\includegraphics{lc-sscyg.eps}}
\caption[]{{\em Left frame:} A flickering flare of SS~Cyg centred on
4545~\AA\, on 1998, July 8, as observed by Pearson et al.\
(\cite{Pearson}).
{\em Right frame:} $\Delta T_{\rm max}$ as a function of
wavelength
of the comparison band for the observational data (filled
circles) and the preferred fireball model (open circles).
The solid and broken lines
represent a linear
least-squares fit to the data points (shortest wavelength
point excluded). The location of the reference band is
indicated by a cross. The dots with error bars are
the average $\Delta T_{\rm max}$ values of SS~Cyg taken from
Table~\ref{Mean flickering time lags},
with the triangle representing the location
of the respective reference band.}
\label{lc-sscyg}
\end{figure}
For comparison, the average $\Delta T_{\rm max}$ values for SS~Cyg from
Table~\ref{Mean flickering time lags}
are shown in the figure as small dots with error bars (mean errors of
the mean). The triangle
indicates the location of the respective reference band. Unfortunately,
Table~\ref{Mean flickering time lags}
only contains data of SS~Cyg taken in the $UBV$ system, defining thus but
two points in the diagram. Therefore, no strong conclusions can be drawn
from the comparison with the
results extracted from Fig.~14 of Pearson et al.\ (\cite{Pearson}). However,
both data sets appear to be roughly compatible with each other.
Except for the shortest wavelength band centred on 3615~\AA,\, the data points
from Pearson et al. (\cite{Pearson})
follow a linear relationship (solid line) with an
inclination of $(2.3 \pm 0.1) \, 10^{-3}$ sec/\AA. This is very similar to
the behaviour of V603~Aql, TT~Ari, and RS~Oph
(Table~\ref{Restricted results for individual systems}
and Fig.~\ref{individu}).
I also determined the times of maximum light of the preferred model of
Pearson et al.\ (\cite{Pearson}) by fitting a high-order polynomial to the
peak of the model curves extracted from their Fig.~14. The corresponding
$\Delta T_{\rm max}$ values are shown as open
circles in Fig.~\ref{lc-sscyg}. They follow a linear relationship
(disregarding the data point at 3615~\AA; broken line) with an inclination
of $(4.86 \pm 0.01) \, 10^{-3}$ sec/\AA. This is twice the value found
for the observational data, indicating a slight disagreement between model
and observations.
Although the fireball model is thus quite attractive to explain the time lag
of the flickering as a function of wavelength, a vexing question remains:
Why does the data point at Pearson et al.'s (\cite{Pearson}) 3615~\AA\, band
deviate from the linear relationship between $\Delta T_{\rm max}$ and
$\lambda$? Or, more generally, how can the tendency for a deviation from
such a relationship at wavelengths shorter than the Balmer limit
(Fig.~\ref{d-deltat}) be
explained? This can be due to optical-depth effects in the flaring
light source. Pearson et al.\ (\cite{Pearson}) found that their fireball is
initially dominated by flux coming from optically thick regions. This
changes in the course of its evolution, so that at late times, optically
thin emission dominates. Thus, more Balmer continuum emission can escape from
the light source at later phases, strengthening the flare at ultraviolet
wavelengths and reversing the $\Delta T_{\rm max}$ -- $\lambda$ relationship
observed at longer wavelengths.
The balance between optically thick and thin radiation will depend on the
detailed structure of the fireball. Therefore, the late dominance of the
Balmer continuum emission may be seen in some, but not all cases, explaining
the different behaviour of the flickering time lag at very short wavelength
in individual systems (e.g.\ V603~Aql and TT~Ari vs.\ RS~Oph and SS~Cyg).
\subsection{Alternative scenarios}
\label{Alternative scenarios}
Scaringi et al.\ (\cite{Scaringi13}) sought to explain the time lags along
lines different from fireballs (Sect.~\ref{Fireballs and the case of SS Cyg})
or an evolving black-body-like light source, as
investigated in Sect.~\ref{A heuristic model}. They discarded explanations
in the context of models of flickering caused by the inward propagation of
perturbations within the accretions disk [see e.g.\
Lyubarskii (\cite{Lyubarskii}), Pavlidou et al.\ (\cite{Pavlidou}),
Kotov et al.\ (\cite{Kotov}) and Ar\'evalo \& Uttley (\cite{Arevalo}) for
detailed models of this kind for X-ray binaries and galactic
nuclei, and Yonehara et al.\ (\cite{Yonehara}),
Dobrotka et al.\ (\cite{Dobrotka10}) and Scaringi (\cite{Scaringi14}) for CVs], which assume that the observed optical variations are
directly related to the inward propagation of matter in the accretion disk
because then flickering variations in the blue should lag those in the red.
This is contrary to what is observed.
They also rejected a scenario of (practically) instantaneous reprocessing of
light of a variable continuum source close to the centre of the accretion
disk by some structure farther out. The time lag should then basically be
equal to the light travel time. However, the typical dimensions of CVs are
such that only the smallest observed time lags would not exceed the light
travel time from the centre to the periphery of the disk. Moreover, the
emission at long wavelengths should then be dominated by reprocessed light,
which may not be likely.
If, however, reprocessing does not occur instantaneously, but on the local
thermal time scale Scaringi et al.\ (\cite{Scaringi13}) are able to
qualitatively reconcile the observed time lags with reprocessing sites not
too far out, provided that only the surface layer of the accretion disk
reprocesses photons from the variable light source. As a final scenario,
Scaringi et al.\ (\cite{Scaringi13}) mentioned
reverse shocks in the accretions disk (Krauland et al.\ \cite{Krauland})
that may originated in the boundary layer and deposit energy first in the
hotter inner parts of the disk and then in the cooler outer regions.
None of scenarios mentioned in the previous paragraph
is backed by a more detailed physical model. Therefore,
it is hard to say if they lead to time scales compatible with the observed
lags and are in accordance with other characteristics of the flickering such
as, for example, the temporal and spectral development of individual flares.
But it would be premature to discard them before an attempt has
been made to predict the implied emission properties in some detail and to
compare them with observations. It appears that the fireball model
is at present the only physical model that has been advanced to a stage
where definite predictions can be made
about at least some flickering properties and that realistic parameter
combinations can be found such that these predictions comply with
observations. It must be recognized, however, that the model does not
explain what makes the fireball explode in the first place. Moreover,
it is not known to which degree a (linear or nonlinear) superposition of
individual fireball events can reproduce the statistical properties of
observed flickering light curves. Thus, it is
by no means clear, which of the mentioned scenarios, if any, is realized
in nature.
\section{Conclusions}
\label{Conclusions}
Confirming earlier results of Scaringi et al.\ (\cite{Scaringi13}),
it has been shown conclusively in the present study that
flickering in cataclysmic variables does not occur exactly simultaneously in
different photometric bands across the optical range. Instead, there is a
time delay in the sense that individual events of the flickering develop
slightly later at red than at blue wavelengths. In many of the
investigated CVs, this effect can only be seen statistically in the ensemble
of the available data because most of the light curves used are not of
sufficient quality to resolve the time delay well enough
in particular data sets. However, in the systems with the best data,
V603~Aql and TT~Ari, the effect can convincingly be measured individually.
This is also true for RS~Oph and SS~Cyg. To these systems MV~Lyr and
LU~Cam, observed by Scaringi et al.\ (\cite{Scaringi13}), can be added.
The time lag of the flickering as a function of wavelengths is of the
order of a couple of milliseconds per \AA ngstrom. The exact value may
depend on the particular system, but appears not to change much within the
same star on time scales of several years (i.e.\ the period spanned
by the observational data; this does, of course, not preclude time lag
changes when systems like TT~Ari go into a low state).
The observed effect can be explained if the evolution of the
radiation characteristics of the light source(s) responsible for the
flickering is such that individual flares reach their peak emission
slightly earlier at blue than at red wavelengths. It is not difficult to
construct simple scenarios that are able to reproduce the observed time
lag not only qualitatively, but also quantitatively, and which also agree
with other properties that are generally observed in the flickering.
While such scenarios lead to the observed effects, they can by no means
substitute a realistic physical mechanism. The fireball model of Pearson
et al. (\cite{Pearson}) is most advanced when it comes to the actual
radiation emission process that may be seen as flickering. It can reasonably
well reproduce a flickering flare observed in SS~Cyg in various wavelengths
ranges not only with respect to the wavelength dependent time lag, but also
with respect to the spectral and temporal development. However,
the detailed emission processes implied in the context of other scenarios
deserve to be investigated as well. This hold true especially for models
where, contrary to the fireball model, a better understanding of the origin
of underlying disturbance exists, for instance,\ the fluctuating disk model.
It may
be interesting to investigate whether the basic fireball radiation physics
can be applied to these models as well.
\begin{acknowledgements}
I am deeply indebted to all those colleagues who put their light cuves of
cataclysmic variables at my disposal. For the current study I used data
contributed by N. Beskrovanaya, A.\ Hollander, R.E.\ Nather, M.\ Niehues,
T.\ Schimpke, and N.M.\ Shakovskoy.
I am particularly grateful to R.E.\ Robinson and E.-H.\ Zhang for their
excellent light curves observed in the Stiening system, which were decisive
for the success of this work. I thank the referee,\ Simone Scaringi,
for many critical comments that helped to improve this publication.
\end{acknowledgements}
|
\section{Introduction}
Let $p$ be a prime and $\f_{q_1},\f_{q_2}\subset\overline\f_p$, where $\overline\f_p$ is the algebraic closure of $\f_p$. A $q_1$-linearized polynomial over $\f_{q_2}$ is a polynomial of the form
\[
f=a_0\x^{q_1^0}+a_1\x^{q_1^1}+\cdots+a_n\x^{q_1^n}\in\f_{q_2}[\x].
\]
If $f\in\f_{q_2}[\x]$ is $q_1$-linearized and $g\in\f_{q_1}[\x]$ is $q_2$-linearized, then $f\circ g=g\circ f$.
We are interested in the following question. Let $\f_r\subset \overline\f_p$ and $q=r^m$. Assume that $z\in\overline\f_p$ satisfies an equation
\begin{equation}\label{1.1}
\sum_{i=0}^{m-1}a_if_i(z)^{r^i}=0,
\end{equation}
where $a_i\in\f_q$ and $f_i\in\f_r[\x]$ is $q$-linearized. Note that \eqref{1.1} is an $r$-linearized equation with coefficients in $\f_q$. Is it possible to derive from \eqref{1.1} a $q$-linearized equation with coefficients in $\f_r$? Such an equation indeed exists and will be given in Section 2; see \eqref{2.3}. The $q$-linearized equation \eqref{2.3} is concise in a determinant form. However, the expansion of the determinant invokes certain questions concerning permutations of $\Bbb Z_m$ and partitions of $m$; these related questions will also be discussed in Section 2.
What is achieved is a transition from an $r$-linearized equation with coefficients in $\f_q$ to a $q$-linearized equation with coefficients in $\f_r$. We mention that for the applications in the present paper and other possible applications elsewhere, a $q$-linearized equation with coefficients in $\f_q$ would suffice. The advantage of a $q$-linearized polynomial $f\in\f_q[\x]$ over an $r$-linearized polynomial $g\in\f_q[\x]$ resides in a folklore in the study of finite fields. The conventional associate of $f$ is a polynomial in $\f_q[\x]$ from which information about the roots of $f$ can be easily extracted. On the other hand, the conventional associate of $g$ is a skew polynomial over $\f_q$ which is not as convenient to use as the counterpart of $f$. For more background of skew polynomials over finite fields, see \cite{Caruso-LeBorhne-arXiv1212.3582, Giesbrecht-JSC-1998}.
Let $f=\sum_{i=0}^na_i\x^{q^i}\in\overline\f_p[\x]$ be a $q$-linearized polynomial. The conventional associate of $f$ is the polynomial $\widetilde f=\sum_{i=0}^na_ix^i\in\overline\f_p[\x]$. The folklore mentioned above is the following theorem which can be derived from \cite[Theorems~ 3.62 and 3.65]{Lidl-Niederreiter-97}.
\begin{thm}\label{T1.1}
Let $f,g\in\f_q[\x]$ be $q$-linearized polynomials. Then $\text{\rm gcd}(f,g)$ is a $q$-linearized polynomial over $\f_q[\x]$ with $\widetilde{\text{\rm gcd}(f,g)}=\text{\rm gcd}(\widetilde f,\widetilde g)$.
\end{thm}
In Section 3, we use the result of Section 2 to answer certain questions arising from a recent study of permutation polynomials \cite{Fernando-Hou-Lappano-FFA-2013,Hou-FFA-2012}. For each integer $n\ge 0$, let $g_{n,q}\in\f_p[\x]$ be the polynomial defined by the functional equation $\sum_{c\in\f_q}(\x+c)^n=g_{n,q}(\x^q-\x)$. The study in \cite{Fernando-Hou-Lappano-FFA-2013,Hou-FFA-2012} aims at the determination of the triples $(n,e;q)$ of positive integers for which $g_{n,q}$ is a permutation polynomial (PP) of $\f_{q^e}$. When $n$ is related to $q$ and $e$ in certain ways, the result of Section 2 combined with Theorem~\ref{T1.1} produces sufficient conditions for $g_{n,q}$ to be a PP of $\f_{q^e}$. These conditions explain numerous findings from the computer searches in \cite{Fernando-Hou-Lappano-FFA-2013}.
\section{$r$-Linearized and $r^m$-Linearized Equations}
Let $p$ be a prime, $\f_r\subset\overline\f_p$ and $q=r^m$. Let $R_q$ denote the set of all $q$-linearized polynomials over $\f_q$. $(R_q,+,\circ)$ is a commutative ring, where $+$ is the ordinary addition and $\circ$ is the composition. The mapping that sends $f\in R_q$ to its conventional associate $\widetilde f$ is a ring isomorphism from $R_q$ to $\f_q[\x]$. Throughout the paper, the $i$-fold composition of a polynomial $f$ is denoted by $f^{[i]}$ while $f^i$ stands for the $i$th power of $f$.
Assume that for $0\le i\le m-1$, $a_i\in\f_q$ and $f_i\in\f_r[\x]$ is $q$-linearized. Define
\begin{equation}\label{2.1}
M=\left[
\begin{matrix}
a_0f_0& a_1f_1&\cdots& a_{m-1}f_{m-1}\cr
a_{m-1}^rf_{m-1}\circ\x^q& a_0^rf_0&\cdots&a_{m-2}^rf_{m-2}\cr
\vdots&\vdots&&\vdots\cr
a_1^{r^{m-1}}f_1\circ\x^q& a_2^{r^{m-1}}f_2\circ\x^q&\cdots&a_0^{r^{m-1}}f_0
\end{matrix}\right]\in{\rm M}_{n\times n}(R_q).
\end{equation}
\begin{thm}\label{T2.1}
In the above notation, assume that $z\in\overline\f_p$ satisfies the equation
\begin{equation}\label{2.2}
\sum_{i=0}^{m-1}a_if_i^{r^i}(z)=0.
\end{equation}
Then we have
\begin{equation}\label{2.3}
(\det M)(z)=0,
\end{equation}
where $\det M$ is a $q$-linearized polynomial over $\f_r$.
\end{thm}
\begin{proof}
Raise the left side of \eqref{2.2} to the power of $r^j$, $0\le j\le m-1$, and express the results in a matrix form. We have
\begin{equation}\label{2.4}
\Bigl(\sum_{j=0}^{m-1}\left[
\begin{matrix}
a_jf_j^{r^j}\cr
a_{j-1}^rf_{j-1}^{r^j}\cr
\vdots\cr
a_0^{r^j}f_0^{r^j}\cr
a_{m-1}^{r^{j+1}}f_{m-1}^{r^j}\circ\x^q\cr
\vdots\cr
a_{j+1}^{r^{m-1}}f_{j+1}^{r^j}\circ\x^q
\end{matrix}\right]\Bigr)(z)=0.
\end{equation}
Label the rows and columns of $M$ from $0$ through $m-1$. Let $M_0$ be the submatrix $M$ with its $0$th column deleted, and, for $0\le i\le m-1$, let $M_{i,0}$ be the submatrix of $M_0$ with its $i$th row deleted. Put $D_i=(-1)^i\det M_{i,0}$, $0\le i\le m-1$. Then for each $1\le j\le m-1$,
\begin{equation}\label{2.5}
\begin{split}
&\Bigl([D_0,\cdots,D_{m-1}]\circ \left[
\begin{matrix}
a_jf_j^{r^j}\cr
\vdots\cr
a_0^{r^j}f_0^{r^j}\cr
a_{m-1}^{r^{j+1}}f_{m-1}^{r^j}\circ\x^q\cr
\vdots\cr
a_{j+1}^{r^{m-1}}f_{j+1}^{r^j}\circ\x^q
\end{matrix}\right]\Bigr)(z)\cr
=\,&\Bigl([D_0,\cdots,D_{m-1}]\circ \left[
\begin{matrix}
a_jf_j\cr
\vdots\cr
a_0^{r^j}f_0\cr
a_{m-1}^{r^{j+1}}f_{m-1}\circ\x^q\cr
\vdots\cr
a_{j+1}^{r^{m-1}}f_{j+1}\circ\x^q
\end{matrix}\right]\Bigr)(z^{r^j})\cr
=\,&\Bigl(\det\left.\left[
\begin{matrix}
a_jf_j\cr
\vdots\cr
a_0^{r^j}f_0\cr
a_{m-1}^{r^{j+1}}f_{m-1}\circ\x^q\cr
\vdots\cr
a_{j+1}^{r^{m-1}}f_{j+1}\circ\x^q
\end{matrix}\ \right|\ M_0\ \right]\Bigr)(z^{r^j})=0.
\end{split}
\end{equation}
Combining \eqref{2.4} and \eqref{2.5} gives
\[
0=\Bigl([D_0,\cdots,D_{m-1}]\circ \left[
\begin{matrix}
a_0f_0\cr
a_{m-1}^rf_{m-1}\circ\x^q\cr
\vdots\cr
a_1^{r^{m-1}}f_1\circ\x^q
\end{matrix}\right]\Bigr)(z)=(\det M)(z).
\]
The claim that the coefficients of $\det M$ are all in $\f_r$ will be proved shortly; see \eqref{2.11} and \eqref{2.12}.
\end{proof}
For $i,j\in\{0,\dots,m-1\}$, define
\[
\delta(i,j)=
\begin{cases}
0&\text{if}\ i\le j,\cr
m&\text{if}\ i>j.
\end{cases}
\]
The $(i,j)$ entry of $M$ is
\[
a_{-i+j}^{r^i}f_{-i+j}\circ\x^{r^{\delta(i,j)}},
\]
where the subscripts are taken modulo $m$. Let $\text{Sym}(\Bbb Z_m)$ denote the group of all permutations of $\Bbb Z_m$. We have
\begin{equation}\label{2.6}
\det M=\sum_{\sigma\in\text{Sym}(\Bbb Z_m)}(-1)^{\text{sgn}(\sigma)}\Bigl(\prod_{i=0}^{m-1}a_{\sigma(i)-i}^{r^i}f_{\sigma(i)-i}\Bigr)\circ\x^{r^{\sum_{i=0}^{m-1}\delta(i,\sigma(i))}}.
\end{equation}
Let $0^{\mu_0}1^{\mu_1}\cdots(m-1)^{\mu_{m-1}}$ denote the multiset consisting elements $0,\dots,m-1$ with respective multiplicities $\mu_0,\dots,\mu_{m-1}$. If there exists $\sigma\in\text{Sym}(\Bbb Z_m)$ such that $\{\sigma(i)-i:i\in\Bbb Z_m\}= 0^{\mu_0}1^{\mu_1}\cdots(m-1)^{\mu_{m-1}}$, then we must have $\sum_{i=0}^{m-1}i\mu_i\equiv 0\pmod m$. (In fact, the converse is also true; see Remark~\ref{R2.2}.) Let
\begin{equation}\label{2.7}
\frak M=\Bigl\{(\mu_0,\mu_1,\dots,\mu_{m-1}):\mu_i\in\Bbb Z,\ \mu_i\ge 0,\ \sum_{i=0}^{m-1}\mu_i=m,\ \sum_{i=0}^{m-1}i\mu_i\equiv 0\pmod m\Bigr\},
\end{equation}
and for each $(\mu_0,\dots,\mu_{m-1})\in\frak M$, let
\begin{equation}\label{2.8}
\frak S_{(\mu_0,\dots,\mu_{m-1})}=\bigl\{\sigma\in\text{Sym}(\Bbb Z_m):\{\sigma(i)-i:i\in\Bbb Z_m\}=0^{\mu_0}\cdots(m-1)^{\mu_{m-1}}\bigr\}.
\end{equation}
Let $\sigma\in \frak S_{(\mu_0,\dots,\mu_{m-1})}$ and $i\in\{0,\dots,m-1\}$. Then $\delta(i,\sigma(i))+\sigma(i)-i$ is the integer in $\{0,\dots,m-1\}$ that is $\equiv\sigma(i)-i\pmod m$. Hence in $\Bbb Z$, we have
\begin{equation}\label{2.9}
\sum_{i=0}^{m-1}\delta(i,\sigma(i))=\sum_{i=0}^{m-1}\bigl[\delta(i,\sigma(i))+\sigma(i)-i\bigr]=\sum_{i=0}^{m-1}i\mu_i.
\end{equation}
Thus we can rewrite \eqref{2.6} as
\begin{equation}\label{2.10}
\det M=\sum_{(\mu_0,\dots,\mu_{m-1})\in\frak M}c_{\mu_0\dots\mu_{m-1}}f_0^{[\mu_0]}\circ\cdots\circ f_{m-1}^{[\mu_{m-1}]}\circ\x^{r^{0\mu_0+\cdots+(m-1)\mu_{m-1}}},
\end{equation}
where
\begin{equation}\label{2.11}
c_{\mu_0\dots\mu_{m-1}}=\sum_{\sigma\in\frak S_{(\mu_0,\dots,\mu_{m-1})}}(-1)^{\text{sgn}(\sigma)}\prod_{i=0}^{m-1}a_i^{\sum_{j,\,\sigma(j)-j\equiv i}r^j}.
\end{equation}
Let $\alpha\in\text{Sym}(\Bbb Z_m)$ be defined by $\alpha(i)=i-1$. We have $\alpha^{-1}\frak S_{(\mu_0,\dots,\mu_{m-1})}\alpha=\frak S_{(\mu_0,\dots,\mu_{m-1})}$. Moreover, for $\sigma\in \frak S_{(\mu_0,\dots,\mu_{m-1})}$,
\begin{equation}\label{2.12}
\Bigl(\prod_{i=0}^{m-1}a_i^{\sum_{j,\,\sigma(j)-j\equiv i}r^j}\Bigr)^r=\prod_{i=0}^{m-1}a_i^{\sum_{j,\,(\alpha^{-1}\sigma\alpha)(j)-j\equiv i}r^j}.
\end{equation}
\eqref{2.11} and \eqref{2.12} imply that $c_{\mu_0\cdots\mu_{m-1}}^r=c_{\mu_0\cdots\mu_{m-1}}$, i.e., $c_{\mu_0\cdots\mu_{m-1}}\in\f_r$.
\begin{rmk}\label{R2.2}\rm
In fact, $\frak S_{(\mu_0,\dots,\mu_{m-1})}\ne\emptyset$ for all $(\mu_0,\dots,\mu_{m-1})\in\frak M$. This follows immediately from the following result by M. Hall.
\end{rmk}
\begin{thm}[\cite{Hall-PAMS-1952}]\label{T2.3}
Let $A$ be a finite abelian group and $f:A\to A$ a function. $f$ can be represented as a difference of two permutations of $A$ if and only if $\sum_{x\in A}f(x)=0$.
\end{thm}
We conclude this section with the explicit expansion of $\det M$ for $1\le m\le 5$. For simplicity, we write $f_{i_0}\circ\cdots\circ f_{i_{m-1}}\circ \x^{i_0+\cdots+i_{m-1}}$ as $f_{i_0\dots i_{m-1}}$.
$m=1$,
\[
\det M=a_0f_0.
\]
$m=2$,
\[
\det M=\text{N}_{r^2/r}(a_0)f_{00}-\text{N}_{r^2/r}(a_1)f_{11}.
\]
$m=3$,
\[
\det M=\text{N}_{r^3/r}(a_0)f_{000}+\text{N}_{r^3/r}(a_1)f_{111}+\text{N}_{r^3/r}(a_2)f_{222}-\text{Tr}_{r^3/r}(a_0a_1^ra_2^{r^2})f_{012}.
\]
$m=4$,
\[
\begin{split}
&\det M=\cr
&\text{N}_{r^4/r}(a_0)f_{0000}-\text{N}_{r^4/r}(a_1)f_{1111}+\text{N}_{r^4/r}(a_2)f_{2222}-\text{N}_{r^4/r}(a_3)f_{3333}\cr
&-\text{Tr}_{r^2/r}\bigl(\text{N}_{r^4/r^2}(a_0)\text{N}_{r^4/r^2}(a_2)^r\bigr)f_{0022}+\text{Tr}_{r^2/r}\bigl(\text{N}_{r^4/r^2}(a_1)\text{N}_{r^4/r^2}(a_3)^r\bigr)f_{1133}\cr
&-\text{Tr}_{r^4/r}(a_0^{1+r}a_1^{r^2}a_3^{r^2})f_{0013}+\text{Tr}_{r^4/r}(a_0a_1^{r+r^2}a_2^{r^3})f_{0112}\cr
&+\text{Tr}_{r^4/r}(a_0a_2^ra_3^{r^2+r^3})f_{0233}-\text{Tr}_{r^4/r}(a_1a_2^{r+r^2}a_3^{r^3})f_{1223}.
\end{split}
\]
$m=5$,
\[
\begin{split}
&\det M=\cr
&\text{N}_{r^5/r}(a_0)f_{00000}-\text{Tr}_{r^5/r}(a_0^{1+r+r^2}a_1^{r^3}a_4^{r^4})f_{00014}\cr
&-\text{Tr}_{r^5/r}(a_0^{1+r+r^3}a_2^{r^2}a_3^{r^4})f_{00023}+\text{Tr}_{r^5/r}(a_0^{1+r}a_1^{r^2+r^3}a_3^{r^4})f_{00113}\cr
&+\text{Tr}_{r^5/r}(a_0^{1+r^2}a_1^{r^3}a_2^{r+r^4})f_{00122}+\text{Tr}_{r^5/r}(a_0^{1+r}a_2^{r^2}a_4^{r^3+r^4})f_{00244}\cr
&+\text{Tr}_{r^5/r}(a_0^{1+r^2}a_3^{r+r^3}a_4^{r^4})f_{00334}-\text{Tr}_{r^5/r}(a_0a_1^{r+r^2+r^3}a_2^{r^4})f_{01112}\cr
&+\text{Tr}_{r^5/r}(a_0a_1^{r+r^3}a_4^{r^2+r^4})f_{01144}\cr
&+\text{Tr}_{r^5/r}(-a_0a_1^ra_2^{r^2}a_3^{r^3}a_4^{r^4}+a_0a_1^{r^2}a_2^{r^4}a_3^{r}a_4^{r^3}-a_0a_1^{r^3}a_2^{r}a_3^{r^4}a_4^{r^2})f_{01234}\cr
&-\text{Tr}_{r^5/r}(a_0a_1^{r^2}a_3^{r+r^3+r^4})f_{01333}-\text{Tr}_{r^5/r}(a_0a_2^{r+r^2+r^4}a_4^{r^3})f_{02224}\cr
&+\text{Tr}_{r^5/r}(a_0a_2^{r+r^2}a_3^{r^3+r^4})f_{02233}-\text{Tr}_{r^5/r}(a_0a_3^{r}a_4^{r^2+r^3+r^4})f_{03444}\cr
&+\text{N}_{r^5/r}(a_1)f_{11111}-\text{Tr}_{r^5/r}(a_1^{1+r+r^3}a_3^{r^2}a_4^{r^4})f_{11134}\cr
&+\text{Tr}_{r^5/r}(a_1^{1+r}a_2^{r^2+r^3}a_4^{r^4})f_{11224}+\text{Tr}_{r^5/r}(a_1^{1+r^2}a_2^{r^3}a_3^{r+r^4})f_{11233}\cr
&-\text{Tr}_{r^5/r}(a_1a_2^{r+r^2+r^3}a_3^{r^4})f_{12223}-\text{Tr}_{r^5/r}(a_1a_2^{r^2}a_4^{r+r^3+r^4})f_{12444}\cr
&+\text{Tr}_{r^5/r}(a_1a_3^{r+r^2}a_4^{r^3+r^4})f_{13344}+\text{N}_{r^5/r}(a_2)f_{22222}\cr
&+\text{Tr}_{r^5/r}(a_2^{1+r^2}a_3^{r^3}a_4^{r+r^4})f_{22344}-\text{Tr}_{r^5/r}(a_2a_3^{r+r^2+r^3}a_4^{r^4})f_{23334}\cr
&+\text{N}_{r^5/r}(a_3)f_{33333}+\text{N}_{r^5/r}(a_4)f_{44444}.
\end{split}
\]
\section{Applications to Permutation Polynomials}
\subsection{A criterion}\
\begin{prop}\label{P3.1}
Let $m$ and $e$ be positive integers, $r$ a prime power and $q=r^m$. Define $S_e=\x^{q^0}+\x^{q^1}+\cdots+\x^{q^{e-1}}$. A polynomial $f\in\f_{q^e}[\x]$ is a PP of $\f_{q^e}$ if the following conditions are all satisfied.
\begin{itemize}
\item [(i)] There exists a PP $\bar f\in\f_q[\x]$ of $\f_q$ such that the diagram
\[
\beginpicture
\setcoordinatesystem units <3mm,3mm> point at 0 0
\ar{1 0}{5 0}
\ar{1 6}{5 6}
\ar{0 5}{0 1}
\ar{6 5}{6 1}
\put {$\f_{q^e}$} at 0 6
\put {$\kern1mm\f_{q^e}$} at 6 6
\put {$\f_q$} at 0 0
\put {$\f_q$} at 6 0
\put {$\scriptstyle f$} [b] at 3 6.2
\put {$\scriptstyle \bar f$} [t] at 3 -0.2
\put {$\scriptstyle S_e$} [r] at -0.2 3
\put {$\scriptstyle S_e$} [l] at 6.2 3
\endpicture
\]
commutes.
\item[(ii)] For each $c\in\f_q$, there exist $q$-linearized polynomials $f_{c,i}\in\f_r[\x]$ and $a_{c,i}\in\f_q$, $0\le i\le m-1$, and $b_c\in\f_{q^e}$ such that
\begin{equation}\label{3.1}
f(x)=f_c(x)+b_c\quad\text{for all}\ x\in S_e^{-1}(c),
\end{equation}
where
\begin{equation}\label{3.2}
f_c=\sum_{i=0}^{m-1}a_{c,i}f_{c,i}^{r^i}.
\end{equation}
\item[(iii)] For each $c\in\f_q$,
\begin{equation}\label{3.3}
\text{\rm gcd}\bigl(\det A_c,\ (\x^e-1)/(\x-1)\bigr)=1,
\end{equation}
where
\begin{equation}\label{3.4}
A_c=\left[
\begin{matrix}
a_{c,0}\widetilde f_{c,0} & a_{c,1}\widetilde f_{c,1} &\cdots& a_{c,m-1}\widetilde f_{c,m-1}\cr
a_{c,m-1}^r\widetilde f_{c,m-1}\x & a_{c,0}^r\widetilde f_{c,0} &\cdots& a_{c,m-2}^r\widetilde f_{c,m-2}\cr
\vdots&\vdots&&\vdots\cr
a_{c,1}^{r^{m-1}}\widetilde f_{c,1}\x & a_{c,2}^{r^{m-1}}\widetilde f_{c,2}\x &\cdots& a_{c,0}^{r^{m-1}}\widetilde f_{c,0}
\end{matrix}\right],
\end{equation}
\end{itemize}
and $\widetilde{(\ )}$ denotes the conventional associate of a $q$-linearized polynomial over $\f_q$.
\end{prop}
\begin{proof}
By \cite[Lemma~1.2]{Akbary-Ghioca-Wang-FFA-2011}, it suffices to show that for every $c\in\f_q$, $f_c$ is 1-1 on $S_e^{-1}(c)$. Since $f_c$ is a linearized polynomial, it suffices to show that $0$ is the only common root of $f_c$ and $S_e$. By Theorem~\ref{T2.1}, a root of $f_c$ is also a root of $\det M_c$, where
\begin{equation}\label{3.5}
M_c=\left[
\begin{matrix}
a_{c,0}f_{c,0} & a_{c,1}f_{c,1} &\cdots& a_{c,m-1}f_{c,m-1}\cr
a_{c,m-1}^rf_{c,m-1}\circ\x^q & a_{c,0}^rf_{c,0} &\cdots& a_{c,m-2}^rf_{c,m-2}\cr
\vdots&\vdots&&\vdots\cr
a_{c,1}^{r^{m-1}}f_{c,1}\circ\x^q & a_{c,2}^{r^{m-1}}f_{c,2}\circ\x^q &\cdots& a_{c,0}^{r^{m-1}}f_{c,0}
\end{matrix}\right].
\end{equation}
Therefore, it suffices to show that $\text{gcd}(\det M_c,\,S_e)=\x$. We have
\[
\text{gcd}(\widetilde{\det M_c},\ \widetilde S_e)=\text{gcd}\bigl(\det A_c,\ (\x^e-1)/(\x-1)\bigr)=1,
\]
and by Theorem~\ref{T1.1}, $\text{gcd}(\det M_c,\,S_e)=\x$.
\end{proof}
\subsection{The polynomial $g_{n,q}$}\
Let $p=\text{char}\,\f_q$. For each integer $n\ge 0$, there is a polynomial $g_{n,q}\in\f_p[\x]$ defined by the functional equation
\begin{equation}\label{3.6}
\sum_{c\in\f_q}(\x+c)^n=g_{n,q}(\x^q-\x).
\end{equation}
The polynomial $g_{n,q}$ was introduced in \cite{Hou-JCTA-2011}, and its permutation property was studied in \cite{Fernando-Hou-Lappano-FFA-2013,Hou-FFA-2012}. The objective is to determine the triples $(n,e;q)$ of positive integers for which $g_{n,q}$ is a PP of $\f_{q^e}$, and this question, as a whole, appears to be difficult. For each integer $a\ge 0$, define $S_a=\x+\x^q+\cdots+\x^{q^{a-1}}$. The polynomial $g_{n,q}$ is related to $S_a$ by the following lemma.
\begin{lem}\label{L3.2}
\begin{itemize}
\item [(i)] {\rm (\cite[Eq.~(3.5)]{Fernando-Hou-Lappano-FFA-2013})} For integers $n\ge 0$ and $a\ge b\ge 0$, we have
\begin{equation}\label{3.7}
g_{n+q^a,q}-g_{n+q^b,q}=(S_a-S_b)\cdot g_{n,q}.
\end{equation}
\item[(ii)] {\rm (\cite[Lemma~2.2]{Fernando-Hou-Lappano-DM-2014})} Let $n=1+q^{a_1}+\cdots+q^{a_{q+t}}$, where $-1\le t\le q-4$ and $a_1,\dots,a_{q+t}\ge 0$. Then
\begin{equation}\label{3.8}
g_{n,q}=-\sum_{1\le i_1<\cdots<i_{t+2}\le q+t}S_{a_{i_1}}\cdots S_{a_{i_{t+2}}}.
\end{equation}
\end{itemize}
\end{lem}
\subsection{Applications to $g_{n,q}$}\
We call a triple of positive integers $(n,e;q)$ {\em desirable} if $g_{n,q}$ is a PP of $\f_{q^e}$. Computer searches for desirable triples with small values of $q$ and $e$ were conducted in \cite{Fernando-Hou-Lappano-FFA-2013,Hou-FFA-2012}; the results for $q=4$ and $e\le 6$ were given in \cite[Table~3]{Fernando-Hou-Lappano-FFA-2013}. Proposition~\ref{P3.1} provides an explanation for several entries of \cite[Table~3]{Fernando-Hou-Lappano-FFA-2013}; in each case, a new class of PPs is discovered.
\begin{prop}\label{P3.3}
Let $q=4$ and $n=1+q^a+q^b+q^e+q^{e+k}$, where $a,b,e,$ and $k$ are positive integers. Then
\begin{equation}\label{3.9}
g_{n,q}\equiv S_aS_b+(S_a+S_b+S_e)S_k+S_e^2\pmod{\x^{q^e}-\x}.
\end{equation}
If $\text{\rm gcd}(e,2k)=1$ and $a=k$ or $b=k$, then $g_{n,q}$ is a PP of $\f_{q^e}$.
\end{prop}
\begin{proof}
Eq.~\eqref{3.9} follows from Lemma~\ref{L3.2} easily. It remains to prove the second claim. Since $a=k$ or $b=k$, \eqref{3.9} gives
\[
g_{n,q}\equiv S_k^2+S_e^2+S_eS_k\pmod{\x^{q^e}-\x}.
\]
We show that conditions (i) -- (iii) in Proposition~\ref{P3.1} are satisfied with $r=2$, $m=2$, $q=4$ and $f=S_k^2+S_e^2+S_eS_k$. Condition (i) is satisfied with $\bar f=\x^2$. For each $c\in\f_q$ and $x\in S_e^{-1}(c)$,
\[
f(x)=cS_k(x)+S_k^2(x)+c^2.
\]
Hence (ii) is satisfied with
\[
f_{c,0}=f_{c,1}=S_k,\ a_{c,0}=c,\ a_{c,1}=1,\ b_c=c^2.
\]
We have
\[
\det A_c=\left|\begin{matrix} c(1+\x+\cdots+\x^{k-1})&1+\x+\cdots+\x^{k-1}\cr \x(1+\x+\cdots+\x^{k-1})& c^2(1+\x+\cdots+\x^{k-1})\end{matrix}\right|=\frac{(\x^{2k}+1)(\x+c^3)}{(\x+1)^2}.
\]
Since $\text{gcd}(e,2k)=1$, $\text{gcd}(\det A_c,\ (\x^e-1)/(\x-1))=1$, and hence (iii) is also satisfied.
\end{proof}
The proofs of the next three propositions are similar to that of Proposition~\ref{P3.3}. For these proofs, we only give $f$, $f_{c,i}$, $a_{c,i}$, $b_c$ and $\det A_c$ and leave the details for the reader.
\begin{prop}\label{P3.4}
Let $q=4$ and $n=1+3q^a+q^e+2q^{e+a}$, where $e$ and $a$ are positive integers. Then
\[
g_{n,q}\equiv\x^{q^a}+ S_e+S_a^2S_e^2+S_aS_e^3\pmod{\x^{q^e}-\x}.
\]
If $2\mid e$ and $\text{\rm gcd}(e,2a+1)=1$, then $g_{n,q}$ is a PP of $\f_{q^e}$.
\end{prop}
\begin{proof}
$f=\x^{q^a}+ S_e+S_a^2S_e^2+S_aS_e^3$, $\bar f=\x$, $f_{c,0}=\x^{q^a}+c^3S_a$, $f_{c,1}=S_a$, $a_{c,0}=1$, $a_{c,1}=c^2$, $b_c=c$,
\[
\det A_c=\x^{2a}+c^3\frac{\x^{2a}+1}{\x+1}.
\]
\end{proof}
\begin{prop}\label{P3.5}
Let $q=4$ and $n=1+2q^1+2q^{e-1}+2q^{e+1}$, where $e>1$ is an integer. Then
\[
g_{n,q}\equiv S_2+\x^2 S_e^2+S_{e-1}^2S_e^2\pmod{\x^{q^e}-\x}.
\]
If $e$ is odd, then $g_{n,q}$ is a PP of $\f_{q^e}$.
\end{prop}
\begin{proof}
$f=S_2+\x^2 S_e^2+S_{e-1}^2S_e^2$, $\bar f=\x$, $f_{c,0}=S_2$, $f_{c,1}=\x+S_{e-1}$, $a_{c,0}=1$, $a_{c,1}=c^2$, $b_c=0$,
\[
\det A_c=\frac1{(1+\x)^2}(1+c^3\x^3+\x^4+c^3\x^{2e-1}).
\]
\end{proof}
\begin{prop}\label{P3.6}
Let $q=4$ and $n=1+2q^1+q^3+q^e+2q^{e+1}$, where $e$ is a positive integer. Then
\[
g_{n,q}\equiv \x^{q^2}+S_e+\x^2S_e^2+S_3S_e^3\pmod{\x^{q^e}-\x}.
\]
If $2\mid e$ but $3\nmid e$, then $g_{n,q}$ is a PP of $\f_{q^e}$.
\end{prop}
\begin{proof}
$f=\x^{q^2}+S_e+\x^2S_e^2+S_3S_e^3$, $\bar f=\x$, $f_{c,0}=\x^{q^2}+c^3S_3$, $f_{c,1}=\x$, $a_{c,0}=1$, $a_{c,1}=c^2$, $b_c=c$, $\det A_c=\x^4+c^3(1+\x+\x^2+\x^4)$.
\end{proof}
\begin{prop}\label{P3.7}
Let $q=4$ and $n=1+2q^2+q^4+q^e+2q^{e+2}$, where $e$ is a positive integer. Then
\[
g_{n,q}\equiv \x^{q^3}+S_e+S_2^2S_e^2+S_4S_e^3 \pmod{\x^{q^e}-\x}.
\]
If $2\mid e$ but $5\nmid e$, then $g_{n,q}$ is a PP of $\f_{q^e}$.
\end{prop}
\begin{proof}
$f=\x^{q^3}+S_e+S_2^2S_e^2+S_4S_e^3$, $\bar f=\x$, $f_{c,0}=\x^{q^3}+c^3S_4$, $f_{c,1}=S_2$, $a_{c,0}=1$, $a_{c,1}=c^2$, $b_c=c$, $\det A_c=\x^6+c^3(1+\x+\x^2+\x^3+\x^4+\x^6)$.
\end{proof}
Examples of Propositions~\ref{P3.3} -- \ref{P3.7} provide explanations for several entries in \cite[Table 3]{Fernando-Hou-Lappano-FFA-2013}; an update of the table is included in the apendix of the present paper.
By Proposition~\ref{P3.3}:
\[
\begin{split}
&q=4,\ e=5,\ n=17429=1+q^1+q^2+q^5+q^7;\\
&q=4,\ e=5,\ n=17489=1+q^2+q^3+q^5+q^7;\\
&q=4,\ e=5,\ n=17681=1+q^2+q^4+q^5+q^7.
\end{split}
\]
By Proposition~\ref{P3.4}:
\[
\begin{split}
&q=4,\ e=4,\ n=2317=1+3q^1+q^4+2q^5;\\
&q=4,\ e=6,\ n=135217=1+3q^2+q^6+2q^8.
\end{split}
\]
By Proposition~\ref{P3.5}:
\[
q=4,\ e=5,\ n=8713=1+2q^1+2q^4+2q^6.
\]
By Proposition~\ref{P3.6}:
\[
q=4,\ e=4,\ n=2377=1+2q^1+q^3+q^4+2q^5.
\]
By Proposition~\ref{P3.7}:
\[
q=4,\ e=6,\ n=135457=1+2q^2+q^4+q^6+2q^8.
\]
\begin{rmk}\label{R3.8}\rm
The statements in the above propositions can be made more general when they are not restricted to the class $g_{n,q}$. For example, Propositions~\ref{P3.5} and \ref{P3.6} allow the following generalizations whose proofs require no additional work.
\begin{itemize}
\item A generalization of Proposition~\ref{P3.5}: Let $q=4$ and $f=S_a+\x^2S_e^2+S_b^2S_e^2$, where $a$, $b$, and $e$ are positive integers. Then $f$ is a PP of $\f_{q^e}$ if $2\mid(a+b)$, $\text{gcd}(e,a)=1$, and
\[
\text{gcd}\Bigl(\frac{\x^{2a}+1+\x^{2b+1}+\x^3}{(\x+1)^2},\, \frac{\x^e+1}{\x+1}\Bigr)=1.
\]
In this case, $\bar f=\x$, $f_{c,0}=S_a$, $f_{c,1}=\x+S_b$, $a_{c,0}=1$, $a_{c,1}=c^2$, $b_c=0$, and
\[
\det A_c=\frac 1{(1+\x)^2}\bigl[\x^{2a}+1+c^3(\x^{2b+1}+\x^3)\bigr].
\]
\item A generalization of Proposition~\ref{P3.6}: Let $q=4$ and $f=S_a+S_b+S_e+\x^2S_e^2+S_bS_e^3$, where $a$, $b$, and $e$ are positive integers. Then $f$ is a PP of $\f_{q^e}$ if $2\mid(a+e)$, $\text{gcd}(e,a-b)=1$, and
\[
\text{gcd}\Bigl(\frac{\x^{2a}+\x^3+\x+1}{(\x+1)^2},\ \frac{\x^e+1}{\x+1}\Bigr)=1.
\]
In this case, $\bar f=\x$, $f_{c,0}=S_a+(1+c^3)S_b$, $f_{c,1}=\x$, $a_{c,0}=1$, $a_{c,1}=c^2$, $b_c=c$, and
\[
\det A_c=\frac 1{(1+\x)^2}\bigl[\x^{2a}+\x^{2b}+c^3(1+\x+\x^3+\x^{2b})\bigr].
\]
\end{itemize}
\end{rmk}
|
\section{Introduction}
In heavy ion collisions at ultra-relativistic energies nuclear matter can be studied under extreme conditions. At zero baryo-chemical potential and finite temperature, a cross-over between the hadron gas at low temperatures and the quark gluon plasma at high temperatures has been established by lattice QCD \cite{Borsanyi:2013bia, Bazavov:2014pvz} and also been confirmed by experimental measurements at the Large Hadron Collider (LHC) and the Relativistic Heavy Ion Collider (RHIC) in combination with dynamical models \cite{Pratt:2015zsa}. At finite net baryon densities, current lattice calculations are not applicable and insights about the structure of the QCD phase diagram can only be achieved by experimental exploration in combination with detailed dynamical modeling and phenomenological approaches providing input on the equation of state in the whole temperature ($T$)-baryo-chemical potential ($\mu_B$) plane.
The previous and currently running programs at CERN-SPS and the recent beam energy scan (BES) program at RHIC provide a comprehensive data set on general bulk observables and selected fluctuation/correlation measurements. In the future, the second generation of experiments with improved capabilities for rare probes will meticulously explore the exciting energy regime accessible at FAIR, NICA, and stage 2 of the BES@RHIC. The second goal, besides obtaining insight on the type of the phase transition of strongly interacting matter, is to understand the properties of the state of matter that is formed, the QGP.
One of the most promising observables is the anisotropic flow, that is measured in terms of Fourier coefficients of the azimuthal distribution of produced particles. Directed flow $v_1$ and elliptic flow $v_2$ are analysed extensively to gain insights on the transport properties and the geometry of heavy ion collisions. Within the last 5 years the whole plethora of odd higher coefficients has been studied to investigate the initial state and its fluctuations in more detail. The flow coefficients - as indicated by the name - are a sign of collective behaviour and quantify the response of the system to spatial anisotropies, translated to momentum space by pressure gradients.
To connect the final state particle distributions to the quantities of interest, e.g. the equation of state or the viscosity to entropy ratio of QCD matter, detailed dynamical models are necessary. The current state of the art is to use a combination of 3+1 dimensional (viscous) hydrodynamic evolution and hadronic transport for the non-equilibrium evolution in the late stages of the reaction. There are various ways to model/parametrize the initial state for these calculation ranging from Glauber/CGC type models, to classical Yang-Mills evolution, AdS/CFT or transport approaches. At lower beam energies the initial non-equilibrium evolution takes a non-negligible time and one needs to certainly pay attention to the type of initial dynamics that is required.
In the present study, the calculations are performed within the Ultra-relativistic Quantum Molecular Dynamics (UrQMD) approach including a 3+1 dimensional ideal hydrodynamic evolution where appropriate. The main ingredients of the model are described in Section \ref{hybrid_model}. After that, Section \ref{flow} contains a summary of the expectations for the beam energy dependence of anisotropic flow in heavy ion collisions and the three subsections \ref{elliptic_flow}, \ref{tri_flow} and \ref{directed_flow} are devoted to $v_1$ to $v_3$ in more detail. The last Section \ref{summary} summarizes the main conclusions.
\section{Hybrid Approach at non zero $\mu_B$}
\label{hybrid_model}
Heavy ion collisions are processes that do not create a system at fixed temperature and density that can be studied over an extended period of time, but rather span a fluctuating finite region in the phase diagram evolving dynamically. Hybrid approaches, based on (viscous) hydrodynamics for the hot and dense stage of the evolution, and non-equilibrium transport for the dilute stages of the reaction are successfully applied to describe bulk observables at high RHIC and LHC energies. These approaches combine the advantages of hydrodynamics and transport and apply both approximations within their respective regions of validity. The equation of state and transport coefficients of hot and dense QCD matter are direct inputs to the hydrodynamic equations, which allows for a controlled modeling of the phase transition. Microscopic transport on the other hand describes all particles and their interactions and provides the full phase-space distribution of produced particles in the final state, which allows for an apples-to-apples comparison to experimental data.
At higher beam energies the standard picture of the dynamical evolution of a heavy ion reaction consists of non-equilibrium initial state dynamics providing the initial state for the viscous hydrodynamics equations and a hadronic afterburner to account for the separation of chemical and kinetic freeze-out. At lower beam energies, it is not clear how close the produced system really gets to local equilibration and if it thermalizes in the whole phase-space. In general, it is expected that dissipative effects are larger for low beam energies and the hadronic interactions gain importance. In addition, the widely used assumption of boost invariance breaks down and a 3+1 dimensional calculation is inevitable. Also, the equation of state and the transport coefficients are required not only as a function of temperature, but in the whole $T-\mu_B$ - plane. The finite net baryon current needs to be conserved during the evolution. Going beyond the standard hybrid approach by exploring the non-equilibrium dynamics in the vicinity of a critical endpoint or a first order phase transition is the ultimate goal, where exploratory work has been performed so far \cite{Nahrgang:2011mg}.
The dynamical approach that is used to calculate the results presented below is the Ultra-relativistic Quantum Molecular Dynamics (UrQMD) approach coupled to ideal relativistic fluid dynamics for the hot and dense stage (SHASTA) \cite{Bass:1998ca, Bleicher:1999xi, Petersen:2008dd}. This approach has recently been improved to include the finite shear viscosity during the hydrodynamic evolution as presented in another contribution to this conference and in \cite{Karpenko:2015xea}. The initial state is produced by generating nucleons according to Woods-Saxon distributions and computing dynamically the first binary interactions until the two nuclei have geometrically passed through each other. Event -by-event fluctuations of the positions of the nucleons and of the energy deposition per collision are naturally included. At $t_{\rm start} = \frac{2R}{\gamma v}$ the energy, momentum and net baryon density distributions are calculated by representing the individual particles with a three dimensional Gaussian distribution. Based on these initial conditions including fluctuating velocity profiles the ideal relativistic hydrodynamic equations are solved. There are different options for the equation of state, e.g. a bag model equation of state matched to a hadron gas at low temperatures to explore the sensitivity of observables to a strong first order phase transition (BM) or a more realistic equation of state provided by a chiral model that is fitted to lattice QCD at zero $\mu_B$, reproduces the nuclear ground state properties and incorporates constraints from neutron star properties. At a constant energy density a hypersurface is constructed and particles are sampled according to the Cooper-Frye formula. The hadronic rescattering and decays are treated by the hadronic transport approach. This approach incorporates most of the above mentioned ingredients that are necessary for a realistic dynamical description of heavy ion reactions at lower beam energies. The current strategy can be seen as a first step to take the well-established picture at high energies and explore how well it works at lower beam energies.
\section{Anisotropic Flow Observables}
\label{flow}
\begin{figure}[hbt]
\centering
\includegraphics[width=7cm]{v2exc_ch_withpot}
\includegraphics[width=8cm]{v2epsilon}
\caption{(Color online) Left: The calculated beam energy excitation function of elliptic flow of charged particles in Au+Au/Pb+Pb
collisions in mid-central collisions (b=5-9 fm) with $|y|<0.1$(full line). This curve is compared to data from
different experiments for mid-central collisions (see \cite{Petersen:2009vx} for refs). The dotted line
in the low energy regime depicts UrQMD calculations with the mean field \cite{Li:2006ez}. Fig. from \cite{Petersen:2006vm}. Right: $v_2/\epsilon$ as a function of $(1/S)dN_{\rm ch}/dy$ for different energies and centralities for Pb+Pb/Au+Au collisions compared to data \cite{Alt:2003ab}. The results from mid-central collisions (b=5-9 fm) calculated within the hybrid model with different freeze-out transitions and different definitions of the eccentricity are shown by black lines. The green full lines correspond to the previously calculated hydrodynamic limits \cite{Kolb:2000sd}. Fig. taken from \cite{Petersen:2009vx}.}
\label{figv2exc}
\end{figure}
Anisotropic flow is the collective response to geometrical structures in the initial state distribution of heavy ion collisions. If there are strong enough interactions spatial anisotropies are getting translated to momentum space throughout the evolution. Therefore, anisotropic flow is supposed to be very sensitive to the equation of state and to the transport properties of the produced matter. It is quantified by the Fourier coefficients of the azimuthal distribution of the produced particles
\begin{equation}
v_n = \cos (n\phi -\Psi_n)
\end{equation}
where $\Psi_n$ is the event plane and $n$ specifies the coefficient of interest. There are many different two-particle and many-particle methods to measure the anisotropic flow coefficients with different sensitivities to non-flow and fluctuations. In this article, either the event plane method (some $v_2$ results and $v_3$) or the theoretical definition with respect to the known reaction plane ($v_1 = \langle \frac{p_x}{p_x^2+p_y^2}\rangle$ and $v_2 = \langle \frac{p_x^2-p_y^2}{p_x^2+p_y^2}\rangle$) have been applied to calculate the flow observables.
Elliptic flow is one of the main observables to support the claim that the quark gluon plasma behaves like a nearly perfect liquid. Therefore, the beam energy dependence of elliptic flow is supposed to be sensitive to the phase transition and the changing transport properties at lower collision energies. Triangular flow is the first odd flow component that is mainly sensitive to fluctuations and would be zero, if the initial state is taken as average over many events. Therefore, the hope is that measuring the beam energy dependence of triangular flow helps to disentangle the 'trivial' initial state fluctuations from the interesting fluctuations that arise due to the critical dynamics. Last but not least directed flow is first of all a measure of the initial angular momentum in the heavy ion collision. The beam energy dependence was predicted to show a dip structure when a first order phase transition occurs during the evolution. In the following Sections, the beam energy dependence of each of these observables will be discussed in more detail.
\begin{figure}
\centering
\includegraphics[width=6.5cm]{Gv2hydrofraction8294}
\includegraphics[width=6.5cm]{Bv2evob8294}
\caption{(Color online) Left: Fraction of elliptic flow generated during hydrodynamic evolution as a function of beam energy. Right: Magnitude of $v_2\{\textrm{EP}\}$ at the beginning of
hydrodynamical evolution (squares), immediately after particlization (diamonds) and after
the full simulation (circles) in midcentral collisions. Fig. from \cite{Auvinen:2013sba}.}
\label{figv2phases}
\end{figure}
\subsection{Elliptic flow - a measure of the perfect fluid}
\label{elliptic_flow}
\begin{figure}[hb]
\centering
\includegraphics[width=7cm]{v2_ptpi}
\includegraphics[width= 7cm]{v2_imp_eos_sps}
\caption{(Color online) Left: Elliptic flow of pions in mid-central (b=5-9 fm) Pb+Pb collisions at $E_{\rm lab}=40A~$GeV and $E_{\rm lab}=160A~$GeV. The full and dashed black lines depict the hybrid model calculation, while the pure transport calculation is shown as the black dotted line. The colored symbols display experimental data obtained with different measurement methods by NA49 \cite{Alt:2003ab}. Fig. taken from \cite{Petersen:2009vx}.
Right: Centrality dependence of elliptic flow of charged
particles at midrapidity ($|y|<0.5$) for Pb+Pb collisions at $E_{\rm lab}=40A$
GeV (a) and $E_{\rm lab}=160A$ GeV (b). The horizontal lines indicate the
results for averaged initial conditions using the hadron gas EoS (blue full
line) and the bag model EoS (black dotted line) while the symbols (full circles for
HG-EoS and open squares for BM-EoS) depict the results for the event-by-event
calculation. Fig. taken from \cite{Petersen:2010md}.}
\label{figv2_ptpi}
\end{figure}
The second Fourier coefficient, the so called elliptic flow, is the response of the system to the initial almond shaped overlap region in non-central collisions. Due to fluctuations $v_2$ is also non-zero in central collisions, but it is mainly generated due to the overall shape deformation of the fireball. Only if the mean free path is small enough the initial spatial anisotropy is converted to the corresponding momentum space anisotropy. The beam energy dependence of elliptic flow is quite interesting. At very low energies, the spectator nuclei are blocking the interaction region and the so called 'squeeze-out' leads to negative elliptic flow values with respect to the reaction plane. In this region the nuclear interactions are important and a hadron transport approach including mean fields can describe the elliptic flow rather well (see Fig. \ref{figv2exc}, left). At high beam energies the elliptic flow turns positive and grows as a function of beam energy. In the intermediate region around 10-40A GeV the hadron transport approach reaches the right "ball park" values while the underestimation at high beam energies is a sign for the importance of partonic interactions starting at around 160A GeV.
Fig. \ref{figv2exc} (right) shows the response function $v_2/\epsilon_2$ as a function of the charged particle density. The green lines indicate the expectations from ideal hydrodynamic calculations and that this line meets the experimental data at $\sqrt{s_{\rm NN}} = 200$ GeV was the basis for the claim that a perfect fluid has been created. The black lines show various UrQMD hybrid calculations to demonstrate that the generic behaviour of this curve can be reproduced once the non-equilibrium stages of the reaction are treated properly.
To investigate how this energy dependence of elliptic flow is generated in the hybrid approach the contribution to the final integrated elliptic flow of charged particles has been calculated and the percentage is shown in Fig. \ref{figv2phases} (left). At lower energies, the hydrodynamic evolution lasts only a few fm/c, therefore almost all the elliptic flow is generated by the hadronic transport approach. At higher energies the hydrodynamic evolution is necessary to reach the high elliptic flow values as discussed above. Fig. \ref{figv2phases} (right) shows that the contribution from the late hadronic rescattering is constant at about 10 \% throughout the whole energy regime investigated.
Fig. \ref{figv2_ptpi} (left) shows the transverse momentum dependent elliptic flow of pions. NA49 measurements are compared to hybrid and pure hadronic transport calculations and again it is clearly visible that at 40 AGeV the hydrodynamic evolution does not play a role while at 160 AGeV the hydrodynamic evolution is crucial to reproduce the elliptic flow. The dependence on initial state fluctuations and the equation of state has been studied in Fig. \ref{figv2_ptpi} (right), where the centrality dependence of elliptic flow of charged particles is shown. There is no difference visible between the fluctuating events (symbols) and the averaged smooth initial conditions (horizontal lines) as expected for elliptic flow. The difference between the bag model equation of state with a strong first order phase transition and the hadron gas equation of state is not observable either.
\subsection{Triangular flow - a measure of fluctuations}
\label{tri_flow}
Since 2010 event by event fluctuations in the initial state have been studied as the source of higher odd anisotropic flow coefficients \cite{Adare:2012kf, Luzum:2013yya}. These odd coefficients average to zero without event by event fluctuations. To demonstrate that $v_3$ is really sensitive to fluctuations the beam energy dependence of triangular flow of charged particles is compared to the dynamical fluctuations of elliptic flow $\sigma_{v_2}$ in Fig. \ref{fig_v3} (left). In central collisions both have roughly the same value whereas in non-central collisions they have a similar magnitude only at high energies. At lower beam energies triangular flow decreases.
\begin{figure}
\centering
\includegraphics[width=6.5cm]{Bv2fluctuations}
\includegraphics[width=6.5cm]{Bscaledvnvshydrot}
\caption{(Color online) Left: $v_3\{\textrm{EP}\}$ compared with
initial state fluctuations' contribution to $v_2$,
$\sigma_{v2}=\sqrt{\frac{1}{2}(v_2\{\textrm{EP}\}^2-v_2\{\textrm{RP}\}^2)}$
(squares). Right: Scaled flow coefficients $v_2/ \langle \epsilon_2 \rangle$ and
$v_3/ \langle \epsilon_3 \rangle$ with respect to the
average total hydro duration for impact parameter ranges $b = 0-3.4$ fm, $6.7-8.2$ fm
and $8.2-9.4$ fm. Figs taken from \cite{Auvinen:2013sba}.}
\label{fig_v3}
\end{figure}
If non-zero triangular flow is measured at lower beam energies and it is generated by initial spatial fluctuations it serves as a nice probe to disentangle these 'trivial' fluctuations from more interesting fluctuations due to the phase transition or critical endpoint. This option relies of course on the fact that a finite triangular flow needs to be present. As Fig. \ref{fig_v3} (right) shows, $v_3$ scales for different energies and centralities with the duration of the hydrodynamic evolution in the hybrid approach. The circles show that for elliptic flow the influence of the hadronic non-equilibrium evolution is much larger. Therefore, triangular flow is much more sensitive to a finite viscosity and might even vanish at lower beam energies. On the positive side, this disappearance of triangular flow can be seen as a more direct evidence that the perfect fluid QGP does not persist for a long enough time.
\subsection{Directed flow - a measure of the phase transition}
\label{directed_flow}
Last but not least, let us come to one more anisotropic flow observable. The directed flow measures the stream of particles within the reaction plane and has opposite sign in the forward and backward hemisphere. Nowadays, a rapidity-even $v_1$ dipole component has also been observed, but here we concentrate on the traditional rapidity odd measure. To plot the beam energy dependence one extracts the slope of directed flow at midrapidity.
\begin{figure}[t]
\includegraphics[width=6.5cm]{dv1dy_hydro_IC}
\includegraphics[width=6.5cm]{dv1dy_hydro_IT}
\caption{[Color online] Left: Slope of $v_1$ of protons and pions around mid rapidity extracted from the ideal 1-fluid calculations with a bag model and crossover EoS. For particle production we applied a Cooper-Frye prescription on a iso-chronous hypersurface. Right: Slope of $v_1$ of protons and pions around mid rapidity extracted from the ideal 1-fluid calculations with a bag model and crossover EoS. For particle production we applied a Cooper-Frye prescription on a iso-energy density hypersurface. Figs. taken from \cite{Steinheimer:2014pfa}.
}\label{fig_fluid}
\end{figure}
Based on the AGS and NA49 measurements there have been predicitions within hydrodynamic calculations that the slope of the directed flow of protons has a dip as a function of beam energy, signaling the phase transition to the quark gluon plasma \cite{Brachmann:1999xt, Csernai:1999nf}. In light of the new STAR measurements from the RHIC beam energy scan program \cite{Adamczyk:2014ipa}, this old observable has been recalculated with modern theoretical techniques. In Fig. \ref{fig_fluid} a one fluid calculation has been performed. This resembles the previous predictions enhanced by actual particle sampling and a $v_1$ calculation that can be compared to experimental measurements. On the left hand side, the rather unphysical scenario of a constant time freeze-out is shown, whereas the right hand side corresponds to a more realistic iso-energy density transition scenario. The two figures demonstrate, that the freeze-out dynamics has a significant effect on the structures in the excitation function of the slope of the directed flow of protons and pions. The large difference between the two different scenarios - with and without first order phase transition - disappears.
\begin{figure}[t]
\includegraphics[width=6.5cm]{protonv1}
\includegraphics[width=6.5cm]{dv1dy_hybrid_proton}
\caption{[Color online] Left: Comparison of proton $v_1(y)$ for the different model applied. Right: Slope of $v_1$ of protons and anti-protons around mid rapidity extracted from the hybrid model calculations with a bag model and crossover EoS. We compare with standard UrQMD and experimental data.Figs. taken from \cite{Steinheimer:2014pfa}.
}\label{fig_v1proton}
\end{figure}
Figs. \ref{fig_v1proton} and \ref{fig_v1pion} both show on the left hand side the directed flow as a function of rapidity at two different beam energies and on the right hand side the slope around mid rapidity that has been extracted to display the full beam energy dependence. Especially for the proton directed flow it is clear that the signal is very small around zero. Even though all the different calculations capture the positive slope it is obvious that pure one fluid calculations lead to too much directed flow, while surprisingly the pure hadron transport shows the best agreement with experimental data at this point.
\begin{figure}[t]
\includegraphics[width=6.5cm]{pionv1}
\includegraphics[width=6.5cm]{dv1dy_hybrid_pion}
\caption{[Color online] Left: Comparison of pion $v_1(y)$ for the different model applied. Right: Slope of $v_1$ of negatively charged pions around mid rapidity extracted from the hybrid model calculations with a bag model and crossover EoS. We compare with standard UrQMD and experimental data.Figs. taken from \cite{Steinheimer:2014pfa}.
}\label{fig_v1pion}
\end{figure}
For pions the negative slope arises because of a shadowing effect. The newly produced particles are blocked by the spectators and therefore the slope of directed flow is negative at low beam energies. Again the fluid calculation overestimates the flow at low beam energies and for pions the hybrid UrQMD approach seems to describe the measurements rather well. More important than a quantitative comparison to experimental data is the conclusion that the hybrid calculations with different assumptions about the order of the phase transition do not show any sensitivity on these observables. Detailed measurements of the centrality dependence of directed flow are going to allow for better conclusions on the stopping mechanism, the nuclear mean fields and the equation of state that are all crucial to understand the dynamics of heavy ion collisions at low beam energies.
\section{Summary and Conclusions}
\label{summary}
Collective flow is one of the main observables to understand the properties of the quark gluon plasma and the dynamics of heavy ion reactions. Especially the beam energy dependence of anisotropic flow provides high potential for insights on the equation of state and the dependence of transport coefficients on the baryo-chemical potential. To connect the final state observables to the quantities of interest, hybrid approaches based on relativistic hydrodynamics and non-equilibrium transport provide a realistic description of the dynamics. The main conclusions are that the small change of elliptic flow as a function of the beam energy can be explained by the transport dynamics that gains importance at lower beam energies and replaces the diminished hydrodynamic evolution. Triangular flow is much more sensitive to the viscosity and cannot be generated by hadronic transport alone. The beam energy dependence of the slope of the directed flow around midrapidity has been studied and even though the old predictions are reproduced, there is no difference between a modern hybrid calculation with and without a strong first order phase transition. To reach final conclusions on this subject, more detailed studies including nuclear potentials and an improved treatment of the baryon stopping are necessary.
\section{Acknowledgements}
H. P. acknowledges funding of a Helmholtz Young Investigator Group VH-NG-822 from
the Helmholtz Association and GSI. This work was supported by the Helmholtz International
Center for the Facility for Antiproton and Ion Research (HIC for FAIR) within the
framework of the Landes-Offensive zur Entwicklung Wissenschaftlich-Oekonomischer Exzellenz (LOEWE) program launched by the State of Hesse. Computational resources have been provided by the Center
for Scientific Computing (CSC) at the Goethe-University of Frankfurt.
|
\section{Motivation}
\label{sec:1}
Situations in which one is concerned with uncertainty quantification often come in the following guise: we are investigating physical models where inputs are not given precisely, but instead are random quantities or random fields, or depend on a set of parameters.
A classical example is the Darcy flow model with a random diffusion coefficient,
\begin{equation}
\label{eq:elliptic}
-\nabla \kappa(x,\omega) \nabla u(x,\omega) = f(x,\omega), \quad x \in D\subset \mathbb{R}^d,
\end{equation}
where $d$ is the spatial dimension, $\kappa(x,\omega)$ is a random field dependent on a random
parameter $\omega$ in a probability space $(\Omega,\mathcal{A},\mathcal{P})$.
The solution $u(x,\omega)$ belongs to $H^1(D)$ w.r.t. $x$ and the same probability space w.r.t. $\omega$.
There is an established theory about the existence and uniqueness of the solution to \eqref{eq:elliptic} under various assumptions on $\kappa$ and $f$; see, for example,
\cite{babuska2004galerkin,Sarkis09,GITTELSON10,matthiesKeese05cmame,mugler2011elliptic}.
In \cite{Sarkis09, GITTELSON10} it is shown that under additional
assumptions on the right-hand side $f$ and special choices of the test space the problem \eqref{eq:elliptic} is well-posed.
The case where the Lax-Milgram theorem is not applicable (e. g. upper and lower constants $\underline{\kappa}$, $\overline{\kappa}$ in
$0<\underline{\kappa}<\kappa<\overline{\kappa}<\infty$ do not exist)
is also considered in \cite{mugler2011elliptic}.
In \cite{Sarkis09} the authors analyze assumptions on $\kappa$ which were made in \cite{babuska2004galerkin} to guarantee the uniqueness and the existence of the solution and to
offer a new method with much weakened assumptions. Additionally, in \cite{Sarkis09}, the authors point out that after truncating terms in the expansion for $\kappa$,
as is done in \cite{matthiesKeese05cmame}, it is not guaranteed that the truncated $\kappa$ will be strictly bounded from zero. As a result the existence of
the approximate solution to \eqref{eq:elliptic} is questionable. The remarkable difference of the approach in \cite{Sarkis09} from approaches in
\cite{babuska2004galerkin, matthiesKeese05cmame} is that the permeability coefficient $\kappa$ is not truncated and, as a result, the ellipticity condition is maintained.
To solve \eqref{eq:elliptic} we need to discretize the random field $\kappa$.
This requires some knowledge on the probability space $\Omega$ and the probability measure.
A widely used approach relies on two assumptions: $\kappa$ is defined as an invertible smooth function of another random field with known distribution (e.g. normal), and the covariance function of either of those fields is given.
After that $\kappa(x,\omega)$ can be \emph{expanded} to a series of functions, depending on $x$ and a set of new parameters $\bm{\theta}=(\theta_1,\theta_2,\ldots)$.
Typically used are polynomials of $\bm{\theta}$, hence called the \emph{Polynomial Chaos Expansion} (PCE) \cite{wiener38, xiuKarniadakis02a} family of discretizations.
Another approach is the collocation on some grid in $\bm{\theta}$ \cite{Babushka_Colloc,back2011stochastic}.
Each of $\bm{\theta}$ is a random quantity depending on $\omega$, but the domain of $\bm{\theta}$ is known deterministically.
Introducing a discretization for $\bm{\theta}$, we turn the stochastic problem \eqref{eq:elliptic} into a high-dimensional deterministic one.
However, its straightforward numerical solution suffers from the \emph{curse of dimensionality}: even if $\bm{\theta}$ contains a finite amount of variables (say, $M$), the number of discrete degrees of freedom grows exponentially with $M$.
To surmount this issue, a data-sparse approximation is needed.
During the last years, low-rank tensor product techniques were successfully applied to the solution of high-dimensional stochastic and parametric PDEs.
A recent literature survey of low-rank tensor approximation techniques can be found in \cite{Grasedyck_Kressner}.
The tensor product approach involves a \emph{format}, in which the data are represented, and a set of \emph{algorithms} to manipulate the data in the format.
The algorithms can be roughly separated into three classes:
methods performing basic approximate algebra (additions, products, simple iterative methods) of data within the format;
methods constructing the formatted representation from a few entries of the initial data;
and methods aiming to solve equations, e.g. linear systems, ODEs or eigenvalue problems, keeping all data in the format.
To some extent, these methods have already been applied to parametric problems.
Non-intrusive (black box) tensor methods for multi-parametric problems, i.e. ``class 2'', were developed in \cite{Ballan_Grasedyck14, Ballan_Grasedyck10, doostan-non-intrusive-2013}.
In particular, in \cite{Ballan_Grasedyck14} the authors follow the stochastic collocation approach and compute functionals of the solution of multi-parametric PDEs.
Since the stochastic collocation allows to solve uncoupled deterministic problems for different collocation points, the functional of the solution (e.g. the average value) can be approximated straightforwardly via the black box hierarchical tensor interpolation algorithm.
To compute the whole stochastic solution is a more difficult problem, especially in the stochastic Galerkin framework, where deterministic problems are coupled.
In \cite{KressTobler11, KressTobler11HighDim, matthieszander-lowrank-2012, Sousedik2014truncated, ZanderDiss} the authors develop iterative methods and preconditioners to solve numerically discretized multi-parametric problems.
Several manipulations of the PCE with a low-rank approximation have been considered.
In \cite{litvinenko-spde-2013} the authors assume that the solution has a low-rank \emph{canonical} (CP) tensor format and develop methods for the CP-formatted computation of level sets.
In \cite{UQLitvinenko12,wahnert-stochgalerkin-2014} the authors analyzed tensor ranks of the stochastic operator.
The proper generalized decomposition was applied for solving high dimensional stochastic problems in \cite{Nouy07,nouy-pgd-stoch-2010}.
In \cite{khos-pde-2010, khos-qtt-2010, KhSch-Galerkin-SPDE-2011} the authors employed newer tensor formats, the Tensor Train (TT) and Quantized TT (QTT), for the
approximation of coefficients and the solution of stochastic elliptic PDEs.
The theoretical study of the complexity of the stochastic equation was provided, for example, by means of the analytic regularity and (generalized) PC approximation \cite{xiuKarniadakis02a} for control problems constrained by linear parametric elliptic and parabolic PDEs \cite{kunoth-2013}.
Other classical techniques to cope with high-dimensional problems are sparse grids
\cite{Griebel,Griebel_Bungartz, nobile2008sparse}
and (quasi) Monte Carlo methods \cite{graham-QMC-2011,scheichl-mlmc-further-2013,Schwab-MLQMC-2015}.
Nevertheless, tensor product methods are more flexible than sparse grids, as they allow to avoid severe reductions of the model from the very beginning, and
adapt a suitable structure on the discrete level.
Compared to Monte Carlo methods, tensor techniques work implicitly with the whole solution, and even a construction of a tensor format for entry-wise given data in a black box manner uses less randomness than the MC approach.
In this article we approximate the PCE of the input coefficient $\kappa(x,\omega)$ in the TT tensor format. After that we compute the solution $u(x,\omega)$ and perform all post-processing in the same TT format.
The first stage, computation of the PCE of $\kappa$, involves a lengthy formula, defining each entry of the discretized coefficient.
To perform this computation efficiently, we develop a new \emph{block cross} approximation algorithm, which constructs the TT format for $\kappa$ from a few evaluations of the entry-wise formula.
This formula delivers several tensors that are to be summed and approximated in a TT format.
We show that the new algorithm is more efficient than several runs of a previously existing cross method \cite{so-dmrgi-2011proc} for each tensor separately.
As soon as the coefficient is given in the TT format, it becomes very easy to construct the stiffness matrix, derived from the stochastic Galerkin discretization of \eqref{eq:elliptic}.
We apply the alternating iterative tensor algorithm to solve a large linear system arising from \eqref{eq:elliptic}, and finally use the cross algorithm again to compute the exceedance probability from the solution.
In the next section, we outline the general Galerkin, Polynomial Chaos Expansion (PCE) and Karhunen-Lo\`{e}ve Expansion (KLE) discretization schemes for a random field.
An introduction to the TT methods and the new block cross interpolation algorithm are presented in Section \ref{sec:TTlow}.
Some details of how to apply the block cross algorithm to the PCE calculations are given in Section \ref{sec:tt-kle-pce}.
We start with the TT approximation of the multi-dimensional input coefficient $\kappa$.
After that, in Section \ref{sec:stiff_tt}, we construct the stochastic Galerkin matrix in the TT format.
Section \ref{subsec:levelSets} is devoted to the efficient post-processing (computation of the mean value, covariance, and probability of a particular event) in the TT format.
Numerical results in Section \ref{sec:numerics} demonstrate the practical performance of the TT approach in the outlined tasks.
\section{Discretisation and computation}
\label{SS:discret}
For brevity, we follow \cite{Matthies_encycl}, where more references may be found.
See also the recent monograph \cite{LeMatre10} to study more about KLE, PCE and multiindices.
In \cite{ ErnstU10, Sousedik2014truncated, Ullmann10, ullmann2012efficient} the authors discuss the stochastic Galerkin matrix, its sparsity and preconditioners.
To discretize \eqref{eq:elliptic}, we follow the Galerkin approach.
The Finite Element method (for example, with piecewise linear functions) is a natural way to discretize the spatial part.
We choose a finite dimensional subspace
\begin{equation}
\mathcal{V}_N = \mathrm{span}\{\varphi_1(x),\ldots,\varphi_N(x)\} \subset \mathcal{V},
\label{eq:V_N}
\end{equation}
where $\mathcal{V} = H^1(D) \cap L_2^0(\bar D)$ is the Hilbert space of functions on $x$.
For simplicity, we impose the homogeneous Dirichlet boundary conditions.
Discretization in the probability space $(\Omega, \mathcal{A},\mathcal{P})$ is less trivial.
We use the Karhunen-Lo\`{e}ve Expansion (KLE) to determine a finite set of independent parameters, defining $\kappa(x,\omega)$.
However, the distribution of these parameters might be unknown, and special efforts are needed to resolve this.
\subsection{Discretization of the input random field}
We assume that $\kappa(x,\omega)$ may be seen as a smooth transformation $\kappa = \phi(\gamma)$ of the Gaussian random field $\gamma(x,\omega)$.
In this Section, we explain how to compute KLE of $\gamma$ if the covariance of $\kappa$ is given.
For more details see \cite[Sections 3.4.1, 3.4.2]{ZanderDiss} or \cite{KeeseDiss}.
A typical example is the log-normal field with $\phi(\gamma) = \exp(\gamma)$.
Expanding $\phi$ in a series of the Hermite polynomials gives
\begin{equation}
\phi(\gamma) = \sum\limits_{i=0}^{\infty} \phi_i h_i(\gamma) \approx \sum\limits_{i=0}^{Q} \phi_i h_i(\gamma), \quad \phi_i = \int\limits_{-\infty}^{+\infty} \phi(z) \frac{1}{i!} h_i(z) \exp(-z^2/2) dz,
\label{eq:phi_tranform}
\end{equation}
where $h_i(\cdot)$ is the $i$-th Hermite polynomial, and $Q$ is the number of terms after the truncation.
The Gaussian field $\gamma(x,\omega)$ can be written as the Karhunen-Lo\`{e}ve expansion.
First, given the covariance function of $\kappa(x,\omega)$, we may relate it with the covariance function of $\gamma(x,\omega)$ as follows (see details in \cite[Sections 3.4.1, 3.4.2]{ZanderDiss}),
\begin{equation}
\mbox{cov}_{\kappa}(x,y) = \int_{\Omega} \left(\kappa(x,\omega)-\bar{\kappa}(x)\right) \left(\kappa(y,\omega)-\bar{\kappa}(y)\right) dP(\omega) \approx \sum_{i=0}^{Q} i!
\phi_i^2 \mbox{cov}_{\gamma}^i (x,y),
\label{eq:cov_transform}
\end{equation}
where $\bar{\kappa}(x)$ is the expectation of $\kappa(x,\omega)$.
Solving this implicit $Q$-order equation, we derive $\mbox{cov}_{\gamma} (x,y)$ \cite{ZanderDiss}.
Now, the KLE can be computed as follows:
\begin{equation}
\gamma(x,\omega) = \sum_{m=1}^{\infty} g_m(x) \theta_m(\omega), \quad \mbox{where} \quad \int\limits_{D} \mbox{cov}_{\gamma}(x,y) g_m(y)dy = \lambda_m g_m(x).
\end{equation}
Here we assume that the eigenfunctions $g_m$ absorb the square roots of the KL eigenvalues. The stochastic variables $\theta_m$ are normalized (they are uncorrelated and jointly Gaussian).
The initial coefficient $\kappa$ depends nonlinearly on $\theta_m$.
In the discrete setting, we truncate PCE and write it for $M$ random variables,
\begin{equation}
\kappa(x,\omega) \approx \sum\limits_{\bm{\alpha} \in \mathcal{J}_M} \kappa(x,\bm{\alpha}) H_{\bm{\alpha}}(\bm{\theta}(\omega)), \quad \text{where}\quad H_{\bm{\alpha}}(\bm{\theta}):=h_{\alpha_1}(\theta_1) \cdots h_{\alpha_M}(\theta_M)
\label{eq:pce_k}
\end{equation}
is the multivariate Hermite polynomial, $\bm{\alpha} = (\alpha_1,\ldots,\alpha_M)$ is a multiindex (a tuple of multinomial orders), $h_{\alpha_m}(\theta_m)$ is the univariate Hermite polynomial, $\bm{\theta} = (\theta_1,\ldots,\theta_M)$ is a tuple of random variables,
and $\mathcal{J}_M$ is a set of all multinomial orders (see definition below).
The Galerkin coefficients $\kappa(x,\bm{\alpha})$ are evaluated as follows:
\begin{equation}
\kappa(x,\bm{\alpha}) = \dfrac{(\alpha_1+\cdots+\alpha_M)!}{\alpha_1! \cdots \alpha_M! } \phi_{\alpha_1+\cdots+\alpha_M} \prod_{m=1}^{M} g_m^{\alpha_m}(x),
\label{eq:pce_k_coeffs}
\end{equation}
where $\phi_{\alpha_1+\cdots+\alpha_M}$ is the Galerkin coefficient of the transform function in \eqref{eq:phi_tranform}
and $g_m^{\alpha_m}(x)$ means just the $\alpha_m$-th power of the KLE function value $g_m(x)$.
In practice, we restrict the polynomial orders in \eqref{eq:pce_k} to finite limits, which can be done in different ways.
\begin{definition}
\label{def:tensor_product}
The \emph{full} multi-index set is defined by restricting each component independently,
$$
\mathcal{J}_{M,\bm{p}} = \left\{\bm{\alpha}=(\alpha_1,\ldots,\alpha_M):~~\alpha_m = 0,\ldots,p_m,~~m=1,\ldots,M \right\},
$$
where $\bm{p}=(p_1,\ldots,p_M)$ is a shortcut for the tuple of order limits.
\end{definition}
The full set provides high flexibility for the resolution of stochastic variables
\cite{wahnert-stochgalerkin-2014, litvinenko-spde-2013, Sousedik2014truncated}.
However, its cardinality is equal to $\prod_{m=1}^M (p_m+1) \le (p+1)^M$, if $p_m\le p$.
This may become a enormous number even if $p$ and $M$ are moderate ($p \sim 3$, $M \sim 20$ is typical).
In this paper, we do not store all $(p+1)^M$ values explicitly, but instead approximate them via a \emph{tensor product} representation.
Another approach is to preselect the set of polynomials with the moderate cardinality \cite{TodorSCA, matthiesKeese05cmame, Sudret_sparsePCE, chkifa-adapt-stochfem-2015}.
\begin{definition}
The \emph{sparse} multi-index set is defined by restricting the sum of components,
$$
\mathcal{J}_{M,p}^{sp} = \left\{\bm{\alpha}=(\alpha_1,\ldots,\alpha_M):\quad \bm{\alpha} \ge 0,~~\alpha_1+\cdots+\alpha_M \le p \right\}.
$$
\end{definition}
The sparse set contains $\mathcal{O}\left(\frac{M!}{p!(M-p)!}\right) = \mathcal{O}(M^p)$ values if $M \gg p$,
which is definitely less than $(p+1)^M$.
However, the negative side is that for a fixed $p$ some variables are resolved worse than others, and the approximation accuracy may suffer.
It may also be harmful to increase $p$ since in the sparse set it contributes exponentially to the complexity.
The low-rank tensor approximation allows to reduce the storage of the coefficient with the full set $\mathcal{J}_{M,\bm{p}}$ from $\mathcal{O}(p^M)$ to $\mathcal{O}(M p r^2)$, where $r$ is a data-dependent \emph{tensor rank}.
Theoretical estimates of $r$ are under development; numerical studies reflect that often $r$ does not depend on $p$ and depends linearly (or even milder) on $M$ \cite{khos-pde-2010}.
When $M$ is moderate, and $p$ is relatively large, the low-rank approach with the full index set (Def. \ref{def:tensor_product}) becomes preferable.
Besides, as soon as the coefficient $\kappa$ is computed in a tensor product form, to assemble the Galerkin matrix of the stochastic PDE is a much easier task than with the sparse set.
Some complexity reduction in Formula \eqref{eq:pce_k_coeffs} can be achieved with the help of the KLE for the initial field $\kappa(x,\omega)$.
Consider the expansion
\begin{equation}
\kappa(x,\omega) = \bar\kappa(x) + \sum_{\ell=1}^{\infty} \sqrt{\mu_\ell} v_\ell(x) \eta_\ell(\omega)
\approx \bar\kappa(x) + \sum_{\ell=1}^{L} \sqrt{\mu_\ell} v_\ell(x) \eta_\ell(\omega),
\label{eq:k_kle}
\end{equation}
where $v_\ell(x)$ are eigenfunctions of the integral operator with the covariance as the kernel (see e.g. \cite{ZanderDiss, KeeseDiss, matthiesKeese05cmame}).
It's hard to work with \eqref{eq:k_kle} straightforwardly, since the distribution of $\eta_\ell$ is generally unknown.
But we know that the set $V(x) = \{v_\ell(x)\}_{\ell=1}^{L}$, where $L$ is the number of KLE terms after the truncation, serves as an optimally reduced basis.
Therefore, instead of using \eqref{eq:pce_k_coeffs} directly, we project it onto $V(x)$:
\begin{equation}
\bm{\tilde \kappa}_{\ell}(\bm{\alpha}) = \dfrac{(\alpha_1+\cdots+\alpha_M)!}{\alpha_1! \cdots \alpha_M! } \phi_{\alpha_1+\cdots+\alpha_M} \int\limits_{D} \prod_{m=1}^{M} g_m^{\alpha_m}(x) v_\ell(x) dx.
\label{eq:k_kle_pce_coeff}
\end{equation}
Note that the range $\ell=1,\ldots,L$ may be much smaller than the number of discretization points $N$ in $D$.
After that, we restore the approximate coefficients \eqref{eq:pce_k_coeffs}:
\begin{equation}
\kappa(x,\bm{\alpha}) \approx \bar\kappa(x) + \sum\limits_{\ell=1}^{L} v_\ell(x) \bm{\tilde \kappa}_{\ell}(\bm{\alpha}).
\label{eq:k_kle_pce}
\end{equation}
\subsection{Discretization of the stochastic operator equation}
The same PCE ansatz of the coefficient \eqref{eq:pce_k} may be adopted to discretize the solution $u$ and ultimately the whole initial problem \eqref{eq:elliptic},
see \cite{wahnert-stochgalerkin-2014, litvinenko-spde-2013}.
For brevity, we illustrate the procedure on the deterministic right-hand side $f=f(x)$.
Given the KLE components \eqref{eq:k_kle} and the spatial discretization basis \eqref{eq:V_N},
we first assemble the spatial Galerkin matrices,
\begin{equation}
\bm{K}_0^{(x)}(i,j) = \int\limits_{D} \bar\kappa(x) \nabla \varphi_i(x) \cdot \nabla \varphi_j(x) dx, \quad \bm{K}_\ell^{(x)}(i,j) = \int\limits_{D} v_\ell(x) \nabla \varphi_i(x) \cdot \nabla \varphi_j(x) dx,
\label{eq:K_l}
\end{equation}
for $i,j=1,\ldots,N$, $\ell=1,\ldots,L$.
Now we take into account the PCE part $\bm{\tilde \kappa}_{\bm{\alpha}}$.
Assuming that $u$ is decomposed in the same way as \eqref{eq:pce_k} with the same $\mathcal{J}_{M,\bm{p}}$ or $\mathcal{J}_{M,p}^{sp}$, we compute the integral in \eqref{eq:K_p} over stochastic coordinates $\bm{\theta}$ and
compute the stochastic parts $\bm{K}_{\ell}^{(\omega)}\in \mathbb{R}^{\#\mathcal{J}_M \times \# \mathcal{J}_M}$ of the Galerkin matrix as follows (see also \cite{KeeseDiss, ZanderDiss, Matthies_encycl}),
\begin{equation}
\bm{K}_{\ell}^{(\omega)}(\bm{\alpha},\bm{\beta}) = \int\limits_{\mathbb{R}^M} H_{\bm{\alpha}}(\bm{\theta}) H_{\bm{\beta}}(\bm{\theta}) \sum\limits_{\bm{\nu} \in \mathcal{J}_{M}}
\bm{\tilde\kappa}_{\ell}(\bm{\nu}) H_{\nu}(\theta) \rho(\bm{\theta})d\bm{\theta} = \sum\limits_{\nu \in \mathcal{J}_{M}} \bm{\Delta}_{\bm{\alpha},\bm{\beta},\bm{\nu}} \bm{\tilde\kappa}_{\ell}(\bm{\nu}),
\label{eq:K_p}
\end{equation}
where $\rho(\bm{\theta}) = \rho(\theta_1) \cdots \rho(\theta_M)$ is the probability density with $\rho(\theta_m) = \frac{1}{\sqrt{2\pi}} \exp(-\theta_m^2/2)$, and
\begin{equation}
\bm{\Delta}_{\bm{\alpha},\bm{\beta},\bm{\nu}} = \Delta_{\alpha_1,\beta_1,\nu_1} \cdots \Delta_{\alpha_M,\beta_M,\nu_M}, \quad \Delta_{\alpha_m,\beta_m,\nu_m} =
\int\limits_{\mathbb{R}} h_{\alpha_m}(\theta) h_{\beta_m}(\theta) h_{\nu_m}(\theta) \rho(\theta) d\theta,
\label{eq:Delta}
\end{equation}
is the triple product of the Hermite polynomials, and $\bm{\tilde\kappa}_{\ell}(\bm{\nu})$ is according to \eqref{eq:k_kle_pce_coeff}.
Let us denote $\bm{\Delta}_0(\bm{\alpha},\bm{\beta}) = \Delta_{\alpha_1,\beta_1,0} \cdots \Delta_{\alpha_M,\beta_M,0}$, i.e. $\bm{\Delta}_0 \in \mathbb{R}^{\#J_M \times \# J_M}$.
Putting together \eqref{eq:k_kle_pce}, \eqref{eq:K_l} and \eqref{eq:K_p}, we obtain the whole discrete stochastic Galerkin matrix,
\begin{equation}
\mathbf{K} = \bm{K}_0^{(x)} \otimes \bm{\Delta}_0 + \sum\limits_{\ell=1}^L \bm{K}_\ell^{(x)} \otimes \bm{K}_{\ell}^{(\omega)},
\label{eq:K_gen}
\end{equation}
which is a square matrix of size $N \cdot \#\mathcal{J}_{M}$, and $\otimes$ is the Kronecker product.
For the sparse index set, we need to compute $\mathcal{O}\left((\#\mathcal{J}_{M,p}^{sp})^3\right)$ entries of $\bm{\Delta}$ explicitly.
For the full index set and $\bm{\tilde \kappa}_{\bm{\nu}}$ given in the tensor format, the direct product in $\bm{\Delta}$ \eqref{eq:Delta} allows to exploit the same format for \eqref{eq:K_gen} and to simplify the procedure, see Sec. \ref{sec:stiff_tt}.
The deterministic right-hand side is extended to the size of $\mathbf{K}$ easily,
\begin{equation}
\mathbf{f} = \bm{f}_0 \otimes \bm{e}_{0}, \qquad \bm{f}_0(i) = \int\limits_{D} \varphi_i(x) f(x)dx, \quad i=1,\ldots,N,
\label{eq:f}
\end{equation}
and $\bm{e}_0$ is the first identity vector of size $\#\mathcal{J}_{M}$, $\bm{e}_0 = (1,0,\ldots,0)^\top$, which assigns the deterministic $f(x)$ to the zeroth-order Hermite polynomial in the parametric space.
\subsection{Solution of the stochastic equation}
Now the discrete equation writes as a linear system
\begin{equation}
\mathbf{K} \mathbf{u} = \mathbf{f},
\label{eq:linsystem}
\end{equation}
where $\mathbf{u}\in\mathbb{R}^{N \cdot \#\mathcal{J}_{M}}$ is the vector of the Galerkin coefficients for the solution, with elements enumerated by $u(i,\bm{\alpha})$.
In \cite{Sousedik2014truncated} the authors propose two new strategies for constructing preconditioners for the stochastic Galerkin system to be used with Krylov subspace
iterative solvers. The authors also research the block sparsity structure of the corresponding coefficient matrix as well as the stochastic Galerkin matrix.
In \cite{Ullmann10, ullmann2012efficient} the authors develop and analyze a Kronecker-product preconditioner.
In many cases it is sufficient to use the mean-field preconditioner $\bm{K}_0^{(x)} \otimes \bm{\Delta}_0$, which is easy to invert due to the Kronecker form.
We follow this approach.
To solve the system in a tensor product format, we employ alternating optimization methods \cite{ds-amen-2014}.
\section{Tensor product formats and low-rank data compression}
\label{sec:TTlow}
We see that the cardinality of the full polynomial set $\mathcal{J}_{M,\bm{p}}$ may rise to prohibitively large values $(p+1)^M$.
In this paper, we study two ways to alleviate this problem.
First, we can fix the basis set a priori.
This is the case with the sparse set $\mathcal{J}_{M,p}^{sp}$.
Due to particular properties of the stochastic elliptic equation, it is possible to derive a posteriori error indicators and build the sparse set adaptively \cite{eigel-adapt-stoch-fem-2014,chkifa-adapt-stochfem-2015}.
Another approach, which is applicable to a wider class of problems, is to use the full discretization set, but to approximate already discrete data, via a low-rank tensor product format.
For stochastic PDEs, low-rank approximations were used in e.g. \cite{beylkin-high-2005,doostan-non-intrusive-2013,litvinenko-spde-2013,khos-pde-2010, matthieszander-lowrank-2012, nouy-pgd-stoch-2010}.
This approach requires specific computational algorithms, since the data are represented in a nonlinear fashion.
In this section we suggest such an algorithm to construct a data-sparse format of the stochastic coefficient and quantities of interest.
\subsection{Tensor Train decomposition}
To show the techniques in the briefest way, we choose the so-called \emph{matrix product states} (MPS) formalism~\cite{fannes-mps-1992}, which introduces the following representation of a multi-variate array, or \emph{tensor}:
\begin{equation}\label{eq:tt}
\mathbf{u}(\alpha_1,\ldots,\alpha_M) = \sum_{s_1=1}^{r_1} \sum_{s_2=1}^{r_2} \cdots \sum_{s_{M-1}=1}^{r_{M-1}}
\bm{u}^{(1)}_{s_0,s_1}(\alpha_1) \bm{u}^{(2)}_{s_1,s_2}(\alpha_2) \cdots \bm{u}^{(M)}_{s_{M-1},s_M}(\alpha_M).
\end{equation}
Surely, in the same form we may write $\bm{\kappa}(\bm{\alpha})$.
In numerical linear algebra this format is known as~``\emph{tensor train} (TT)'' representation \cite{osel-tt-2011, ot-tt-2009}.
Each TT core (or \emph{block}) is a three-dimensional array, $\bm{u}^{(m)} \in \mathbb{R}^{r_{m-1} \times (p_m+1) \times r_m}$, $m=1,\ldots,M$, where
$p_m$ denotes the \emph{mode size}, the polynomial order in the variable $\alpha_m$, and $r_m=r_m(\mathbf{u})$ is the \emph{TT rank}.
The total number of entries scales as $\mathcal{O}(Mpr^2)$, which is tractable as long as $r=\max\{r_m\}$ is moderate.
We still have not specified the \emph{border} rank indices $s_0$ and $s_M$.
In the classical TT definition \cite{osel-tt-2011} and in \eqref{eq:tt}, there is only one tensor $\mathbf{u}(\bm{\alpha})$ in the left-hand side, therefore $s_0=s_M=1$, and also $r_0=r_M=1$.
However, the left-hand side might contain several tensors, such as $\bm{\tilde\kappa}_{\ell}(\bm{\alpha})$ \eqref{eq:k_kle_pce_coeff}.
Then we can associate $s_0=\ell$ or $s_M=\ell$, and approximate all tensors for different $\ell$ in the same \emph{shared} TT representation.
High-dimensional matrices (cf. $\mathbf{K}$ in \eqref{eq:K_gen}) can be also presented in the TT format,
$$
\mathbf{K} = \sum_{s_0,\ldots,s_{M-1}} \bm{K}^{(0)}_{s_0} \otimes \bm{K}^{(1)}_{s_0,s_1} \otimes \cdots \otimes \bm{K}^{(M)}_{s_{M-1}}.
$$
The matrix by vector product $\mathbf{K} \mathbf{u}$ is then recast to TT blocks of $\mathbf{K}$ and $\mathbf{u}$.
Similarly, using the multilinearity of the TT format, we can cast linear combinations of initial tensors to concatenations of TT blocks.
The principal favor of the TT format comparing to the Canonical Polyadic (CP) decomposition (which is also popular, see e.g. \cite{wahnert-stochgalerkin-2014, litvinenko-spde-2013}) is a stable quasi-optimal rank reduction procedure \cite{osel-tt-2011}, based on singular value decompositions.
The complexity scales as $\mathcal{O}(Mpr^3)$, i.e. it is free from the curse of dimensionality,
while the full accuracy control takes place.
This procedure can be used to reduce unnecessarily large ranks after the matrix by vector product or the linear combination in the TT format, and to compress a tensor if it is fully given as a multidimensional array and fits into the memory.
However, in our situation this is not the case, and we need another approach to construct a TT format.
\subsection{Cross interpolation of matrices}
If $M=2$, the left-hand side of \eqref{eq:tt} can be seen as a matrix. For simplicity we consider this case first.
The basic assumption is that any entry of a matrix (or tensor) can be evaluated.
However, to construct a TT approximation, we do not want to compute all elements, but only few of them.
The principal ingredient for this is based on the efficiency of an incomplete Gaussian elimination in an approximation of a low-rank matrix, also known as the Adaptive Cross Approximation (ACA) \cite{bebe-2000,bebe-aca-2011}.
Given is a matrix $\bm{U}=[\bm{U}(i,j)]\in\mathbb{R}^{p \times q}$, we select some ``good'' columns and rows to approximate the whole matrix,
\begin{equation}
\bm{U}(i,j) \approx \bm{\tilde U}(i,j)= \bm{U}(i,\mathbb{J}) \cdot \widehat \bm{U}^{-1} \cdot \bm{U}(\mathbb{I},j),
\label{eq:2dcross}
\end{equation}
where $\widehat \bm{U} = \bm{U}(\mathbb{I}, \mathbb{J})$, and $\mathbb{I} \subset \{1,\ldots,p\}$, $\mathbb{J} \subset \{1,\ldots,q\}$ are sets of indices of cardinality $r$, so e.g. $\bm{U}(i,\mathbb{J}) \in \mathbb{R}^{1 \times r}$.
It is known that there exists a quasi-optimal set of interpolating indices $\mathbb{I},\mathbb{J}$.
\begin{lemma}[Maximum volume (\emph{maxvol}) principle \cite{gt-maxvol-2001} ]
$\;$ If $\mathbb{I}$ and $\mathbb{J}$ are such that $\mathrm{det}~\bm{U}(\mathbb{I}, \mathbb{J})$ is maximal among all $r \times r$ submatrices of $\bm{U}$, then
$$
\|\bm{U}-\bm{\tilde U}\|_C \le (r+1) \min_{\mathrm{rank}(\bm{V}) = r} \|\bm{U}-\bm{V}\|_2,
$$
where $\| \cdot \|_C$ is the Chebyshev norm, $\|\bm{X}\|_C = \max_{i,j}|\bm{X}_{i,j}|$.
\label{lem:maxvol}
\end{lemma}
In practice, however, the computation of the true maximum volume submatrix is infeasible, since it is an NP-hard problem.
Instead one performs a heuristic iteration in an alternating fashion \cite{gostz-maxvol-2010}:
we start with some (e.g. random) low-rank factor $\bm{U}^{(1)} \in \mathbb{R}^{p \times r}$, determine indices $\mathbb{I}$ yielding a quasi-maxvol $r \times r$ submatrix in $\bm{U}^{(1)}$, and compute $\bm{U}^{(2)}$ as $r$ columns of $\bm{U}$ of the indices $\mathbb{I}$.
Vice versa, in the next step we find quasi-maxvol column indices in $\bm{U}^{(2)}$ and calculate corresponding $pr$ elements, collecting them into the newer $\bm{U}^{(1)}$, which hopefully approximates the true low-rank factor better than the initial guess.
This process continues until the convergence, which appears to be quite satisfactory in practice.
\subsection{TT-cross interpolation of tensors}
In higher dimensions we recurrently proceed in the same way, since the TT format constitutes a recurrent matrix low-rank factorization.
Let us merge the first $m$ and the last $M-m$ indices from $\bm{\alpha}$.
The corresponding TT blocks will induce the following matrices:
\begin{definition}\label{def:interface}
Given a TT format \eqref{eq:tt} and an index $m$.
Define the \emph{left interface matrix} $\bm{U}^{<m} \in \mathbb{R}^{(p_1+1)\cdots(p_{m-1}+1) \times r_{m-1}}$ and the \emph{right interface} matrix $\bm{U}^{>m} \in \mathbb{R}^{r_m \times (p_{m+1}+1)\cdots(p_M+1)}$ as follows,
\begin{equation*}
\begin{split}
\bm{U}^{<m}_{s_{m-1}}(\alpha_1,\ldots,\alpha_{m-1}) & = \sum\limits_{s_1=1}^{r_1} \cdots \sum\limits_{s_{m-2}=1}^{r_{m-2}} \bm{u}^{(1)}_{s_1}(\alpha_1) \cdots \bm{u}^{(m-1)}_{s_{m-2},s_{m-1}}(\alpha_{m-1}), \\
\bm{U}^{>m}_{s_{m}}(\alpha_{m+1},\ldots,\alpha_M) & = \sum\limits_{s_{m+1}=1}^{r_{m+1}} \cdots \sum\limits_{s_{M-1}=1}^{r_{M-1}} \bm{u}^{(m+1)}_{s_m,s_{m+1}}(\alpha_{m+1}) \cdots \bm{u}^{(M)}_{s_{M-1}}(\alpha_M).
\end{split}
\end{equation*}
\end{definition}
Such matrices are convenient to relate high-dimensional TT expressions with their two-dimensional analogs.
For example, the TT format \eqref{eq:tt} can be written in the following form,
\begin{equation}
\mathbf{u}(\bm{\alpha}) \approx \bm{U}^{<m}(\alpha_1,\ldots,\alpha_{m-1}) \cdot \bm{u}^{(m)}(\alpha_m) \cdot \bm{U}^{>m}(\alpha_{m+1},\ldots,\alpha_M).
\label{eq:tt-l-b-r}
\end{equation}
The \emph{TT-cross} algorithm \cite{ot-ttcross-2010} assumes that the above expansion is valid on some $r_{m-1}(p_m+1)r_m$ indices, and can thus be seen as a system of equations on elements of $u^{(m)}$.
Let us be given $r_{m-1}$ \emph{left} indices $(\hat\alpha_1,\ldots,\hat\alpha_{m-1}) \in \mathbb{I}^{(m-1)}$ and $r_m$ \emph{right} indices $(\hat\alpha_{m+1},\ldots,\hat\alpha_{M}) \in \mathbb{J}^{(m+1)}$.
For the single index $\alpha_m$ we allow its full range $\{\alpha_m\} = \{0,1,\ldots,p_m\}$.
Requiring that \eqref{eq:tt-l-b-r} is valid on the combined indices
$$
\left(\hat\alpha_1,\ldots,\hat\alpha_{m-1},\alpha_m,\hat\alpha_{m+1},\ldots,\hat\alpha_{M}\right) \equiv \left(\mathbb{I}^{(m-1)}, \alpha_m, \mathbb{J}^{(m+1)}\right),
$$
we obtain the following computational rule for $\bm{u}^{(m)}$,
\begin{equation}
\bm{u}^{(m)}(\alpha_m) = \widehat \bm{U}_{<m}^{-1} \cdot \mathbf{u}\left(\mathbb{I}^{(m-1)}, \alpha_m, \mathbb{J}^{(m+1)}\right) \cdot \widehat \bm{U}_{>m}^{-1},
\label{eq:ttcross}
\end{equation}
where $\widehat \bm{U}_{<m} = \bm{U}^{<m}(\mathbb{I}^{(m-1)})$ and $\widehat \bm{U}_{>m} = \bm{U}^{>m}(\mathbb{J}^{(m+1)})$ are the submatrices of $\bm{U}^{<m}$ and $\bm{U}^{>m}$ at indices $\mathbb{I}^{(m-1)}$ and $\mathbb{J}^{(m+1)}$, respectively.
Of course, these submatrices must be nonsingular.
In the previous subsection we saw a good strategy: if indices $\mathbb{I}^{(m-1)}$ and $\mathbb{J}^{(m+1)}$ are chosen in accordance with the maxvol principle for $\bm{U}^{<m}$ and $\bm{U}^{>m}$, the submatrices are not only nonsingular, but also
provide a good approximation for \eqref{eq:tt-l-b-r}.
However, in practice $\bm{U}^{<m}$ and $\bm{U}^{>m}$ are too large to be treated directly.
Instead, we use the \emph{alternating} iteration over the dimensions and build nested index sets.
The alternating iteration means that we loop over $m=1,2,\ldots,M$ (the so-called \emph{forward} iteration) and $m=M,M-1,\ldots,1$ (\emph{backward} iteration).
Consider first the forward transition, $m-1 \rightarrow m$.
Given the set $\mathbb{I}^{(m-1)}$, the \emph{nested} indices $\mathbb{I}^{(m)}$ are defined as follows.
We concatenate $\mathbb{I}^{(m-1)}$ and $\alpha_m$, i.e. consider $r_{m-1} (p_m+1)$ indices $(\hat \alpha_{1},\ldots,\hat\alpha_{m-1},\alpha_m) \in \{\mathbb{I}^{(m-1)}, \alpha_m\}$, and select $r_m$ indices $\mathbb{I}^{(m)}$ only from $\{\mathbb{I}^{(m-1)}, \alpha_m\}$, not from all $\mathcal{O}(p^m)$ possibilities.
It can be seen that the next interface $\bm{U}^{<m+1}$, restricted to $\{\mathbb{I}^{(m-1)}, \alpha_m\}$, can be computed as a product of the previous submatrix $\widehat \bm{U}_{<m}$ and the current TT block $\bm{u}^{(m)}$.
To formalize this, we need the following definition.
\begin{definition}\label{def:folded}
Given a three-dimensional tensor $\bm{u}^{(m)} \in \mathbb{R}^{r_{m-1} \times (p_m+1) \times r_m}$, introduce the following reshapes, both pointing to the same data stored in $\bm{u}^{(m)}$:
\begin{itemize}
\item \emph{left folding}: $\bm{u}^{|m\rangle}(s_{m-1},\alpha_m;~s_m) = \bm{u}^{(m)}_{s_{m-1},s_m}(\alpha_m), \quad \bm{u}^{|m\rangle} \in \mathbb{R}^{r_{m-1}(p_m+1) \times r_m},$ and
\item \emph{right folding}: $\bm{u}^{\langle m |}(s_{m-1};~\alpha_m,s_m) = \bm{u}^{(m)}_{s_{m-1},s_m}(\alpha_m), \quad \bm{u}^{\langle m |} \in \mathbb{R}^{r_{m-1}\times (p_m+1) r_m}.$
\end{itemize}
\end{definition}
Then the restriction of $\bm{U}^{<m+1}$ writes as follows,
$$
\bm{V}^{\langle m|} = \widehat \bm{U}_{<m} \bm{u}^{\langle m |}, \quad \mbox{and} \quad \bm{U}^{<m+1}\left(\mathbb{I}^{(m-1)}, \alpha_m\right) = \bm{V}^{| m \rangle} \in \mathbb{R}^{r_{m-1}(p_m+1) \times r_m}.
$$
Thus, it is enough to apply the maximum volume algorithm \cite{gostz-maxvol-2010} to the matrix $\bm{V}^{| m \rangle}$, deriving \emph{local} maxvol indices $\hat i_m \subset \{1,\ldots, r_{m-1}(p_m+1)\}$, and obtain both $\mathbb{I}^{(m)}$ and $\widehat \bm{U}_{<m+1}$ by the restriction
$$
\mathbb{I}^{(m)} = \{\mathbb{I}^{(m-1)}, \alpha_m\}(\hat i_m), \quad \widehat \bm{U}_{<m+1} = \bm{V}^{| m \rangle}(\hat i_k) \in \mathbb{R}^{r_m \times r_m}.
$$
The backward transition $m+1 \rightarrow m$ for $\mathbb{J}^{(m)}$ and $\widehat \bm{U}_{>m}$ can be written analogously.
We show it directly in Alg.\ref{alg:cross} below.
In total, we need only $\mathcal{O}(n_{it} M p r^2)$ entries of $\mathbf{u}$ to be evaluated, where $n_{it}$ is the number of alternating iterations, typically of the order of $10$.
\subsection{Rank-adaptive DMRG-cross algorithm}
A drawback of the TT-cross method is that the TT ranks are fixed; they have to be guessed a priori, which is also a problem of exponential complexity in $M$.
A remedy is to consider larger portions of data in each step.
The Density Matrix Renormalization Group (DMRG) algorithm was developed in the quantum physics community (\cite{white-dmrg-1993}, see also the review \cite{schollwock-2011} and
the references therein) to solve high-dimensional eigenvalue problems coming from the stationary spin Schroedinger equation.
It is written in a similar alternating fashion as the TT-cross procedure described above.
The crucial difference is that instead of one TT block as in \eqref{eq:ttcross} we calculate two neighboring TT blocks at once.
In the \emph{DMRG-cross} \cite{so-dmrgi-2011proc} interpolation algorithm, this is performed as follows.
Given $\mathbb{I}^{(m-1)}$, $\mathbb{J}^{(m+2)}$, we compute
$$
\bm{u}^{(m,m+1)}(\alpha_m,\alpha_{m+1}) = \widehat \bm{U}_{<m}^{-1} \cdot \mathbf{u}\left(\mathbb{I}^{(m-1)}, \alpha_m, \alpha_{m+1}, \mathbb{J}^{(m+2)}\right) \cdot \widehat \bm{U}_{>m+1}^{-1}.
$$
Then we need to separate indices $\alpha_m$ and $\alpha_{m+1}$ to recover the initial TT structure.
This can be done via the singular value decomposition.
The four-dimensional array $\bm{u}^{(m,m+1)}$ is reshaped to a matrix $\bm{U}^{(m,m+1)} \in \mathbb{R}^{r_{m-1}(p_m+1) \times (p_{m+1}+1) r_{m+1}}$, and the truncated SVD is computed,
$$
\bm{U}^{(m,m+1)} \approx \bm{V}\bm{\Sigma} \bm{W}^\top, \quad \mbox{s.t.} \quad \left\|\bm{U}^{(m,m+1)} - \bm{V}\bm{\Sigma} \bm{W}^\top\right\|_F\le \varepsilon \left\|\bm{U}^{(m,m+1)}\right\|_F,
$$
where $\bm{V} \in \mathbb{R}^{r_{m-1} (p_m+1) \times \hat r_m}$, $\bm{\Sigma} \in \mathbb{R}^{\hat r_m \times \hat r_m}$, $\bm{W} \in \mathbb{R}^{(p_{m+1}+1) r_{m+1} \times \hat r_m}$, and $\|\cdot\|_F$ is the Frobenius norm.
The new TT rank $\hat r_m$ is likely to differ from the old one $r_m$.
After that, we rewrite $\bm{u}^{(m)}$ and $\bm{u}^{(m+1)}$ with $\bm{V}$ and $\bm{W}$, respectively, replace $r_m=\hat r_m$, and continue the iteration.
To ensure that the perturbations introduced to $\bm{u}^{(m,m+1)}$ and the whole tensor $\mathbf{u}$ coincide, we need to ensure that the interfaces $\bm{U}^{<m}$ and $\bm{U}^{>m+1}$ have orthonormal columns and rows, respectively.
Fortunately, this requires only small-sized QR decompositions of matrices $\bm{u}^{|m\rangle}$ and $\bm{u}^{\langle m |}$ \cite{osel-tt-2011,schollwock-2011} in the first iteration.
Later, the SVD will provide orthogonal factors in $\bm{u}^{(m)}$ and $\bm{u}^{(m+1)}$ automatically.
This algorithm allows a fast adaptation of TT ranks towards the requested accuracy threshold.
The price is, however, a larger degree of $p$ in the complexity, since we perform a full search in both $\alpha_m$ and $\alpha_{m+1}$ in each step.
The greedy-cross method \cite{sav-qott-2014} avoids this problem by maximizing the error over only $\mathcal{O}(rp)$ random entries among all $\mathcal{O}(r^2 p^2)$ elements in $\bm{u}^{(m,m+1)}$.
Here, instead of the neighboring block $\bm{u}^{(m+1)}$, it is the error that provides additional information and improves the approximation.
However, for the KLE-PCE coefficient \eqref{eq:k_kle_pce_coeff}, we have other candidates for such auxiliary data.
And we have reasons to consider them prior to the error.
\subsection{Block TT-Cross interpolation algorithm}
Note that each call of \eqref{eq:k_kle_pce_coeff} throws $L$ values, corresponding to different $\ell=1,\ldots,L$.
We may account for this in various ways.
Since $\ell$ has the meaning of the reduced spatial variable, we may feature it as an additional dimension.
But when we restrict the indices $\left\{\mathbb{I}^{(m-1)}, \alpha_m\right\}(\hat i_m) = \mathbb{I}^{(m)}$, we will remove some values of $\ell$ from consideration.
Therefore, a vast majority of information cannot be used: we evaluate $L$ values, but only a few of them will be
used to improve the approximation.
Another way is to run $L$ independent cross (e.g. DMRG-cross) algorithms to approximate each $\bm{\tilde \kappa}_{\ell}(\bm{\alpha})$ in its own TT format.
Yet, this is also not very desirable.
First, the TT ranks for the whole $\bm{\kappa}$ are usually comparable to the ranks of individual TT formats.
Therefore, $L$ cross algorithms consume almost $L$ times more time than the single run.
Secondly, we will have to add $L$ TT formats to each other by summing their ranks, which is to be followed by the TT approximation procedure.
The asymptotic complexity thus scales as $\mathcal{O}(L^3)$, which was found to be too expensive in practice.
A better approach is to store all $\bm{\tilde \kappa}_{\ell}(\bm{\alpha})$ in the same TT representation, employing the idea of the \emph{block} TT format \cite{dkos-eigb-2014}.
The resulting method has a threefold advantage: all data in each call of \eqref{eq:k_kle_pce_coeff} are assimilated, the algorithm adjusts the TT ranks automatically according to the given accuracy, and the output is returned as a single optimally-compressed TT format, convenient for further processing.
\begin{algorithm}[t]
\caption{Block cross approximation of tensors in the TT format} \label{alg:cross}
\begin{algorithmic}[1]
\REQUIRE A function to evaluate $\mathbf{u}_\ell(\alpha_1,\ldots,\alpha_M)$, initial TT guess $\bm{u}^{(1)}(\alpha_1)\cdots \bm{u}^{(M)}(\alpha_M)$, relative accuracy threshold $\varepsilon$.
\ENSURE Improved TT approximation $\bm{u}^{(1)}_{\ell}(\alpha_1) \bm{u}^{(2)}(\alpha_2) \cdots \bm{u}^{(M)}(\alpha_M)$.
\STATE Initialize $\mathbb{I}^{(0)}=[]$, \quad $\widehat \bm{U}_{<1}=1$, \quad $\mathbb{J}^{(M+1)} = []$, \quad $\widehat \bm{U}_{>M} = 1$.
\FOR{$\mathrm{iteration}=1,2,\ldots,n_{it}$ or until convergence}
\FOR[Forward sweep]{$m=1,2,\ldots,M-1$}
\IF[All indices are available, assimilate the information]{$\mathrm{iteration}>1$}
\STATE Evaluate the tensors at cross indices and compute the common block by \eqref{eq:ttb_cross}.
\STATE Compute the truncated SVD \eqref{eq:btt_svd} with accuracy $\varepsilon$.
\ELSE[Warmup sweep: the indices are yet to be built]
\STATE Find QR decomposition $\bm{u}^{|m\rangle} = \bm{q}^{|m\rangle} \bm{R}$, $\left(\bm{q}^{|m\rangle}\right)^* \bm{q}^{|m\rangle } =\bm{I}$.
\STATE Replace $\bm{u}^{\langle m+1 |} = \bm{R} \bm{u}^{\langle m+1 |}$, \quad $\bm{u}^{|m\rangle}=\bm{q}^{|m\rangle}$.
\ENDIF
\STATE Compute the pre-restricted interface $\bm{V}^{\langle m |} = \widehat \bm{U}_{<m} \bm{u}^{\langle m |}$.
\STATE Find \emph{local} maxvol indices $\hat i_m = \mathtt{maxvol}\left(\bm{V}^{|m\rangle}\right)$.
\STATE New indices $\mathbb{I}^{(m)} = \left\{\mathbb{I}^{(m-1)}, \alpha_m\right\}(\hat i_m)$, interface $\widehat \bm{U}_{<m+1} = \bm{V}^{|m\rangle}(\hat i_m) \in \mathbb{R}^{r_{m} \times r_{m}}$.
\ENDFOR
\FOR[Backward sweep]{$m=M,M-1,\ldots,2$}
\STATE Evaluate the tensors at cross indices and compute the common block by \eqref{eq:ttb_cross}.
\STATE Compute the truncated SVD \eqref{eq:btt_svd_back} with accuracy $\varepsilon$.
\STATE Compute the pre-restricted interface $\bm{V}^{|m\rangle} = \bm{u}^{| m \rangle} \widehat \bm{U}_{>m}$
\STATE Find \emph{local} maxvol indices $\hat j_m = \mathtt{maxvol}\left(\left(\bm{V}^{\langle m |}\right)^\top\right)$.
\STATE Restrict $\mathbb{J}^{(m)} = \left\{\alpha_m, \mathbb{J}^{(m+1)}\right\}(\hat j_m)$, \quad $\widehat \bm{U}_{>m-1} = \bm{V}^{\langle m |}(\hat j_m) \in \mathbb{R}^{r_{m-1} \times r_{m-1}}$.
\ENDFOR
\STATE Evaluate the first TT block $\bm{u}^{(1)}_{\ell}(\alpha_1) = \mathbf{u}_\ell \left(\alpha_1, \mathbb{J}^{(2)} \right) \widehat \bm{U}_{>1}^{-1}$.
\ENDFOR
\end{algorithmic}
\end{algorithm}
Assume that we have a procedure that, given an index $\alpha_1,\ldots,\alpha_M$, throws $L$ values $\mathbf{u}_\ell(\bm{\alpha})$, $\ell=1,\ldots,L$.
When the tensor entries $\mathbf{u}_{\ell}(\mathbb{I}^{(m-1)},\alpha_m,\mathbb{J}^{(m+1)})$ are evaluated, we modify \eqref{eq:ttcross} as follows:
\begin{equation}
\bm{y}^{(m)}(\alpha_m,\ell) = \widehat \bm{U}_{<m}^{-1} \cdot \mathbf{u}_{\ell}\left(\mathbb{I}^{(m-1)},\alpha_m,\mathbb{J}^{(m+1)}\right) \cdot \widehat \bm{U}_{>m}^{-1}.
\label{eq:ttb_cross}
\end{equation}
Now $\bm{y}^{(m)}$ is a four-dimensional tensor, in the same way as in the DMRG-cross.
We need to find a basis in $\bm{\alpha}$ that is best suitable for all $\mathbf{u}_{\ell}$.
Hence, we reshape $\bm{y}^{(m)}$ to the matrix $\bm{Y}^{(m)} \in \mathbb{R}^{r_{m-1}(p_m+1) \times L r_m}$ and compute the truncated singular value decomposition
\begin{equation}
\bm{Y}^{(m)} \approx \bm{u}^{|m\rangle} \bm{\Sigma} \bm{W}^\top, \qquad \bm{u}^{|m\rangle} \in \mathbb{R}^{r_{m-1}(p_m+1) \times \hat r_m}, \quad \bm{\Sigma} \in \mathbb{R}^{\hat r_m \times \hat r_m}, \quad \bm{W} \in \mathbb{R}^{L r_m \times \hat r_m}.
\label{eq:btt_svd}
\end{equation}
Again, the new rank $\hat r_m$ satisfies the Frobenius-norm error criterion, and replaces $r_m$ for the next iteration.
In the backward iteration, we reshape $\bm{y}^{(m)}$ to $\bm{Y}^{(m)} \in \mathbb{R}^{L r_{m-1} \times (p_m+1) r_m}$ and take the right singular vectors to the new TT block,
\begin{equation}
\bm{Y}^{(m)} \approx \bm{W} \bm{\Sigma} \bm{u}^{\langle m |}, \qquad \bm{W} \in \mathbb{R}^{L r_{m-1} \times \hat r_m}, \quad \bm{\Sigma} \in \mathbb{R}^{\hat r_m \times \hat r_m}, \quad \bm{u}^{\langle m |} \in \mathbb{R}^{\hat r_m \times (p_m+1) r_m}.
\label{eq:btt_svd_back}
\end{equation}
The whole procedure is summarized in the \emph{Block TT-Cross} Algorithm \ref{alg:cross}.
\section{TT-structured calculations with PCE}
\subsection{Computation of the PCE in the TT format via the cross interpolation}\label{sec:tt-kle-pce}
Equipped with Alg. \ref{alg:cross}, we may apply it to the PCE approximation, passing Formula \eqref{eq:k_kle_pce_coeff} as a function $u_\ell(\bm{\alpha})$ that evaluates tensor values on demand.
The initial guess may be even a rank-1 TT tensor with all blocks populated by random numbers, since the cross iterations will adapt both the representation and TT ranks.
Considering the sizes of the involved matrices, the complexity estimate can be written straightforwardly.
\begin{statement}
The cost to compute $\bm{\tilde\kappa}_{\ell}(\bm{\alpha})$ via the block TT-Cross Algorithm \ref{alg:cross} is
$$
\mathcal{O}(r^2 p (MN+NL) + r^3 pL + r^3 p L \cdot \min\{p,L\}).
$$
\end{statement}
The first two terms come from Formula \eqref{eq:k_kle_pce_coeff}, and the last one is the complexity of SVDs \eqref{eq:btt_svd} and \eqref{eq:btt_svd_back}.
\begin{remark}
$\;$It is unclear in general which term will dominate.
For large $N$, we are typically expecting that it is the evaluation \eqref{eq:ttb_cross}.
However, if $N$ is moderate (below $1000$), but the rank is large ($\sim 100$), the SVD consumes most of the time.
For the whole algorithm, assuming also $L \sim M$, we can thus expect the $\mathcal{O}(n_{it} M^2 N p r^3)$ complexity,
which is lower than $\mathcal{O}(M p L^3)$, which we could receive if we run independent cross algorithms for each $\ell$.
\end{remark}
As soon as the reduced PCE coefficients $\bm{\tilde\kappa}_{\ell}(\bm{\alpha})$ are computed, the initial expansion \eqref{eq:k_kle_pce} comes easily.
Indeed, after the backward cross iteration, the $\ell$ index belongs to the first TT block, and we may let it play the role of the ``zeroth'' TT rank index,
\begin{equation}
\bm{\tilde\kappa}_{\ell}(\bm{\alpha}) = \sum_{s_1,\ldots,s_{M-1}} \bm{\kappa}^{(1)}_{\ell,s_1}(\alpha_1) \bm{\kappa}^{(2)}_{s_1,s_2}(\alpha_2) \cdots \bm{\kappa}^{(M)}_{s_{M-1}}(\alpha_M).
\end{equation}
For $\ell=0$ we extend this formula such that $\bm{\tilde\kappa}_0(\bm{\alpha})$ is the first identity vector $\bm{e}_0$, cf. \eqref{eq:f}.
Now we collect the spatial components from \eqref{eq:k_kle} into the ``zeroth'' TT block,
\begin{equation}
\kappa^{(0)}(x) = \begin{bmatrix} \kappa^{(0)}_\ell(x) \end{bmatrix}_{\ell=0}^{L} = \begin{bmatrix}\bar\kappa(x) & v_1(x) & \cdots & v_L(x) \end{bmatrix},
\end{equation}
then the PCE \eqref{eq:pce_k} writes as the following TT format,
\begin{equation}
\kappa(x,\bm{\alpha}) = \sum_{\ell,s_1,\ldots,s_{M-1}} \kappa^{(0)}_{\ell}(x) \bm{\kappa}^{(1)}_{\ell,s_1}(\alpha_1) \cdots \bm{\kappa}^{(M)}_{s_{M-1}}(\alpha_M).
\label{eq:k_pce_tt}
\end{equation}
\subsection{Stiffness Galerkin operator in the TT format}\label{sec:stiff_tt}
With the full set $\mathcal{J}_{M,\bm{p}}$, we may benefit from the rank-1 separability of $\bm{\Delta}$, since each index $\alpha_m,\beta_m,\nu_m$ varies independently on the others.
\begin{lemma}
Given the PCE TT format \eqref{eq:k_pce_tt} for the coefficient $\kappa$ with the TT ranks $r(\kappa)$, the Galerkin operator \eqref{eq:K_gen}
can be constructed in the TT format with the same ranks.
\end{lemma}
\begin{proof}
Given the PCE \eqref{eq:k_pce_tt} in the TT format, we split the whole sum over $\bm{\nu}$ in \eqref{eq:K_p} to the individual variables,
\begin{equation}
\begin{split}
\sum\limits_{\nu \in \mathcal{J}_{M,\bm{p}}} \bm{\Delta}_{\bm{\alpha},\bm{\beta},\bm{\nu}} \bm{\tilde \kappa}_{\ell}(\bm{\nu}) & = \sum_{s_1,\ldots,s_{M-1}} \bm{K}^{(1)}_{\ell,s_1}(\alpha_1,\beta_1) \bm{K}^{(2)}_{s_1,s_2}(\alpha_2,\beta_2) \cdots \bm{K}^{(M)}_{s_{M-1}}(\alpha_M,\beta_M), \\
\bm{K}^{(m)}(\alpha_m,\beta_m) & = \sum_{\nu_m=0}^{p_m} \Delta_{\alpha_m,\beta_m,\nu_m} \bm{\kappa}^{(m)}(\nu_m), \quad m=1,\ldots,M.
\end{split}
\label{eq:deltakappa_rank1}
\end{equation}
A similar reduction of a large summation to one-dimensional operations arises also in quantum chemistry \cite{vekh-lattice-2014}.
Agglomerate $\bm{K}^{(x)}_{\ell}(i,j)$ from \eqref{eq:K_l} to the ``zeroth'' TT block for the operator,
then the TT representation for the operator writes with the same TT ranks as in $\bm{\tilde \kappa}$,
\begin{equation}
\mathbf{K} = \sum_{\ell,s_1,\ldots,s_{M-1}} \bm{K}^{(x)}_{\ell} \otimes \bm{K}^{(1)}_{\ell,s_1} \otimes \cdots \otimes \bm{K}^{(M)}_{s_{M-1}} \in \mathbb{R}^{(N \cdot \#\mathcal{J}_{M,\bm{p}}) \times (N \cdot \#\mathcal{J}_{M,\bm{p}})}.
\label{eq:K_tt}
\end{equation}
\end{proof}
One interesting property of the Hermite triples is that $\Delta_{\alpha,\beta,\nu}=0$ if e.g. $\nu>\alpha+\beta$.
That is, if we set the same $p$ for $\alpha$, $\beta$ and $\nu$, in the assembly of \eqref{eq:K_gen} we may miss some components, corresponding to $\alpha>p/2$, $\beta>p/2$.
To compute the Galerkin operator exactly, it is reasonable to vary $\nu_m$ in the range $\{0,\ldots,2p\}$, and hence assemble $\tilde \kappa$ in the set $\mathcal{J}_{M,2\bm{p}}$.
While in the sparse set it would inflate the storage of $\bm{\Delta}$ and $\mathbf{K}$ significantly,
in the TT format it is feasible: the TT ranks do not depend on $p$, and the storage grows only linearly with $p$.
\subsection{Computation of the solution function}
Having solved the system \eqref{eq:linsystem}, we obtain the PCE coefficients of the solution in the TT format,
\begin{equation}
u(x,\bm{\alpha}) = \sum_{s_0,\ldots,s_{M-1}} u^{(0)}_{s_0}(x) \bm{u}^{(1)}_{s_0,s_1}(\alpha_1) \cdots \bm{u}^{(M)}_{s_{M-1}}(\alpha_M).
\label{eq:tt_u}
\end{equation}
Some statistics are computable directly from $u(x,\bm{\alpha})$, but generally we need a function in the initial random variables, $u(x,\bm{\theta})$.
Since $\mathcal{J}_{M,\bm{p}}$ is a tensor product set, $u(x,\bm{\alpha})$ can be turned into $u(x,\bm{\theta})$ without changing the TT ranks, similarly to the construction of the
Galerkin matrix in the previous subsection:
\begin{equation}
u(x,\bm{\theta}) = \sum_{s_0,\ldots,s_{M-1}} u^{(0)}_{s_0}(x) \left(\sum_{\alpha_1=0}^{p} h_{\alpha_1}(\theta_1) \bm{u}^{(1)}_{s_0,s_1}(\alpha_1)\right) \cdots \left(\sum_{\alpha_M=0}^{p} h_{\alpha_M}(\theta_M) \bm{u}^{(M)}_{s_{M-1}}(\alpha_M)\right).
\label{eq:alpha_to_theta}
\end{equation}
\subsection{Computation of statistics}\label{subsec:levelSets}
In this section we discuss how to calculate some statistical outputs from the solution in the TT format, such as the mean, the (co)variance and the probability of an event.
The \emph{mean} value of $u$, in the same way as in $\kappa$, can be derived as the PCE coefficient at $\bm{\alpha}=(0,\ldots,0)$, $\bar{u}(x) = u(x,0,\ldots,0)$.
It requires no additional calculations.
The \emph{covariance} is more complicated and requires both multiplication (squaring) and summation over $\bm{\alpha}$.
By definition, the covariance reads
\begin{equation*}
\begin{split}
\mbox{cov}_u(x,y) & = \int\limits_{\mathbb{R}^M} \left(u(x,\bm{\theta}) - \bar u(x)\right) \left(u(y,\bm{\theta}) - \bar u(y)\right) \rho(\bm{\theta}) d\bm{\theta} \\
& = \sum\limits_{\substack{\bm{\alpha},\bm{\beta} \neq (0,\ldots,0), \\ \bm{\alpha}, \bm{\beta} \in \mathcal{J}_{M,\bm{p}}}} u(x,\bm{\alpha}) u(y,\bm{\beta}) \int\limits_{\mathbb{R}^M} H_{\bm{\alpha}}(\bm{\theta}) H_{\bm{\beta}}(\bm{\theta}) \rho(\bm{\theta}) d\bm{\theta}.
\end{split}
\end{equation*}
Knowing that $\int H_{\bm{\alpha}}(\bm{\theta}) H_{\bm{\beta}}(\bm{\theta}) \rho(\bm{\theta}) d\bm{\theta} = \bm{\alpha}! \delta_{\bm{\alpha},\bm{\beta}}$,
we take $\bm{u}_{s_0}(\bm{\alpha}) = \bm{u}^{(1)}_{s_0}(\alpha_1) \cdots \bm{u}^{(M)}(\alpha_M)$ from \eqref{eq:tt_u} and multiply it with the Hermite mass matrix (TT ranks do not change),
\begin{equation}
\bm{w}_{s_0}(\bm{\alpha}) := \bm{u}_{s_0}(\bm{\alpha}) \sqrt{\bm{\alpha}!} = \sum_{s_1,\ldots,s_{M-1}} \left(\bm{u}^{(1)}_{s_0,s_1}(\alpha_1) \sqrt{\alpha_1!} \right) \cdots \left(\bm{u}^{(M)}_{s_{M-1}}(\alpha_M) \sqrt{\alpha_M!} \right),
\label{eq:param_chunk_masscorr}
\end{equation}
and then take the scalar product $\bm{C} = \left[\bm{C}_{s_0,s_0'}\right]$, where $\bm{C}_{s_0,s_0'} = \dotprod{\bm{w}_{s_0}}{\bm{w}_{s_0'}}$ with $\bm{w}$ defined in \eqref{eq:param_chunk_masscorr}.
Given the TT rank bound $r$ for $u(x,\bm{\alpha})$, we deduce the $\mathcal{O}(Mpr^3)$ complexity of this step.
After that, the covariance is given by the product of $\bm{C}$ with the spatial TT blocks,
\begin{equation}
\mbox{cov}_u(x,y) = \sum\limits_{s_0,s_0'=0}^{r_0} u^{(0)}_{s_0}(x) \bm{C}_{s_0,s_0'} u^{(0)}_{s_0'}(y),
\label{eq:cov_tt}
\end{equation}
where $u^{(0)}_{s_0}$ is the ``zeroth'' (spatial) TT block of the decomposition \eqref{eq:tt_u}.
Given $N$ degrees of freedom for $x$, the complexity of this step is $\mathcal{O}(N^2 r_0^2)$.
Note that a low-rank tensor approximation of a large covariance matrix is very important, for example, in Kriging \cite{LitvNowak13}.
The \emph{variance} is nothing else than the diagonal of the covariance, $\mbox{var}_u(x) = \mbox{cov}_u(x,x)$.
Other important outputs are the characteristic, level set functions, and the probability of a particular event \cite{litvinenko-spde-2013}.
\begin{definition}\label{defn:LevelSetFrequency}
Let $\mathbb{S} \subset \mathbb{R}$ be a subset of real numbers.
\begin{itemize}
\item The \emph{characteristic} function of $u$ at $\mathbb{S}$ is defined pointwise for all $\bm{\theta} \in \mathbb{R}^{M}$ as follows,
\begin{equation}\label{equ:chi}
\chi_{\mathbb{S}}(x,\bm{\theta}):=\left\{
\begin{array}{ll}
1, & u(x,\bm{\theta}) \in \mathbb{S}, \\
0, & u(x,\bm{\theta}) \nin \mathbb{S}.
\end{array}
\right.
\end{equation}
\item The \emph{level set} function reads $\mathcal{L}_{\mathbb{S}}(x,\bm{\theta}):= u(x,\bm{\theta}) \chi_{\mathbb{S}}(x,\bm{\theta})$.
\item The \emph{probability} of $\mathbb{S}$ reads $\mathbb{P}_{\mathbb{S}}(x) = \int_{\mathbb{R}^{M}} \chi_{\mathbb{S}}(x,\bm{\theta}) \rho(\bm{\theta}) d\bm{\theta}$.
\end{itemize}
\end{definition}
The characteristic function can be computed using either the cross Algorithm \ref{alg:cross} (see also \cite{Ballan_Grasedyck10, Ballan_Grasedyck14}), which takes Formula \eqref{equ:chi} as the function that evaluates a high-dimensional array $\chi$ at the indices in $x,\bm{\theta}$, or the Newton method for the sign function \cite{litvinenko-spde-2013}.
In both cases we may face a rapid growth of TT ranks during the cross or Newton iterations: the characteristic function is likely to have a discontinuity that
is not aligned to the coordinate axes, and some of
the singular values in the TT decomposition will decay very slowly.
We face the same problem with increasing ranks during computing the level set functions and exceedance probabilities.
However, the probability is easier to compute, especially if it is relatively small.
Using the same cross algorithm, we can compute directly the product $\hat\chi_{\mathbb{S}}(x,\bm{\theta}) = \chi_{\mathbb{S}}(x,\bm{\theta}) \rho(\bm{\theta})$.
The probability (at a fixed $x$) is then computed as a scalar product with the all-ones vector in the TT format, $\mathbb{P}_{\mathbb{S}}(x) = \dotprod{\hat\chi}{\mathtt{1}}$.
But if $\mathbb{P}$ is small, it means that most of the entries in $\hat\chi$ are small, and do not inflate TT ranks, which might be the case for $\chi$.
Typically, the computation of small probabilities is used to predict the failure risk of a technical system.
The event set has the form $\mathbb{S} = \left\{z\in\mathbb{R}:~z>\tau\right\}$, and the probability is called the \emph{exceedance} probability.
This will be studied in numerical experiments.
\section{Numerical Experiments}
\label{sec:numerics}
We verify the approach on the elliptic stochastic equation \eqref{eq:elliptic} in a two-dimensional $L$-shape domain, $x=(x_1,x_2) \in D = [-1,1]^2 \backslash [0,1]^2$.
We pose zero Dirichlet boundary conditions and use the deterministic right-hand side $f = f(x) = 1$.
The stochasticity is introduced in the diffusion coefficient $\kappa(x,\omega)$; we investigate log-normal and beta distributions for $\kappa$.
To generate the spatial mesh, we use the standard PDE Toolbox in MATLAB.
We consider from $1$ to $3$ refinement levels of the spatial grid, denoted by $R$.
The first refinement $R=1$ yields $557$ degrees of freedom, $R=2$ gives $2145$ points, and $R=3$ corresponds to $8417$ points.
Since we have to store the stiffness matrices in a dense form, we cannot refine the grid further.
The KLE of both $\kappa$ and $\gamma$ is truncated to the same number of terms $L=M$.
All utilities related to the Hermite PCE were taken from the \emph{sglib} \cite{sglib}, including discretization and solution routines in the sparse polynomial set $\mathcal{J}_{M,p}^{sp}$.
However, to work with the TT format (for full $\mathcal{J}_{M,\bm{p}}$), we employ the \emph{TT-Toolbox} \cite{tt-toolbox}.
The same polynomial orders are chosen in all variables, $\bm{p}=(p,\ldots,p)$.
We use the modules of \emph{sglib} for low-dimensional stages and replace the parts corresponding to high-dimensional calculations by the TT algorithms.
The block TT-cross Alg. \ref{alg:cross} is implemented in the procedure \texttt{amen\_cross.m} from the TT-Toolbox, and the linear system \eqref{eq:linsystem} was solved via the Alternating Minimal Energy (AMEn) method \cite{ds-amen-2014}, the procedure \texttt{amen\_solve.m} from the companion package tAMEn \cite{tamen}.
Computations were conducted in MATLAB R2012a on a single core of the Intel Xeon X5650 CPU at 2.67GHz, provided by the Max- Planck-Institute, Magdeburg.
The accuracy of the coefficient and the solution was estimated using the Monte Carlo method with $N_{mc}$ simulations.
We approximate the average $L_2$-error as follows,
\begin{equation*}
E_{\kappa} = \frac{1}{N_{mc}} \sum\limits_{z=1}^{N_{mc}} \dfrac{\sqrt{\sum_{i=1}^{N} \left(\kappa(x_i,\bm{\theta}_z) - \kappa_*(x_i,\bm{\theta}_z)\right)^2}}{\sqrt{\sum_{i=1}^{N} \kappa_*^2(x_i,\bm{\theta}_z)}} \approx \int\limits_{\mathbb{R}^M} \dfrac{\|\kappa(x,\bm{\theta}) - \kappa_*(x,\bm{\theta})\|_{L_2(D)}}{\|\kappa_*(x,\bm{\theta})\|_{L_2(D)}} \rho(\bm{\theta}) d\bm{\theta},
\end{equation*}
where $\{\bm{\theta}_z\}_{z=1}^{N_{mc}}$ are normally distributed random samples, $\{x_i\}_{i=1}^{N}$ are the spatial grid points,
and $\kappa_*(x_i,\bm{\theta}_z) = \phi\left(\gamma(x_i,\bm{\theta}_z)\right)$ is the reference coefficient computed without using the PCE for $\phi$.
The same definition is used for the solution $u$, with $u_*(x,\bm{\theta}_z)$ being the solution of the deterministic PDE with the coefficient $\kappa_*(x,\bm{\theta}_z)$.
Besides that we compare the statistics obtained with our approaches and the Monte Carlo method.
For the mean and variance we use the same discrete approximation to the $L_2$-norm,
$$
E_{\bar u} = \dfrac{\|\bar u - \bar u_*\|_{L_2(D)}}{\|\bar u_*\|_{L_2(D)}}, \qquad E_{\mbox{var}_u} = \dfrac{\|\mbox{var}_u - \mbox{var}_{u*}\|_{L_2(D)}}{\|\mbox{var}_{u*}\|_{L_2(D)}}.
$$
To examine the computation of probabilities, we compute the exceedance probability.
This task can be simplified by taking into account the maximum principle: the solution is convex w.r.t. $x$.
We compute the maximizer of the mean solution, $x_{\max}:~\bar u(x_{\max}) \ge \bar u(x)~\forall x \in D$.
Fix $x$ to $x_{\max}$ and consider only the stochastic part $\mathbf{u}_{\max}(\bm{\theta}) = u(x_{\max},\bm{\theta})$, and $\hat u = \bar u(x_{\max})$.
Now, taking some $\tau>1$, we compute
\begin{equation}
\mathbb{P} = \mathbb{P}\left(\mathbf{u}_{\max}(\bm{\theta})>\tau \hat u \right) = \int\limits_{\mathbb{R}^M} \chi_{\mathbf{u}_{\max}(\bm{\theta})>\tau \hat u} (\bm{\theta}) \rho(\bm{\theta}) d\bm{\theta}.
\label{eq:rare_prob}
\end{equation}
By $\mathbb{P}_*$ we will also denote the probability computed from the Monte Carlo method, and estimate the error as
$E_{\mathbb{P}} = \left|\mathbb{P}-\mathbb{P}_*\right|/\mathbb{P}_*$.
\subsection{Log-normal distribution}
Let
$$
\kappa(x,\omega) = \exp\left(1+\sigma \frac{\gamma(x,\omega)}{2}\right)+10,
$$
where $\gamma \sim \mathcal{N}(0,1)$ is the standard normally distributed random field.
The covariance function is taken Gaussian, $\mbox{cov}_{\kappa}(x,y) = \exp\left(-\frac{\|x-y\|^2}{2 l_c^2} \right)$, where $l_c$ is the (isotropic) correlation length.
The default parameters are the following: number of KLE terms $M=20$, polynomial order $p=3$, correlation length $l_c=1$, dispersion $\sigma=0.5$, refinement level of the spatial grid $R = 1$, and the tensor approximation accuracy $\varepsilon=10^{-4}$.
Below we will vary each of these parameters, keeping the others fixed.
For the computation of the probability \eqref{eq:rare_prob} we use $\tau=1.2$.
\subsubsection{Verification of the block cross algorithm}
Formula \eqref{eq:k_kle_pce_coeff} can be evaluated for each KLE index $\ell$ independently, using existing cross approximation algorithms.
We compare with the so-called DMRG cross method \cite{so-dmrgi-2011proc}, which is conceptually the closest approach to our Algorithm \ref{alg:cross}.
In Table \ref{tab:crosses} we show the performance of the single run of Algorithm \ref{alg:cross} (which gives the coefficient for all $\ell$ simultaneously) and of $L$ DMRG crosses, followed by the summation of individual terms to the common representation.
We see that even if the TT ranks of the output are exactly the same, the times differ dramatically.
This is because the ranks of individual components and of the whole coefficient are comparatively the same, and the DMRG approach requires roughly $L$ times more operations.
The last column in Table \ref{tab:crosses} confirms that both approaches deliver the same data up to the approximation tolerance.
\begin{table}[t]
\centering
\caption{Performance of the block cross Alg. \ref{alg:cross} versus the DMRG cross \cite{so-dmrgi-2011proc}}
\label{tab:crosses}
\begin{tabular}{c|cc|cc|c}
$M$ & \multicolumn{2}{c|}{Block cross} & \multicolumn{2}{c|}{DMRG crosses} & $\underline{\|\kappa_{DMRG}-\kappa_{Block}\|}$ \\
& CPU time, sec. & $r_{\kappa}$ & CPU time, sec. & $r_{\kappa}$ & $\|\kappa_{Block}\|$ \\ \hline
10 & 4.908 & 20 & 31.29 & 20 & 2.77e-5 \\
15 & 10.36 & 27 & 114.9 & 27 & 2.24e-5 \\
20 & 19.11 & 32 & 286.2 & 33 & 1.83e-4 \\
30 & 49.89 & 39 & 1372.0 & 50 & 2.52e-4 \\
\end{tabular}
\end{table}
\subsubsection{Experiment with the polynomial order $p$}
First, we provide a detailed study of the computational time of each of the stages in the TT and Sparse methods: construction of the coefficient ($T_{\kappa}$), construction of the operator ($T_{op}$), and the solution of the system ($T_{u}$).
The results are shown in Table \ref{tab:log-normal-ttimes-p}, and times are measured in seconds.
The complexity of the cross algorithm, employed for the computation of $\kappa$ in the TT format, grows linearly with $p$, since the TT ranks are stable w.r.t. $p$ (see Table \ref{tab:log-normal-p}).
However, it is much slower than the direct evaluation of the coefficients in the sparse set.
This is mainly due to the singular value decompositions, involving matrices of sizes hundreds.
Nevertheless, for the computation of the Galerkin matrix the situation is the opposite.
In the TT format, the computations via formula \eqref{eq:deltakappa_rank1} are very efficient, since they involve $M$ products of $p^2 \times 2p$ matrices.
In the sparse representation, we have to perform all $(\#\mathcal{J}_{M,p}^{sp})^3$ operations, which is very time-consuming.
Since $\#\mathcal{J}_{M,p}^{sp}$ grows exponentially with $p$, we had to skip the cases with large $p$.
The solution stage is more simple, since the mean value preconditioner is quite efficient, both for the standard CG method with the sparse set and the AMEn method for the TT format.
Again, the complexity of the TT solver grows linearly with $p$.
The sparse solver works reasonably fast as well, but it cannot be run before the matrix elements are computed\footnote{It is sometimes advocated to
avoid a construction of the matrix and to compute its elements only when they are needed in the
matrix-vector product. It saves memory, but the computational time will be comparatively the same, since it is proportional to the number of operations.}; hence it is also skipped for $p=4,5$.
\begin{table}[t]
\centering
\caption{Detailed CPU times (sec.) versus $p$, log-normal distribution}
\label{tab:log-normal-ttimes-p}
\begin{tabular}{c|ccc|ccc}
& \multicolumn{3}{c|}{TT (full index set $\mathcal{J}_{M,\bm{p}}$)} & \multicolumn{3}{c}{Sparse (index set $\mathcal{J}_{M,p}^{sp}$)} \\
$p$ & $T_{\kappa}$ & $T_{op}$ & $T_{u}$ & $T_{\kappa}$ & $T_{op}$ & $T_{u}$ \\ \hline
1 & 9.6166 & 0.1875 & 1.7381 & 0.4525 & 0.2830 & 0.6485 \\
2 & 14.6635 & 0.1945 & 2.9584 & 0.4954 & 3.2475 & 1.4046 \\
3 & 19.1182 & 0.1944 & 3.4162 & 0.6887 & 1027.7 & 18.1263 \\
4 & 24.4345 & 0.1953 & 4.2228 & 2.1597 & --- & --- \\
5 & 30.9220 & 0.3155 & 5.3426 & 9.8382 & --- & --- \\
\end{tabular}
\end{table}
In Table \ref{tab:log-normal-p} we present the total CPU times required in both methods to find the solution $u$, the time for computing $\hat \chi$, maximal TT ranks of the coefficient ($r_{\kappa}$),
the solution ($r_u$) and the weighted characteristic function ($r_{\hat \chi}$), as well as statistical errors in the coefficient ($E_{\kappa}$) and the solution ($E_{u}$).
The probability $\mathbb{P}$ is presented only for the TT calculation.
Since $\mathbb{P} \sim 10^{-4}$, $10000$ simulations may be not enough to compute $\mathbb{P}$ with the Monte Carlo method.
Below we present a dedicated test of the Monte Carlo approach.
\begin{table}[t]
\centering
\caption{Performance versus $p$, log-normal distribution}
\label{tab:log-normal-p}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$p$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
1 & 11.54 & 1.38 & 0.23 & 32 & 42 & 1 & 3.75e-3 & 1.69e-1 & 9.58e-3 & 1.37e-1 & 0 \\
2 & 17.81 & 5.14 & 0.32 & 32 & 49 & 1 & 1.35e-4 & 1.10e-1 & 4.94e-4 & 4.81e-2 & 0 \\
3 & 22.72 & 1046 & 83.12 & 32 & 49 & 462 & 6.21e-5 & 2.00e-3 & 2.99e-4 & 5.29e-4 & 2.75e-4 \\
4 & 28.85 & --- & 69.74 & 32 & 50 & 416 & 6.24e-5 & --- & 9.85e-5 & --- & 1.21e-4 \\
5 & 36.58 & --- & 102.5 & 32 & 49 & 410 & 6.27e-5 & --- & 9.36e-5 & --- & 6.20e-4 \\
\end{tabular}
\end{table}
\subsubsection{Experiment with the KLE dimension $M$}
The length of the truncated KLE is another crucial parameter of the stochastic PDE.
In Table \ref{tab:log-normal-M} we compare the TT and Sparse procedures.
\begin{table}[t]
\centering
\caption{Performance versus $M$, log-normal distribution}
\label{tab:log-normal-M}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$M$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
10 & 6.401 & 6.143 & 1.297 & 20 & 39 & 70 & 2.00e-4 & 1.71e-1 & 3.26e-4 & 1.45e-1 & 2.86e-4 \\
15 & 12.15 & 92.38 & 22.99 & 27 & 42 & 381 & 7.56e-5 & 1.97e-3 & 3.09e-4 & 5.41e-4 & 2.99e-4 \\
20 & 21.82 & 1005 & 67.34 & 32 & 50 & 422 & 6.25e-5 & 1.99e-3 & 2.96e-4 & 5.33e-4 & 2.96e-4 \\
30 & 52.92 & 48961 & 136.5 & 39 & 50 & 452 & 6.13e-5 & 9.26e-2 & 3.06e-4 & 5.51e-2 & 2.78e-4
\end{tabular}
\end{table}
We see that the accuracy of the TT approach is stable w.r.t. $M$, and the complexity grows mildly.
Note, however, that the correlation length $l_c=1$ is quite large and yields a fast decay of the KLE, such that $M=20$ is actually enough for the accuracy $10^{-4}$.
The TT approach demonstrates stability w.r.t. the overapproximation at $M=30$.
This is not the case for the sparse approach: at high $M$ the accuracy is lost.
This is because $p=3$ is not enough to transform the covariance \eqref{eq:cov_transform} accurately.
Since eigenvalues of the covariance decay rapidly, higher eigenpairs become strongly perturbed (the eigenvalues can even become negative), and a large error propagates to the PCE.
In the full set, the maximal polynomial order is equal to $pM$, and the error of the covariance transform is negligible.
\subsubsection{Experiment with the correlation length $l_c$}
The Gaussian covariance function yields an exponential decay of the KLE coefficients \cite{SCHWAB2006,Bieri2009},
but the actual rate is highly dependent on the correlation length \cite{Hcovariance, litvinenko2009sparse}.
In this experiment, we study the range of lengths from $0.1$ to $1.5$.
In order to have a sufficient accuracy for all values of $l_c$, we fix the dimension $M=30$.
The results are presented in the same layout as before in Table \ref{tab:log-normal-l}.
\begin{table}[t]
\centering
\caption{Performance versus $l_c$, log-normal distribution}
\label{tab:log-normal-l}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$l_c$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
0.1 & 216 & 55826 & 0.91 & 70 & 50 & 1 & 1.98e-2 & 1.98e-2 & 1.84e-2 & 1.84e-2 & 0 \\
0.3 & 317 & 52361 & 41.8 & 87 & 74 & 297 & 3.08e-3 & 3.51e-3 & 2.64e-3 & 2.62e-3 & 8.59e-31 \\
0.5 & 195 & 51678 & 58.1 & 67 & 74 & 375 & 1.49e-4 & 2.00e-3 & 2.58e-4 & 3.10e-4 & 6.50e-33 \\
1.0 & 57.3 & 55178 & 97.3 & 39 & 50 & 417 & 6.12e-5 & 9.37e-2 & 3.18e-4 & 5.59e-2 & 2.95e-04 \\
1.5 & 32.4 & 49790 & 121 & 31 & 34 & 424 & 3.24e-5 & 2.05e-1 & 4.99e-4 & 1.73e-1 & 7.50e-04 \\
\end{tabular}
\end{table}
In the TT approach, we see a clear decrease of the computational complexity and the error with growing covariance length.
This is because the SVD approximation in the TT format automatically reduces the storage w.r.t. the latter (less important) variables, if the KLE decay is fast enough.
The TT errors reflect the amount of information discarded in the truncated KLE tail, which is large for small $l_c$ and small otherwise.
The errors in the sparse approach behave in the same way until $l_c=0.5$, but for $l_c=1$ and $1.5$ the dimension $M=30$ is too large, and the instability w.r.t. the overapproximation takes place.
With fixed $M$, the exceedance probability is very sensitive to the correlation length.
Truncating the KLE, we reduce the total variance of the random field. For a (quasi)-Gaussian distribution, a small perturbation of the variance has
a small effect on the integral over the peak region,
but may have a very large relative effect on the tail region, which corresponds to the small exceedance probability.
\subsubsection{Experiment with the dispersion $\sigma$}
The variance of the normally distributed field $\gamma(x,\omega)$ is equal to $\sigma^2$.
Since it enters $\kappa$ under the exponential, it influences the variance of $\kappa$ significantly.
In Table \ref{tab:log-normal-sigma} we vary $\sigma$ from $0.2$ to $1$ and track the performance of the methods.
As expected, the computational complexity grows with $\sigma$, as does the contrast in the coefficient.
However, we were able to perform all computations for each value of $\sigma$.
\begin{table}[t]
\centering
\caption{Performance versus $\sigma$, log-normal distribution}
\label{tab:log-normal-sigma}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$\sigma$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
0.2 & 15.93 & 1008 & 0.348 & 21& 31 & 1 & 5.69e-5 & 4.76e-5 & 4.19e-5 & 1.30e-5 & 0 \\
0.4 & 18.72 & 968.3 & 0.341 & 29& 42 & 1 & 6.88e-5 & 8.04e-4 & 1.40e-4 & 2.14e-4 & 0 \\
0.5 & 21.23 & 970.1 & 79.96 & 32& 49 & 456 & 6.19e-5 & 2.02e-3 & 3.05e-4 & 5.45e-4 & 2.95e-4 \\
0.6 & 24.08 & 961.5 & 24.72 & 34& 57 & 272 & 9.12e-5 & 4.42e-3 & 6.14e-4 & 1.16e-3 & 2.30e-3 \\
0.8 & 31.69 & 969.1 & 67.93 & 39& 66 & 411 & 4.40e-4 & 8.33e-2 & 2.02e-3 & 2.90e-2 & 8.02e-2 \\
1.0 & 50.67 & 1071 & 48.44 & 44& 82 & 363 & 1.73e-3 & 4.10e-1 & 4.96e-3 & 3.08e-1 & 9.17e-2
\end{tabular}
\end{table}
\subsubsection{Experiment with the spatial grid refinement $R$}
Since the efforts of dealing with the full spatial matrix grow significantly with the grid refinement,
in this test we limit ourselves to $M=10$.
The principal observations from Table \ref{tab:log-normal-R_l} are that the TT rank and the accuracy\footnote{Note that the errors are estimated via the Monte Carlo method on the same grids, thus they show the accuracy of the PCE approximation, not the spatial discretization.} are stable w.r.t. the grid refinement.
Therefore, we may expect that the TT approach will also be efficient for finer grids, if we find an efficient way to deal with the spatial dimension.
A research on \emph{non-intrusive} stochastic Galerkin methods, addressing this issue, has begun recently \cite{matthies-to-be-intrusive-1-2014}, and we plan to adopt it in the TT framework in future.
\begin{table}[t]
\centering
\caption{Performance versus $R$, log-normal distribution. The left column shows the number of spatial degrees of freedom ($\#$DoF) for $R=1,2,3$.}
\label{tab:log-normal-R_l}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$\#$DoF & \multicolumn{3}{c|}{CPU time, sec} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
557 & 6.40 & 6.09 & 1.29 & 20 & 39 & 71 & 2.00e-4 & 1.71e-1 & 3.26e-4 & 1.45e-1 & 2.86e-4 \\
2145 & 8.98 & 13.7 & 1.17 & 20 & 39 & 76 & 1.74e-4 & 1.89e-3 & 3.33e-4 & 5.69e-4 & 2.90e-4 \\
8417 & 357 & 171 & 0.84 & 20 & 40 & 69 & 1.65e-4 & 1.88e-3 & 3.24e-4 & 5.64e-4 & 2.93e-4 \\
\end{tabular}
\end{table}
\subsubsection{Comparison with the Monte Carlo Method}
For the Monte Carlo test, we prepare the TT solution with parameters $p=5$ and $M=30$.
The results are presented in Table \ref{tab:log-normal-mc}.
In the left part of the table we show the performance of the Monte Carlo method with different numbers of simulations: total CPU time ($T_{MC}$),
errors in the mean and variance of $u$, and a small exceedance probability with its error.
The right part contains the results of the TT approach: the aggregated CPU time of construction of the coefficient, operator and solution ($T_{solve}$), the time to compute the weighted characteristic function $(T_{\hat\chi})$,
TT ranks of all data and the probability calculated from $\hat\chi$.
\begin{table}[t]
\centering
\caption{Verification of the Monte Carlo method, log-normal distribution}
\label{tab:log-normal-mc}
\begin{tabular}{c|l|cc|cc||c|c}
$N_{mc}$ & $T_{MC}$, sec. & $E_{\bar u}$ & $E_{\mbox{var}_u}$ & $\mathbb{P}_*$ & $E_{\mathbb{P}}$ & \multicolumn{2}{c}{TT results} \\ \hline
$10^2$ & 0.6398 & 9.23e-3 & 1.49e-1 & 0 & $\infty$ & $T_{solve}$ & 96.89 sec. \\
$10^3$ & 6.1867 & 1.69e-3 & 5.97e-2 & 0 & $\infty$ & $T_{\hat\chi}$ & 157.0 sec. \\
$10^4$ & 6.1801$\cdot 10^1$ & 5.81e-4 & 7.12e-3 & 4.00e-4 & 5.53e-1 & $r_{\kappa}$ & 39 \\
$10^5$ & 6.2319$\cdot 10^2$ & 2.91e-4 & 2.91e-3 & 4.10e-4 & 5.15e-1 & $r_{u}$ & 50 \\
$10^6$ & 6.3071$\cdot 10^3$ & 1.23e-4 & 9.76e-4 & 4.60e-4 & 3.51e-1 & $r_{\hat\chi}$ & 432 \\
& & & & & & $\mathbb{P}$ & 6.214e-4
\end{tabular}
\end{table}
We see that the cost of the TT method is comparable with the cost of $40000$ Monte Carlo tests.
That many realizations already provide a good approximation of the mean, a bit less accurate for the variance, but it is by far not sufficient for a confident estimate of the exceedance probability.
Therefore, the tensor product methods can be recommended as a competitive alternative to classical techniques for computing exceedance probabilities.
\subsection{Beta distribution}
The Hermite expansion \eqref{eq:phi_tranform} of the $\exp$ function in the log-normal case yields the coefficients of the form $\phi_i = \frac{c}{i!}$.
Therefore, the PCE coefficient formula \eqref{eq:pce_k_coeffs} resolves to a direct product of univariate functions of $\alpha_1,\ldots,\alpha_M$,
and the tensor format of the PCE can be constructed explicitly \cite{wahnert-stochgalerkin-2014}.
To demonstrate the generality of the cross algorithm, we also consider a more exotic beta-distributed coefficient,
$$
\kappa(x,\omega) = \mathcal{B}^{-1}_{5,2}\left(\frac{1+\mathrm{erf}\left(\frac{\gamma(x,\omega)}{\sqrt{2}}\right)}{2}\right)+1, \qquad \mbox{where} \quad \mathcal{B}_{a,b}(z) = \frac{1}{B(a,b)} \int\limits_{0}^{z} t^{a-1} (1-t)^{b-1} dt.
$$
For the purpose of computing $\phi_i$, the function $\mathcal{B}_{5,2}(z)$ is inverted by the Newton method.
Again, the covariance function is $\mbox{cov}_{\kappa}(x,y) = \exp\left(-\frac{\|x-y\|^2}{2 l_c^2} \right)$.
Since this distribution varies stronger than the log-normal one, for the computation of the probability \eqref{eq:rare_prob} we use larger $\tau=1.4$.
All other parameters are the same as in the experiments with the log-normal coefficient.
The performance of both TT and Sparse approach in case of the beta distribution is shown in Tables \ref{tab:beta-p}, \ref{tab:beta-M}, \ref{tab:beta-l},
\ref{tab:beta-R_l}, \ref{tab:beta-mc} for $p$, $M$, $l_c$, the spatial grid level and
the Monte Carlo tests, respectively.
We see the same behavior as in the log-normal case.
The only significant difference is the lower error of the Sparse method in the case $M=10$, $R_l=1$, which is 1.08e-3 for the beta distribution and 1.45e-1 for the log-normal one.
\begin{table}[t]
\centering
\caption{Performance versus $p$, beta distribution}
\label{tab:beta-p}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$p$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
1 & 21.40 & 1.382 & 0.059 & 64 & 49 & 1 & 2.24e-3 & 5.13e-2 & 1.14e-2 & 2.50e-2 & 0 \\
2 & 39.87 & 5.301 & 0.100 & 65 & 50 & 1 & 1.92e-4 & 5.50e-3 & 7.67e-4 & 1.28e-3 & 0 \\
3 & 57.16 & 1000 & 70.98 & 65 & 50 & 445 & 9.07e-5 & 1.76e-3 & 5.01e-4 & 1.06e-3 & 1.88e-4 \\
4 & 76.22 & --- & 21.18 & 65 & 50 & 416 & 8.81e-5 & --- & 1.41e-4 & --- & 9.84e-5 \\
5 & 100.6 & --- & 119.7 & 65 & 50 & 428 & 8.89e-5 & --- & 1.10e-4 & --- & 1.23e-4 \\
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\caption{Performance versus $M$, beta distribution}
\label{tab:beta-M}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$M$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
10 & 9.777 & 5.796 & 0.942 & 34 & 40 & 39 & 1.70e-4 & 1.65e-3 & 5.18e-4 & 1.08e-3 & 1.95e-4 \\
15 & 26.46 & 90.34 & 25.16 & 50 & 48 & 374 & 1.03e-4 & 1.73e-3 & 4.96e-4 & 1.08e-3 & 1.94e-4 \\
20 & 56.92 & 986.2 & 59.57 & 65 & 50 & 413 & 9.15e-5 & 1.80e-3 & 5.08e-4 & 1.09e-3 & 1.88e-4 \\
30 & 156.7 & 55859 & 147.9 & 92 & 50 & 452 & 7.75e-5 & 7.01e-2 & 5.12e-4 & 4.53e-2 & 1.85e-4
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\caption{Performance versus $l_c$, beta distribution}
\label{tab:beta-l}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$l_c$ & \multicolumn{3}{c|}{CPU time, sec.} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
0.1 & 665.8 & 55923 & 0.91 & 90 & 48 & 1 & 8.7e-3 & 8.77e-3 & 7.9e-3 & 7.92e-3 & 0 \\
0.3 & 2983 & 53783 & 1.49 & 177 & 74 & 1 & 1.5e-3 & 2.02e-3 & 1.2e-3 & 1.30e-3 & 0 \\
0.5 & 1138 & 54297 & 100 & 132 & 74 & 403 & 1.5e-4 & 1.71e-3 & 2.9e-4 & 8.21e-4 & 2.47e-23 \\
1.0 & 158.8 & 56545 & 153 & 92 & 50 & 463 & 7.8e-5 & 6.92e-2 & 5.1e-4 & 4.47e-2 & 1.96e-04 \\
1.5 & 62.20 & 55848 & 89.5 & 75 & 42 & 409 & 6.9e-5 & 7.85e-2 & 8.3e-4 & 4.56e-2 & 2.20e-03 \\
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\caption{Performance versus $R$, beta distribution}
\label{tab:beta-R_l}
\begin{tabular}{c|ccc|ccc|cc|cc|c}
$\#$DoF & \multicolumn{3}{c|}{CPU time, sec} & $r_{\kappa}$ & $r_u$ & $r_{\hat \chi}$ & \multicolumn{2}{c|}{$E_{\kappa}$} & \multicolumn{2}{c|}{$E_{u}$} & $\mathbb{P}$ \\ \hline
& TT & Sparse & $\hat\chi$ & & & & TT & Sparse & TT & Sparse & TT \\ \hline
557 & 9.73 & 5.94 & 0.94 & 34 & 40 & 39 & 1.70e-4 & 1.65e-3 & 5.21e-4 & 1.08e-3 & 1.95e-4 \\
2145 & 36.2 & 12.7 & 0.77 & 34 & 41 & 41 & 1.56e-4 & 1.64e-3 & 5.19e-4 & 1.08e-3 & 1.97e-4 \\
8417 & 378 & 162 & 1.12 & 34 & 40 & 43 & 1.53e-4 & 1.62e-3 & 5.07e-4 & 1.06e-3 & 1.96e-4 \\
\end{tabular}
\end{table}
\begin{table}[t]
\centering
\caption{Verification of the Monte Carlo method, beta distribution}
\label{tab:beta-mc}
\begin{tabular}{c|l|cc|cc||c|c}
$N_{mc}$ & $T_{MC}$, sec. & $E_{\bar u}$ & $E_{\mbox{var}_u}$ & $\mathbb{P}_*$ & $E_{\mathbb{P}}$ & \multicolumn{2}{c}{TT results} \\ \hline
$10^2$ & 0.9428 & 9.12e-3 & 1.65e-1 & 0 & $\infty$ & $T_{solve}$ & 278.4014 sec. \\
$10^3$ & 9.5612 & 1.04e-3 & 6.04e-2 & 0 & $\infty$ & $T_{\hat\chi}$ & 179.4764 sec. \\
$10^4$ & 8.849$\cdot 10^1$ & 4.38e-4 & 5.56e-3 & 0 & $\infty$ & $r_{\kappa}$ & 92 \\
$10^5$ & 8.870$\cdot 10^2$ & 2.49e-4 & 3.06e-3 & 7.00e-5 & 6.80e-1 & $r_{u}$ & 50 \\
$10^6$ & 8.883$\cdot 10^3$ & 8.16e-5 & 8.56e-4 & 1.07e-4 & 9.94e-2 & $r_{\hat\chi}$ & 406 \\
& & & & & & $\mathbb{P}$ & 1.1765e-04
\end{tabular}
\end{table}
\section{Conclusion}
We have developed the new block TT cross algorithm to compute the TT approximation of the polynomial chaos expansion of a random field with the tensor product set of polynomials, where the polynomial degrees are bounded individually for each random variable.
The random field can be given as a transformation of a Gaussian field by an arbitrary smooth function.
The new algorithm builds the TT approximation of the PCE in a black box manner.
Compared to the previously existing cross methods, the new approach assimilates all KLE terms simultaneously, which reduces the computational cost significantly.
The uncertain (diffusion) coefficient in the elliptic PDE is approximated via PCE.
We show that the tensor product polynomial set allows a very efficient construction of the stochastic Galerkin matrix in the TT format, provided the coefficient is precomputed in the TT format.
Interestingly, we can even compute the Galerkin matrix exactly by preparing the coefficient with two times larger polynomial orders than those employed for the solution.
In the TT format, we can store the Galerkin matrix in the dense form, since the number of the TT elements $\mathcal{O}(M p^2 r^2)$ is feasible for $p \sim 10$.
This also means that other polynomial families, such as the Chebyshev or Laguerre, may be used straightforwardly.
The Galerkin matrix defines a large linear system on the PCE coefficients of the solution of the stochastic equation.
We solve this system in the TT format via the alternating minimal energy algorithm and calculate the post-processing of the solution, such as the mean, variance and exceedance probabilities.
We demonstrate that with the new TT approach we can go to a larger number of random variables (e.g. $M=30$) used in the PCE (larger stochastic dimension)
and take a larger order of the polynomial approximation in the stochastic space ($p=5$) on a usual PC desktop computer.
For all stages of numerical experiments (computation of the coefficient, operator, solution and statistical functionals)
we report the computational times and the storage costs (TT ranks), and show that they stay moderate in the investigated range of parameters.
In particular, the TT ranks do not grow with the polynomial degree $p$.
This remains in sharp contrast to the traditional sparse polynomial approximation, where the total polynomial degree is bounded.
The cardinality of this sparse polynomial set grows exponentially with $p$, but the tensor product decomposition is not possible anymore.
This renders the total computational cost of the sparse PCE approach higher than the cost of the TT method.
Besides, the tensor product PCE is more accurate than the expansion in the sparse set due to a larger total polynomial degree.
Comparison with the classical Monte Carlo method shows that the TT methods can compute the exceedance probabilities more accurately,
since the TT format approximates the whole stochastic solution implicitly.
Several directions of research can be pursued in the future.
Currently, we store both the matrix and the inverted mean-field preconditioner in the dense form.
This imposes rather severe restrictions on the spatial discretization.
The spatial part of the Galerkin matrix must be dealt with in a more efficient way.
With the tensor product methods the stochastic collocation approach seems very attractive \cite{khos-pde-2010}.
We may introduce quite large discretization grids in each random variable $\theta_m$: additional data compression can be achieved with the QTT approximation \cite{khor-qtt-2011}.
It is important that the deterministic problems are decoupled in the stochastic collocation.
The cross algorithms can become an efficient non-intrusive approach to stochastic equations.
\subsection*{Acknowledgments}
We would like to thank Elmar Zander for his assistance and help in the usage of the Stochastic Galerkin library \emph{sglib}.
A. Litvinenko was supported by the SRI-UQ Center at King Abdullah University of Science and Technology.
A part of this work was done during his stay at Technische Universit\"at Braunschweig and was supported by the German DFG Project CODECS "Effective approaches and solution techniques for conditioning, robust design and control in the subsurface".
S. Dolgov was partially supported by RFBR grants 13-01-12061-ofi-m, 14-01-00804-A, RSCF grants 14-11-00806, 14-11-00659, and the Stipend of the President of Russia during his stay at the Institute of Numerical Mathematics of the Russian Academy of Sciences.
|
\section{Introduction}
\label{intro}
The level density of atomic nuclei provides important information on the heated nucleonic many-particle system. In the pioneering work of
Hans Bethe in 1936~\cite{bethe36}, the level density was described as a gas of non-interacting fermions moving in equally
spaced single-particle orbitals. Despite the scarce nuclear structure knowledge at that time, this Fermi-gas model contained
the essential components apart from the influence of the pairing force between nucleons in time-reversed orbitals. The
understanding of such Cooper pairs and the nuclear structure consequences, was first realized twenty years
later through the Bardeen-Cooper-Schrieffer (BCS) theory~\cite{BCS}.
Unfortunately, the new knowledge of the pairing force was implemented in an oversimplified way by
still assuming a gas of fermions.
The original Fermi-gas level density curve was simply shifted up in energy by $\Delta$ and $2\Delta$
for the description of odd-mass and even-even mass nuclei, respectively. Later, this energy correction was found to be too large,
and the shift was somewhat back-shifted again.
The resulting back-shifted Fermi-gas model has since then been very
popular and has been the most common description of nuclear level densities for decades~\cite{capote2009}.
These manipulations maintained the typical $\exp(2\sqrt{aE})$ Fermi-gas level density shape with excitation energy $E$.
There is a growing interest in the nuclear science community for the study of nuclear level densities. The introduction of novel
theoretical approaches and fast computers has opened the way for a microscopical description of heavier nuclei up to high
excitation energy~\cite{goriely2009}.
Thus, experimental level densities represent a basic testing ground for many-particle nuclear structure models. They
also have an increasing importance for various nuclear applications. Nuclear level densities play an essential role in the calculation of
reaction cross sections applied to astrophysical nucleosynthesis, nuclear energy production and transmutation
of nuclear waste.
The number of levels is exponentially increasing with excitation energy and typically for the rare earth region,
the level density increases by a factor of one million when going from the ground state up to the neutron binding energy.
This tremendous change has implications for the theoretical interpretations since levels start to overlap in energy at
high excitation energies, typically around 10-15 MeV. Thus, nuclear excitations are
often divided into three regimes \cite{Ericson}, classified by the average level spacings $D$ compared to the
$\gamma$-decay width $\Gamma$. These key numbers are connected to the level density and life time by $\rho=1/D$ and $\tau=\hbar/\Gamma$, respectively.
Figure~\ref{fig:regimes} illustrates the connections between the three regimes i.e.~the discrete, quasi-continuum and continuum regions.
Experimentally, the quasi-continuum
region is often defined as the region where the experimental resolution $\Delta E_{\rm res}$ prevents
the separation into individual levels. This appears when $\Delta E_{\rm res}\geq D$. One
should be aware that this definition depends on the selectivity
of the reactions used and the specific experimental detectors and conditions. In this work,
the experimental quasi-continuum region starts when $\Delta E_{\rm res}\approx D \approx 100$~keV, i.e. when the level density
exceeds $\rho\approx 10$~MeV$^{-1}$.
\begin{figure}[]
\begin{center}
\includegraphics[width=\columnwidth]{regimes.pdf}
\caption{(Color online) The energy regimes of nuclear excitations.}
\label{fig:regimes}
\end{center}
\end{figure}
Today various experimental techniques are used to determine nuclear level densities. The obvious method
in the discrete region of Fig.~\ref{fig:regimes} is to count the number of
levels $\Delta N$ per excitation energy bin $\Delta E$, readily giving the level density as $\rho(E)=\Delta N/\Delta E$.
This requires that all levels are known. Making a short review on known discrete levels,
e.g.~in the compilation of Ref.~\cite{NNDC}, we
find that the data set starts to be incomplete where the first Cooper pairs are broken,
i.e. at excitation energies $E \approx 2\Delta$. This is
simply because the high density of levels makes conventional spectroscopy difficult.
A rule of thumb is that when $\rho \approx 50- 100$~MeV$^{-1}$, the number of missing levels is dramatically increasing.
Level density information at the neutron separation energy $S_n$
can be extracted from neutron capture resonance spacings $D$~\cite{RIPL3}. The
spin selection of these resonances depends on the target ground-state spin and the neutron spin transfer,
giving level densities for a narrow spin window, only. In particular, thermal neutrons with angular momentum
transfer $\ell=0$ results in a spin window of $I_{\rm target}\pm 1/2$.
At higher excitation energies the
method of Ericson fluctuations can be used~\cite{Ericson}. A well known technique~\cite{Vonach} is
to extract the level density from particle evaporation spectra.
The measurement must be carried out at backwards center-of-mass angles to avoid direct reaction contributions.
The method also requires well determined optical potential model parameters.
Furthermore,
the level density can be extracted in the two-step cascade method~\cite{Hoogenboom} from simulations
provided that the $\gamma$-ray strength function is known.
An upcoming technique, see e.g. Ref.~\cite{parity2007}, is to use high-energy light-ion reactions
like $(p, p')$, $(e, e')$ and $(^3$He, $t)$ at small angles with respect to the beam direction. These reactions select
the population of discrete levels with certain spin/parity assignments, information
that is essential in the detailed understanding of level densities.
This method requires that the spacings of the selected levels
are larger or comparable to the experimental particle resolution.
In the present work we will show level densities of several nuclei measured at the Oslo Cyclotron Laboratory (OCL).
The technique used, the Oslo method~\cite{Schiller00,Lars11}, is unique in the sense that it
allows for a simultaneous determination of the level density and the $\gamma$-ray
strength function without assuming any models for these functions.
The structure of the manuscript is as follows: Section 2 describes briefly the experimental set-up and the Oslo method,
and previously measured level densities will be discussed in Sect.~3. Finally, concluding remarks and outlooks are
presented in Sect.~4.
\section{The Oslo Method}
\label{sec:1}
The Oslo method is based on a set of $\gamma$ spectra as a function of excitation energy. The
excitation of the nucleus is performed by light ion reactions, e.g. $(d, p\gamma)$, $(p,p'\gamma)$ and $(^3$He, $\alpha\gamma)$
where the energy of the charged ejectile determines the excitation energy.
A schematic drawing of the set-up is shown in Fig.~\ref{fig:set-up}. A silicon particle detection system (SiRi)~\cite{siri}, which
consists of 64 telescopes, is used for the selection of a certain ejectile type and to determine its energies. The
front $\Delta E$ and back $E$ detectors have thicknesses of 130 $\mu$m and 1550 $\mu$m, respectively. SiRi
is usually placed in backward angles covering $\theta = 126^\circ$ to $140^\circ$
relative to the beam axis. Coincidences with $\gamma$ rays are performed with the CACTUS array~\cite{CACTUS},
consisting of 26 collimated $5" \times 5"$ NaI(Tl) detectors with a total efficiency of $14.1$\% at $E_\gamma = 1.33$~MeV.
\begin{figure}[]
\begin{center}
\includegraphics[width=\columnwidth]{set-up.pdf}
\caption{(Color online) Typical particle-$\gamma$ coincidence set-up for the Oslo method.
The 64 silicon particle telescopes of SiRi are placed in the vacuum chamber at the center of CACTUS. }
\label{fig:set-up}
\end{center}
\end{figure}
Figure~\ref{fig:matrices} demonstrates the previously measured\\
$^{237}$Np$(d, p \gamma )^{238}$Np reaction~\cite{238Np} and
how one can proceed from the raw particle-$\gamma$ coincidences to the first generation or primary $\gamma$-ray spectra. The first step
is to sort the events into a raw particle-$\gamma$ matrix $R(E, E_{\gamma})$ with proper subtraction of
random coincidences. Then, for all initial excitation energies $E$, the $\gamma$ spectra are unfolded with the NaI response
functions giving the matrix $U(E, E_{\gamma})$~\cite{gutt1996}. The procedure is iterative and stops
when the folding ${\cal F}$ of the unfolded matrix equals the raw matrix within the statistical fluctuations, i.e.~when
${\cal F}(U)\approx R$.
\begin{figure}[]
\begin{center}
\includegraphics[width=0.9\columnwidth]{238Np_matrices.pdf}
\caption{(Color online) Initial excitation energy $E$ versus $\gamma$-ray
energy $E _{\gamma}$ from the $^{237}$Np$(d, p \gamma )^{238}$Np reaction~\cite{238Np}. The raw $\gamma$-ray spectra
(a) are first unfolded by the NaI response function (b) and then
the primary $\gamma$-ray spectra are extracted (c).}
\label{fig:matrices}
\end{center}
\end{figure}
The primary $\gamma$-ray spectra can be extracted from the unfolded total $\gamma$ spectra
$U$ of Fig.~\ref{fig:matrices} (b).
The primary $\gamma$ spectrum at an initial excitation energy $E$ is obtained by subtracting a weighted sum of
$U(E',E_{\gamma})$ spectra below excitation energy $E$:
\begin{equation}
P(E,E_{\gamma})=U(E,E_{\gamma}) - \sum_{E' < E}W(E,E')U(E',E_{\gamma}).
\end{equation}
The weighting coefficients $W(E, E')$ are determined in an iterative way described in Ref.~\cite{Gut87}.
After a few iterations, $W(E,E')$ converges to $P(E,E_{\gamma})$, where we have normalized
each $\gamma$ spectrum by $\sum_{E_{\gamma}}P(E,E_{\gamma})=1$.
This equality $P\approx W$ is exactly what is expected, namely that
the primary $\gamma$-ray spectrum
equals to the weighting function. The validity of the
procedure rests on the assumption that the $\gamma$-energy
distribution is the same whether the levels were populated directly by
the nuclear reaction or by $\gamma$ decay from
higher-lying states.
This is illustrated in Fig.~\ref{fig:firstgen}.
\begin{figure}[]
\begin{center}
\includegraphics[width=0.8\columnwidth]{firstgen.pdf}
\caption{(Color online) The assumption of the extraction of primary $\gamma$ spectra
is that we obtain the same energy distribution (yellow) whether the levels of the energy bin
$E_2$ are populated directly by e.g. the $(d, p)$ reaction or by $\gamma$ transitions from
excitation bin $E_1$.}
\label{fig:firstgen}
\end{center}
\end{figure}
The statistical part of this
landscape of probability, $P(E,E_{\gamma})$, is then assumed
to be described by the product of two vectors
\begin{equation}
P(E, E_{\gamma}) \propto \rho(E-E_{\gamma}){\cal{T}}(E_{\gamma}) ,\
\label{eqn:rhoT}
\end{equation}
where the decay probability should be proportional to the
level density at the final energy $\rho(E-E_{\gamma})$ according
to Fermi's golden rule~\cite{dirac,fermi}. The decay is also proportional
to the $\gamma$-ray transmission coefficient ${\cal{T}}$, which
according to the Brink hypothesis~\cite{brink}, is independent of excitation energy;
only the transitional energy $E_{\gamma}$ plays a role.
\begin{figure*}[]
\begin{center}
\includegraphics[width=1.9\columnwidth]{does_it_work_238Np.pdf}
\caption{(Color online) Primary $\gamma$-ray spectra from
the $^{237}$Np$(d, p \gamma )^{238}$Np reaction~\cite{238Np}
gated on various initial excitation energies $E$ (crosses).
The spectra are compared to the product of the level density and transmission coefficient functions
i.e.~$\rho(E-E_{\gamma}) {\mathcal{T}}(E_{\gamma})$ (blue lines).
The excitation energy gates are 121 keV broad.}
\label{fig:doesitwork}
\end{center}
\end{figure*}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\columnwidth]{countingx.pdf}
\caption{Normalization procedure for the experimental level density.
The experimental data are normalized to known discrete levels
at low excitation energy and to the level density
extracted at $S_n$ from neutron capture resonance spacings $D_0$.
The two set of arrows indicate where the data are normalized by means of Eq.~(\ref{eq:array1}).
}
\label{fig:countingx}
\end{center}
\end{figure}
\begin{figure*}[tbh]
\begin{center}
\includegraphics[width=2\columnwidth]{rhotot_all.pdf}
\caption{(Color online) Experimental level densities obtained with the Oslo method.
For references, see Table~\ref{tab:parameters}.
}
\label{fig:rhotot_all}
\end{center}
\end{figure*}
The values of the vector elements of $\rho$ and $\cal{T}$ are varied
freely to obtain a least-$\chi ^2$ fit to the experimental $P$ matrix.
Figure~\ref{fig:doesitwork} shows that the product of $\rho(E-E_{\gamma}){\cal{T}}(E_{\gamma})$ of Eq.~(\ref{eqn:rhoT})
fits very well the primary spectra of $^{238}$Np at the six excitation energy bins shown.
There are actually four times more spectra and the same quality of fit is obtained for all spectra.
However, the nice fit to the data does not mean that $\rho$ and $\cal{T}$ are uniquely determined.
There are infinitely many functions that make identical
fits to the experimental $P$ matrix.
These $\tilde{\rho}$ and $\tilde{\cal{T}}$ functions can be generated by the transformations~\cite{Schiller00}
\begin{eqnarray}
\tilde{\rho}(E_i-E_\gamma)&=&A\exp[\alpha(E_i-E_\gamma)]\,\rho(E_i-E_\gamma),
\label{eq:array1}\\
\tilde{{\mathcal{T}}}(E_\gamma)&=&B\exp(\alpha E_\gamma){\mathcal{T}} (E_\gamma).
\label{eq:array2}
\end{eqnarray}
Thus, the normalization of the two functions requires information to fix
the parameters $A$, $\alpha$, and $B$, which is not available from our experiment.
For the level density, we use two anchor points to determine $A$ and $\alpha$.
The procedure is demonstrated for the case of $^{238}$Np in Fig.~\ref{fig:countingx}.
The level density is normalized to known discrete levels at low excitation energy.
At high excitation energy, we exploit the neutron resonance spacings $D_0$
at the neutron separation energy $S_n$. In order to translate the spacing
to level density, we use the spin distribution~\cite{GC}
\begin{equation}
g(E,I) \simeq \frac{2I+1}{2\sigma^2}\exp\left[-(I+1/2)^2/2\sigma^2\right],
\label{eq:spindist}
\end{equation}
where $E$ is excitation energy and $I$ is spin. Here, the spin cut-off parameter
$\sigma$ is rather uncertain, and may introduce an uncertainty of a factor of 2 in $\rho(S_n)$.
For cases where $D_0$ is unknown, one has to use systematics to estimate $\rho(S_n)$.
Since the $\gamma$ strength is not discussed in this work, the
determination of the $B$ parameter is irrelevant.
Further description and tests of the Oslo method are given in Refs.~\cite{Schiller00,Lars11}.
\section{Experimental level densities}
\label{sec:2}
\begin{table*}[]
\begin{center}
\caption{Parameters to describe the level density.}
\begin{tabular}{l|cccc|r}
\hline
\hline
Reaction &$S_n$ & $T_{\rm CT}$ & $E_0$ & $\rho(S_n)$ & Refs. \\
&(MeV) & (MeV) & (MeV) & ($10^6$MeV$^{-1}$)& \\
\hline
$(p, p')^{45}$Ti & 9.530& 1.55 & -2.33 & 0.0014(7) &\cite{45Ti} \\
$(d, p)^{46}$Ti &13.189& 1.70 & -2.07 & 0.0047(10) &\cite{46Ti} \\
$(p, p')^{89}$Y &11.478& 1.00 & 0.48 & 0.060(20) &\cite{89Y} \\
$(d, p)^{90}$Y & 6.857& 1.00 & -1.38 & 0.0038(8) &\cite{89Y} \\
($^3$He,$\alpha)^{116}$Sn& 9.563& 0.76 & -0.03 & 0.40(20) &\cite{117SnPRL,116117Snnld,116117Sn,121122Sn} \\
($^3$He,$^3$He$'$)$^{117}$Sn & 6.944& 0.67 & -0.40 & 0.091(27) &\cite{117SnPRL,116117Snnld,116117Sn,121122Sn} \\
($^3$He,$\alpha$)$^{163}$Dy& 6.271& 0.59 & -1.55 & 0.96(12) &\cite{nldDy} \\
($^3$He,$^3$He$'$)$^{164}$Dy & 7.658& 0.60 & -0.66 & 1.74(21) &\cite{nldDy} \\
($d, p)^{238}$Np & 5.488& 0.41 & -1.35 & 43.0(78) &\cite{238Np} \\
\hline
\hline
\end{tabular}
\label{tab:parameters}
\end{center}
\end{table*}
The Oslo group has now published more than 70 level densities.
Some of these are for one and the same nucleus, but populated with different reactions.
The full compilation of level densities (and $\gamma$-ray strength functions)
is found in Ref.~\cite{compilation}. We will here focus on
some selected nuclei in order to convey to the reader a general picture
of how the level density behaves
as function of mass region and excitation energy. The references and key
numbers are given in Table~\ref{tab:parameters}.
There is a tremendous span in level densities when going from mass
around $A \approx 40$ to 240. Figure~\ref{fig:rhotot_all} shows that the
level density at excitation energy $E \approx 6$~MeV for $^{238}$Np
is almost ten million times higher than for $^{46}$Ti. It seems obvious
that more nucleons $A$ produce higher level density, but this conclusion is
not strictly true.
Indeed, we see that the level density for $^{90}$Y is larger than for $^{89}$Y.
Also $^{117}$Sn has more levels than $^{116}$Sn.
However, $^{45}$Ti has more levels than $^{46}$Ti and the
same is true for $^{163}$Dy versus $^{164}$Dy. The pieces fall into place
if pairing is taken into account: For neighboring isotopes, the nucleus
with either one or more unpaired protons or neutrons displays $5 - 7$ times more levels.
An increase in $A$ only gives negligible effects between neighboring nuclei,
as e.g.~found for the even-even $^{160,162,164}$Dy isotopes~\cite{luciano2014}.
The number
of quasi-particles is the basis building blocks for the level density.
However, the number of active particles making the level density is not that
different for $^{238}$Np and $^{46}$Ti. The odd-odd $^{238}$Np has about 6 quasi-particles at $E \approx 6$~MeV,
while $^{46}$Ti has about 2 quasi-particles.
The second factor needed to understand the general behaviour of the level density is the density of
available single-particle orbitals in
the vicinity of the Fermi level where the quasi-particles can play around.
In summary, the large span in level density
between $A = 40$ and 240 is a combined result of the number of quasi-particles and the density of
near lying single-particle orbitals.
Another striking feature of Fig.~\ref{fig:rhotot_all} is that,
above a certain excitation energy $E$, the level densities are linear
in a log-plot, which corresponds to an exponential function.
The $^{45}$Ti has an exponential behavior for $E > 4$~MeV,
whereas $^{238}$Np has it for $E > 1$~MeV.
Generally, the exponential behavior sets in when the first nucleon pairs
are broken, i.e. $E > 2 \Delta$. With a pairing gap of
$\Delta \approx 12A^{-1/2}$~MeV, we find for the two cases $E= 3.6$ and $1.6$ MeV, respectively.
The level density is not only dependent on the atomic mass number $A$,
but also on the number of unpaired nucleons around the ground state.
As an example, the exponential behavior is delayed for even-even nuclei compared to odd-even nuclei.
This is e.g. apparent for $^{116,117}$Sn and $^{163,164}$Dy.
The interpretation of the pure exponential growth above $E > 2 \Delta$ has recently been discussed in Ref.~\cite{luciano2014}.
For the micro-canonical ensemble, the entropy is related to the level density by
\begin{equation}
S(E) \approx \ln \rho(E),
\end{equation}
with
\begin{equation}
T(E)=[\partial S(E)/\partial E]^{-1}.
\label{eqn:t}
\end{equation}
Thus, the temperature is constant when $\ln \rho(E)$ is linear in $E$. The common
expression for the constant-temperature level density is~\cite{Ericson}
\begin{equation}
\rho_{\rm CT}(E)=\frac{1}{T_{\rm CT}}\exp{\frac{E-E_0}{T_{\rm CT}}},
\label{eq:ct}
\end{equation}
where $T_{\rm CT}$ is determined by the slope of $\ln \rho(E)$.
This is the key characteristic of a first-order phase transition.
Our interpretation of the data is that the energy goes into breaking
Cooper pairs and thus the temperature remains constant. The
obvious analogy is the melting of ice to water at constant temperature.
The fit parameters $T_{\rm CT}$ and $E_0$ for
the nuclei shown in Fig.~\ref{fig:rhotot_all} are listed in Table~\ref{tab:parameters}.
Figure \ref{fig:ctfg} shows that the constant-temperature level density model fits
the data of $^{164}$Dy much better than the Fermi gas model. The reduced $\chi ^2$ fit value
evaluated between excitation energies $E= 1.9$ and $6.2$ MeV is about a factor of 10 better
for the CT model. One should of course mention that the CT model is not a perfect model.
There are some oscillations in the experimental log-plot mainly due to the onset of the
two and four quasi-particle regimes.
Recommendations are available for the Fermi-gas spin-cut off
parameter $\sigma (A,E)$ \cite{Egidy2009}. However, for the CT-model there are
no clear recommendations. According to the original expression of Ericson~\cite{Ericson}:
\begin{equation}
\sigma^2= \frac{\Theta T}{\hbar^2},
\end{equation}
we see that for the CT model, only the moment of inertia $\Theta$ will depend on $E$,
since $T=T_{CT}$ is a constant. There are indications~\cite{Alhassid2007} that the rigid moment
of inertia is reached at the neutron separation energy, i.e.~$\Theta(E=S_n)=\Theta_{\rm rigid}$.
However, how to parameterize $\sigma$ to lower excitation energies is still uncertain.
In addition to the uncertain spin distribution, the parity distribution is important for
lighter nuclei and nuclei close to magic numbers. There are only few experimental
data available in order to pin down the spin and parity distributions and this
represents perhaps the most important issue for future research on nuclear level densities.
The pair breaking process will continue at least up to the binding energy, but probably
to much higher excitations. In several evaporation experiments the
constant temperature behavior has been observed up to $E \approx 15$~MeV \cite{Voinov2009,Byun2014}.
However, at the higher excitation energies the temperature will start increasing with
excitation energies, and the Fermi gas description~\cite{bethe36,capote2009} will to an increasing degree describe the data.
\section{Concluding remarks and outlook}
\begin{figure}[b]
\begin{center}
\includegraphics[width=1\columnwidth]{counting.pdf}
\caption{(Color online) Experimental level densities of $^{164}$Dy compared
with the constant-temperature (CT) and the Fermi gas (FG) level density formulas.
The reduced least-$\chi ^2$ fit is best for the CT model.}
\label{fig:ctfg}
\end{center}
\end{figure}
The present work describes the Oslo method and results using light-ion reactions on stable targets.
The method is demonstrated by means of the $^{237}$Np$(d, p \gamma )^{238}$Np reaction, which provides almost ideal
conditions for the Oslo method. One important aim of these investigations is to obtain more
predictive power for extrapolation of the level densities to nuclei outside the valley of $\beta$-stability.
From light ($^{45}$Ti) to the heavy elements ($^{238}$Np) we find certain common features of the level densities.
It turns out that the pairing force is responsible for the mechanism behind these features.
The level density above the excitation energy of 2$\Delta$ behaves as an exponential curve
characterized by the nuclear temperature. Furthermore, neighboring nuclei have the same
temperature and their respective level densities appear parallel in a log plot.
Odd-mass nuclei exhibit about $5 - 7$ times more levels than their even-even partners.
There are indications that each unpaired nucleon carry approximately an entropy of $1.6 - 1.9 k_B$,
corresponding to the multiplicative factor above.
In the field of experimental level densities, there are several skeletons in the closet.
In particular, it is crucial to get more information on the true spin and parity distributions
as function of excitation energy and mass. The few experimental data on these quantities
also have implications for applications when e.g. calculating reaction rates.
Today, such information is often taken from theoretical models
without verifications based on experimental observations.
In the energy regime from the ground state up to the energy where the breaking
of the first Cooper pairs set in, the level density depends strongly on the
specific nuclear structure properties of the nucleus studied. For even-even nuclei,
the level density is determined by the collective modes like rotation and vibrations.
For the odd-mass nuclei the unpaired valence nucleon already carries a quasi-particle
entropy that adds to these collective degrees of freedom.
A great challence is to obtain data also for nuclei far away from the $\beta$-stability line,
nuclei that are essential for nucleo-synthesis networks and other applications.
Since it is a rather uncertain approach to extrapolate results from more stable nuclei
to these nuclei, novel experimental techniques are needed.
The Oslo group has already initiated experiments to measure level densities for
short-lived nuclei. The so-called $\beta$-Oslo method is based on the population of
rare nuclei by means of $\beta$ decay. The first experiment, which was performed at
the National Superconducting Cyclotron Laboratory at Michigan State University (MSU)
using the SuN detector has already been published~\cite{spyrou}, and several new experiments
have been carried out. Also inverse kinematic experiments with radioactive beams are planned at HIE-ISOLDE.
The constant-temperature level density behavior is expected also for nuclei far
from the $\beta$-stability line. However, when the particle separation energies
approaches the paring energy of $E = 2\Delta$, the individual low-lying discrete levels will
play an increasing role. Furthest away from the valley of stability, discrete spectroscopy of
low-lying levels will probably be the way to proceed. In such scenarios the concept
of level density will lose its importance and be replaced by discrete spectroscopy.
\label{sec:3}
|
\section{Introduction}
Long back $CP$ violation has been found in the quark sector of the Standard Model of Particle Physics. But so far there is no evidence of $CP$ violation in the leptonic sector. One way to search for such $CP$ violation is through neutrino oscillation experiments in which one flavor of neutrino could oscillate to other flavors of neutrinos. Neutrino oscillation probability depends on various oscillation parameters present in the $3 \times 3$ mixing matrix - the PMNS matrix \cite{pmns,bilen,part} and the neutrino mass squared differences. Two of the three angles - $\theta_{12}$ and $\theta_{23}$ present in PMNS matrix have been known with certain accuracy for some time. Recently several experiments like Double CHOOZ, Daya Bay, RENO and T2K collaboration \cite{double,daya,reno,t2k} found non-zero value of $\sin^2 2 \theta_{13}$ corresponding to the third mixing angle of PMNS matrix with a global level of significance which is well above conventional $5\sigma$ discovery threshold. As three angles are non-zero there could be
non-zero $CP$ violating phase $\delta $ in the PMNS matrix and as such there could be $CP$ violation in the leptonic sector. The mass squared differences - $|\Delta m^2_{31}|$ and $\Delta m^2_{21}$ is known (where $\Delta m^2_{ij} = m_i^2 -m_j^2$) but the sign of $\Delta m^2_{31}$ and as such the hierarchy (whether it is normal (NH) or inverted (IH)) of neutrino masses is still unknown. Various neutrino oscillation experiments like superbeam, neutrino factory, beta beams and reactor experiments are focussing on determining these unknown parameters corresponding to neutrino oscillations \cite{lbne,lbno,beta,Edgecock:2011ova,others,others5,Coloma:2012ma,others4,Itow,Choubey,Zucc,Bouch}.
There could be various kind of non-standard interactions of neutrinos with matter. This could have some effect on
finding $CP$ violation in neutrino oscillation experiments. Such interactions could be the non-standard
interactions of neutrinos with matter fermions ($u, d$ and $e$) during propagation of neutrinos only. This could affect oscillations
of different flavors of neutrinos as sub-leading effect. We have discussed it in further detail in the next section. There could be
other different kind of interactions beyond Standard model leading to non-unitarity of
$3\times 3$ PMNS neutrino mixing matrix \cite{nu}. Considering non-standard interactions of neutrinos at the source and the detector in
neutrino oscillation experiments also lead to such possibility. However, such NSI at the source and detector have highly stringent
constraints \cite{Ohlsson} and as such the effect on neutrino oscillation is expected to be lesser affected than that due to NSI
in matter during propagation of neutrinos. We shall not consider NSI at the source and detector in this work.
To study $CP$ violation in neutrino oscillation experiments we have considered superbeam, neutrino factory and monoenergetic
neutrino beam from electron capture process. There are some earlier studies for short and long baselines for
standard interactions \cite{raut,Dighe,Coloma:2012ma,Agarwalla:2011hh,others4,Coloma:2012ut,betabeam}
for neutrinos coming from superbeam. We consider neutrino superbeam (which mainly contains $\nu_{\mu}$ and ${\bar \nu_{\mu}}$ ) coming from CERN travelling a baseline length of 2300 km to Pyh\"asalmi (Finland) and another baseline of 130 km to Fr\'ejus (France). We have also considered superbeam coming
from Tokai to Hyper-Kamiokande travelling a baseline of 295 km.
We do a comparative study in the discovery potentials of the $CP$ violating phase $\delta$ for these three baselines in presence of both standard and non-standard interactions. Main emphasis of our work is on exploring
$CP$ violation due to $\delta$ in the light of recent experimental findings of large value of $\theta_{13}$ and how NSI could affect the discovery reach of such $CP$ violation. Also there are some studies on $CP$ violation
in presence of non-standard interactions \cite{nsi08p,Ota} before Daya Bay experiment when $\theta_{13}$ was not so precisely known.
The discovery of $CP$ violation for different parent muon energy in neutrino factory and different length of baselines
have been
studied earlier \cite{Ballett:2012rz,Agarwalla:2010hk}. In the analysis done in
\cite{Ballett:2012rz} three different kinds of magnetized detectors: a Totally-Active Scintillator(TASD) and
two kinds of Liquid Argon detectors have been considered for baselines ranging from 1000 km to 4000 km and parent muon energies from 4 GeV to 25 GeV, whereas
in the analysis done in \cite{Agarwalla:2010hk} MIND (Toroidal
magnetized iron neutrino detector) detector has been considered.
As recent reactor neutrino experiments indicates
large value of $\theta_{13}$ it is important to study the discovery potential of $CP$ violation in neutrino oscillation
experiments in low energy neutrino factory \cite{lownufact2}.
There are some studies on the performance of low energy neutrino factories \cite{Choubey} in the context of standard model
interactions of neutrinos with matter
\cite{Ballett:2012rz,lownufact2,lownufact1,pas2,pas3,pas4,pas5,pas6,pas7} and also in presence of NSI \cite{lowenergy3,nsi3,fried,Coloma:2011rq,nsi4}.
In the works related to NSI \cite{lowenergy3,nsi3,fried,Coloma:2011rq,nsi4} longer baselines like 4000 and 7500 kms has been considered to have larger NSI effect on neutrino oscillation experiments
so as to find better NSI sensitivity and in that context $CP$ violation has also been studied.
Our work is complementary to these earlier analysis in the sense that we focus here on the better discovery reach of $CP$ violation
and as such we consider relatively shorter baselines where the matter effect and as such NSI effect will be lesser in
the neutrino oscillation experiments. However, in the context of such experimental set-up in neutrino factory which is better
optimized for $CP$ violation discovery (considering only SM interactions of neutrinos with matter) we have also addressed the
question of $CP$ violation coming due to NSI phases.
We have considered lower parent muon energy within 10 GeV which is found to be more favourable for the discovery of
$CP$ violation. The optimization for $CP$ violation discovery in the context of SM interactions has been done earlier \cite{Ballett:2012rz,Agarwalla:2010hk}. However, in this work we re-analysed this optimization with the updated detector
characteristics for the MIND detector \cite{mind1}. Based on this optimization
we have chosen a few relatively shorter baselines (730, 1290 and 1500 kms) and low parent muon energy of 10 GeV.
In recent years oscillation experiment using a neutrino beam with neutrinos emitted from an electron capture process is proposed \cite{Zucc,rolinec,mon,mono2,mono3,mono9,mono11,mono12}. Such beam can be produced using an accelerated nuclei that decay by electron capture. In this process an electron is captured by a proton releasing a neutron and an electron neutrino. So
the beam is purely of one flavor. In the rest frame of the mother nuclei the electron neutrino that is released from such process, has a definite energy Q. Since the idea of using a neutrino beam emitted from an electron capture process is based on the acceleration and storage of radioactive isotopes that decays to daughter nuclei, one may get the suitable neutrino energy by accelerating the mother nuclei with suitable Lorentz boost factor $\gamma$. One can control the neutrino energy by choosing the appropriate Lorentz boost factor as the energy that has been boosted by an appropriate boost factor towards the detector is given as $E=2 \gamma Q$. Hence for certain mother nuclei to get the required neutrino energy the boost factor have to be chosen appropriately with respect to Q. Due to the almost monoenergetic nature of such beam one can appropriately choose the neutrino energy for which the probability of oscillation could be large and sensitive to certain unknown neutrino oscillation parameters.
In this work we consider such a flavor pure electron neutrino beam emitted from electron capture process for a suitable $\gamma$ value where the beam is targeted towards a Water Cherenkov detector and perform numerical simulation to study the discovery reach of $CP$ violation in oscillation experiments.
In section \ref{sec:prob} we have mentioned various model dependent and independent bounds of NSIs and have discussed the
neutrino oscillation probability in presence of NSI. In section \ref{sec:sim}, we have discussed the procedure of numerical simulation
and in its different subsections related to superbeam, neutrino factory and monoenergetic neutrino beam we have discussed various experimental set-ups and results of our analysis on discovery of $CP$ violation. In section \ref{sec:con} we have concluded with
some remarks on the possibility of knowing the presence of NSIs in the context of superbeam and neutrino factory. For monoenergetic beam
we have discussed the suitable nuclei for discovering $CP$ violation and related technical constraints.
\section{NSI and its effect on neutrino oscillation probabilities}
\label{sec:prob}
We consider the non-standard interactions of neutrinos which could be outcome of effective theory at low energy after integrating out the heavy mediator fields at the energy scale of neutrino oscillation experiments. Apart from Standard Model (SM) Lagrangian density we consider the following non-standard fermion-neutrino interaction in matter defined by the Lagrangian:
\begin{eqnarray}
\label{eq:Lang}
\mathcal{L}_{NSI}^{M}=-2\sqrt{2}G_F\varepsilon_{\alpha \beta}^{f P}[\bar{f}\gamma_{\mu}Pf][\bar{\nu}_{\beta}\gamma^{\mu}L\nu_{\alpha}]
\end{eqnarray}
where $P \in (L,R)$, $L=\frac{(1-\gamma^5)}{2}$, $R=\frac{(1+\gamma^5)}{2}, $ $f=e, u, d$ and $\varepsilon_{\alpha \beta}^{fP}$ are termed as non-standard interactions (NSIs)
parameters signifying the deviation from SM interactions. These are non-renormalizable as well as not gauge invariant and are dimension-6 operators after heavy fields are integrated out \cite{Ohlsson}. Although at low energy NSI look like this but at high energy scale
where actually such interactions originate there they have different form. These NSI parameters can be reduced to the effective parameters and can be written as:
\begin{equation}
\label{eq:nsieffec}
\varepsilon_{\alpha \beta}=\sum_{f,P} \varepsilon_{\alpha \beta}^{fP}\frac{n_f}{n_e}
\end{equation}
where $n_f$ and $n_e$ are the fermion and the electron number density respectively in matter.
As these NSIs modify the interactions with matter from the Standard Model interactions the effective mass matrix for the neutrinos
are changed and as such there will be change in the oscillation probability of different flavor of neutrinos. Although NSIs could be present at the source of neutrinos, during the propagation of neutrinos and also during detection of neutrinos \cite{ra} but as those effects are expected to be smaller at the source and
detector due to their stringent constraints \cite{nsi1,Ohlsson}, we consider the NSI effect during the propagation of neutrinos only.
Model dependent \cite{nsi01p,nsi02p,nsi03p,nsi04p,nsi05p,nsi06p,nsi09p,nsi010p,nsi011p, indbound1,depbound2, 1loop, sknsi,Ohlsson} and independent \cite{nsi1,68,Es} bounds are obtained for these matter NSI parameters and are shown in the following table. In obtaining model dependent bounds on matter NSI the experiments with neutrinos and charged leptons - LSND, CHARM, CHARM-II, NuTeV and also LEP-II have been considered. Bounds coming from loop effect have been used for model dependent bounds. However, model independent bounds are less stringent and could be larger than the model dependent bounds
by several orders and have been discussed in \cite{nsi1,Ohlsson}.
\begin{table}[ht]
\centering
\begin{tabular}{|c |c |c|}
\hline
NSI & Model dependent & Model independent \\
& bound on NSI [Reference \cite{Ohlsson}]& Bound on NSI [Reference \cite{nsi1}]\\
\hline
$\varepsilon_{ee}$ & $> -0.9; < 0.75$ & $ < 4.2$ \\
\hline
$|\varepsilon_{e \mu}|$ & $ \mathrel{\mathpalette\@versim<} 3.8 \times 10^{-4}$ & $ < 0.33$ \\
\hline
$|\varepsilon_{e \tau}|$ & $\mathrel{\mathpalette\@versim<} 0.25$ & $< 3.0$ \\
\hline
$\varepsilon_{\mu \mu}$ & $ > -0.05 ; < 0.08$ & $ < 0.068$ \\
\hline
$|\varepsilon_{\mu \tau}|$ & $ \mathrel{\mathpalette\@versim<} 0.25$ & $ < 0.33$ \\
\hline
$\varepsilon_{\tau \tau}$ & $ \mathrel{\mathpalette\@versim<} 0.4$ & $ < 21 $ \\
\hline
\end{tabular}
\caption{Strength of Non standard interaction terms used for our Analysis.}
\label{table:bound}
\end{table}
Considering recent results from experiments in IceCube-79 and DeepCore more stringent bound on $\varepsilon_{\mu \mu}$,
$|\varepsilon_{\mu \tau}|$ and $\varepsilon_{\tau \tau}$ have been obtained in \cite{Es}.
In section IV, we shall consider both model dependent and independent allowed range of values of different NSIs as shown in the table above for earth like matter while showing discovery reach for $CP$ violation in presence of NSI.
\subsection{\bf $\nu_e \rightarrow \nu_\mu$ oscillation probabilities with NSI}
The flavor eigenstates $\nu_\alpha$ is related to mass eigenstates of neutrinos $\nu_i$ as
\begin{equation}
\nu_\alpha =\sum_{i} U_{\alpha i} \nu_i
\; ;\;\quad
\qquad i=1, 2, 3,
\end{equation}
in vacuum where $U$ is $3 \times 3$ unitary matrix parametrized by
three mixing angles $\theta_{12}$, $\theta_{23}$ and $\theta_{13}$ and by one
$CP$ violating phase $\delta$ for Dirac neutrino as given in \cite{bilen,part}. However, for Majorana neutrino apart from $\delta$ there are two more $CP$ violating phases
which do not play role in neutrino oscillation \cite{bilen}.
In the flavor basis the total Hamiltonian consisting both standard ($H_{SM} $) and non-standard interactions ($H_{NSI}$) of neutrinos interacting with matter during propagation can be written
as:
\begin{eqnarray}
\label{eq:hamil}
H &=& H_{SM} + H_{NSI}
\end{eqnarray}
where
\begin{eqnarray}
\label{eq:H}
H_{SM} = \frac{\Delta m^2_{31}}{2E}\left[U\begin{pmatrix}0 & 0 & 0 \cr 0 & \alpha & 0 \cr
0
& 0 & 1\end{pmatrix} U^{\dag}+\begin{pmatrix}A & 0 & 0 \cr 0 & 0 & 0 \cr
0
& 0 & 0\end{pmatrix} \right], \nonumber \\
\end{eqnarray}
\begin{eqnarray}
\label{eq:H1}
&&H_{\text{NSI}} =
\frac{\Delta m^2_{31}}{2E}\;A \begin{pmatrix}\varepsilon_{e e} & \varepsilon_{e \mu} & \varepsilon_{e \tau} \cr
\varepsilon_{e \mu}^* & \varepsilon_{\mu \mu} & \varepsilon_{\mu \tau} \cr
\varepsilon_{e \tau}^* & \varepsilon_{\mu \tau}^* & \varepsilon_{\tau \tau}\end{pmatrix}
\end{eqnarray}
In equations (\ref{eq:H}) and (\ref{eq:H1})
\begin{eqnarray}
\label{eq:matternsi}
A = \frac{2E\sqrt{2}G_{F}n_{e}}{\Delta m_{31}^2} ; \;
\alpha = \frac{\Delta m^2_{21}}{\Delta m^2_{31}} ; \;
\Delta m^2_{i j} = m^2_i - m^2_j \nonumber \\
\end{eqnarray}
where $m_i$ is the mass of the $i$-th neutrino, $A$ corresponds to the interaction of neutrinos with
matter in SM and $G_{F}$ is the Fermi constant. $\varepsilon_{ee}$, $\varepsilon_{e\mu}$ , $\varepsilon_{e\tau}$, $\varepsilon_{\mu\mu}$, $\varepsilon_{\mu\tau}$
and $\varepsilon_{\tau\tau}$ correspond to the non-standard interactions (NSIs) of neutrinos with matter. In equation (\ref{eq:H1}), ($\; ^{*} \;$)
denotes complex conjugation. The NSIs - $\varepsilon_{e\mu}$, $\varepsilon_{e\tau}$ and $\varepsilon_{\mu\tau}$ could be complex. Later on, in the expressions of probability of oscillation we have
expressed these NSIs as $\varepsilon_{ij}=|\varepsilon_{ij}|e^{i \phi_{ij}}$.
In our numerical analysis we have considered the NSIs - $\varepsilon_{e\mu}$, $\varepsilon_{e\tau}$ and $\varepsilon_{\mu\tau}$ as both real as well as complex.
For longer baselines following the perturbation method adopted in references \cite{m1,m2} for the oscillation probability $P_{\nu_{e}\rightarrow \nu_{\mu}}$ upto order $\alpha^{2}$ (considering $\sin\theta_{13} \sim \sqrt{\alpha}$ as follows from recent reactor experiments) and the matter effect parameter $A$ in the
leading order of perturbation (which happens for longer baselines) and NSI parameters $\varepsilon_{\alpha\beta} $ of the order of $\alpha$ one obtains \cite{m3}
\begin{eqnarray}
\label{eq:prob}
P_{\nu_{e}\rightarrow \nu_\mu} &=& P^{SM}_{\nu_{e}\rightarrow \nu_\mu} + P^{NSI}_{\nu_{e}\rightarrow \nu_\mu}
\end{eqnarray}
where
\begin{widetext}
\begin{eqnarray}
\label{eq:prob1}
&&P^{SM}_{\nu_{e}\rightarrow \nu_\mu} = 4\sin \frac{(A-1)\Delta m^2_{31}L}{4E}\frac{s^2_{13}s^2_{23}}{(A-1)^4}\bigg(((A-1)^2-(1+A)^2s^2_{13})\sin\frac{(A-1)\Delta m^2_{31}L}{4E}\nonumber \\
&+&A(A-1)\frac{\Delta m^2_{31}L}{E}s^2_{13}\cos\frac{(A-1)\Delta m^2_{31}L}{4E}\bigg)
+ \frac{\alpha^2c^2_{23}}{A^2}\sin^2 2\theta_{12}\sin^2\frac{\Delta m^2_{31}AL}{4E}\nonumber \\
&+&\frac{\alpha s^2_{12}s^2_{13}s^2_{23}}{(A-1)^3}\bigg(\frac{(A-1)\Delta m^2_{31}L}{E}\sin \frac{(A-1)\Delta m^2_{31}L}{2E}-8A\sin^2 \frac{(A-1)\Delta m^2_{31}L}{4E}\bigg)\nonumber \\
&+& \frac{\alpha s_{13}s_{2 \times 12}s_{2 \times 23}}{A(A-1)}\bigg(2\cos\bigg(\delta - \frac{\Delta m^2_{31}L}{4E}\bigg)\sin \frac{(A-1)\Delta m^2_{31}L}{4E}\sin\frac{A\Delta m^2_{31}L}{4E}\bigg)
\end{eqnarray}
\end{widetext}
\begin{widetext}
\begin{eqnarray}
\label{eq:pro2}
P^{NSI}_{\nu_{e}\rightarrow \nu_\mu} &&= \frac{4|a_2|s_{2\times 23}s_{13}}{A(A-1)}\sin \frac{A\Delta m^2_{31}L}{4E}\sin\frac{(A-1)\Delta m^2_{31}L}{4E}\cos \big(\delta - \frac{\Delta m^2_{31}L}{4E}+\phi_{a_2}\big)\nonumber \\
&+& \frac{4|a_3|s^2_{23}}{(A-1)^2}\sin^2 \frac{(A-1)\Delta m^2_{31}L}{4E}(|a_3|+2\cos(\delta + \phi_{a_3})s_{13})\nonumber \\
&+&\frac{s^2_{13}s^2_{23}(|a_5|-|a_1|)}{(A-1)^3E}\bigg(8E\sin^2 \frac{(A-1)\Delta m^2_{31}L}{4E}-(A-1)\Delta m^2_{31}L\sin \frac{(A-1)\Delta m^2_{31}L}{2E}\bigg)\nonumber \\
&+&\frac{4 |a_2|c_{23}}{(A-1)A^2}\sin \frac{A\Delta m^2_{31}L}{4E}\bigg((A-1)c_{23}\sin \frac{A\Delta m^2_{31}L}{4E}(|a_2|+\alpha\cos \phi_{a_2}\sin 2\theta_{12})\bigg)\nonumber \\
&-&\frac{4|a_2| |a_3|\sin 2\theta_{23}}{A(A-1)}\cos\bigg[\frac{\Delta m^2_{31}L}{4E}-\phi _{a_2}+\phi_{a_3}\bigg]\sin \frac{(1-A)\Delta m^2_{31}L}{4E}\sin \frac{A\Delta m^2_{31}L}{4E}\nonumber \\
&+& \frac{4|a_3|s_{23}}{(A-1)^2A}\sin \frac{(A-1)\Delta m^2_{31}L}{4E}(A-1)\alpha c_{23}\cos\bigg[\frac{\Delta m^2_{31}L}{4E}-\phi _{a_3}\bigg]\sin \frac{A\Delta m^2_{31}L}{4E}\sin 2\theta_{12}\nonumber \\
&+&\frac{|a_4|s^2_{13}\sin 2 \theta_{23}}{(A-1)^2A}\sin \frac{(A-1)\Delta m^2_{31}L}{4E}\bigg(-4A\cos\frac{A\Delta m^2_{31}L}{4E} \cos \phi_{a_4}\sin \frac{\Delta m^2_{31}L}{4E}\nonumber\\
&+&4\sin \frac{A\Delta m^2_{31}L}{4E}\bigg(\cos\frac{\Delta m^2_{31}L}{4E}\cos \phi_{a_4}+(A-1)\sin\frac{\Delta m^2_{31}L}{4E}\sin\phi_{a_4}\bigg)\bigg)
\end{eqnarray}
\end{widetext}
\begin{comment}
\begin{widetext}
\begin{eqnarray}
&&P_{\nu_\mu \rightarrow \nu_\mu} = 1 - s^{2}_{2\times 23} \sin^2\Big(\frac{L\Delta m^{2}_{31}}{4E}\Big) + \alpha c^{2}_{12}s^{2}_{2\times 23}\Big(\frac{L\Delta m^{2}_{31}}{4E}\Big)\sin \Big(\frac{L\Delta m^{2}_{31}}{2E}\Big)
+ 4\frac{s^{2}_{13}s^{2}_{23}}{(A-1)^2}\Big[A(A-1)\Big(\frac{L\Delta m^{2}_{31}}{4E}\Big)
\nonumber \\
&\times & c^{2}_{23}\sin \Big(\frac{L\Delta m^{2}_{31}}{2E}\Big)
+\cos \Big(\frac{AL\Delta m^{2}_{31}}{4E}\Big)\sin \Big(\frac{L\Delta m^{2}_{31}}{4E}\Big) + c_{2\times 23}\sin \Big(\frac{AL\Delta m^{2}_{31}}{4E}\Big)\Big]
+ (|a_6| - |a_5|)
\Big(\frac{L\Delta m^{2}_{31}}{4E}\Big)\nonumber \\
&\times &
s^{2}_{2\times 23}\sin \Big(\frac{L\Delta m^{2}_{31}}{2E} \Big) -
2|a_4| \cos \phi_{a_4} s_{4\times 23}\sin \Big(\frac{L\Delta m^{2}_{31}}{4E}\Big)
\end{eqnarray}
\end{widetext}
\end{comment}
where
\begin{eqnarray}
&&a_1 = A \varepsilon_{e e} \nonumber \\
&&|a_2|e^{i \phi_{a_2}} = A \bigg(e^{i\phi_{e\mu}}|\varepsilon_{e\mu}|c_{23}-e^{i\phi_{e\tau}}|\varepsilon_{e\tau}|s_{23}\bigg) \nonumber \\
&&|a_3|e^{i \phi_{a_3}} = A \bigg(e^{i\phi_{e\tau}}|\varepsilon_{e\tau}|c_{23}+e^{i\phi_{e\mu}}|\varepsilon_{e\mu}|s_{23}\bigg) \nonumber \\
&&|a_4|e^{i \phi_{a_4}} = A \bigg(|\varepsilon_{\mu \tau}|e^{i\phi_{\mu \tau}} - 2 |\varepsilon_{\mu \tau}|s^2_{23} + (\varepsilon_{\mu \mu}-\varepsilon_{\tau \tau})c_{23}s_{23}\bigg) \nonumber \\
&&a_5 = A \bigg( \varepsilon_{\tau \tau} c^2_{23} + \varepsilon_{\mu \mu} s^2_{23} + |\varepsilon_{\mu \tau}| \cos \phi_{\mu \tau}s_{2 \times 23}\bigg) \nonumber \\
\end{eqnarray}
and
\begin{eqnarray}
\phi_{a_2}&=&\tan^{-1}\left[\frac{\vert\varepsilon_{e \mu}\vert c_{23} \sin \phi_{e\mu} -\vert\varepsilon_{e \tau}\vert s_{23} \sin \phi_{e \tau}}{\vert\varepsilon_{e \mu}\vert c_{23} \cos \phi_{e\mu}]-\vert\varepsilon_{e \tau}\vert \cos \phi_{e \tau}] s_{23} }\right]\nonumber \\
\phi_{a_3}&=&\tan^{-1}\left[\frac{\vert\varepsilon_{e\mu}\vert s_{23} \sin \phi_{e \mu}+\vert\varepsilon_{e \tau}\vert c_{23} \sin \phi_{e \tau}}{\vert\varepsilon_{e \tau}\vert c_{23} \cos \phi_{e \tau} +\vert\varepsilon_{e\mu}\vert \cos \phi_{e \mu} s_{23} }\right] \nonumber \\
\phi_{a_4} &=& \tan^{-1}\left( \frac{|\varepsilon_{\mu \tau}|\sin(\phi_{\mu \tau})}{|\varepsilon_{\mu \tau}|c_{2\times 23}\cos(\phi_{\mu \tau})+(\varepsilon_{\mu \mu} - \varepsilon_{\tau \tau})c_{23} s_{23}}\right) \nonumber \\
\end{eqnarray}
where $s_{ij}=\sin \theta_{ij}$, $c_{ij}=\cos \theta_{ij}$, $s_{2 \times ij}=\sin 2\theta_{ij}$, $c_{2 \times ij}=\cos 2\theta_{ij}$.
The oscillation channel $ \nu_e \rightarrow \nu_\mu $ is important for neutrino factory. For superbeam, $\nu_\mu \rightarrow \nu_e $
oscillation channel is important which can be obtained by the following transformation
\begin{equation}
P_{{\alpha}{\beta}} (A, \delta, \phi_{ij})= P_{ \beta \alpha}(A, -\delta, -\phi_{ij}).
\label{nufacttosb}
\end{equation}
The oscillation probabilities for antineutrinos can be obtained from the oscillation probabilities given for neutrinos above
by using the following relation:
\begin{equation}
P_{\bar{\alpha}\bar{\beta}}(\delta ,\; {A}, \; \phi_{ij} )= P_{\alpha \beta}(-\delta ,\; {-A}, \; -\phi_{ij} ).
\label{probsd}
\end{equation}
To estimate the order of magnitude of $\delta $ dependent and $\delta $ independent but matter dependent ( i.e., $A $ dependent) part in the above two oscillation probability,
following results of reactor experiments we shall consider $\sin\theta_{13} \sim \sqrt{\alpha}$. For only SM interactions, (i.e $\varepsilon_{\alpha\beta} \rightarrow 0$) from last term of Eq.\eqref{eq:prob1} for oscillation probabilities one finds that the
$\delta$ dependence occurs at order of $ \alpha s_{13} \sim \alpha^{3/2}$ for both neutrino oscillation and anti-neutrino oscillation probabilities.
In presence of NSI of order $\alpha$, in the first two leading terms of Eq.\eqref{eq:pro2}, the $\delta$ dependence occurs at order of $ a_2 s_{13} \sim a_3 s_{13} \sim \alpha^{3/2}$ through terms containing $a_2$ and $a_3$ (which are $\varepsilon_{e\mu}$ and $\varepsilon_{e\tau}$ dependent).
Due to this (if we assume all NSI of same order $\sim \alpha $), these two NSI ($\varepsilon_{e\mu}$ and $\varepsilon_{e\tau}$) in contrast to other NSI could have greater effect on oscillation probability in $\nu_e \rightarrow \nu_\mu$ or the reverse oscillation
channel .
\begin{comment}
We have checked that for slightly higher NSIs of the order of $ \sqrt{\alpha}$ using perturbation method the
same $\delta$ dependent terms appear with $a_2$ and $a_3$ in the oscillation probability for long baseline as given in Eq.\eqref{eq:pro2}
and this slightly higher NSI could make these terms at the order of $\alpha$ which could compete with the $\delta$ independent but matter
dependent part (which could mimic $CP$ violation) for long baseline as that could be also at the order of $\alpha$. So presence of slightly
higher NSIs of order $\sqrt{\alpha}$ present in $a_2$ and $a_3$ could improve the discovery reach of $CP$ violation for longer baseline.
These remarks are based on one oscillation channel. However, the detailed numerical analysis of $CP$ violation discovery in our work also involves consideration of other different neutrino oscillation channels as mentioned later.
\end{comment}
For shorter baselines the above oscillation probability expression becomes much simpler. Following the perturbation method \cite{ra} and by considering the standard model matter effect $A \sim \alpha$ for neutrino energy $E$ around 1 GeV
and $\sin\theta_{13} \sim \sqrt{\alpha}$ as obtained from reactor oscillation data, the probability of oscillation $P_{\nu_e \rightarrow \nu_{\mu}}$ upto order $\alpha^{2}$ is given by
\begin{widetext}
\begin{eqnarray}
P_{\nu_e \rightarrow \nu_{\mu}} = P_{\nu_e \rightarrow \nu_{\mu}}(\alpha) + P_{\nu_e \rightarrow \nu_{\mu}}(\alpha^{3/2}) + P_{\nu_e \rightarrow \nu_{\mu}}(\alpha^2)
\label{eq:p250}
\end{eqnarray}
\end{widetext}
where
\begin{widetext}
\begin{eqnarray}
P_{\nu_e \rightarrow \nu_{\mu}}(\alpha)&=& \sin^2 2\theta_{13}\sin^2 \theta_{23}\sin^2\bigg(\frac{\Delta m^2_{31}L}{4E}\bigg) \nonumber \\
P_{\nu_e \rightarrow \nu_{\mu}}(\alpha^{3/2}) &= &\alpha\bigg(\frac{\Delta m^2_{31}L}{2E}\bigg) \sin \theta_{13}\sin 2\theta_{23}\sin 2 \theta_{12}\sin \bigg(\frac{\Delta m^2_{31}L}{4E}\bigg)
\cos \bigg(\delta - \frac{\Delta m^2_{31}L}{4E}\bigg) \nonumber \\
P_{\nu_e \rightarrow \nu_{\mu}}(\alpha^2)&=& \alpha^2 \cos^2\theta_{23}\sin^2 2\theta_{12}\bigg(\frac{\Delta m^2_{31}L }{4E}\bigg)^2-2\alpha \sin^2 \theta_{13}\sin^2 \theta_{12}\sin^2 \theta_{23}\frac{\Delta m^2_{31}L}{2E}
\sin \frac{\Delta m ^2_{31}L}{2E} \nonumber \\ &+& 8A\sin^2 \theta_{13}\sin^2 \theta_{23}\bigg(\sin^2 \frac{\Delta m ^2_{31}L}{4E}-\frac{\Delta m ^2_{31}L}{8E} \sin \frac{\Delta m ^2_{31}L}{2E}\bigg)
\label{eq:p250}
\end{eqnarray}
\end{widetext}
One may note that this expression of oscillation probability is little bit different from that presented by earlier authors \cite{akh} because they have considered the perturbative approach for relatively longer baseline and small $\sin\theta_{13} \sim \alpha$.
For shorter baseline as presented here, the $\delta$ dependence appears to be more and instead of being at the order of $\alpha^2$ it appears at the order $\alpha^{3/2}$ as $ \alpha \sin\theta_{13} \sim {\alpha}^{3/2}$ although the expression is same. However at order $\alpha^2$ some extra terms appear in comparison to that presented by earlier authors although their contribution is relatively smaller being at the order of $\alpha^2$.
The matter effect is occurring at order $\alpha^2$ through term containing parameter $A$.
\section{Numerical simulation}
\label{sec:sim}
There is detailed global analysis of three flavor neutrino oscillation data \cite{Fogli:2012ua} presented at the \emph{Neutrino 2012} conference. In their analysis the correlation between various oscillation parameters has been taken into account.
Very recently there is another such global analysis \cite{Gonzalez-Garcia:2014bfa} presented at the \emph{Neutrino 2014} conference which
has been considered in our analysis and best fit values of neutrino mixing parameters and their respective errors at $3\sigma$ confidence level are shown
in table
\ref{table:mix}.
For earth matter
density the PREM profile \cite{prem} has been considered. Also we have considered an error of $2\%$ on matter density profile.
\begin{table}[ht]
\centering
\begin{tabular}{|c |c |c |}
\hline
Oscillation Parameters & Central Values & Error ($3\sigma$) \\
\hline
$\frac{\Delta m^2_{21}}{10^{-5}} \; \textrm{eV}^2$ & $7.45 $ & 7.02 - 8.09 \\
\hline
\multirow{2}{*}{$\frac{\Delta m^2_{31}}{10^{-3}} \; \textrm{eV}^2$} &\multirow{2}{*}{ $2.457$} & 2.325 - 2.599 (NH) \\
& & (-2.59)- (-2.307) (IH)\\
\hline
$\sin^2 \theta_{12}$ & 0.304 & 0.270 - 0.344 \\
\hline
$ \sin^2 \theta_{23}$ & 0.452 & 0.385 - 0.644 \\
\hline
$\sin^2 \theta_{13}$ & 0.0218 & 0.0188 - 0.0251 \\
\hline
\end{tabular}
\caption{Central values of the oscillation parameters with errors.}
\label{table:mix}
\end{table}
The numerical simulation has been done by using GLoBES \cite{globes1,globes11}. GLoBES uses poissonian $\chi^2$ for the
oscillation parameters ${\bf \lambda}$ and for nuisance parameters $\xi_i$. To implement the systematic errors the
`pull method' \cite{pull} has been used by GLoBES and $\chi^2_{pull}$ is given as \cite{globes11}
\begin{align}
\chi^2_{pull}({\bf \lambda}) &:= \min_{\{\xi_i\}}\bigg(\chi^2({\bf \lambda},\xi_1,\cdots,\xi_k)+\sum_{j=1}^k\frac{\xi^2_j}{\sigma^2_{\xi_j}}\bigg)
\end{align}
where $\chi^2({\bf \lambda},\xi_1,\cdots,\xi_n)$ is the usual Poissonian $\chi^2$ depending on the neutrino oscillation parameters
$\lambda$ and the nuisance parameter $\xi_i$ and it has been summed over different energy binning on the theoretical and observed event rates. The oscillation parameters ${\bf \lambda}$ is
\begin{align}
{\bf \lambda} &= (\theta_{12},\theta_{13},\theta_{23},\Delta m^2_{21},\Delta m^2_{31},\rho,\delta_{CP},\varepsilon_{\alpha \beta},\phi_{\alpha \beta})\; ;
\end{align}
The nuisance parameters $\xi_i$ are the signal and background normalization and the calibration errors. In order to implement the error $\sigma_{\xi_i}$, the $\chi^2$ (in which all
event rates of all bins are included) is minimized over the different nuisance parameters $\xi_i$ independently.
\begin{comment}
\begin{align}
\chi^2({\bf \lambda},\xi_1,\cdots,\xi_n) &= \sum_{i=1}^{\# of bins} 2\bigg((1+\sum_{j=1}^n\xi_j)T_i({\bf \lambda}) - O_i({\bf \lambda}) \log\frac{O_i({\bf \lambda}) }{(1+\sum_{j=1}^n\xi_j)T_i({\bf \lambda}) } \bigg) + \sum_{j=1}^n\frac{\xi_j^2}{\sigma^2_{\xi_{j}}}\; ;
\end{align}
the oscillation parameters ${\bf \lambda}$ is
\begin{align}
{\bf \lambda} &= (\theta_{12},\theta_{13},\theta_{23},\delta_{CP},\Delta m^2_{12},\Delta m^2_{13},\rho,\varepsilon^m_{\alpha \beta},\phi^m_{\alpha \beta})\; ;
\end{align}
$O_i$ are the observed rates (corresponding to the true values) and $T_i$ are the theoretical fit (corresponding to test
values).
\end{comment}
We have also marginalized over ${\bf \lambda^\prime}$ where
\begin{align}
{\bf \lambda^\prime} &= (\theta_{12},\theta_{13},\theta_{23},\Delta m^2_{21},\Delta m^2_{31},\rho)\; .
\end{align}
The final projected $\chi_F^2$ is given as
\begin{align}
\chi_F^2 &= \min_{\lambda^\prime} \left( \chi^2_{pull}(\lambda) + priors (\lambda^\prime)\right)
\end{align}
where
\begin{align}
priors(\lambda^\prime) &= \bigg(\frac{\rho - \rho^0}{\sigma_{\rho}}\bigg)^2 + \sum_{i\neq j} \bigg(\frac{\sin^2\theta_{ij} - \sin^2\theta_{ij}^0}{\sigma_{\sin^2\theta_{ij}}}\bigg)^2 + \bigg(\frac{\Delta m^2_{21} - (\Delta m^{2}_{21})^0}{\sigma_{\Delta m^2_{21}}}\bigg)^2 + \bigg(\frac{|\Delta m^2_{31}| - | (\Delta m^{2}_{31})^0|}{\sigma_{|\Delta m^2_{31}|}}\bigg)^2
\end{align}
and $\rho^0,\; \sin^2\theta^0_{12},\; \sin^2\theta^0_{13},\;\sin^2\theta^0_{23},\; (\Delta m^2_{21})^0 \text{and} | (\Delta m_{31}^2)^0|)$ are the
central values of the corresponding parameters. We have marginalized over two different hierarchies of neutrino masses also.
For taking into account NSI in GLoBES we have followed the method described in \cite{globes11} and modified the source file "probability.c" in GLoBES appropriately
by inserting the NSI's in the subroutine where the hamiltonian for matter interaction is defined. Then we have included the new probability program as instructed by the manual of GLoBES. We have not used the approximate analytical expression of oscillation probability for NSI of order $\alpha$ in our numerical
analysis.
At present we do not know the hierarchy of neutrino masses and the value of $\delta$. In the test values we have considered both
possibilities of the hierarchy. However, we shall show the results in Figures only for true normal hierarchy. Before we get the knowledge of specific value of $\delta$ it could be possible to know
whether nature admits $CP$ violation or not which could be possible for a wide range of values of $\delta$. So what we are going to analyse
is not the measurement of $\delta$ values but finding out all possible $\delta$ (true) values which could be distinguishable from
the $CP$ conserving $\delta $ values at certain confidence level through oscillation experiments.
For that it is useful to calculate the $CP$ fraction due to $\delta $ denoted as $F(\delta )$ which may be defined as the fraction of the total allowed
range (0 to 2 $\pi$) for the $\delta$(true) over which $CP$ violation can be discovered \cite{nsi3,Ballett:2012rz}. While calculating the fraction due to
$\delta$ for the discovery of $CP$ violation, we consider the exclusion of all parameter sets with $\delta \in \{0,\pi \}$ which means
$\delta$(test) is fixed in its $CP$ conserving values $\{0,\pi\}$. For no NSIs, the degrees of freedom in $\chi^2$ analysis is one due to
due to $\delta$.
We have calculated the minimum of those $\chi^2$ for each $\delta$(true) and have calculated the fraction of the $\delta$(true) over its entire range which satisfies $\Delta \chi^2 \geq 9 ( 25) \; \text{at} \; 3 (5) \sigma$ confidence level.
In studying the NSI effect (one at a time) on $CP$ violation we have fixed absolute values of NSI(test) to their true values
which implies that the absolute values of NSI are assumed to be known from some other experiments.
In
Figures we have shown the effect of real NSI's on
the $CP$ fraction discovery. In this case, the degrees of freedom in $\chi^2$ analysis is one due to $\delta$.
The off-diagonal
NSI in NSI matrix could be complex and the corresponding complex phases could be new sources of $CP$ violation. If the $CP$ violation
is only due to NSI phase, then the $CP$ fraction due to NSI phases ${\phi_{ ij}}$ (denoted as $F({\phi_{ ij}})$) is calculated in the same way as described above for $\delta$ by replacing $\delta$ by the corresponding $\phi_{ ij}$. Here the degrees of freedom is one
due to $\phi_{ij}$. In
Figures we have shown the effect of absolute values of non-diagonal NSI's (one at a time) on
$F({\phi_{ ij}})$. In order to calculate the the total $CP$ violation discovery due to two phases $\phi_{ ij}$ and $\delta$ we have
set $\phi_{ ij}$(test) and $\delta$(test) to their respective $CP$ conserving values at $(\delta,\phi_{ij})\in \{(0,0),(0,\pi),(\pi,0),(\pi,\pi)\}$ and also both of the mass hierarchy in the test values have been taken into account. Here, the degrees of freedom is two due to $\delta $ and one of the NSI phases $\phi_{ij}$. Then we have calculated the above $\chi^2$ minimum for each true values of
$\phi_{ ij}$ and $\delta$ and have shown the allowed region for discovery in $\phi_{ ij}$-$\delta$ space for $\Delta \chi^2 \geq 11.83
$ at $3 \sigma$ confidence level.
In all Figures true hierarchy has been considered to be normal.
So by numerical simulation
we are studying the different possibilities of finding $CP$ violation in different experimental set-ups due to $\delta$ or NSI phase $\phi_{ ij}$ or total
$CP$ violation due to $\delta $ and NSI phase $\phi_{ ij}$ for certain absolute value of NSI ($\varepsilon_{ij}$ (true)) assumed to be present in nature. If we assume that any allowed values of $\delta $ has equal probability to be true $\delta$ value of nature then
lower (higher) $CP$
fraction which we present in the result sections will indicate lower (higher) probability of finding $CP$ violation in particular
experimental-set-up discussed below. One may note that NSI effect and the $\delta $ effect in the hamiltonian $H$ has different neutrino energy dependence (see equation Eq.\eqref{eq:H} for $\delta$ present in $U$ and Eq.\eqref{eq:H1} for NSI). Also
in the shorter (longer) baseline there should be lesser (larger) NSI effect. If the experimental data is available over certain range of neutrino energy from the shorter and longer baseline experiments then only the multi-parameter fit
will help in disentangling effects due to different unknown parameters and hence improving bounds on NSI parameters or
discovering it.
In the following we present three subsections based on three different sources of neutrino flux : Superbeam in \ref{sec:sb},
Neutrino Factory in \ref{sec:nu} and monoenergetic neutrino beam in \ref{sec:mono}. Particularly for Superbeam and
Neutrino factory, we have studied the NSI effect
in $CP$ violation discovery in the leptonic sector in neutrino oscillation experiments. Monoenergetic neutrino beam requires shorter baseline due to technical reasons as discussed later. Due to that the NSI effect in the monoenergetic neutrino beam studies is insignificant and we will be discussing only $CP$ violation discovery reach in absence of NSI. Two different nuclei have been considered for
monoenergetic beam and to study $CP$ violation.
\subsection{\bf Superbeam}
\label{sec:sb}
We discuss below the experimental set-up and systematic errors for superbeam. We have presented our results for discovery
reach of $CP$ violation
due to $\delta $ and other off-diagonal NSI phases $\phi_{ij}$ for different experimental set-up with
different baselines. We have shown also the effect of the absolute values of different $\varepsilon_{\alpha\beta}$ in $CP$ fraction.
\subsubsection{\bf Experimental set-ups and systemetic errors}
We consider three experimental set-ups:
(a) A Superbeam set-up originating in CERN and reaching a 500 kt Water Cherenkov detector
\cite{Agostino:2012fd} placed at a distance of 130 km at Fr\'ejus (France), (b) A Superbeam set-up originating in Tokai and reaching a 500 kt Water Cherenkov detector \cite{Huber:2002mx,Itow,Ishitsuka:2005qi} placed at a distance of 295 km at Kamiokande (Japan) and (c) A Superbeam set-up which originates in CERN and reaches a 100 kt Liquid Argon detector placed at a distance of 2300 km at Pyh\"asalmi (Finland).
The observable channels that we have considered are
$\nu_{\mu}\rightarrow \nu_e$, $\bar{\nu}_{\mu}\rightarrow \bar{\nu}_e$, $\nu_{\mu}\rightarrow \nu_{\mu}$ and $\bar{\nu}_{\mu}\rightarrow \bar{\nu}_{\mu} $.
The flux considered for set-up (a) has mean energy $\sim 0.3$ GeV, which are computed for 3.5 GeV protons and $10^{23}$ protons on target per year. For our analysis the beam power has been considered of about 4 MW per year and the time period has been taken to be 2 years for neutrinos and 4 years for anti-neutrinos.
We consider the same flux as in \cite{others5,splflux}.
In the case of set-up(a) the efficiencies for the signal and background are
included in the migration matrices based on \cite{Agostino:2012fd} except for the channels $\nu_\mu$ disappearance, $\bar{\nu}_\mu$
disappearance and $\nu_\mu$ (NC) which are $64\%$, $81\%$ and $11.7\%$ efficiencies respectively. We have considered systematic uncertainties of $2\%$ on signal and background channels.
The experimental set-up (b) has been considered as given in GLoBES \cite{hkflux} and
the detector specification is
mentioned in table \ref{table:hk}:
\begin{table}[h!]
\centering
\begin{tabular}{|c|c|}
\hline
Target Power & 4 Mega-Watt \\
\hline
Fiducial Mass & 500 kt \\
\hline
Data Taking & 4 yrs $\nu$ and 4 yrs $\overline{\nu}$ \\
\hline
Baseline & 295 km \\
\hline
Energy Resolution & 0.085 \\
\hline
\end{tabular}
\caption{Detector characteristics for Hyper-Kamiokande.}
\label{table:hk}
\end{table}
The flux considered for set-up (c) has mean energy $\sim 5$ GeV, which are computed for 50 GeV protons and $3\times 10^{21}$ protons
on target per year. For our analysis the beam power is about 0.8 MW per year and the time period has been taken to be 5 years each for neutrinos and anti-neutrinos.
We consider the same flux as in \cite{Coloma:2012ma}. The detector characteristics for set-up (c) is given in table \ref{table:charac} \cite{Adams:2013qkq}.
The correlation between the visible energy of background NC events and the neutrino energy is
implemented by migration matrices which has been provided by L. Whitehead \cite{white}.
\begin{table}[ht]
\centering
\begin{tabular}{|c |c |c |c|}
\hline
\hline
Signal Studies& $\nu_e$ CC appearance Studies & $\nu_\mu$ CC Disappearance Studies \\
\hline
Signal efficiency & 80\% & 85\% \\
\hline
$\nu_\mu$ NC mis-identification rate (Background) & 1\% & 0.5\% \\
\hline
$\nu_\mu$ CC mis-identification rate (Background) & 1\% & 0\% \\
\hline
Signal Normalization error & 5\% & 10\% \\
\hline
Background Normalization Error & 15\% & 20\% \\
\hline
\multicolumn{3}{|c|}{Neutrino Energy Resolution} \\
\hline
\hline
$\nu_e$ CC energy resolution & \multicolumn{2}{|c|}{0.15$\sqrt{E(\textrm{GeV})}$} \\
\hline
$\nu_\mu$ CC energy resolution & \multicolumn{2}{|c|}{0.2$\sqrt{E(\textrm{GeV})}$} \\
\hline
$E_{\nu_\mu}$ scale uncertainty & \multicolumn{2}{|c|}{2\%} \\
\hline
$E_{\nu_e}$ scale uncertainty & \multicolumn{2}{|c|}{0.01\%} \\
\hline
\end{tabular}
\caption{Detector characteristics for set-up (c).}
\label{table:charac}
\end{table}
\begin{table}
\begin{tabular}{|c|c|c|c|}
\hline
\hline
& {CERN-Fr\'ejus } & {T2HK } & {CERN-Pyh\"asalmi } \\
&(130 km) &(295 km) & (2300 km) \\
\hline
SM & 4533 & 6098 & 421 \\
\hline
$\varepsilon_{ee}$ & 4536 & 6110 & 426 \\
\hline
$\varepsilon_{e\mu}$ &4634 & 6272 & 502 \\
\hline
$\varepsilon_{e\tau}$ & 4543 & 6143 & 470 \\
\hline
$\varepsilon_{\mu\tau}$ & 4529 & 6081 & 414 \\
\hline
$\varepsilon_{\mu\mu}$ & 4533 & 6099 & 425 \\
\hline
$\varepsilon_{\tau\tau}$ & 4530 & 6085 & 413 \\
\hline
\end{tabular}
\caption{The number of $\nu_{\mu}\rightarrow \nu_e$ events for CERN-Fr\'ejus (130 km), T2HK (295 km) and CERN-Pyh\"asalmi (2300 km) for SM and for different NSI (one at a time each of which equals to $\alpha$).}
\label{table:eventssb}
\end{table}
In table \ref{table:eventssb} for $\delta =0$ we have shown the expected number of events for three experimental set-ups, for no NSIs and also for different
real NSIs , each of which equal to $\alpha$ (one NSI at a time). We have considered the central values of various parameters
as shown in table \ref{table:mix} and have used PREM profile \cite{prem} for matter densities.
One can see that there is significant effect of $\varepsilon_{e\mu}$ and $\varepsilon_{e\tau}$ on the number of $\nu_{\mu}\rightarrow \nu_e$ events.
Possibilities of significant effect of these NSIs can be seen in the expression of oscillation probabilities in Eq.\eqref{eq:pro2}.
\subsubsection{\bf Results}
\begin{figure}[h!]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.5\textwidth]{dn.pdf}
\end{tabular}
\caption[] {{\small Discovery reach of $CP$ violation due to $\delta$ for three different experimental set-up : CERN-Fr\'ejus, T2HK and CERN-Pyh\"asalmi set-up considering only SM interactions for normal hierarchy. }}
\label{fig:delta1}
\end{figure}
\begin{figure}[htb]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{eed_super.pdf}&
\includegraphics[width=0.45\textwidth]{emd_super.pdf}\\
\includegraphics[width=0.45\textwidth]{etd_super.pdf}&
\includegraphics[width=0.45\textwidth]{mmd_super.pdf}\\
\includegraphics[width=0.45\textwidth]{mtd_super.pdf}&
\includegraphics[width=0.45\textwidth]{ttd_super.pdf}
\end{tabular}
\caption[] {{\small $F(\delta)$ for three different experimental set-up: CERN-Fr\'ejus, T2HK and CERN-Pyh\"asalmi set-up at $3\sigma$ considering real NSIs $\varepsilon_{ee}$, $\varepsilon_{e\mu}$ , $\varepsilon_{e\tau}$, $\varepsilon_{\mu\mu}$, $\varepsilon_{\mu\tau}$ and $\varepsilon_{\tau\tau}$. }}
\label{fig:delta2}
\end{figure}
\begin{figure*}[htb]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{emph_super.pdf}&
\includegraphics[width=0.45\textwidth]{etph_super.pdf}
\end{tabular}
\includegraphics[width=0.45\textwidth]{mtph_super.pdf}
\caption[] {{\small $F({\phi_{e\mu}})$, $F({\phi_{e\tau}})$ and $F({\phi_{\mu\tau}})$ (one NSI at a time) for $\delta_{CP}=0$ at 3$\sigma$ confidence level for three different experimental set-up. }}
\label{fig:sbnsid0}
\end{figure*}
\begin{figure}[htb]
\centering
\begin{tabular}{ccc}
\includegraphics[width=0.35\textwidth]{300emphnd.pdf}&
\includegraphics[width=0.35\textwidth]{300etphnd.pdf}&
\includegraphics[width=0.35\textwidth]{300mtphnd.pdf}\\
\includegraphics[width=0.35\textwidth]{2300emphnd.pdf}&
\includegraphics[width=0.35\textwidth]{2300etphnd.pdf}&
\includegraphics[width=0.35\textwidth]{2300mtphnd.pdf}
\end{tabular}
\caption[] {{\small Allowed region (unshaded) for total $CP$ violation discovery reach $F(\delta)$ for two different baselines T2HK (top panel) and CERN-Pyh\"asalmi set-up (bottom panel) considering NSI phases $\phi_{e\mu}$, $\phi_{e\tau}$ and $\phi_{\mu\tau}$.
}}
\label{fig:sbnsiphase}
\end{figure}
Here we first discuss the $CP$ violation discovery reach without any NSI for the three experimental set-ups (a), (b) and (c).
Then we make a comparative study between no NSI and with real NSI. We also consider off-diagonal NSIs with phases.
The Figures are given only for normal hierarchy (true). However, there is no significant departures in the results for inverted hierarchy from that
with normal hierarchy.
In Figure \ref{fig:delta1} we have shown the $\Delta \chi^2$ values versus $\delta$(true) for SM interactions of neutrinos with matter from which the discovery
reach can be obtained at different confidence levels. We have fixed $\delta$(test) to 0 and $\pi$ and have marginalized over hierarchy for every $\delta$(true) value. To calculate $CP$ fraction from Figure \ref{fig:delta1} at $3 \sigma $ one is required to find the fraction of the total allowed range of $\delta $ for which $\Delta \chi^2$ values are above the solid horizontal line in Figure \ref{fig:delta1}. Particularly at $3 \sigma $ confidence level for CERN-Fr\'ejus set-up the $CP$ violation could be discovered
with $F(\delta)$ of around 0.58 of the possible $\delta$ values for normal hierarchy (true value) whereas for
T2HK set-up these values are around 0.67 and for CERN-Pyh\"asalmi set-up these values are about 0.21. For longer baseline for normal hierarchy the discovery reach is little bit better than the inverted hierarchy (not shown in the Figure). The discovery reach for longer baseline of 2300 km was shown earlier by Coloma {\it et al} \cite{Coloma:2012ma}. So with only SM interactions of neutrinos with matter the short baseline like T2HK set-up seems to be better for good discovery reach of $CP$ violation. This was observed earlier by different authors \cite{Agarwalla:2011hh,others4,Coloma:2012ma,Coloma:2012ut,betabeam}.
\paragraph{\bf Possibilities of Discovery of $CP$ violation due to $\delta$ for real NSI:}
If we compare the $F (\delta)$ obtained from Figure \ref{fig:delta1} at $3 \sigma $ with that shown in Figure \ref{fig:delta2} at $3 \sigma $ we find that for smaller values of NSIs satisfying model dependent bound for short baseline of CERN-Fr\'ejus and T2HK set-up , there is
insignificant effect of NSI on the the discovery reach of $CP$ violation due to $\delta$. However, for CERN-Pyh\"asalmi set-up even
for smaller NSI values one finds the effect on $CP$ violation discovery reach.
One interesting feature is found from these figures which is that if one does not observe $CP$ violation due to $\delta $ in shorter baselines (say CERN-Fr\'ejus set-up) there is still a possibility of observing $CP$ violation
in longer baselines (see CERN-Pyh\"asalmi set-up with off-diagonal NSI like $\varepsilon_{e\mu}$ and $\varepsilon_{e\tau}$) which could be the signal of NSI with significant strength. In Figure \ref{fig:delta2} one can see that except $\varepsilon_{\mu\mu}$ for other NSIs the $F(\delta )$ changes considerably, with increase
in the values of different NSIs for CERN-Pyh\"asalmi set-up.
\paragraph{\bf Possibilities of Discovery of $CP$ violation due to complex NSI phases :}
There could be $CP$ violation only due to off-diagonal NSI phases for which $\delta =0$. Such cases have been shown in Figures
\ref{fig:sbnsid0}. We observe that for certain NSI value the $CP$ fraction for longer baselines is more in comparison to that for the shorter baselines. With the increase of $|\varepsilon_{\alpha \beta}|$ there
is increase in discovery of $CP$ fraction in general. If $CP$ violation is not observed in short baselines (say CERN-Fr\'ejus) but it is observed at relatively longer baselines (say CERN-Pyh\"asalmi set-up) then it would signal the presence of NSI as can be seen from Figure \ref{fig:sbnsid0}.
\paragraph{\bf Possibilities of Discovery of $CP$ violation due to both $\delta$ and complex NSI phases :}
In Figure \ref{fig:sbnsiphase} in the un-shaded region discovery of total $CP$ violation could be possible at $3 \sigma$ confidence level.
Here, we are assuming that nature has two sources of $CP$ violation - one due to $\delta $ phase and other due to one NSI phase. For this we ask whether there is ($\delta - \phi_{ij}$) parameter space where both can be distinguishable from their $CP$ conserving points as discussed earlier.
From the Figure it is seen that for T2HK set-up there is more allowed region and as such better discovery reach in the $\delta - \phi_{ij}$ parameter space particularly for NSI - $\varepsilon_{e\mu}$ in comparison to CERN-Pyh\"asalmi set-up . Particularly for NSI - $\varepsilon_{e\tau}$ and $\varepsilon_{\mu\tau}$, the discovery reach of total $CP$ violation is better for both longer and shorter baselines (like CERN-Pyh\"asalmi set-up and T2HK set-up). Simultaneous consideration of experimental data from both these set-ups could improve the discovery reach further because of presence of non-overlapping un-shaded discovery regions in the
upper panel two right Figures and lower panel two right Figures corresponding to two different NSIs.
From Figure \ref{fig:delta1} it is observed that for certain range of values of $\delta $ (about $0^\circ \mbox{to} \;25^\circ$, $160^\circ \mbox{to} \;208^\circ$, $340^\circ \mbox{to} \; 360^\circ$), the $CP$ violation due to $\delta $ could remain unobservable.
However, even if nature has such values of $\delta $, from Figure \ref{fig:sbnsiphase} one can see that total $CP$ violation could be observed in such cases in presence of
NSI for a wide range of values of different NSI phases. However, disentangling the sources of two types of $CP$ violating phases (one from NSI phase and other from $\delta$) could be difficult. Here the total $CP$ fraction can be thought of as the ratio of the un-shaded region divided by the total region covered by $\delta - \phi_{ij}$ parameter space. As for example, in the upper panel extreme left Figure
such total $CP$ fraction is about $75 \%$. For both the set-ups for $\varepsilon_{e\tau}$ such $CP$ fraction is even more. For CERN-Pyh\"asalmi set-up for $\varepsilon_{e\mu}$ the total $CP$ fraction is not good.
\subsection{\bf Neutrino factory}
\label{sec:nu}
We discuss below the experimental set-up and systematic errors for neutrino factory. We have presented our results for discovery
reach of $CP$ violation
for different baselines of length 730 km (FNAL-Soudan),1290 km (FNAL-Homestake) and 1500 km (FNAL-Henderson) due to $\delta $ and other off-diagonal NSI phases $\phi_{ij}$. We have shown the effect of the absolute values of different $\varepsilon_{\alpha\beta}$ in that.
For this analysis we have considered one NSI at a time.
\subsubsection{\bf Experimental set-ups and systemetic errors}
We have used a large magnetised iron neutrino detector(MIND) \cite{mind1} with a toroidal magnetic field having a mass of 100 KTon. MIND can also be described as an iron-scintillator calorimeter. This detector has the capability of excellent reconstruction and
charge detection efficiency. In this section we have considered muons in a storage ring consisting of both $\mu^+$ and $\mu^-$ which decay with energies of 10 GeV. We consider $5\times 10^{21}$ stored muons. The golden channel ($\nu_e \rightarrow \nu_{\mu}$ oscillation channel) where the charged current interactions of the $\nu_{\mu}$ produce muons of the opposite
charge to those stored in the storage ring (generally known as wrong-sign muons), is the most promising channel to explore $CP$ violation at a neutrino factory. The detector that we are considering in this work - MIND is optimized to exploit the golden channel
oscillation as this detector has the capacity to easily identify signal i.e. a muon with a sign opposite to that in the muon storage ring.
Different
oscillation channels which have been considered as signals and backgrounds \cite{mind1} in the analysis are
shown in table \ref{table:channelnufact}.
We have taken the migration matrices for the true and reconstructed neutrino energies as given in reference \cite{mind1}. The signal and background efficiencies are taken into account in those matrices. We have taken into account
the reconstruction of $\tau$ contamination coming from the \emph{silver channel} $\nu_e \rightarrow \nu_\tau$ as background.
We have considered systematic errors to be $1\%$. In this work we have considered a running time of 5 years for both $\mu^+$ and $\mu^-$ .
\begin{table}[ht]
\begin{tabular}{|l|l|l|l|}
\hline
& Channel Name & $\mu^+$ & $\mu^-$ \\ \hline
\multirow{1}{*}{Signal} & Golden Channel & $\nu_e \rightarrow \nu_\mu$ & $\overline{\nu}_e \rightarrow \overline{\nu}_\mu$ \\ \hline
\multirow{5}{*}{Background} & $\nu_e$ disappearance channel & $\nu_\mu \rightarrow \nu_e$ & $\overline{\nu}_\mu \rightarrow \overline{\nu}_{e}$ \\
& Silver Channel & $\nu_e \rightarrow \nu_\tau$ & $\overline{\nu}_e \rightarrow
\overline{\nu}_\tau$ \\
& $\nu_\mu$ disappearance channel & $\overline{\nu}_\mu \rightarrow
\overline{\nu}_\mu$ & $\nu_\mu \rightarrow \nu_\mu$ \\
& Platinum Channel & $\overline{\nu}_\mu \rightarrow
\overline{\nu}_e$& $\nu_\mu \rightarrow \nu_e$ \\
& Dominant Channel& $\overline{\nu}_\mu \rightarrow
\overline{\nu}_\tau$& $\nu_\mu \rightarrow \nu_\tau$\\
\hline
\end{tabular}
\caption{ Different oscillation channels considered as signals and backgrounds in the analysis.}
\label{table:channelnufact}
\end{table}
\begin{table*}
\begin{tabular}{|c|c|c|c|}
\hline
\hline
&730 km & 1290 km & 1500 km \\
\hline
SM & 248491 & 222097 & 210819\\
\hline
$\varepsilon_{ee}$ & 248715 & 222644 & 211527\\
\hline
$\varepsilon_{e\mu}$& 316049 & 283709& 269412 \\
\hline
$\varepsilon_{e\tau}$ & 251331 & 229148 & 220134 \\
\hline
$\varepsilon_{\mu\tau}$& 248049 & 220985 & 209394\\
\hline
$\varepsilon_{\mu\mu}$& 248682 & 222667 & 211597 \\
\hline
$\varepsilon_{\tau\tau}$& 248072 & 220964 & 209313\\
\hline
\end{tabular}
\caption{ Number of ${\nu_{e}\rightarrow\nu_\mu}$ events for 730 km (FNAL-Soudan),1290 km (FNAL-Homestake) and 1500 km (FNAL-Henderson) for SM and for different NSI (one at a time each of which equals to $\alpha$).}
\label{table:eventnufact}
\end{table*}
In table \ref{table:eventnufact} for $\delta =0$ we have shown the expected number of events for three baselines in neutrino factory, for no NSI and also for different
real NSI , each of which equal to $\alpha$ (one NSI at a time). We have considered the central values of various parameters
as shown in table \ref{table:mix} and have used PREM profile \cite{prem} for matter densities. Like superbeam case here also for neutrino factory one can see that there is significant effect of $\varepsilon_{e\mu}$ and $\varepsilon_{e\tau}$ on the number of $\nu_e \rightarrow \nu_{\mu}$ events.
Possibilities of significant effect of these NSIs can be seen in the expression of oscillation probabilities ${\nu_{e}\rightarrow\nu_\mu}$ (which can be obtained using Eq.\eqref{nufacttosb} in Eq.\eqref{eq:pro2}).
\subsubsection{\bf Results}
\begin{figure}[ht]
\centering
\includegraphics[width=0.7\textwidth]{EL.pdf}
\caption[] {{\small $F(\delta)$ with only SM interactions of neutrinos with matter for baselines of different lengths and different muon energies $E_\mu$ at 5$\sigma$ confidence level.}}
\label{fig:elsm}
\end{figure}
The optimization of $CP$ violation discovery reach has been done earlier \cite{Ballett:2012rz,Agarwalla:2010hk} for different baselines and different parent muon energy when only SM interactions of neutrinos with matter during propagation is present.
We have re-analysed this optimization based on the updated MIND detector characteristics
and the migration matrices as given in \cite{mind1}. We present the result for neutrino factory at $5 \sigma $ confidence level unlike superbeam case where we have given the results at $3 \sigma $ confidence level. This is because MIND detector for neutrino factory is found to give quite good $CP$ violation discovery reach. It is found that at 5$\sigma$ confidence level the $CP$ fraction (shown as $F(\delta)$
in Figure \ref{fig:elsm})
of about ($0.8 \leq F(\delta) < 0.9$ ) is possible for baselines ranging from about 300 to 800 km for energies lesser than 5 GeV and for baselines ranging from about 700 to 2000 km for energies 6-10 GeV respectively. Based
on high $CP$ fraction discovery potential as found in this Figure we have chosen 10 GeV muon energy
and a few baselines which are : 730 km(FNAL-Soudan),
1290 km (FNAL-Homestake) and 1500 km (FNAL-Henderson).
\begin{figure*}[h!]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{eed_nufact.pdf}&
\includegraphics[width=0.45\textwidth]{emd_nufact.pdf}\\
\includegraphics[width=0.45\textwidth]{etd_nufact.pdf}&
\includegraphics[width=0.45\textwidth]{mtd_nufact.pdf}\\
\includegraphics[width=0.45\textwidth]{mmd_nufact.pdf}&
\includegraphics[width=0.45\textwidth]{ttd_nufact.pdf}
\end{tabular}
\caption[] {{\small $F(\delta)$ versus real NSI ($\varepsilon_{ij}$) at $5\sigma$ confidence levels.}}
\label{fig:cpfrac}
\end{figure*}
Next we have tried to answer the question that had there been non-Standard
interactions what could have been their effect on the $CP$ violation discovery reach for such experimental set-ups.
While taking into account NSI effect, for off-diagonal NSI we have also considered the effect of NSI phases over $\delta_{CP}$ violation.
In contrast to the results on superbeam we have presented results for neutrino factory at $5\sigma$ confidence level in Figures \ref{fig:cpfrac} and \ref{fig:nsid0} because in most of
the cases the discovery reach of $CP$ violation is far better in neutrino factory set-ups than those in superbeam set-ups particularly
for NSIs $\mathrel{\mathpalette\@versim<} 1$. However, for
comparison of the results of neutrino factory set-ups with that of superbeam set-ups we have considered the $3 \sigma $ results for
both type of set-ups. For brevity we have avoided presenting the results at $3 \sigma $
corresponding to Figures \ref{fig:cpfrac} and \ref{fig:nsid0}.
\paragraph{\bf Possibilities of Discovery of $CP$ violation due to $\delta$ for real NSI:}
In
Figure \ref{fig:cpfrac}, we have studied $F(\delta)$ in the presence of real NSI (NSI phases have been chosen to be zero) for different baselines of length 730 km, 1290 km and 1500 km at $5 \sigma$ confidence level. Here, in plotting the Figures we have considered the NSI values upto the model independent bounds as shown in table \ref{table:bound}. For lower values of NSI there is essentially negligible effect on discovery reach of $CP$ violation which is seen in the Figure as horizontal straight line. This corresponds to $F(\delta)$ due to SM interactions only which can be verified from Figure \ref{fig:elsm} at muon energy of 10 GeV
for the appropriate baseline. For NSI $\varepsilon_{\mu\mu}$ with model dependent bound there is insignificant effect on $F(\delta)$
whereas for other NSIs there are some effect on $F(\delta)$. Except $\varepsilon_{e\mu}, \; \varepsilon_{e\tau}$ and $\varepsilon_{\mu\mu}$
for other NSIs there is considerable decrease in discovery reach of $CP$ violation with the increase of NSI values.
Even if there is $CP$ violating phase $\delta$ but as one
can see that in presence of higher allowed values of NSI - $\varepsilon_{\tau\tau} $ there could be no discovery of $\delta_{CP}$
violation.
The results are shown in Figure \ref{fig:cpfrac} at $5 \sigma $ confidence level. However, at $ 3 \sigma $ also the decreasing features
of $\varepsilon_{ee}$ and $\varepsilon_{\tau\tau}$ $\mathrel{\mathpalette\@versim>} 1$ is present and in comparison to Figure \ref{fig:delta2} it is found that for large NSI values of $\varepsilon_{ee},\varepsilon_{\tau \tau} \mathrel{\mathpalette\@versim>} 1$ superbeam T2HK set-up is found to be better for the discovery reach of $CP$ violation as compared to neutrino factory set-ups.
But for lower values of NSIs particularly $\varepsilon \lesssim 10^{-1}$ the neutrino factory set-ups are found to be much better for the discovery of
$CP$ violation at $3 \sigma $ and even at $5 \sigma $ the $CP$ fraction is more in neutrino factory set-ups.
\begin{figure*}[h!]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\textwidth]{emph_nufact.pdf}&
\includegraphics[width=0.45\textwidth]{etph_nufact.pdf}
\end{tabular}
\includegraphics[width=0.45\textwidth]{mtph_nufact.pdf}
\caption[] {{\small $F({\phi_{e\mu}})$, $F({\phi_{e\tau}})$ and $F({\phi_{\mu\tau}})$ for $\delta_{CP}=0$ at 5$\sigma$ confidence level for three baselines. }}
\label{fig:nsid0}
\end{figure*}
\paragraph{\bf Possibilities of Discovery of $CP$ violation due to complex NSI phases :}
In Figure \ref{fig:nsid0} we have addressed the question of what could be
the $CP$ fraction for discovery of $CP$ violation if Dirac phase $\delta$
is absent in PMNS mixing matrix and $CP$ violation comes from purely NSI phases.
We observe that for certain NSI value the $CP$ fraction for longer baselines could be more in comparison to that for the shorter baselines. With the increase of $|\varepsilon_{\alpha \beta}|$ there
is increase in discovery of $CP$ fraction in general. If $CP$ violation is not observed in short baselines (say 730 km) but it is observed at relatively longer baselines (say 1500 km) then it would signal the presence of NSI as can be seen from our Figures although for
$\varepsilon_{e \mu}$ it is difficult to distinguish the effect between short and long baselines considered by us. Comparing $3\sigma $ result of neutrino factory (not shown here) with superbeam result of Figure \ref{fig:sbnsid0} it is found that in the presence of $\varepsilon_{e\mu}$ the Neutrino factory set-up gives better
$CP$ violation discovery reach whereas for $\varepsilon_{\mu \tau}$ the superbeam CERN-Pyh\"asalmi set-up is better . For $\varepsilon_{e\tau}$ the neutrino factory
set-up is slightly better than superbeam set-up.
\begin{figure*}[t]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.4\textwidth]{730emphd.pdf}&
\includegraphics[width=0.4\textwidth]{730etphd.pdf}\\
\end{tabular}
\includegraphics[width=0.4\textwidth]{730mtphd.pdf}\\
\caption[] {{\small $\delta$ versus phase ($\phi_{ij}$) considering the value of NSIs($\varepsilon_{ij}$) at the uppermost value with model dependent bounds. Unshaded regions correspond to $CP$ violation discovery reach.}}
\label{fig:phcol}
\end{figure*}
\paragraph{\bf Possibilities of Discovery of $CP$ violation due to both $\delta$ and complex NSI phases :}
In Figures \ref{fig:phcol} we have considered the
case for 730 km baseline where the $CP$ violation might come from $\delta_{CP}$ as well as from
NSI phase $\phi_{\alpha \beta}$. For 1290 km and 1500 km baselines the essential feature is almost same (not shown in the Figure).
Here we have chosen the uppermost value of NSI with model dependent bound.
The total $CP$ violation due to $\delta $ and $\phi_{ij}$ could be observable even for those $\delta $ which could remain unobservable
in absence of NSI phases for some ranges of $\delta $ values.
In fact, the $CP$ violation discovery reach could be possible
for the whole range of allowed $\delta $ values for some specific values of $\phi_{ij}$ as can be seen from Figures.
But it would be difficult to conclude whether the observed $CP$ violation is due to $\delta $ or any NSI phase. As compared to
Figure \ref{fig:sbnsiphase} due to different superbeam set-ups, we find better discovery reach of $CP$ violation for
neutrino factory set-up for 730 km baseline in Figures \ref{fig:phcol}. As for example, in the extreme left Figure for $\varepsilon_{e\mu}$
the total $CP$ fraction is around $94 \%$ for neutrino factory in contrast to $75 \%$ for the T2HK superbeam set-up as shown in Figures \ref{fig:sbnsiphase}.
Comparing Figures \ref{fig:phcol} and \ref{fig:sbnsiphase} we find that for $\phi_{e\tau}$ if we combine
experimental data from superbeam set-ups and neutrino factory set-ups, there are scopes to improve the total $CP$ violation discovery
reach further.
\subsection{\bf monoenergetic neutrino beam}
\label{sec:mono}
We discuss below the possible discovery reach of $CP$ violation for two different nuclei, $^{110}_{50}$Sn and $^{152}$Yb
which is considered for the electron capture experiment. We also discuss our procedure
of energy resolution at the detector for numerical simulation.
Earlier in section \ref{sec:prob} we have shown how $\nu_e \rightarrow \nu_\mu$ oscillation probability depends on $\delta$ for
shorter baseline ($A \sim \alpha $).
Based on the oscillation probability and its' variation with respect to $\delta $ we discuss the procedure
for choosing suitable boost factor $\gamma $ numerically for specific baseline and specific nuclei considered for the neutrino beam. We mention
below four experimental set-ups that we have considered for analysis of discovery reach of $CP$ violation with monoenergetic neutrino beam.
\subsubsection{\bf Suitable boost factor, neutrino energy from $\nu_e \rightarrow \nu_\mu$ oscillation probability}
The most suitable candidate for producing neutrino beam from electron capture process would be the one with a low Q value and high boost factor, $\gamma$ \cite{mono3}. Also it would be preferable if the nuclei has a short half life. The reason for these is as follows. We need neutrino energy around the peak of the oscillation probability where variation due to $\delta $ is significant and as such $\frac{\Delta m^2_{31}L}{4E} \approx (2 n + 1) \frac{\pi}{2}$. Considering $E=2\gamma Q$, it follows that $\gamma = \frac{\Delta m^2_{31}L}{4\pi Q}$ as for example, for the first oscillation peak. For sufficiently high $\gamma$ almost all neutrinos are expected to go through the detector. Then to satisfy the above condition we need to lower Q value. Then another condition is, $\gamma \tau < T$ where $T$ is the time considered to perform the experiment so that all the nuclei decay and
$\tau $ is the half life of the nuclei. If $\gamma$ is increased then the half life $\tau$ is required to be small. So the preferable factors considered in choosing the candidate for producing neutrino beam from electron capture process are - low Q value, small half life $\tau$ and high $\gamma$ \cite{mono3}. Although higher $\gamma $ needs technological advancement for the accelerator.
The isotope, $^{110}_{50}$Sn, has
Q = 267 keV in the rest frame and a half life of 4.11 h. As it
has a low Q value so one may consider high $\gamma$ value. However, it has a longer half life as compared to other nuclei like $^{150}Er$, $^{152}$Yb, $^{156}$Yb, $^{150}$Dy, $^{148}$Dy \cite{mono2,mono11} whose half lifes are 18.5 seconds, 3.04 seconds, 261 seconds, 7.2 min and 3.1 min respectively. However, these nuclei have larger Q values of the order of $10^3$ keV. On the other hand, considering effective running time
per year as $10^7$ second all the nuclei for isotope $^{110}_{50}$Sn will not decay. But for $\gamma =500$ or 320 (as considered in our analysis to obtain the suitable neutrino energy $E$
resulting in high oscillation probability) the
useful decays are respectively about 0.608 or 0.768 times the total number of $^{110}_{50}$Sn nuclei considered. So there is not much suppression in numbers of nuclei. Hence although $^{110}_{50}$Sn has a larger half life, due to its lower Q value there is scope to consider higher $\gamma$
for CERN-Fr\'ejus or 250 km baselines. For these reasons we have preferred isotope, $^{110}_{50}$Sn in comparison to other nuclei. However, there is recent study on finding suitable candidate nuclei for electron capture process and it has been found that $^{150}Er$, $^{152}$Yb, $^{156}$Yb nuclei have dominant electron capture decay
to one level. Particularly, $^{152}$Yb has been found to be most suitable one \cite{mono11}.
For this reason, apart from nuclei $^{110}_{50}$Sn we shall consider $^{152}$Yb also for our analysis.
However, as $Q$ value (5435 keV) for $^{152}$Yb is higher, corresponding $\gamma$ value for such nuclei are supposed to be small.
The neutrino beam produced from electron capture process is boosted with certain boost factor, $\gamma$. The boosted neutrino beam produced from such process hits the detector at a baseline of length $L$ at a radial distance $R$ from the beam axis and the energy, $E$ of this beam in rest frame of the detector , i.e., in laboratory frame is given by:
\begin{eqnarray}
E (R)=\frac{Q}{\gamma}\bigg[1-\frac{\beta}{\sqrt{1+(R/L)^2}}\bigg]^{-1}\approx \frac{2\gamma Q}{1+(\gamma R/L)^2}
\label{eq:erest}
\end{eqnarray}
where R is the radial distance at the detector from the beam axis. At beam center, $R=0$. From the above equation (\ref{eq:erest}) the energy window considered for the analysis which is constrained by the size of the detector is given by:
\begin{equation}
\frac{2\gamma Q}{1+(\gamma R _{max}/L)^2} \leq E \leq 2\gamma Q
\label{eq:ewindow}
\end{equation}
From equation (\ref{eq:ewindow}) we can see that once the baseline length L and $\gamma$ is fixed the energy window gets fixed.
However, even considering radius of the detector $R_{max}=100 $ m the energy window is very small as can be seen from Figure \ref{fig:probmono}.
One can see from equation Eq.\eqref{eq:erest} that it is possible to tell precisely the energy from the $R$ value of the Cherenkov ring at the Cherenkov detector instead of measuring directly the neutrino energy. So there is scope to get good energy resolution by measuring position if the vertex resolution is good. The $\sigma(E)$ function corresponding to energy resolution function (as used in running GLoBES \cite{globes1})
in terms of vertex measurement uncertainty $\sigma(R)$ can be written as:
\begin{eqnarray}
\sigma(E)=-\frac{Q R \beta }{L^2 \left(1+\frac{R^2}{L^2}\right)^{3/2} \left(1-\frac{\beta }{\sqrt{1+\frac{R^2}{L^2}}}\right)^2 \gamma }\sigma(R)
\label{eq:ereso}
\end{eqnarray}
where $\beta$ is defined as
\begin{eqnarray}
\beta=\frac{\sqrt{\gamma ^2-1}}{\gamma }
\label{eq:beta}
\end{eqnarray}
Vertex measurement uncertainty for electron (muon) identification at Super-K is around 30 (20) cm \cite{skvertex}. To estimate the energy resolution using position measurement one may consider $\sigma(R) \sim 30 $ cm provided that the beam spreading $\frac{\sigma(R)}{L}$ is negligible ( lesser than
about $1 \mu$rad) \cite{rolinec} which is difficult to achieve experimentally.
If we take into account the beam divergence about 10 $\mu rad$ (which is
almost one order larger than that considered in references \cite{ beam, divergence}), one may consider larger $\sigma(R) \sim 130 $ cm particularly for baseline of 130 km. For baseline like
250 km it would be more but we have considered same $\sigma(R)$ which means the beam divergence has been assumed to be lesser than about 5 $\mu rad$ for the analysis for baseline with length 250 km.
In this work GLoBES\cite{globes1} has been used for doing the simulations.
In order to use this software, the radial binning is replaced by binning in energy and the bins are not equidistant.
If we divide $R^2_{max}$ into $k$ bins the edges of the bins are given as:
\begin{eqnarray}
\label{eq:r}
R^2_i=R^2_{max}-(i-1)\Delta R^2
\end{eqnarray}
with
\begin{eqnarray}
\label{eq:dr}
\Delta R^2=\frac{R^2_{max}}{k}
\end{eqnarray}
We consider $R^2_i>R^2_{i+1}$ so that in GLoBES the respective energy bins are in the correct order as given below
\begin{eqnarray}
\label{eq:ep}
E(R^2_i)<E(R^2_{i+1})
\end{eqnarray}
where
$E$ is the neutrino energy in the lab frame.
The number of events per bin $i$ and channel $c$ (different channels mentioned later in this section) is given by:
\begin{eqnarray}
\label{eq:event}
N_{event} \simeq \frac{N_{norm}}{L^2} \int_{E_i - \Delta E_i/2}^{E_i + \Delta E_i/2} dE' \int_0^\infty
dE \phi (E) P^c(L,E) \sigma^c(E) \epsilon^c(E') R^c(E,E')
\end{eqnarray}
where $N_{norm}$ is the normalization factor for using GLoBES and is related to length of the baseline, area and energy binning related to
flux , number of target nuclei per unit target mass and number of nuclei decaying.
$\epsilon^c $ is the signal efficiency in the respective channel, $P^c(L,E)$ is the neutrino oscillation probability in particular channel, $\sigma^c(E)$ is total cross section for particular flavor of neutrinos and particular interaction corresponding to particular channel. $R(E,E')$ is the energy resolution function of the detector where $E'$ is the reconstructed neutrino energy. $\phi (E)$ has been calculated from the angular neutrino flux
$\frac{dn}{d\Omega}(E)$ as defined below :
\begin{eqnarray}
\label{eq:flux}
\frac{dn}{d\Omega}=\frac{N_{decays}}{4\pi}\left(\frac{E}{Q}\right)^2
\end{eqnarray}
where $N_{decays}$ is number of nuclei actually decaying per year. The detailed derivation of these expressions can be found in \cite{rolinec}. Considering equation (\ref{eq:ewindow})
and ( \ref{eq:flux}) one can see that with increase in $\gamma$ value the angular flux increases.
\begin{figure}[H]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.4\textwidth]{sn-130.pdf}&
\includegraphics[width=0.4\textwidth]{sn250_2.pdf}\\
\includegraphics[width=0.4\textwidth]{yb-130.pdf}&
\includegraphics[width=0.4\textwidth]{yb250_1.pdf}
\end{tabular}
\caption[] {{\small Probability $P(\nu_e\rightarrow \nu_{\mu})$ vs neutrino energy $E$ for
two different nuclei $^{110}_{50}$Sn and $^{152}$Yb and the corresponding energy window satisfying equation (\ref{eq:ewindow}) for different $\gamma$ values.}}
\label{fig:probmono}
\end{figure}
We have plotted numerically the probability $P(\nu_{e}\rightarrow \nu_{\mu})$ with respect to energy for two different baselines of length 130 km (CERN-Fr\'ejus) and 250 km for
two different nuclei $^{110}_{50}$Sn and $^{152}$Yb. We have considered normal hierarchy in plotting Figure \ref{fig:probmono}. For all plots in Figure \ref{fig:probmono} , $\delta $ has been varied over its' entire allowed range (0 to 2 $\pi$) resulting in the shaded region in each plots
showing the significant variation of probability at particular neutrino energies. Corresponding to each of
the nuclei (whose $Q$ values are fixed) we have considered appropriate $\gamma$ value
so that the corresponding energy window (as mentioned in (\ref{eq:ewindow})) overlaps with
the shaded region near the oscillation peaks having significant variation of probability due to $\delta$ variation. In choosing $\gamma$, one also has to keep in mind that the neutrino energy is not too low as otherwise flux will be much lesser. The energy window has been shown by the shaded vertical strips. For our suitable choice of $\gamma$ value, the energy window is larger for 130 km baseline and relatively smaller for 250 km baseline for both the nuclei. Also the energy window for $^{110}_{50}$Sn is larger than $^{152}$Yb.
For finding $\delta$ we shall prefer the maximum variation of the probability with $\delta$ which will occur for neutrino energy satisfying $\frac{\Delta m ^2_{31}L}{4E} \approx (2n+1) \pi/2$ where $n$ is an integer. This has been shown in Figure \ref{fig:probmono} in which the oscillation probability has been evaluated numerically considering the evolution of neutrino flavor states. However, the energy also depends on the $Q$ value of the corresponding nuclei. So we have considered the case of two nuclei separately. As for example, for $^{110}_{50}$Sn nuclei, for baseline of length 130 km we have considered first oscillation maximum and for baseline of length 250 km we have considered second oscillation maximum where the variation of the probability due to $\delta $ is significant. For $^{152}$Yb, for both the baselines we have considered first oscillation peak.
In considering the suitable peak in the oscillation probability one has to keep in mind that the neutrino flux increases with $E^2$
as shown in Eq.\eqref{eq:flux} and so after doing the numerical analysis one can decide which
energy out of various energies near various peaks are suitable. However, one also have to think about feasible boost factor $\gamma $.
For that reason although the first oscillation peak for $^{110}_{50}$Sn nuclei corresponds to higher energy with respect to
2nd oscillation peak but as it also requires higher boost factor $\gamma $ around 900, we have considered the neutrino energy near the second
oscillation peak for $^{110}_{50}$Sn nuclei for baseline of length 250 km. Depending on the energy chosen near a peak one can appropriately choose the boost factor $\gamma$ on which the neutrino energy window as shown in equation (\ref{eq:ewindow}) as well as $\nu_e$ flux as shown in equation (\ref{eq:flux}) depend. This has been illustrated in Figure \ref{fig:probmono}.
\subsubsection{\bf Experimental set-ups and systematic errors}
For doing the analysis we choose four different set-ups:\
\newline
{\bf set-up(a)}: The length of the baseline is taken to be 130 km (CERN-Fr\'ejus baseline) and the boost factor $\gamma$ to be 500 for nuclei $^{110}_{50}$Sn.
\newline
{\bf set-up(b)}: The length of the baseline is taken to be 250 km and the boost factor $\gamma$ to be 320 for nuclei $^{110}_{50}$Sn.
\newline
{\bf set-up(c)}: The length of the baseline is taken to be 130 km (CERN-Fr\'ejus baseline) and the boost factor $\gamma$ to be 26 for nuclei $^{152}$Yb.
\newline
{\bf set-up(d)}: The length of the baseline is taken to be 250 km and the boost factor $\gamma$ to be 40 for nuclei $^{152}$Yb.
We consider a Water Cherenkov detector of fiducial mass 500 kt. Following \cite{signal}, the signal efficiency is considered to be 0.55 for $\nu_\mu$ appearance channel. Background rejection factor coming from neutral current events is considered to be $10^{-4}$ for $\nu_\mu$ appearance channel. Signal error of $2.5\%$ and background error of $5\%$ has been considered.
For quasi-elastic $\nu_\mu$ appearance and $\nu_e$ disappearance we have followed signal efficiency and error as given in reference \cite{signal}.
We have considered the neutrino energy resolution as discussed earlier in (\ref{eq:ereso}) which can be obtained from vertex resolution after taking into account beam spreading. The neutrino energy is known from
Eq.\eqref{eq:erest} and the energy width considered by us is obtained from Eq.\eqref{eq:ewindow} by considering the radius of the detector $R_{max}=100$m. We assume $10^{18}$ electron capture decays per year and the running time of 10 years for accumulating data.
However, depending on the half life of Sn (4.11 hrs), the number of useful decays per effective year ($10^{7}$ seconds ) considered are about
$0.608 \times 10^{18}$ with boost factor ($\gamma = 500$) for 130 km baseline and $0.768 \times 10^{18}$ with boost factor ($\gamma = 320$) for 250 km baseline. Also, depending on the half life of $^{152}$Yb (3.04 seconds) and the boost factor $\gamma = 26$ or $\gamma = 40$, for baselines 130 km or 250 km respectively, the number of useful decays per effective year ($10^{7}$ seconds ) considered are almost equal to the total number of nuclei i.e, $10^{18}$ as half life is much smaller than $^{110}_{50}$Sn.
It is possible to achieve $\gamma$ about 480 at upgraded SPS facility at CERN \cite{sps,gam440,sps1} and $\gamma > 1000$ for LHC based design \cite{sps2}.
We have considered six energy bins keeping in mind the available energy window for different set-ups and the corresponding energy resolution in equation (\ref{eq:ereso}). In considering the energy resolution we have taken into account beam spreading. For that the the energy resolution considered by us is bad in comparison to the
energy resolution considered in reference \cite{rolinec} and we have to consider much lesser number of energy bins.
As the number of events corresponding to all set-ups are quite large, any background due to atmospheric neutrinos are expected to be quite small and we have not considered such background in our analysis.
\subsubsection{\bf Possibilities of Discovery of $CP$ violation}
Here we discuss the discovery of $CP$ violation for four different experimental set-ups (a-d)
mentioned earlier for monoenergetic neutrino beam.
In presenting our analysis we have followed the numerical procedure as discussed at the beginning of section \ref{sec:sim}. We have considered the true hierarchy as normal hierarchy. However, we have considered the uncertainty in the hierarchy of neutrino masses in the test values as it is not known
at present. We have also considered the uncertainties in the other oscillation parameters as mentioned in table \ref{table:mix}. For finding $CP$ violation we have fixed $\delta$(test) at $CP$ conserving $\delta$ values (0,$\pi$).
In Figure \ref{fig2}, $\Delta \chi^2$ versus $\delta$ (true) has been plotted to show the discovery reach of the $CP$ violation for two different set-ups - set-up(a) and set-up(b) for $^{110}_{50}$Sn nuclei. We find that the discovery of $CP$ violation for set-up(a) \& (b)
could be found with $F(\delta)$ for about 51\% and 49\% respectively of the possible $\delta $ values
at $3\sigma$ confidence level.
In Figure \ref{fig3}, $\Delta \chi^2$ versus $\delta$ (true) has been plotted to show the discovery reach of the $CP$ violation for two different set-ups - set-up(c) and set-up(d) for $^{152}$Yb nuclei. We find that the discovery of $CP$ violation for set-up(c) and set-up (d)
are not good
and could not be found even at $1\sigma$ confidence level. For longer than 250 km baselines we have not
presented any plots for $CP$ violation discovery reach. It seems one of the basic problem
for longer baselines will be relatively bad energy resolution because we are trying to use vertex resolution for getting energy resolution but there is beam spreading and as such over longer baseline beam spreading will make the energy resolution poorer.
\begin{figure}[H]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.4\textwidth]{Sn130n}&
\includegraphics[width=0.4\textwidth]{Sn250n}
\end{tabular}
\caption[] {{\small $\Delta \chi^2$ versus $\delta$ (true) for two experimental set-ups
(a) \& (b) with nuclei $^{110}_{50}$Sn for true normal hierarchy.}}
\label{fig2}
\end{figure}
\begin{figure}[H]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.4\textwidth]{Yb130n}&
\includegraphics[width=0.4\textwidth]{Yb250n}
\end{tabular}
\caption[] {{\small $\Delta \chi^2$ versus $\delta $(true) for two experimental set-ups
(c) \& (d) with nuclei $^{152}$Yb for true normal hierarchy.}}
\label{fig3}
\end{figure}
In our analysis we have chosen neutrino energy near the oscillation peak (as shown in Figure \ref{fig:probmono}) which is more $\delta$ sensitive region. This consideration improves the $CP$ violation discovery reach.
However, in \cite{rolinec} (as shown in Figure 7 of that paper) $CP$ violation discovery reach has been shown to be about 81\% of the possible $\delta $ values for their set-up II for 250 km baseline for the presently known
$\theta_{13}$ value. The discovery reach seems to be not so good in our case although we have chosen appropriately the neutrino energy
where probaility of oscillation is $\delta$ sensitive. It is because
in our analysis we have considered more realistic value of $\gamma$ (which is lower than that considered in \cite{rolinec} but could be achievable at present (keeping in mind the possible SPS upgrade at CERN). We have also taken into account the beam spreading in our present analysis in estimating effective energy resolution
and have taken only a few neutrino energy bins for the analysis unlike \cite{rolinec} as it is important to consider the size of the
energy bins larger in comparison to the level of energy resolution. Our results shows that the nuclei $^{110}_{50}$Sn is more suitable than the nuclei
$^{152}$Yb so far $CP$ violation discovery reach is concerned.
\section{Conclusion}
\label{sec:con}
We have studied the possible $CP$ violation discovery reach due to Dirac phase $\delta$ in the leptonic sector through neutrino oscillation experiments with superbeam, neutrino factory and mono-energetic neutrino beam from electron capture process.
Particularly for superbeam and neutrino factory we have studied the NSI effect on the $CP$ violation discovery reach.
However, for mono-energetic neutrino beam we have considered shorter baseline as required technically for lower boost factor $\gamma $. For that NSI effect seems to be insignificant and hence we
have studied discovery of $CP$ violation due to $\delta $ without different NSI in this case.
For short baseline with CERN-Fr\'ejus and T2HK set-up in case of superbeam for different NSI satisfying model dependent bound we find that there is
insignificant effect due to real NSI on the the discovery reach of $CP$ violation due to $\delta$. For longer baseline with
CERN-Pyh\"asalmi set-up such effects are significant. In case of neutrino factory, the baselines considered are slightly longer than those for superbeam. In this case, except $\varepsilon_{\mu\mu}$ other NSIs have significant effect on $F(\delta)$.
As $F(\delta)$ is never found to be 1, obviously for
some values of $\delta $, $CP$ violation may not be observed at certain confidence level for the kind of detectors we have considered
(see Figure \ref{fig:delta1} ).
Even if one observes $CP$ violation due to $\delta $ in shorter baselines there is a possibility of observing no $CP$ violation in
presence of real NSI. This
feature can be seen in Figure \ref{fig:delta2} for superbeam
in presence of real NSIs - $\varepsilon_{ee}$, $\varepsilon_{\tau\tau}$ and in Figure \ref{fig:cpfrac} in presence of real NSI $\varepsilon_{\tau \tau}$ for neutrino factory. For smaller values of NSI, the $CP$ violation discovery reach is much better for neutrino factory set-ups than those for superbeam set-ups. However, if NSIs like $\varepsilon_{ee},\; \varepsilon_{e\tau}$ and $\varepsilon_{\tau\tau}$ are of significant strength $\mathrel{\mathpalette\@versim>} 0.1$
then $CP$ violation discovery reach at neutrino factory could be very bad in comparison to particularly T2HK set-up in superbeam. One interesting feature is found from Figure \ref{fig:delta2}. If one does not observe $CP$ violation due to $\delta $ in shorter baselines (say CERN-Fr\'ejus set-up) there is still a possibility of observing $CP$ violation
in longer baselines (see CERN-Pyh\"asalmi set-up with off-diagonal NSI like $\varepsilon_{e\mu}$ and $\varepsilon_{e\tau}$) which could be the signal of NSI with significant strength. Using short and long baseline one could conclusively tell about $CP$ violation due to $\delta $
and about NSI under such situation.
NSI - $\varepsilon_{e\mu}$, $\varepsilon_{e\tau}$ and $\varepsilon_{\mu\tau}$ could be complex. We have considered the corresponding phases
$\phi_{e\mu}$, $\phi_{e\tau}$ and $\phi_{\mu\tau}$ respectively in the analysis of the discovery reach of total $CP$ violation
for both in absence and in presence of $\delta $. Even in absence of $\delta $ one may observe $CP$ violation due to the
presence of NSI phases. The possibility of observing $CP$ violation under such scenario is relatively better in general for longer baselines
for both superbeam (see Figure \ref{fig:sbnsid0} ) and neutrino factory (see Figure \ref{fig:nsid0}) provided that absolute values of
NSIs are known. For $\varepsilon_{e\mu}$ there is better $CP$ violation discovery reach in neutrino factory set-up. For $\varepsilon_{e\tau}$
neutrino factory set-up with longer baseline is slightly better than CERN-Pyh\"aslmi set-up in superbeam at $3 \sigma$.
For $\varepsilon_{\mu\tau}$ there is better $CP$ violation discovery reach for CERN-Pyh\"aslmi set-up in superbeam.
Assuming that there are two sources of $CP$ violation simultaneously existing in nature- one due to $\delta $ and the other due to say one of the NSI phases $\phi_{ij}$, the total $CP$ violation discovery reach has been shown as unshaded region in the $\phi_{ij} - \delta $ plane in Figure \ref{fig:sbnsiphase} for superbeam and Figure \ref{fig:phcol} for neutrino factory.
From the Figures it is seen that for NSI - $\varepsilon_{e\mu}$ for T2HK set-up there is more allowed region and as such better discovery reach is possible in the $\delta - \phi_{ij}$ parameter space particularly in comparison to CERN-Pyh\"asalmi set-up in superbeam. For NSI - $\varepsilon_{e\tau}$ and $\varepsilon_{\mu\tau}$, the discovery reach of total $CP$ violation is better for both longer and shorter baselines (like CERN-Pyh\"asalmi set-up and T2HK set-up).
As compared to
Figure \ref{fig:sbnsiphase} due to different superbeam set-ups, we find significantly better discovery reach of $CP$ violation for
neutrino factory set-up for 730 km baseline in Figures \ref{fig:phcol}. Comparing Figures \ref{fig:phcol} and \ref{fig:sbnsiphase} we find that for $\phi_{e\tau}$ if we combine
experimental data from superbeam set-ups and neutrino factory set-ups, there are scopes to improve the total $CP$ violation discovery
reach further.
However, it would be difficult to disentangle the observed $CP$ violation coming due to both $\delta $ and NSI phase.
In presence of some ranges of off-diagonal NSI phase values $\phi_{e\mu}$ and $\phi_{\mu\tau}$ (see upper panel of Figure \ref{fig:sbnsiphase} for superbeam and Figure \ref{fig:phcol} for neutrino factory) there
is possibility of discovering total $CP$ violation for any possible $\delta_{CP}$ value at $3 \sigma$.
The $CP$ violation due to $\delta $ could remain unobservable with present and near future experimental facilities
in superbeam and neutrino factory for certain range of values of $\delta $ (as for example for superbeam with T2HK set-up in Figure \ref{fig:delta1} for about $0^\circ \mbox{to} \;25^\circ$, $160^\circ \mbox{to} \;208^\circ$, $340^\circ \mbox{to} \; 360^\circ$ at 3 $\sigma $ confidence level). However, in presence of NSIs (with or without phases) the $CP$ violation due to $\delta$ or
the total $CP$ violation due to $\delta$ and NSI phase could be observed
even for such values of $\delta $ as can be seen in various Figures shown in Section III-A and B.
Basic strategy to find $CP$ violation in the leptonic sector in presence of NSI for superbeam set-up may be to
consider both shorter baseline
(say T2HK set-up) as well as one longer baseline (say CERN-Pyh\"asalmi set-up) because of their complementary nature
with respect to the discovery reach of $CP$ violation. If NSI values are smaller ($\mathrel{\mathpalette\@versim<} 0.1$) and real then
the $CP$ violation discovery in neutrino factory set-up with MIND detector seems significantly better than superbeam set-up.
For complex NSIs, consideration of both superbeam set-up and neutrino factory set-up could give better $CP$ violation discovery reach.
We have discussed only the possibility of discovering $CP$ violation. However, experimentally disentangling $CP$ violation
coming from $\delta$ present in PMNS matrix, the NSI phases and the absolute value of NSI could be very difficult. One may note that for longer baselines due to higher matter
density NSI effect could be more. Also the effect due to $\delta $ and that due to NSI on oscillation probability in particular channel varies differently with neutrino energy. Only if the experimental data is available over certain range of neutrino energy from the shorter and longer baseline experiments then only the multi-parameter fit
could help in disentangling effects due to different unknown parameters.
For monoenergetic neutrino beam sources we have considered two different
nuclei - one $\nu_e$ source is from electron capture decays of $^{110}_{50}$Sn isotopes and the other $\nu_e$ source is from electron capture decays of isotopes $^{152}$Yb.
For each case we have considered two baselines 130 km and 250 km. Among experimental set-ups
(a-d) in \ref{sec:mono} we find that the set-up (a) with $^{110}_{50}$Sn isotope and 130 km baseline is found to be the most
suitable set-up for discovering $CP$ violation with $F(\delta)$ about 51 \% at $3 \sigma$ confidence level.
When one considers technical issues involved in the accelerator and running the ions through
vacuum tube, isotopes $^{152}$Yb is better candidate than $^{110}_{50}$Sn isotopes because of
much lesser half life. $^{152}$Yb is also better because of the dominant electron capture decay to one energy level. However, as can be seen from Figures \ref{fig2} and \ref{fig3}, the discovery reach of $CP$ violation is found to be better for $^{110}_{50}$Sn isotopes. Out of different baselines for $^{110}_{50}$Sn nuclei, we find slightly better discovery reach for shorter baseline of 130 km with
$\gamma = 500$.
Building up of such monoenergetic neutrino beam facilities will require some technological development and the implementation of it might take some time \cite{Lind}. The existing CERN accelerator complex could be used to study such facility. However, the monoenergetic neutrino flux require a very large number of ions to be stored in the decay ring. It is difficult to control the beam at high intensities because of space charge detuning, intra beam scattering and vacuum loss. With SPS upgrade it could be possible to accelerate the ions to $\gamma =480$ but accelerating above that seems difficult \cite{sps,gam440,sps1}. Depending on the half life
of $^{110}_{50}$Sn we have reduced the total number of useful decays of the ion per effective year from $10^{18}$ but the value considered is
still extreme because of the requirement of acceleration and storage of the partially charged ion. For improving this the vacuum conditions in
SPS would be required to be upgraded. It requires more study on such beam facility. With technological improvement if it is possible to consider monoenergetic beam with $\gamma > 1000$
\cite{sps2}, then the $CP$ violation discovery reach will improve further than what has been
presented in this work.
\hspace*{\fill}
\noindent
\section*{Acknowledgments} ZR thanks University Grants Commission, Govt. of India for providing research fellowships. AD thanks Council of Scientific and Industrial Research, India for financial support through
Senior Research Fellowship (EMR No. 09/466(0125)/2010-EMR-I). We thank L. Whitehead and Luca Agostino for providing migration matrices for Liquid Argon detector and large scale water Cherenkov detector (as studied by MEMPHYS collaboration) respectively and for their other helpful communications. RA thanks R. Gandhi and S. K. Agarwalla for helpful discussion. We thank anonymous referee whose suggestions and comments have helped us immensely in improving the work.
|
\section{Introduction}
In this paper, we continue to study applications of nonassociative algebras to elliptic PDEs started in \cite{Tk14}, \cite{NTVbook}.
Let us consider the $p$-Laplace equation
\begin{equation}\label{plaplace}
\Delta_p u:=|D u|^2\Delta u+\half{p-2}{2}\scal{Du}{D|Du|^2}=0.
\end{equation}
Here $u(x)$ is a function defined on a domain $E\subset \R{n}$, $Du$ is its gradient and $\scal{}{}$ denotes the standard inner product in $\R{n}$. It is well-known that for $p>1$ and $p\ne2$ a weak (in the distributional sense) solution to (\ref{plaplace}) is normally in the class $C^{1,\alpha}(E)$ \cite{Ural68}, \cite{Uhlenbeck}, \cite{Evans82}, but need not to be a H\"older continuous or even continuous in a closed domain with nonregular boundary \cite{KrolMaz72}. On the other hand, if $u(x)$ is a weak solution of (\ref{plaplace}) such that $\mathrm{ess} \sup |Du(x)|>0$ holds locally in a domain $E\subset \R{n}$ then $u(x)$ is in fact a real analytic function in $E$ \cite{Lewis77}.
An interesting problem is whether the converse non-vanishing property holds true. More precisely: is it true that any real analytic solution $u(x)$ to (\ref{plaplace}) for $p>1$, $p\ne2,$ in a domain $E\subset \R{n}$ with vanishing gradient $Du(x_0)=0$ at some $x_0\in E$ must be identically zero? Notice that the analyticity assumption is necessarily because for any $d\ge 2$ and $n\ge2$ there exists plenty non-analytic $C^{d,\alpha}$-solutions $u(x)\not\equiv 0$ to (\ref{plaplace}) in $\R{n}$ for which $Du(x_0)=0$ for some $x_0\in \R{n}$, see \cite{Krol73}, \cite{Aronsson86}, \cite{KSVeron}, \cite{Veron}, \cite{Tk06c}.
The non-vanishing property was first considered and solved in affirmative in $\R{2}$ by John L. Lewis in \cite{Lewis80} as a corollary of the following crucial result (Lemma~2 in \cite{Lewis80}): if $u(x)$ is a real homogeneous polynomial of degree $m=\deg u\ge 2$ in $\R{2}$ and $\Delta_pu(x)=0$ for $p>1$, $p\ne2$ then $u(x)\equiv 0$. Concerning the general case $n\ge 3$, it is not difficult to see (see also Remark~4 in \cite{Lewis80}) that the non-vanishing property for real analytic solutions to (\ref{plaplace}) in $\R{n}$ is equivalent to following conjecture.
\begin{conjecture}
\label{con:1}
Let $u(x)$ be a real homogeneous polynomial of degree $m=\deg u\ge 2$ in $\R{n}$, $n\ge3$. If $\Delta_pu(x)=0$ for $p>1$, $p\ne2$ then $u(x)\equiv 0$.
\end{conjecture}
Notice that a simple analysis shows that Conjecture~\ref{con:1} is true for $m=2$ and any dimension $n\ge2$, therefore the only interesting case is when $m\ge3$. In \cite{Lewis80}, Lewis mentioned that Conjecture~\ref{con:1} holds also true for $n=m=3$ (unpublished). In 2011, J.L.~Lewis asked the author whether Conjecture~\ref{con:1} remains true for any $n\ge3$ and $m\ge3$. In this paper we obtain the following partial result for the cubic polynomial case.
\begin{theorem}\label{th:main}
Conjecture~\ref{con:1} is true for $m=3$ and any $n\ge2$. More precisely, if $u(x)$ is a homogeneous degree three solution of (\ref{plaplace}) in $\R{n}$, $n\ge2$ and $p\not\in\{0,2\}$ then $u(x)\equiv 0$.
\end{theorem}
It follows from the above discussion that the following property holds true.
\begin{corollary}
Let $u(x)\not\equiv0$ be a real analytic solution of (\ref{plaplace}) in a domain $E\subset\R{n}$, $n\ge2$ and $p\not\in\{0,2\}$. If $Du(x_0)=0$ at some point $x_0\in E$ then $D^4u(x_0)=0$.
\end{corollary}
\begin{remark}
Concerning Theorem~\ref{th:main}, notice that for $p=2$, there is a reach class of homogeneous polynomial solutions of (\ref{plaplace}) of any degree $m\ge1$. In other exceptional case, $p=0$ one easily sees that $u(x)=(a_1x_1+\ldots+a_nx_n)^3$ is a cubic polynomial solution to (\ref{plaplace}) in any dimension $n\ge1$.
\end{remark}
\begin{remark}In the limit case $p=\infty$, an elementary argument (see Proposition~\ref{pro:infty} below) yields the non-vanishing property for real analytic solutions of the $\infty$-Laplacian \begin{equation}\label{infplaplace}
\Delta_\infty u:=\scal{Du}{D|Du|^2}=0.
\end{equation}
On the other hand, it is interesting to note that, in contrast to the case $p\ne \infty$, the non-vanishing property holds still true for H\"older continuous $\infty$-harmonic functions. Namely, for $C^2$-solutions of (\ref{infplaplace}) and $n=2$ the non-vanishing property was established by G.~Aronnson \cite{Aronsson68}. In any dimension $n\ge2$ it was proved for $C^4$-solutions by L.~Evans \cite{Evans93} and for $C^2$-solutions by Yifeng Yu \cite{YuY}. The non-vanishing property for $C^2$-smooth $\infty$-harmonic maps was recently established by N.~Katzourakis \cite{Kat14}.
\end{remark}
The proof of Theorem~\ref{th:main} is by contradiction and makes use a nonassociative algebra argument which was earlier applied for an eiconal type equation in \cite{Tk10a}, \cite{Tk14} and study of Hsiang cubic minimal cones \cite{NTVbook}. First, in section~\ref{sec:1} we recall the definition of a metrised algebra and give some preparatory results. In particular, in Proposition~\ref{pro:nonass} we reformulate the original PDE-problem for cubic polynomial solutions as the existence of a metrised non-associative algebra structure on $\R{n}$ satisfying a certain fourth-order identity. Then in Proposition~\ref{pr:equiv}, we show that any such algebra must be zero, thus implying the claim of Theorem~\ref{th:main}.
\section{Preliminaries}\label{sec:1}
\subsection{Metrised algebras}
By an algebra on a vector space $V$ over a field $\mathbb{F}$ we mean an $\mathbb{F}$-bilinear form $(x,y)\to xy\in V$, $x,y\in V$, also called the multiplication and in what follows denoted by juxtaposition. An algebra $V$ is called a zero algebra if $xy=0$ for all $x,y\in V$.
Suppose that $(V,Q)$ is an inner product vector space, i.e. a vector space $V$ over a field $\mathbb{F}$ with a non-degenerate bilinear symmetric form $Q:V\otimes V\to \mathbb{F}{}$. The inner product $Q$ on an algebra $V$ is called {associative} (or invariant) \cite{Bordemann}, \cite[p.~453]{Knus} if
\begin{equation}\label{Qass}
Q(xy,z)=Q(x,yz), \qquad \forall x,y,z\in V.
\end{equation}
An algebra $V$ with an associative inner product is called \textit{metrised} \cite{Bordemann}, \cite[Ch.~6]{NTVbook}.
In what follows, we assume that $\mathbb{F}=\R{}$ and that $(V,Q)$ is a commutative, but may be non-associative metrised algebra. Let us consider the cubic form
$$
u(x):=Q(x^2,x):V\to \R{}.
$$
Then it is easily verified that the multiplication $(x,y)\to xy$ is uniquely determined by the identity
\begin{equation}\label{Qmul}
Q(xy,z)=u(x;y;z),
\end{equation}
where
$$
u(x;y;z):=u(x+y+z)-u(x+y)-u(x+z)-u(y+z)+u(x)+u(y)+u(z)
$$
is a symmetric trilinear form obtained by the linearization of $u$. For further use notice the following corollary of the homogeneity of $u(x)$:
\begin{equation}\label{hom}
u(x;x;y)=2\partial_y u|_{x}.
\end{equation}
In the converse direction, given a cubic form $u(x):V\to \R{}$ on an inner product vector space $(V,Q)$, (\ref{Qmul}) yields a non-associative commutative algebra structure on $V$ called the Freudenthal-Springer algebra of the cubic form $u(x)$ and denoted by $V^{\mathrm{FS}}(Q,u)$, see for instance \cite[Ch.~6]{NTVbook}). According to the definition, $V^{\mathrm{FS}}(Q,u)$ is a metrised algebra with an associative inner product $Q$.
We point out that the multiplication operator $L_x:V\to V$ defined by $L_xy=xy$ is self-adjoint with respect to the inner product $\scal{}{}$. Indeed, it follows from the symmetricity of $u(x,y,z)$ that
$$
Q(L_xy,z)=Q(xy,z)=Q(y,xz)=Q(y,L_xz).
$$
Furthermore, for $k\ge1$ one defines the $k$th principal power of $x\in V$ by
\begin{equation}\label{xk}
x^k=L_x^{k-1}x=\underbrace{x(x(\cdots (xx)\cdots))}_{k \text{ copies of }x}
\end{equation}
In particular, we write $x^2=xx$ and $x^3=xx^2$. Since $V$ is non-associative, in general $x^kx^m\ne x^{k+m}$. However, one easily verifies that the latter power-associativity holds for $k+m\le3$.
\smallskip
We recall that an element $c\in V$ is called an \textit{idempotent} if $c^2=c$. By $\mathscr{I}(V)$ we denote the set of all non-zero idempotents of $V$.
\begin{lemma}\label{Lem:1}
Let $(V,Q)$ be a non-zero commutative metrised algebra with positive definite inner product $Q$. Then $\mathscr{I}(V)\ne \emptyset$.
\end{lemma}
\begin{proof}
First notice that the cubic form $u(x):=Q(x^2,x)\not\equiv 0$, because otherwise the linearization would yield $Q(xy,z)\equiv 0$ for all $x,y,z\in V$, implying $xy\equiv 0$, i.e. $V$ is a zero algebra, a contradiction. Next notice that in virtue of the positive definiteness assumption, the unit sphere $S=\{x\in V:Q(x)=1\}$ is compact in the standard Euclidean topology on $V$. Therefore as $u$ is a continuous function on $S$, it attains its maximum value at some point $y\in S$, $Q(y)=1$. Since $u\not\equiv 0$ is an odd function, the maximum value $u(y)$ must be strictly positive and the stationary equation $\partial_x u|_{y}=0$ holds whenever $x\in V$ satisfies the tangential condition
\begin{equation}\label{tang}
Q(y;x)=0.
\end{equation}
Using (\ref{hom}) and (\ref{Qmul}) we have
$$
0=\partial_x u|_{y}=\frac{1}{2}u(y;y;x)=\half{1}{2}Q(y^2;x)
$$
which implies in virtue of the non-degeneracy of $Q$ and (\ref{tang}) that $y^2=ky$, for some $k\in \R{\times}$. It follows that
$$
kQ(y;y)=Q(y^2;y)=u(y)>0,
$$
which yields $k\ne0$. Then setting $c=y/k$ we obtain $c^2=c$, i.e. $c\in \mathscr{I}(V)$.
\end{proof}
\begin{remark}
In a general finite-dimensional non-associative algebra over $\R{}$, there exist either an idempotent or an absolute nilpotent, see a topological proof, for example, in \cite{Lyubich1}.
\end{remark}
\subsection{Preliminary reductions}
Now suppose that $V=\R{n}$ be the Euclidean space endowed with the standard inner product $Q(x;y)=\scal{x}{y}$. Let $u:V\to \R{}$ be a cubic homogeneous polynomial solution of (\ref{plaplace}) and let $V^{\mathrm{FS}}(u)$ denotes the corresponding Freudenthal-Springer algebra with multiplication $xy$ uniquely defined by
\begin{equation}\label{qscal}
\scal{xy}{z}=u(x;y;z).
\end{equation}
Then the homogeneity of $u(x)$ and (\ref{hom}) yield
\begin{equation}\label{hesu1}
\scal{x^2}{x}=u(x;x;x)=2\partial_x u|_{x}=6u(x).
\end{equation}
Similarly, it follows from (\ref{hom}) that
\begin{equation}\label{hesu3}
\scal{x^2}{y}=u(x;x;y)=2\partial_y u|_x=2\scal{Du(x)}{y}
\end{equation}
which yields the expression for the gradient of $u$ as an element of the Freudenthal-Springer algebra:
\begin{equation}\label{L50}
Du(x)=\half{1}{2}x^2.
\end{equation}
A further polarization of (\ref{hesu3}) yields
$$
\scal{y}{D^2u(x)\, z}=u(x;y;z)=\scal{y}{L_xz},
$$
where $L_xy=xy$ is the multiplication operator by $x$ and $D^2u(x)$ is the Hessian operator of $u$. This implies
\begin{equation}\label{Lxx}
D^2u(x)=L_x,
\end{equation}
\begin{proposition}
\label{pro:nonass}
A cubic form $u:V=\R{n}\to \R{}$ satisfies $(\ref{plaplace})$ if and only if its Freudenthal-Springer algebra $V^{\mathrm{FS}}(u)$ satisfies the following identity:
\begin{equation}\label{L6}
\scal{b}{x}\scal{x^2}{x^2}+\half{p-2}{2}\scal{x^2}{x^3}=0
\end{equation}
where
\begin{equation}\label{L5}
b=b(V):=\sum_{i=1}^ne_i^2,
\end{equation}
and $e_1,\ldots,e_n$ is an arbitrary orthonormal basis of $\R{n}$.
\end{proposition}
\begin{proof}
Using (\ref{L50}) and (\ref{Lxx}), one obtains
$$
\scal{Du}{D^2u|_{x}Du}=\half{1}{4}\scal{L_xx^2}{x^2} =\half{1}{4}\scal{x^3}{x^2},
$$
and similarly,
\begin{equation}\label{L11}
\begin{split}
\Delta u(x)&=\trace D^2u|_x=\trace L_x=\sum_{i=1}^n \scal{L_xe_i}{e_i}=\sum_{i=1}^n \scal{e_i^2}{x}=\scal{b(V)}{x},
\end{split}
\end{equation}
where $b$ is defined by (\ref{L5}). Inserting the found relations into (\ref{plaplace}) yields (\ref{L6}). In the converse direction, if $V$ is a metrised algebra satisfying (\ref{L6}) then $u(x)$ defined by (\ref{hesu1}) is easily seen to satisfy (\ref{plaplace}).
\end{proof}
\section{Proof of Theorem~\ref{th:main}}
Using the introduced above definitions and Proposition~\ref{pro:nonass}, one easily sees that the following property is equivalent to Theorem~\ref{th:main}.
\begin{proposition}\label{pr:equiv}
A commutative metrised algebra $(V,Q)$ with $\dim V\ge2$ and satisfying (\ref{L11}) with $p\not\in\{0,2\}$, is a zero algebra.
\end{proposition}
\begin{proof}
We argue by contradiction and assume that $(V,\scal{}{})$ is a non-zero commutative metrised algebra satisfying (\ref{L6}). Since $p\ne 2$, this identity is equivalent to
\begin{equation}\label{L01}
\scal{q}{x}\scal{x^2}{x^2}+\scal{x^2}{x^3}=0,
\end{equation}
where
\begin{equation}\label{q}
q=\frac{2}{p-2}b(V)\in V.
\end{equation}
Polarizing (\ref{L01}) we obtain in virtue of
$$
\partial_yx^3=\partial_y(x(xx))=yx^2+2x(xy)
$$
and the associativity of the inner product that
$$
\scal{q}{y}\scal{x^2}{x^2}+
4\scal{q}{x}\scal{xy}{x^2}+
4\scal{xy}{x^3}+\scal{x^2}{yx^2}=0,
$$
implying by the arbitrariness of $y$ that
\begin{equation}\label{L02}
\scal{x^2}{x^2}q+4\scal{q}{x}x^3+4x^4+x^2x^2=0,
\end{equation}
we according to (\ref{xk}) $x^4=xx^3$.
A further polarization of (\ref{L02}) yields
$$
4\scal{x^2}{xy}q+4\scal{q}{y}x^3+4\scal{q}{x}(yx^2+2x(xy)) +4yx^3+4x(yx^2+2x(xy)) +4x^2(xy)=0,
$$
which implies an operator identity
\begin{equation}\label{L03}
2L_x^3+L_{x^3}+\scal{q}{x}(L_{x^2}+2L_x^2)+L_xL_{x^2}+L_{x^2}L_x+(q\otimes x^3+x^3\otimes q)=0.
\end{equation}
Here $a\otimes b$ denotes the rank one operator acting by $(a\otimes b)y=a\scal{b}{y}$.
Now, notice that by our assumption and Lemma~\ref{Lem:1}, $\mathscr{I}(V)\ne \emptyset$. Let $c\in \mathscr{I}(V)$ be an arbitrary idempotent. Then setting $x=c$ in (\ref{L01}) we find
$$
|c|^2q+(4\scal{q}{c}+5)c=0.
$$
Taking scalar product of the latter identity with $c$ yields
\begin{equation}\label{C03}
\scal{q}{c}=-1,\qquad q=-\frac1{|c|^2}c
\end{equation}
in particular $q\ne0.$ Furthermore, setting $x=c$ in (\ref{L03}) and applying (\ref{C03}) yields
$$
2L_c^3+L_{c}+\scal{q}{c}(L_{c}+2L_c)+2L_c^2+(q\otimes c+c\otimes q)=2L_c^3-\frac{2}{|c|^2}c\otimes c=0,
$$
therefore
\begin{equation}\label{C04}
L_c^3=\frac{1}{|c|^2}c\otimes c.
\end{equation}
The latter identity, in particular, implies that
\begin{equation}\label{zero}
L_c=0 \text{ on } c^\bot:=\{x\in V: \scal{c}{x}=0\},
\end{equation}
where by the assumption $\dim c^\bot=\dim V-1\ge1$.
We claim that $c^\bot$ is a zero subalgebra of $V$. Indeed, if $x,y\in c^\bot$ then by the associativity of the inner product and (\ref{zero}) we have
$$
\scal{xy}{c}=\scal{x}{cy}=\scal{x}{L_cy}=0,
$$
hence $xy\in c^\bot$ which implies that $c^\bot$ is a subalgebra (in fact, an ideal) of $V$. Suppose that $c^\bot$ is a non-zero subalgebra, then it follows by Lemma~\ref{Lem:1} that there is a nontrivial idempotent in $c^\bot$, say $w$. Then by the second identity in (\ref{C03}) we have $\scal{w}{q}=0$, therefore (\ref{L01}) yields
\begin{equation}\label{C05}
\scal{w^2}{w^3}=|w|^2=0.
\end{equation}
The obtained contradiction proves our claim.
To finish the proof, we consider an arbitrary orthonormal basis $\{e_i\}_{1\le i\le n}$ of $V$ with $e_n=c/|c|$. Then $e_i\in c^\bot$ for all $1\le i\le n-1$, hence by the above zero-algebra property we have $e_i^2=0$. Applying (\ref{L5}) we get
$$
\frac{p-2}{2}q=b(V)=\sum_{i=1}^ne_i^2=\frac{c}{|c|^2}=-q,
$$
which yields in virtue of $q\ne 0$ that $p=0$, a contradiction. The theorem is proved.
\end{proof}
\section{Concluding remarks}
We notice that the appearance of non-associative algebras in the above analysis of the $p$-Laplace equation is not accident and becomes more substantial if one considers the following eigenfunction problem
\begin{equation}\label{radial}
\Delta_p u(x)=\lambda |x|^2u(x),\quad \lambda\in \R{},\quad p\ne2,
\end{equation}
with $u(x)$ being a cubic homogeneous polynomial. Notice that (\ref{plaplace}) correspond to $\lambda=0$ in (\ref{radial}). The problem (\ref{radial}) for $p=1$ has first appeared in Hsiang's study of cubic minimal cones in $\R{n}$ \cite{Hsiang67}. In fact, it follows from recent results in \cite[Ch.~6]{NTVbook} that any cubic polynomial solution of (\ref{radial}) is necessarily harmonic, and thus satisfies (\ref{radial}) for \textit{any} $p\ne2$! The zero-locus of any such solution is an algebraic minimal cone in $\R{n}$ \cite{Hsiang67}. Furthermore, it was shown in \cite{NTVbook} that (\ref{radial}) has a large class of non-trivial cubic solutions for $p=1$ (and thus for any $p\ne 2$) sporadically distributed over dimensions $n\ge2$. It turns out that these solutions have a deep relation to rank 3 formally real Jordan algebras and their classification requires a mush more delicate analysis by using nonassociative algebras, we refer to \cite{Tk15} for more examples of solutions to (\ref{radial}) and their classification.
Finally, below we give an elementary proof of the non-vanishing property for real analytic $\infty$-harmonic functions.
\begin{proposition}\label{pro:infty}
If $v(x)$ is a real analytic solution of the (\ref{infplaplace}) in a domain $D\subset \R{n}$ and $Dv(x_0)=0$ for some $x_0\in D\subset \R{n}$ then $v(x)\equiv v(x_0)$.
\end{proposition}
\begin{proof}
Indeed, we may assume that $x_0=0$ and suppose by contradiction that $v(x)\not\equiv v(0)$. Then a direct generalization of Lewis' argument given in Lemma~1 in \cite{Lewis80} easily yields the existence of a real homogeneous polynomial $u(x)\not\equiv 0$ of order $\deg u=k\ge2$ which also is a solution to (\ref{infplaplace}). Notice that $u(x)$ attains its maximum value on the unit sphere $S=\{x\in \R{n}:|x|=1\}$ at some point $y$. The stationary equation yields $Du(y)=\lambda y$ for some real $\lambda$ and by Euler's homogeneous function theorem
$$
ku(y)=\scal{y}{Du(y)}=\lambda |y|^2=\lambda
$$
and
$$
\scal{Du(y)}{D|Du|^2(y)}=\lambda (2k-2)|Du|^2(y)=2(k-1)\lambda^3,
$$
which yields by (\ref{infplaplace}) that $u(y)=0$, hence
$$
\max_{x\in S} u(x)=\frac{\lambda}{k}=0.
$$
A similar argument applied to the minimum value implies $\min_{x\in S} u(x)=0$, a contradiction with $u\not\equiv 0$ follows.
\end{proof}
\section*{Acknowledgements}
I would like to thank John L. Lewis for helpful discussion and bringing my attention to Conjecture~\ref{con:1}.
\bibliographystyle{amsplain
\def$'${$'$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\subsection{Cost efficiency in Cloud-RAN}
So far, quantitative studies on Cloud-\ac{RAN} have focused on improvements in terms of throughput and energy efficiency. The constraints under which Cloud-RAN is sustainable from a cost perspective are, therefore, unclear. These constraints, however, will play an important role in the decision to deploy Cloud-RAN. In a Cloud-RAN system, many different \textit{interdependent} factors determine the cost of deployment and operation, such as device intensities, equipment cost, capacity cost, infrastructure cost, and data processing cost. This paper is, to the best of our knowledge, the first complete analysis which takes both communication and data processing capabilities as well as their relationship to deployment costs into account. Furthermore, it also considers how the costs (listed above) are related and how different system parameters impact the overall cost. The framework derived in this paper allows identifying operating regimes in which Cloud-\ac{RAN} proves more cost effective. Most importantly, the framework derived is parametrized in a way that permits the evaluation of various deployments and utilization scenarios.
\subsection{Related Work}
Though knowledge of the total cost of networks and methods of modeling them have always been important, there are not many non-proprietary works which show how the cost (irrespective of whether it is capital expenditure or operational expenditure) of an entire network can be computed. The first paper which addressed this topic was \cite{FB'97} and it demonstrated how the deployment costs of a fixed line telecommunication network could be calculated based on a model using homogeneous Poisson processes. Inspired by which, our paper \cite{ICC'14} used a similar model (using homogeneous Poisson processes) to compute the deployment costs of a homogeneous mobile network including the entire backbone infrastructure. This work was extended upon in \cite{WiOpt'14}, which modeled heterogeneous networks (along with their backhaul infrastructure) using various stationary point processes. The above mentioned works, however, take neither the Cloud-\ac{RAN} concept nor the additional information processing costs into account. Now, as far as Cloud-\ac{RAN} is concerned, since its introduction in \cite{Guan.Kolding.Merz.CRAN.2010}, the concept (as a whole) has drawn significant attention. Not long ago, the Next Generation Mobile Networks Alliance published a technical report (see \cite{NGMN.CRAN.TechReport.2013}) which
states that, besides performance improvements through multi-cell processing, improvements in cost- and energy-efficiency are also expected. However, the report does not provide a quantitative analysis of the characteristic benefits of a Cloud-RAN system. In comparison, the white paper \cite{ChinaMobile.CRAN.WhitePaper.2011} states that Cloud-\ac{RAN} reduces capital expenditure by $15$\% and operational expenditure by $50$\% when compared to a (traditionally deployed) distributed network. The report, however, does not detail how these numbers were obtained.
Most of the work on Cloud-\ac{RAN} focuses on fully centralized networks, i.e., all \ac{RAN} functionality is executed at the data center. In contrast, \cite{iJOINref'13} proposed a flexible centralization of \ac{RAN} depending on the actual backhaul network characteristics. This flexibility allows exploiting a part of the centralization gain despite non-ideal connections between small-cells and a data center. However, as detailed in \cite{Wuebben.etal.SPM.2014}, this is accompanied by challenges in the operation of Cloud-\ac{RAN} and the signal processing performed. In addition to which, none of the literature available considers the relationship between the data processing resources required and the mobile communication traffic offered.
In \cite{Bhaumik.Chandrabose.Jataprolu.Kumar.Muralidhar.Polakos.Srinivasan.Woo.MobiCom.2012}, Bhaumik et al. provided a comprehensive quantitative assessment of the computational resources required for a specific configuration of a 3GPP LTE mobile network and showed that the turbo-decoding process requires a majority of the processing resources. However, \cite{Bhaumik.Chandrabose.Jataprolu.Kumar.Muralidhar.Polakos.Srinivasan.Woo.MobiCom.2012} does not provide a model which allows extrapolating these results to different system configurations and nor does it enable quantifying centralization gain in a Cloud-\ac{RAN} system.
The first comprehensive analytical model to assess decoder complexity was by Grover et al. in \cite{Grover.Woyach.Sahai.JSAC.2011}.
In \cite{Rost.Talarico.Valenti.TWC.2014}, this model was extended to include the entire mobile network and to allow quantifying the data processing requirements. This framework allows to dimension the data processing resources of a Cloud-\ac{RAN} system such that a given quality criterion is met; e.g., the probability that the system has insufficient data processing resources to process all incoming transmissions, which is referred to as computational outage \cite{Valenti.Talarico.Rost.GC.2014}.
\subsection{Contributions and organization of the paper}
This paper introduces a model using various spatial point processes for analyzing the deployment costs of a Cloud-RAN system which takes users, base stations (both macro and micro), backhaul (both microwave and optic fiber), and data centers into account. An expression for the average cost of deploying a data center is derived from which the total deployment cost of the network is found. Then, we utilize a data processing model (which helps dimension the data center based on the traffic demands as well as decoder quality) to provide values which can be used in the expression for deployment cost. Using these values, we examine whether Cloud-RAN based networks are more economical than traditional LTE networks and the reasons behind it. Numerical evaluations reveal that Cloud-RAN based networks are indeed more cost effective with respect to deployment costs because they are better at adapting to network load (i.e., the number of active users) and can exploit the fact that the number of processors required does not increase linearly with the load. However, these evaluations also highlight the fact that deployment cost of Cloud-RAN technologies increases when user and data center intensities increase and when the deployment favors a particular backhaul technology.
The paper is structured as follows. Section \ref{Problem_Framework} introduces our spatial point process system model and the underlying data processing model. Section \ref{Cost_Model} introduces the various costs involved and derives the framework used to analyze the deployment cost of Cloud-\ac{RAN} systems. Section \ref{Numerical_Evaluation} evaluates the findings of Section \ref{Cost_Model} numerically and discusses the quantitative results. The paper is concluded in Section \ref{End}.
\section{Introduction}
\input{Introduction}
\section{Problem Framework}\label{Problem_Framework}
\input{Problem_Framework_Intro}
\subsection{Multi-Layer System Model}\label{System_Model}
\input{System_Model}
\subsection{Data Processing Model}\label{Complexity_Model}
\input{Data_Processing_and_Cloud_RAN_Model}
\section{Cost Model}\label{Cost_Model}
\subsection{Components of Deployment Cost}\label{Cost_Description}
\input{Cost_Definitions}
\subsection{Computing the Deployment Cost}\label{Cost_Framework}
\input{Cost_Model}
\subsection{Obtaining Values for Cost Calculation}\label{Obtaining_Cost_Values}
\input{Obtaining_Cost_Values}
\section{Numerical Evaluation}\label{Numerical_Evaluation}
\input{Numerical_Evaluation}
\section{Conclusions}\label{End}
\input{Conclusions}
|
\section{Introduction}
Relation algebras are a natural extension of Boolean algebras. In addition to lattice operations $\land$ and $\lor$, complementation $-$, a FALSE or bottom constant $0$ and TRUE or top constant $\top$, the signature of relation algebras includes a binary operation of composition $\cdot$, a unary operation of converse $\con{}$, and an identity constant $\mathsf{e}$. For a detailed introduction we refer the reader to~\cite{Maddux2006}.
Such an algebra is called \emph{representable} if it is isomorphic to an algebra of binary relations in which composition is represented as relational composition, $\land$ as intersection, $\lor$ as union, $0$ as the empty set, $\top$ as an equivalence relation $W$, $-$ as complement with respect to $W$, $\con{}$ as relational converse, and $\mathsf{e}$ as the identity relation. In this article we will consider representability for weaker signatures in which only some of these operations are respected. A similar problem is \emph{finite representability}, in which the base set of representation is to be a finite set.
In 1941 Tarski~\cite{Tarski1941} offered a class of algebras in the signature above along with a finite set of axioms. Tarski~\cite[p.~88]{Tarski1941} along with J\'onsson~\cite{JonssonTarski1948} asked if every model of these axioms was isomorphic to an algebra of binary relations. In 1950, Lyndon~\cite{Lyndon1950} offered an infinite family of non-representable relation algebras that satisfied Tarski's axioms. Monk~\cite{Monk1964} later showed that no finite axiomatisation is possible. Finally in 2001 Hirsch and Hodkinson~\cite{HirschHodkinson2001} showed that representability is undecidable for finite relation algebras. So in the full signature we have non-finite axiomatisability of representability and undecidability of representability for finite algebras.
A natural question then is whether or not these results hold for reducts of the full signature. A survey on this topic is offered by Schein~\cite{Schein1991}. In particular, Schein remarks that ``it would be interesting to describe `complemented semigroups'\dots This problem may be more treatable for ordered complemented semigroups.'' These are algebras with signature $\langle \cdot, \le, - \rangle$. In this article we show that representability in this signature is undecidable if complements are taken with respect to a universal relation.
Representability over a finite base set is also shown to be undecidable, a result we are able to extend to a weaker notion of complementation. We also prove undecidability of finite representability for lattice-ordered semigroups, which are those with signature $\langle \cdot, \land, \lor \rangle$. Furthermore, these results regarding either representability or finite representability apply to any signature between one of these and that of a Boolean monoid, $\langle \cdot, \land, \lor, -, \mathsf{e}, 0, \top \rangle$.
In order to prove these results we adapt a construction of Boolean monoids used by Hirsch and Jackson~\cite{HirschJackson2012}, correcting issues that arise from weakening the signature.
\section{Partial groups and Boolean monoids}
A \emph{partial group} is a system $\mathbf{A} = \langle A; *, e \rangle$ with a partial binary operation $*$ such that $*$ is associative and $e$ acts as an identity whenever $*$ is defined. That is to say, $\mathbf{A} $ is a group except that some compositions are undefined.
Furthermore, $\mathbf{A} $ is also a \emph{square partial group} if there is a subset $\sqrt{A}$ of $A$ containing the identity $e$ such that the following holds.
\begin{enumerate}
\item $a * b$ is defined if and only if $a,b \in \sqrt{A}$.
\item $\sqrt{A} * \sqrt{A} = A$; that is, for every $c \in A$ there are $a,b \in \sqrt{A}$ such that $a * b = c$.
\end{enumerate}
A partial group $\mathbf{A} $ is \emph{cancellative} if it satisfies the cancellation laws
\begin{align*}
x * y = x * z &\implies y = z \\
\text{and } x * y = z * y &\implies x = z.
\end{align*}
From a finite cancellative square partial group $\mathbf{A} $ the authors of~\cite{HirschJackson2012} construct a finite Boolean monoid $\mathbf{M(A)} $ with signature $\langle \cdot, \land, \lor, -, \mathsf{e}, 0, \top \rangle$. While we will not go into the details of the construction here, it is worth noting that the resultant $\mathbf{M(A)} $ is a \emph{normal} Boolean monoid. That is, if $\dom(a) = (a\top) \land \mathsf{e}$ and $\ran(a) = (\top a) \land e$ then $\dom(a)a = a = a\ran(a)$. In a representation of a normal Boolean monoid, $\dom(a)$ and $\ran(a)$ will be represented as a restriction of the identity relation to the domain and range of $a$, respectively. Note also that $\dom(a)$ and $\ran(a)$ are idempotent in $\mathbf{M(A)} $.
This construction from a partial group references the \emph{partial group embedding problem} for a class of groups $\mathcal{K} $. This problem takes a finite partial group $\mathbf{A} $ and returns YES if there is a group $\mathcal{G} \in \mathcal{K} $ and an injective map $\phi\colon A \to G$ that respects all products defined in $\mathbf{A} $. Evans~\cite{Evans1953} showed that this problem is decidable for a class $\mathcal{K} $ if and only if the uniform word problem for $\mathcal{K} $ is decidable. In particular, this problem is undecidable if $\mathcal{K} $ is either the class of groups or class of finite groups, and $\mathbf{A} $ is a finite cancellative square partial group~\cite[Lemma~3.4]{HirschJackson2012}.
One of the key concepts required to prove undecidability of representability is a formal means of referring to all elements that act as injective partial maps, hereafter called \emph{injective functions}. In the full signature of relation algebras, one can consider a unary relation $\mathtt{i}$ as in Definition~\ref{defn:irelbooleanmonoid} to capture these elements.
\begin{definition}\label{defn:irelbooleanmonoid}
Define a unary relation $\mathtt{i}$ in the language of relation algebras by
\[
x \in \mathtt{i} \iff x\con{x} \le \mathsf{e} \text{ and } \con{x} x \le \mathsf{e}.
\]
\end{definition}
In a representation respecting converse, composition and identity, elements in $\mathtt{i}$ are exactly those relations that would be represented as injective functions. By considering the diversity relation $\div = -\mathsf{e}$, we can also view $\mathtt{i}$ as the set of elements satisfying the following formula in a signature containing $\{\cdot, \land, \div\}$.
\begin{lemma}[{\cite[Lemma~2.12]{HirschJackson2012}}]\label{lemma:irel}
Let $\mathbf{R}$ be a relation algebra. Then $a \in \mathtt{i}^\mathbf{R}$ if and only if
\[
(a\div) \land a = 0 = (\div a) \land a.
\]
\end{lemma}
The final concepts needed are those of domain and range equivalence. Binary relations $r$ and $s$ in an algebra over base set $X$ are domain equivalent, denoted $r \mathrel{\domsquare} s$, if
\[
\{x \in X \mid (\exists y \in X) (x,y) \in r \} = \{x \in X \mid (\exists y \in X) (x,y) \in s \}.
\]
We use the same notation for the abstraction of this concept in a Boolean monoid, with $x \mathrel{\domsquare} y$ if $\dom(x) = \dom(y)$. Range equivalence $\mathrel{\ransquare}$ is defined similarly. Note that in signatures weaker than that of a Boolean monoid a representation may preserve $\mathrel{\domsquare}$ or $\mathrel{\ransquare}$ without necessarily preserving $\dom$ or $\ran$, respectively.
The following theorem is a combination of Propositions 5.1 and 6.3 from~\cite{HirschJackson2012}.
\begin{theorem}\label{thm:mainundecidabilitytheorem}\label{thm:piggyback}
Let $\mathbf{A} $ be a finite cancellative square partial group. The following are equivalent, with the statements in square brackets giving a separate set of equivalences.
\begin{enumerate}
\item $\mathbf{M(A)} $ is representable [over a finite base set].
\item There is a $\{ \cdot, \mathtt{i}, \mathrel{\domsquare}, \mathrel{\ransquare}\}$-embedding of $\mathbf{M(A)} $ into $\wp(X \times X)$ for some [finite] set $X$.
\item $\mathbf{A} $ embeds into a [finite] group $\mathcal{G} $.
\end{enumerate}
\end{theorem}
We have already observed that both versions of (3) are known to be undecidable, and so too are the items in (1) and (2). The goal of this paper is to introduce equivalent statements regarding decidability of representability of signatures weaker than that of a Boolean monoid.
In considering signatures without converse, we cannot be certain that $\top$ is represented as an equivalence relation. It turns out that a representable normal Boolean monoid can always be represented in such a way that $\top$ acts as an equivalence relation~\cite[Lemma~2.2]{HirschJackson2012}. This does not necessarily hold for weaker signatures. While this requirement on $\top$ is not without precedent for reducts of relation algebras (see~\cite{Mikulas2004, Schein1991}) we may wish to remove it, or even consider algebras in which no top element exists, and so we will always state when this assumption is in use. With this in mind, we introduce statements in Theorem~\ref{thm:main} equivalent to those in Theorem~\ref{thm:mainundecidabilitytheorem}, but regarding representability of lattice-ordered semigroups and ordered complemented semigroups, thus proving undecidability of these problems as well.
\begin{theorem}\label{thm:main}
Let $\mathbf{A} $ be a finite, cancellative, square partial group. The following are equivalent.
\begin{enumerate}
\item $\mathbf{M(A)} $ is representable [over a finite base set].
\item $\mathbf{M(A)} $ is representable [over a finite base set] as a lattice-ordered semigroup with $\top$ represented as an equivalence relation.
\item $\mathbf{M(A)} $ is representable [over a finite base set] as an ordered complemented semigroup with $\top$ represented as an equivalence relation.
\item There is a $\{ \cdot, \mathtt{i}, \mathrel{\domsquare}, \mathrel{\ransquare}\}$-embedding of $\mathbf{M(A)} $ into $\wp(X \times X)$ for some [finite] set $X$.
\item $\mathbf{A} $ embeds into a [finite] group $\mathcal{G} $.
\end{enumerate}
\end{theorem}
We note that a representation of $\mathbf{M(A)} $ as a lattice-ordered semigroup would respect the operations in $\{\cdot, \land, \lor\}$, while a representation as an ordered complemented semigroup would respect those in $\{\cdot, \le, -\}$. Since both signatures are weaker than that of a Boolean monoid we can see that $(1) \implies (2)$ and $(1) \implies (3)$. Similarly we note that these results apply to any signature between one of these reducts and that of a Boolean monoid.
The remaining implications are $(2) \implies (4)$ and $(3) \implies (4)$. Since composition is preserved in a representation of a semigroup, this aspect is trivial. We must prove that relations in $\mathtt{i}$ are preserved as injective functions under a representation in either signature. We do this for lattice-ordered semigroups in Lemma~\ref{lemma:latticeordered} and ordered complemented semigroups in Lemma~\ref{lemma:orderedcomplemented}. In Lemma~\ref{lemma:domranequivalence}, we prove that a representation of $\mathbf{M(A)} $ in either reduct preserves domain and range equivalence.
There is some ambiguity here as to the definition of complementation in a reduct. In the full signature of relation algebras complementation is taken with respect to the top element. In the absence of a top element, one can declare that if $x$ is related to $y$ by an element of the algebra, then for all relations $a$ we have that $(x,y)$ belongs to just one of $\{-a,a\}$. This mimics the behaviour of complementation when a top element is present by taking complements with respect to the union of all elements. We call this \emph{relative complementation}. A stronger definition would take complements with respect to a universal relation, demanding that every $(x,y)$ belongs to just one of $\{-a,a\}$. We refer to this as \emph{universal complementation}.
For lattice-ordered semigroups and ordered complemented semigroups with relative complementation, the requirement on $\top$ can be removed if the representation is to be over a finite base set, and we show this in Lemmas~\ref{lemma:latticeorderedtop} and \ref{lemma:orderedcomplementedtop}. A representation of an ordered complemented semigroup with universal complementation will always represent $\top$ as an equivalence relation if it exists, as shown in Lemma~\ref{lemma:strongcomplement}, and so representability of algebras in this signature is undecidable. These results are stated in Theorem~\ref{thm:succinctresults}.
\begin{theorem}\label{thm:succinctresults}
Let $\tau$ be a signature such that $\tau \subseteq \{\cdot, \land, \lor, -, \mathsf{e}, 0, \top\}$. The following problems are undecidable.
\begin{itemize}
\item Finite representability of algebras with signature $\tau$ where $\{\cdot, \land, \lor\} \subseteq \tau$.
\item Finite representability of algebras with signature $\tau$ where $\{\cdot, \le, -\} \subseteq \tau$ and $-$ is to be represented as relative complementation.
\item Representability and finite representability of algebras with signature $\tau$ where $\{\cdot, \le, -\} \subseteq \tau$ and $-$ is to be represented as universal complementation.
\end{itemize}
\end{theorem}
Theorem~\ref{thm:succinctresults} also yields results about non-finite axiomatisability for the same signatures. If there exists a finite set of first-order axioms characterising representability of a class of algebras, then one can consider an algorithm that checks a finite algebra against each of these axioms to determine representability. Hence, finite axiomatisability implies decidability of representability, giving us the results in Corollary~\ref{cor:succinctresults}.
\begin{corollary}\label{cor:succinctresults}
Let $\tau$ be a signature such that $\tau \subseteq \{\cdot, \land, \lor, -, \mathsf{e}, 0, \top\}$. The following classes of algebras are not finitely axiomatisable in first order logic.
\begin{itemize}
\item Any class whose finite members are the finitely representable algebras with signature $\tau$ where $\{\cdot, \land, \lor\} \subseteq \tau$.
\item Any class whose finite members are the finitely representable algebras with signature $\tau$ where $\{\cdot, \le, -\} \subseteq \tau$ and $-$ is to be represented as relative complementation.
\item Representable algebras with signature $\tau$ where $\{\cdot, \le, -\} \subseteq \tau$ and $-$ is to be represented as universal complementation.
\end{itemize}
\end{corollary}
We note that non-finite axiomatisability of representability of algebras with signature $\tau$ where $\{\cdot, \land, \lor\} \subseteq \tau \subseteq \{\cdot, \land, \lor, \con{}, \mathsf{e}, 0, \top\}$ was shown by Andr\'eka~\cite{Andreka1991}; see also Andr\'eka and Mikul\'{a}s~\cite{AndrekaMikulas2011}.
\section{Proofs of main results}
Let $\S = \langle S; \cdot, \mathsf{e} \rangle$ be a monoid. We also consider meet $\land$, join $\lor$, complementation $-$, a partial order relation $\le$, and constants $0$ and $\top$.
Let $h\colon \S \to \wp ( X \times X )$ be a representation of $\S$ on a base set $X$ preserving at least composition. Define an equivalence relation $\sim$ on $X$ such that for all $x,y \in X$, $x \sim y$ if $x = y$ or $\mathsf{e}$ acts as the universal relation on the set $\{x,y\}$, a situation illustrated in Figure~\ref{fig:compressed}. Define a new representation $\hat{h} \colon \S \to \wp( X/\!\!\sim \times\, X/\!\!\sim )$ such that for $a \in S$,
\[
\hat{h} \colon a \mapsto \{ [x], [y] \mid (\exists x' \in [x])(\exists y' \in [y]) \; (x',y') \in h(a) \}.
\]
\begin{figure}[ht]
\centering
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,auto,node distance=3cm,thick]]
\tikzstyle{vertex}=[circle, fill=black, draw=black, inner sep = 0.06cm]
\node[vertex, label = below left: $x$] (x) at (0:0cm) {};
\node[vertex, label = above right: $y$] (y) at (45:2cm) {};
\draw[->] (x) .. controls ([xshift=0mm, yshift=1.44cm] x) and ([xshift=-7.22mm, yshift=0cm] y) .. (y);
\draw[->] (y) .. controls ([xshift=0mm, yshift=-1.44cm] y) and ([xshift=7.22mm, yshift=0cm] x) .. (x);
\Loop[dist=2cm,dir=SOWE,label=$\mathsf{e}$,labelstyle=below left](x);
\Loop[dist=2cm,dir=NOEA,label=$\mathsf{e}$,labelstyle=above right](y);
\node at (1.3cm,0.1cm) {$\mathsf{e}$};
\node at (0cm,1.3cm) {$\mathsf{e}$};
\end{tikzpicture}
\caption{$\mathsf{e}$ acting as the universal relation on $\{x,y\}$}\label{fig:compressed}
\end{figure}
\begin{lemma}\label{lemma:quotient}
If $h$ is a representation of $\S$ preserving composition then so too is $\hat{h}$. Furthermore, $\hat{h}$ preserves Boolean operations and constants $0$, $\top$ and $\mathsf{e}$, if they are correctly represented by $h$.
\end{lemma}
\begin{proof}
Consider $a,b \in S$ such that $a \ne b$. Then, since $h$ is faithful, we may assume without loss of generality that there exists $(x,y) \in h(a) \backslash h(b)$. Then $([x], [y]) \in \hat{h}(a)$. Suppose by way of contradiction that $([x],[y]) \in \hat{h}(b)$. Then there exists $(w,z) \in h(b)$ with $\mathsf{e}$ acting as the universal relation on $\{x,w\}$ and on $\{y,z\}$. That is, $(x,w) \in h(\mathsf{e})$ and $(z,y) \in h(\mathsf{e})$. Since $h$ preserves composition we have that $(x,y) \in h(ebe)$ and so $(x,y) \in h(b)$. But this violates the assumptions on $(x,y)$. So $\hat{h}$ is faithful.
Now we turn our attention to composition under $\hat{h}$. Let $([x],[y]) \in \hat{h}(a)$ and $([y],[z]) \in \hat{h}(b)$. Without loss of generality, assume $(x,y) \in h(a)$ and $(y,z) \in h(b)$, since as before we can always compose elements with $\mathsf{e}$ to move around within equivalence classes. Then, as $h$ preserves composition, $(x,z) \in h(ab)$ and so $([x],[z]) \in \hat{h}(ab)$. Similarly we have that $([x],[z]) \in \hat{h}(ab) \implies ([x],[z]) \in \hat{h}(a)\hat{h}(b)$. So $\hat{h}$ also preserves composition.
We note that $\hat{h}$ only contracts binary relations in $h(\S)$. Hence, Boolean operations and constants $0$, $\top$ and $\mathsf{e}$ are preserved in $\hat{h}$, assuming they were correctly represented by $h$. In particular, if $\mathsf{e}$ is represented correctly then $(x,y) \in h(\mathsf{e}) \iff x = y$, and so $h(\mathsf{e}) = \hat{h}(\mathsf{e})$.
\end{proof}
It is by this quotient that we will ensure that the elements of the Boolean monoid $\mathbf{M(A)} $ that are in $\mathtt{i}$ are represented as injective functions. Recall from Lemma~\ref{lemma:irel} that an element $a \in \mathtt{i}$ if and only if $(a\div) \land a = 0 = (\div a) \land a$.
\begin{lemma}\label{lemma:latticeordered}
Suppose that the Boolean monoid $\mathbf{M(A)} $ is representable in a signature containing $\{\cdot, \land, \lor\}$ in such a way that $\top$ is represented as an equivalence relation. Then there exists a representation in the same signature with the property that if $a \in M(A)$ is such that $(a \div ) \land a = 0 = (\div a) \land a$ then $a$ is represented as an injective function.
\end{lemma}
\begin{proof}
Let $h \colon \mathbf{M(A)} \to \wp(X \times X)$ be such a representation of $\mathbf{M(A)} $ onto some base set $X$ and consider $a \in M(A)$ such that $(a \div ) \land a = 0 = (\div a) \land a$. By applying Lemma~\ref{lemma:quotient} we may work under the assumption that $h = \hat{h}$, and note that this preserves the property that $\top$ is represented as an equivalence relation.
Suppose there exists $x,y,z \in X$ such that $(y,x) \in h(a)$ and $(y,z) \in h(a)$, a situation illustrated in Figure~\ref{fig:anotfunction}. We will show that $x=z$. As $\top$ is acting as the universal relation on $\{x, y, z\}$ and $0\top = \top 0 = 0$, if $h(0)$ relates any two (potentially equal) elements of $\{x, y, z\}$ then it must act as the universal relation on all three. Since $0 \le e$, this would imply that $e$ is also acting as a universal relation, a situation we have precluded unless $x = z$. So assume otherwise, that is, assume that $h(0)$ is not relating any two elements of $\{x, y, z\}$.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,auto,node distance=3cm,thick,scale=0.8]
\tikzstyle{vertex}=[circle, fill=black, draw=black, inner sep = 0.06cm]
\node[vertex, label = left: $y$] (y) at (0:0cm) {};
\node[vertex, label = right: $x$] (x) at (30:3cm) {};
\node[vertex, label = right: $z$] (z) at (-30:3cm) {};
\path[->,font=\normalsize]
(y) edge node [above left] {$h(a)$} (x)
(y) edge node [below left] {$h(a)$} (z);
\end{tikzpicture}
\caption{An element $a$ not represented as a function under $h$.}\label{fig:anotfunction}
\end{figure}
We note also that $(x,z) \notin h(\div)$ as if this were the case then we would have $(y,z) \in h(a\div)$. But $a \land (a\div) = 0$, giving $(y,z) \in h(0)$. Similarly, we have that $(z,x) \notin h(\div)$. As $\div \lor \mathsf{e} = \top$ we therefore have $(x,z)$ and $(z,x)$ in $h(e)$. Hence $(x,x)$ and $(z,z)$ are in $e$. As $\hat{h} = h$ it follows that $x = z$ as required. That is, $a$ is a function under $h$. By symmetry we also have that $a$ is injective under $h$.
\end{proof}
The requirement that $(\div a) \land a = 0 = (a \div) \land a$ simply ensures that $a$ is disjoint from $a \div$ and also from $\div a$. We can restate this with operations in $\{\cdot, \le, -\}$ such that $(\div a) \land a = 0$ if and only if $a \le -(\div a)$, and similarly $(a \div) \land a$ if and only if $a \le -(a\div)$. This allows us to replicate the previous result in the signature of ordered complemented semigroups.
\begin{lemma}\label{lemma:orderedcomplemented}
Suppose that the Boolean monoid $\mathbf{M(A)} $ is representable in a signature containing $\{\cdot, \le, -\}$ in such a way that $\top$ is represented as an equivalence relation. Then there exists a representation in the same signature with the property that if $a \in M(A)$ is such that $a \le -(a\div)$ and $a \le -(\div a)$ then $a$ is represented as an injective function.
\end{lemma}
\begin{proof}
Let $h \colon \mathbf{M(A)} \to \wp(X \times X)$ be a such a representation of $\mathbf{M(A)} $ onto some base set $X$ and consider $a \in M(A)$ such that $a \le -(a\div)$ and $a \le -(\div a)$. Again we work under the assumption that $h = \hat{h}$. Since $0$ is the unique element with the property that $0 \le -0$, we have that $h(0) \subseteq h(\top)\backslash h(0)$, and so $0$ is represented correctly as the empty set.
We take $x,y,z$ as in Figure~\ref{fig:anotfunction} with $(y,x) \in h(a)$ and $(y,z) \in h(a)$. As $a \le -(a\div)$, we cannot have $(x,z) \in h(\div )$, since we could compose to get $(y,z) \in h(a \div )$. Similarly, $(z,x) \notin h(\div)$. Because $\mathsf{e}$ and $\div$ are complementary with respect to $\top$, we have that $(x,z)$ and $(z,x)$ are in $h(\mathsf{e})$. We compose to realise $\mathsf{e}$ acting as a universal relation on $\{x,z\}$, a situation we have precluded unless $x = z$. Hence, $a$ is represented as a function under $h$. By symmetry we also have that $a$ is injective under $h$.
\end{proof}
Hence, the $\mathtt{i}$ relation as given in Definition~\ref{defn:irelbooleanmonoid} can be recovered in the case of signatures containing $\{\cdot, \land, \lor\}$ or $\{\cdot, \le, -\}$, as long as $\top$ is to be represented as an equivalence relation. To complete the $\{\cdot, \mathtt{i}, \mathrel{\domsquare}, \mathrel{\ransquare}\}$-embedding required by Theorem~\ref{thm:main}, we must also check that domain and range equivalence are respected in the representation of a Boolean monoid as either a lattice-ordered semigroup or an ordered complemented semigroup.
\begin{lemma}\label{lemma:domranequivalence}
Suppose that the Boolean monoid $\mathbf{M(A)} $ is representable in a signature containing $\{\cdot, \land, \lor\}$ or in a signature containing $\{\cdot, \le, -\}$, and in either case suppose that $\top$ is represented as an equivalence relation. Then one can define domain and range equivalence of the elements in $\mathbf{M(A)} $ in such a way that they are respected by the representation.
\end{lemma}
\begin{proof}
In either case, take $x \mathrel{\domsquare} y$ if $x\top = y\top$, and $x \mathrel{\ransquare} y$ if $\top x = \top y$.
\end{proof}
This establishes that $(2) \implies (4)$ and $(3) \implies (4)$ in Theorem~\ref{thm:main}, completing the proof. Subject to the assumption that $\top$ is represented as an equivalence relation, we have undecidability of representability and finite representability of finite algebras in either signature. If we restrict our attention to representations over a finite base set then we can remove this assumption. The proofs are largely the same for lattice-ordered semigroups and ordered complemented semigroups, and both involve Lemma~\ref{lemma:idempotentlemma}.
\begin{lemma}\label{lemma:idempotentlemma}
Let $h$ be a representation of a Boolean monoid $\mathbf{M} $ onto a finite base set $X$ respecting composition and order. Then for every nonzero idempotent $f \in M$ there exists an element of $X$ fixed by $h(f)$ but not by $h(0)$.
\end{lemma}
\begin{proof}
By faithfulness there exists $x,y \in X$ such that $(x,y) \in h(f)$ and $(x,y) \notin h(0)$, or such that $(x,y) \notin h(f)$ and $(x,y) \in h(0)$. Since $0$ is the bottom element, we conclude that the latter is not possible and assume that $(x,y) \in h(f)$. Since $f$ is idempotent we must witness an element $z \in X$ such that $(x,z) \in h(f)$ and $(z,y) \in h(f)$. We must continue to witness this for every pair in $h(f)$. But the representation is finite, so we must eventually witness a loop $(x_a,x_a) \in h(f)$. If $(x_a, x_a) \in h(0)$ also, then we could compose to get $(x,y) \in h(0)$, violating our initial assumption. Hence, $f$ but not $0$ fixes $x_a$ in the representation.
\end{proof}
We first remove the assumption that $\top$ be represented as an equivalence relation in finite representations of $\mathbf{M(A)} $ as a lattice-ordered semigroup, although in actuality only composition and meet are required for the proof. Recall that in the Boolean monoid signature we defined $\dom(a) = (a\top) \land \mathsf{e}$ and that $\dom(a)$ is idempotent in $\mathbf{M(A)} $.
\begin{lemma}\label{lemma:latticeorderedtop}
Let $h$ be a representation of the Boolean monoid $\mathbf{M(A)} $ onto a finite base set $X$ respecting the operations in $\{\cdot, \land\}$. Then there exists a representation $h^\circ$ in the same signature but representing the top element $\top$ as an equivalence relation. Furthermore, if $h$ respects the operations in $\{\lor,-,\mathsf{e}, 0\}$ then so too does $h^\circ$.
\end{lemma}
\begin{proof}
For a binary relation $r$ define the symmetric interior
\[
r^\circ := \{ (x,y) \mid (x,y) \in r \text{ and } (y,x) \in r \}.
\]
If $r$ is reflexive and transitive then one can view $r^\circ$ as the largest equivalence relation contained in $r$. Define $h^\circ \colon M(A) \to \wp(X \times X)$ as $h^\circ \colon a \mapsto h(a) \cap h(\top)^\circ$.
Since we are only omitting non-loops in the representation we have that $h^\circ$ preserves any operation in $\{\lor,-,\mathsf{e}, 0\}$, assuming that $h$ does. For composition, consider $(x,y) \in h^\circ(ab)$ for some $a,b \in M(A)$. Then since $h$ respects composition there exists $z \in X$ such that $(x,z) \in h(a)$ and $(z,y) \in h(b)$. As $(x,y)$ is in the image of $h^\circ$ we have that $(y,x) \in h(T)^\circ$. So $(y,z) \in h(\top a)$ and, as $\top$ is the top element, $(y,z) \in h(\top)$. Similarly, $(z,x) \in h(\top)$. We conclude that $(x,z) \in h^\circ(a)$ and $(z,x) \in h^\circ(b)$. By similar composition with $\top$ we have that if $(x,z) \in h^\circ(a)$ and $(z,y) \in h^\circ(b)$ then $(x,y) \in h^\circ(ab)$, and so composition is respected by $h^\circ$.
Now we must prove that $h^\circ$ is faithful. Let $a,b \in M(A)$ be distinct and assume without loss of generality that $b \nleq a$, so that $b\land(-a) \ne 0$. Note that we are only considering $-a$ as an element of $M(A)$, and do not require complementation to be represented in any way. As $\mathbf{M(A)} $ is normal we have that $\dom(b \land(-a))(b \land(-a)) = (b \land(-a))$, and so $\dom(b\land(-a)) \ne 0$. We established in Lemma~\ref{lemma:idempotentlemma} that nonzero idempotents under $h$ fix points in $X$ that are not fixed by $0$. As such, $h^\circ(\dom(b\land(-a))) \ne h^\circ(0)$. But clearly $h^\circ(\dom(a\land(-a))) = h^\circ(0)$. As such $h^\circ(a) \ne h^\circ(b)$, and so $h^\circ$ is faithful.
Since $h$ respects composition and order, $h^\circ$ represents $\top$ as transitive and symmetric, and hence reflexive on a subset of $X$. Since $h^\circ$ is faithful, this subset is nonempty. Hence, $h^\circ$ represents $\top$ as an equivalence relation over a nonempty subset of $X$.
\end{proof}
If we have a representation of $\mathbf{M(A)} $ as a lattice-ordered semigroup, then we can take the symmetric interior and then the quotient used in Lemma~\ref{lemma:quotient} to obtain a similar representation preserving the $\mathtt{i}$ relation and representing $\top$ as an equivalence relation. This allows us to remove the requirement in Lemma~\ref{lemma:latticeordered} that the top element of $\mathbf{M(A)} $ be represented as an equivalence relation, if the representation is to be taken over a finite set. The following lemma permits us to do the same in the case of Lemma~\ref{lemma:orderedcomplemented}, which deals with ordered complemented semigroups.
\begin{lemma}\label{lemma:orderedcomplementedtop}
Let $h$ be a representation of the Boolean monoid $\mathbf{M(A)} $ onto a finite base set $X$ respecting the operations in $\{\cdot, \le, -\}$. Then there exists a representation $h^\circ$ in the same signature but representing the top element $\top$ as an equivalence relation. Furthermore, if $h$ respects the operations in $\{\land, \lor, \mathsf{e}, 0\}$ then so too does $h^\circ$.
\end{lemma}
\begin{proof}
The proof is largely the same as for Lemma~\ref{lemma:latticeorderedtop}, though we must recover faithfulness with a different approach. Recall that $0$ is forced to be represented as the empty set since $0 \le -0$, and so any representation preserving $\{\cdot, \le -\}$ also trivially preserves $0$. Again we define $h^\circ \colon \mathbf{M(A)} \to \wp(X \times X)$ as $h^\circ \colon a \mapsto h(a) \cap h(T)^\circ$ and take distinct $a,b \in M(A)$ with the assumption that $b \nleq a$. Then there exists a nonzero $c$ such that $c \le b$ and $c \le -a$. Hence we can distinguish between $a$ and $b$ if we witness a nonempty $h^\circ(c)$.
Since $c$ is nonzero we use Lemma~\ref{lemma:idempotentlemma} to conclude that $\dom(c)$ fixes a point $x \in X$ under $h$. Now we note that, since $\top$ has maximum domain and range and composition on the right cannot restrict domain, $D(c) \le D(c)T = cT$. We must witness this composition as in Figure~\ref{fig:witnesscT} and so $h^\circ(D(c)) \le h^\circ(cT) = h^\circ(c)h^\circ(T)$. That is, $h^\circ(c)$ is nonempty.
\begin{figure}[ht]
\centering
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,auto,node distance=3cm,thick]]
\tikzstyle{vertex}=[circle, fill=black, draw=black, inner sep = 0.06cm]
\node[vertex, label = below left: $x$] (x) at (0:0cm) {};
\node[vertex, label = above right: $y$] (y) at (45:2cm) {};
\draw[->] (x) .. controls ([xshift=0mm, yshift=1.44cm] x) and ([xshift=-7.22mm, yshift=0cm] y) .. (y);
\draw[->] (y) .. controls ([xshift=0mm, yshift=-1.44cm] y) and ([xshift=7.22mm, yshift=0cm] x) .. (x);
\Loop[dist=2cm,dir=SOWE,label=$\dom(c)$,labelstyle=below left](x);
\node at (1.4cm,0.1cm) {$\top$};
\node at (0cm,1.3cm) {$c$};
\end{tikzpicture}
\caption{Witnessing the composition $D(c) \le cT$.}\label{fig:witnesscT}
\end{figure}
\end{proof}
We noted before that the definition of complementation requires care in the absence of a top element. We used here relative complementation which mimics the definition of complementation when $\top$ is present: that if $x$ is related to $y$ by an element of the algebra, then for all relations $a$ we have that $(x,y)$ belongs to just one of $\{-a,a\}$. Under this weaker definition and without $\top$ acting as the universal relation we could have, for example, a situation as in Figure~\ref{fig:anotfunction} such that no element relates $x$ to $z$ or $z$ to $x$. If this occurs, we cannot take the complement to reason that $\mathsf{e}$ acts as the universal relation on these points, as we did in Lemma~\ref{lemma:orderedcomplemented}. Under the weaker definition of relative complementation, these proofs require that an element already relates these two points.
Alternatively, we can represent complementation as universal complementation in which the complement is taken with respect to $\wp(X \times X)$, where $X$ is the base set of the representation. That is, in the absence of a top element we can take complements with respect to a universal relation. Under this interpretation it will turn out that if $\top$ does exist then it must act as an equivalence relation in any representation. We thank Marcel Jackson for the following observation.
\begin{lemma}\label{lemma:strongcomplement}
Let $\S$ be a complemented semigroup of binary relations with complement taken with respect to a universal relation. If there exists an idempotent $f$ such that $f(-f) = -f = (-f)f$ and $-f$ is also idempotent, then $f$ is the universal relation.
\end{lemma}
\begin{proof}
Suppose $f$ relates $x$ to $y$ and, for contradiction, $f$ does not relate $y$ to $x$. Then $y(-f)x$ and, by assumption, $x(-f)x$. Then composing we get that $x(-f)y$, a contradiction. Hence, $f$ is a symmetric, and so a reflexive binary relation.
We now know that $f$ is an equivalence relation on its domain. Suppose that the domain of $f$ is not full, and so we have that $x(-f)y$ and $y(-f)x$. Since $-f$ is idempotent we compose to get $x(-f)x$. If $x$ is in the domain of $f$ then we have a contradiction. If not, take $y$ to be in the domain of $f$ to reach a similar contradiction. Hence, $f$ is an equivalence relation with full domain.
\end{proof}
Hence, any representation of $\mathbf{M(A)} $ respecting the operations in $\{\cdot, \le, -\}$ in which $-$ is represented as universal relation will always represent $\top$ as that universal relation. This extends the undecidability of representability of finite algebras in this signature to include infinite representations.
It would be interesting to see if we can do the same for infinite representability of lattice-ordered semigroups.
\begin{problem}
We know that finite representability is undecidable for lattice-ordered semigroups. Can the same be said of representability in general?
\end{problem}
Another problem to consider is semigroups with either form of complementation but no order. As far as we can determine this problem remains unexplored in the literature, and is mentioned by Schein~\cite{Schein1991}.
\begin{problem}
Is representability or finite representability decidable in the signature $\langle \cdot, - \rangle$, with either relative or universal complementation? Are representable algebras in this signature finitely axiomatisable?
\end{problem}
|
\section{Introduction}
\label{intro}
A central question today in physics is to understand the subatomic structure of matter in
terms of fundamental degrees of freedom, i.e, quarks and gluons~\cite{PDG2014}.
The standard model for the strong interaction, namely, quantum chromodynamics~(QCD),
with quarks and gluons~\cite{Yndurain,Smilga}, exhibit confinement, beyond a perturbative expansion. Calculations with Lattice QCD and
Schwinger-Dyson approaches are performed in Euclidian space, while it is still a challenge to extract information of the hadron in
Minkowski space. In this sense, it is still useful to describe the composite states in constituent quark models~(CQM),
which are defined in Minkowski space, and driven towards experiments (see e.g.~\cite{Close1980}).
A vast literature is devoted to investigate the hadronic electromagnetic
structure ( see e.g. \cite{Roberts96,Pacheco97,deMelo99,Choi99,
deMelo2003,Pacheco2002,deMelo2004,Otoniel2012}) since it provides a laboratory to deeper our understanding of QCD.
A natural framework to combine the quantum field theory~(QFT) and CQM to describe bound states is the
light-front quantization~(LFQCM)~\cite{Brodsky98,Harindranath2000,Beane2013}.
One can resort to the usual covariant formulation of quantum field theory and the light-front
quantisation to test the nonperturbative properties of QCD.
The aim of this contribution is to study the sensitivity of static electroweak properties
of the $\rho$-meson with the model parameters of an ansatz of the covariant Bethe-Salpeter
amplitude with constituent quarks proposed in Ref. \cite{Pacheco97}. The variables used
to evaluate numerically the loop integrals is the light-front momentum, and the integrals
are reduced to three-dimensional ones, by analytical integration
on the minus component of the loop momentum, relating the calculations
of the electroweak observables with the present model
performed in a light-front framework considers only valence contributions. In the case of the electromagnetic observables, we use a
prescription for computing
the elastic form factors which was shown to be free of non-valence contributions in the Drell-Yan frame (see e.g. \cite{Pacheco2012}). We also
found that the vector meson decay constant in the present model is free of a zero-mode when computed with the plus component of the current operator.
In our systematic study we address
the dependence of the static electromagnetic observables and decay constant with the two model parameters, namely the
constituent quark and the regulator masses. In addition, we compare our results with the ones obtained with different models.
\section{Light-Front model, electromagnetic form factors and decay constant }
The general framework for the light-front constituent quark model adopted here has been used to study
several properties of hadronic states, like in the case of mesons and baryons in the vacuum~\cite{Jaus1990,Pacheco2009}
or in nuclear matter \cite{Kazuo2014}. Electromagnetic form factors of composite vector particles within the light-front framework
have been addressed in many works
\cite{Frankfurt79,Inna84,Chung88,Inna89,Brodsky92,Frankfurt93,Cardarelli95,Ji2001,Bakker2001},
where a
specific frame with momentum transfer $q^+=q^0+q^3=0$ (Drell-Yan condition)
is chosen to compute the matrix elements of the plus component of the current operator.
Here, we will use this frame and current
component, with the choice of matrix elements suggested in
\cite{Inna89}, which has been shown to be free of zero modes \cite{Melo2004,Choi2004,Pacheco2012}.
In this case, the computation is performed
only in the valence region. We will return to this point later in this section.
The $\rho-q\bar{q}$ vertex model for an on-mass-shell meson, is
the same one proposed in reference~\cite{Pacheco97}:
\begin{eqnarray}
\Gamma^\mu = \left[ \gamma^\mu -
\frac{m_{\rho}}{2}~
\frac{2 k^{\mu} -P^{\mu}}{(P^\mu k_\mu + m\, m_\rho) -\imath \epsilon }
\right]~,
\end{eqnarray}
where the quark momentum is $k^{\mu}$, the constituent quark mass is $m$,
and the $\rho$ meson mass is $m_\rho.$
The electromagnetic form factors
of a spin-1 particle are computed from a linear combination of the
matrix elements of the current, $J^{\mu}_{ji}$ with $i$ and $j$ indexing the polarization states. The
general covariant expression of the current is~\cite{Frankfurt79}:
\begin{equation}
J_{\alpha \beta}^{\mu}=\left[F_1(q^2)g_{\alpha \beta} -F_2(q^2)
\frac{q_{\alpha}q_{\beta}}{2 m_{\rho}^2}\right] (p^\mu + p^{\prime \mu})
- F_3(q^2)
(q_\alpha g_\beta^\mu- q_\beta g_\alpha^\mu) \ ,
\label{eq:curr1}
\end{equation}
where $m_\rho$ is the rho meson mass, $q^\mu$ is the momentum transfer,
and $P^\mu$ is the sum of the initial and final momentum. In the impulse approximation, the plus component of the
electromagnetic current,~${ J}^+_{ji}$, is:
\begin{equation}
J^+_{ji} = \imath \int\frac{d^4k}{(2\pi)^4}
\frac{ Tr[\epsilon^{'\nu}_j \Gamma_{\beta}(k,k-p_f)
(\sla{k}-\sla{p_f} +m)
\gamma^{+}
(\sla{k}-\sla{p_i}+m) \epsilon^\mu_i \Gamma_{\alpha}(k,k-p_i)
(\sla{k}+m)]
\Lambda(k,p_f)\Lambda(k,p_i) }
{(k^2 - m^2+\imath \epsilon)
((k-p_i)^2 - m^2+\imath\epsilon)
((k-p_f)^2 - m^2+\imath \epsilon )}~,
\label{current}
\end{equation}
where ${\epsilon^\prime_j}$ and ${\epsilon_i}$ are the
polarization four-vectors of the final and initial states,
respectively. In the cartesian instant form spin basis the initial polarization state is given by:
\begin{equation}
\epsilon^\mu_x=(-\sqrt{\eta},\sqrt{1+\eta},0,0),~\epsilon^\mu_y=(0,0,1,0),
~\epsilon^\mu_z=(0,0,0,1)~, \label{polar}
\end{equation}
and for the final state is given by:
\begin{equation}
\epsilon^\mu_x=(\sqrt{\eta},\sqrt{1+\eta},0,0),~\epsilon^{\prime\mu}_y=(0,0,1,0),
~\epsilon^{\prime\mu}_z=(0,0,0,1)~,
\end{equation}
where $\eta=-q^2/4m^2_\rho$.
The function~$\Lambda(k,p)=N/((p-k)^2-m^2_R+\imath \epsilon)^2$
reguralizes the loop integral in~Eq.(\ref{current}). The regulator mass is $m_R$
and the normalisation factor $N$ is fixed by the condition of unit charge.
The Breit-frame is used in the numerical calculations, with the choice $p^\mu=(p^0,-q_x/2,0,0)$ for the initial state and
$p^{\prime\mu}=(p^0,q_x/2,0,0)$ for the final state, such that the momentum transfer $q^\mu=(p^{\prime^\mu}-p^\mu)$ satisfies the
the Drell-Yan condition,~$q^+=0$. In this case four matrix elements are independent and
only three electromagnetic form factors exists. The physical constraint is expressed by the angular condition, which in the light-front spin basis is written as:
\begin{equation}
\Delta(q^2)=(1+2 \eta) I^{+}_{11}+I^{+}_{1-1} -
\sqrt{8 \eta} I^{+}_{10} -
I^{+}_{00} \ = \ (1 + \eta)(J^+_{yy}-J^+_{zz})=0 \ .
\label{eq:ang}
\end{equation}
This relation is violated in the present model by the presence of zero-modes not accounted by the naive integration in the loop momentum $k^-$,
which leads only to the valence region contribution (see e.g. \cite{Pacheco2012}). In principle, if the angular condition is satisfied,
we have the freedom to extract in different ways the form factors from the matrix elements of the current
~\cite{Inna84,Chung88,Brodsky92,Frankfurt93}.
In the present work we calculate the electromagnetic observables,
~$G_0$,~$G_1$ and $G_2$, without ambiguities due to zero-mode contributions
by using the prescription proposed in \cite{Inna84,Inna89}, which can be also computed in the
instant form basis, as well as in the light-front spin basis:
\begin{eqnarray}
&&G_0^{GK} = \frac{1}{3}[(3-2 \eta) I^{+}_{11}+ 2 \sqrt{2 \eta} I^{+}_{10}
+ I^{+}_{1-1}]
= \frac{1}{3}[J_{xx}^{+} +(2 -\eta) J_{yy}^{+}
+ \eta J_{zz}^{+}], \nonumber \\
&&G_1^{GK} = 2 [I^{+}_{11}-\frac{1}{ \sqrt{2 \eta}} I^{+}_{10}]
=J_{yy}^{+} - J_{zz}^{+} - \frac{J_{zx}^{+}}{\sqrt{\eta}},
\nonumber \\
&&G_2^{GK}=\frac{2 \sqrt{2}}{3}[- \eta I^{+}_{11}+
\sqrt{2 \eta} I^{+}_{10} - I^{+}_{1-1}]
= \frac{\sqrt{2}}{3}[J_{xx}^{+}-(1+\eta) J_{yy}^{+}
+ \eta J_{zz}^{+}] \ .
\label{inna}
\end{eqnarray}
The prescription \cite{Inna84,Inna89} amounts to eliminate the matrix element $I_{00}^+$ in the computation of the form factors.
In \cite{Melo2004,Choi2004,Pacheco2012}, it was demonstrated that the zero-modes or non-valence contributions are present only in $I^+_{00}$, and
therefore the combinations expressed in Eq. (\ref{inna}) can be computed relying only on the valence region.
The low-energy electromagnetic observables considered here are
the radius,~$\langle r^2 \rangle$,
the magnetic moment $\mu$ and the quadrupole moment $Q_2$, which are given by \cite{Pacheco97,Cardarelli95}:
\begin{eqnarray}
\langle r^2 \rangle = \lim_{q^2 \rightarrow 0}
\dfrac{6 [G_0(q^2)-1] }{q^2} \ , \
\mu = \lim_{q^2 \rightarrow 0} G_1 (q^2) \ , \
Q_2 = \lim_{q^2 \rightarrow 0}
\dfrac{3 \sqrt{2} G_2 (q^2)}{q^2}~.
\end{eqnarray}
In addition we calculate the $\rho$-meson decay constant $f_\rho$ defined by the matrix element ~\cite{Zuber1980}:
\begin{eqnarray}
\langle 0| \bar{d}\Gamma^\mu u |\rho; i \rangle \dfrac{Q_u + Q_{\bar{d}} }{\sqrt{2}} = i
\sqrt{2} f_\rho m_\rho \epsilon^\mu_i \ .
\end{eqnarray}
where $Q_{u}=2/3$ and $Q_{\bar{d}}=1/3$
are the charges of quark and antiquark. The matrix element of the electromagnetic current between the vacuum and $\rho-$meson states is computed from
a loop integral, and we choose the plus momentum component and the polarisation $\epsilon_z$ defined in (\ref{polar}). After integrating in $k^-$
the result is:
\begin{eqnarray}
f_\rho = \dfrac{N_c N}{m_\rho} \int \dfrac{d^2 k_\perp}
{(2 \pi)^3} \int_0^1 dx\dfrac{ 4 \left( - x ( p^{+} + p^{+^2})+ k_{\perp}^2
+ m^2 \right) -
\dfrac{m_{\rho}}{2}\dfrac{ 4m p^+(2x
- 1)(k^-_{on} - x p^+)}{ \left[
p^+ (k^-_{on} + x p^+)/2 + m_{\rho}m \right] }}
{x (1 - x)^3 (m^2_\rho - M^2_0 ) (m^2_\rho-M^2_R )^2}
\label{cc26} \ ,
\end{eqnarray}
where $p^{\mu}=(m_\rho,\vec{0})$, $N_c=3$, is number of colors
and $k^-_{on}=(k^2_{\perp}+m^2)/x$ is the particle on-mass-shell relation. The other quantities in (\ref{cc26}) are:
\begin{eqnarray}
M^2_0 = \dfrac{\vec{k}^2_{\perp}+m^2}{x} + \dfrac{(\vec{p}_{\perp}-\vec{k}_{\perp})^2+ m^2}{1 - x} - p^2_\perp \ ; \
M^2_{_R} = \dfrac{ \vec{k}^2_{\perp}+m^2}{x} +
\dfrac{(\vec{p}_{\perp}-\vec{k}_{\perp})^2+ m_R^2}{1 - x} - \vec{p}^2_\perp ~,
\end{eqnarray}
In order to have bound state, the
following conditions are necessary,~$m> m_{\rho}/2$~e~$m_{R}+m > m_{\rho}$,~where $m_R$ is the
regulator mass. It is worthwhile to observe that the expression for the decay constant is free of zero-mode contributions, which are suppressed by
our particular covariant choice of the vertex function.
\begin{table}[tbh!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Model & $ f_{\rho}~[GeV]$ & $\langle r^2 \rangle~[fm^2]$ & $\mu~[e/2m_{\rho}]$ & $Q_2~[e/m^2_{\rho}]$ \\
\hline
our work & 0.154 & 0.268 & 2.21 & -0.882 \\
\cite{Bakker2002} & 0.134 & 0.296 & 2.10 & -0.910 \\
\cite{Roberts2011} & 0.130 & 0.312 & 2.11 & -0.850 \\
\cite{Bhagwat2008} & 0.207 & 0.540 & 2.01 & -0.410 \\
\hline
\end{tabular}
\caption{$\rho$ meson electroweak static observables. We compare our model
with others ones from the literature~\cite{Bakker2002,Roberts2011,Bhagwat2008}.
In our model the quark mass is 0.430 GeV and the regulator mass is $3$ GeV.
}
\end{center}
\end{table}
\section{Numerical Results}
The input for the numerical calculations of the $\rho$ meson electroweak observables are
the quarks mass and regulator mass. The experimental decay
constant~$f^{exp}_{\rho}=0.152\pm0.008$~GeV~\cite{PDG2014} is fitted
with the Eq.~(\ref{cc26}) for $m=0.430$~GeV and $m_R=3$~GeV. In table I,
we compare our results with other models ~\cite{Bakker2002,Roberts2011,Bhagwat2008}.
The qualitative trend for the static electromagnetic observables seen in the table,
indicates that for increasing charge radius the magnetic moment decreases, while the quadrupole moment increases.
It is suggestive that the contribution of the orbital motion to the magnetic and
quadrupole moments is in part due to relativistic effects present in the quark spin coupling to form the composite state.
It is natural that relativistic effects decrease as the charge radius increases and the quarks slow down.
This general trend is further explored by computing the static observables for
different regulator and quark masses in figures \ref{fig1} and \ref{fig2}.
\vspace{0.5cm}
\begin{figure}[tbh!]
\epsfig{figure=fig21v1.eps,width=7.50cm,height=7.50cm}
\epsfig{figure=fig22v1.eps,width=7.50cm,height=7.50cm}
\begin{center}
\caption{$\rho$ meson electromagnetic radius (left frame)
and magnetic moment (right frame). Variation with the regulator mass with fixed
quark mass (dashed line) and variation with the quark mass with fixed regulator mass (solid line). } \label{fig1}
\end{center}
\end{figure}
In figure \ref{fig1}, we show the dependence of the charge radius and
magnetic moment with the regulator and quark masses. We varied each mass
independently while keeping the other one fixed. In the right frame of figure~\ref{fig1},
we observe that the electromagnetic radius
decreases by increasing any of the mass scales. The effect is sharp for variations
in the quark mass, because in our model the quarks forms a
bound state, with minimum quark mass of $m_\rho/2$, when the meson is
dissolved in the continuum. Towards this limit the size grows to infinity and $\mu\to 2$, as clearly seen
in the figure. The results for the quadrupole moment presented in figure \ref{fig2} corroborate the
general trend already discussed together
with table I. As the mass scales decrease the quadrupole moment tends to have its magnitude decreased.
\begin{figure}[tbh!]
\begin{center}
\epsfig{figure=fig3v1.eps,width=7.50cm,height=7.50cm}
\end{center}
\caption{Quadrupole momentum of the $\rho$ meson.
Variation with the regulator mass with fixed quark mass (dashed line) and variation with the quark mass with fixed regulator mass (solid line).} \label{fig2}
\end{figure}
\section{Summary}
In the present work, we adopted the prescription proposed in
Refs.~\cite{Inna84,Inna89} to compute the $\rho$-meson electromagnetic form factors. The matrix element
$I^+_{00}$ in the light-front spin basis is not used in the calculation of the form factors, and due to that the contribution of
non-valence or zero-mode terms are not present in these observables \cite{Choi2004,Pacheco2012}.
Using the Bethe-Salpeter amplitude model for the $\rho$ meson proposed in~\cite{Pacheco97},
we calculate the decay constant, charge radius, magnetic and quadrupole moments.
We perform a quantitative analysis of the static electromagnetic observables with the variation of the model parameters, namely
the constituent quark and regulator masses. The results obtained for the charge radius in
the present model are in agreement with those expected for a composite state of nonrelativistic constituent quarks:
the radius decreases by increasing the quark mass, the same behaviour is also found when the regulator mass increases.
The values of constituent quark mass~$m=0.430$~GeV and regulator
mass~$m_R=3.0$~GeV give $f_{\rho}=0.154$~GeV~(see also table I), which is close
to the experimental decay constant of $f^{exp}_{\rho}=0.152 \pm 0.008$~GeV \cite{PDG2014}.
In this case we compare our results for the charge radius, magnetic and quadrupole moments to the outcome of other models found in the
literature~\cite{Bakker2002, Roberts2011, Bhagwat2008}. We note that the magnetic moment is pretty much close to 2 in all calculations,
while a larger variation is found for the decay constant, charge radius and quadrupole moment. The magnetic moment is more robust to model variations, as it approaches the expected sum of individual quark magnetic moments, while the other observables depends crucially on the spatial distribution of the quarks, which is more sensitive to the details of the model.
{\bf Acknowledgments.}
This work was supported in part by the
Brazilian agencies FAPESP (Funda\c{c}\~ao de Amparo \`a Pesquisa do
Estado de S\~ao Paulo), CNPq (Conselho Nacional de
Desenvolvimento Cient\'\i fico e Tecnol\'ogico) and CAPES
(Coordena\c c \~ao de Aperfei\c c oamento de Pessoal de
N\'ivel Superior). Also, we thank the {\it Centro
Nacional de Supercomputa\c c \~ao~(CESUP),
~Universidade Federal do Rio Grande do Sul} by the
use of computacional facilities.
|
\section{Introduction}
The simple gravity pendulum, a massless rod with a fixed end and a mass attached to the free end, is one of the simplest integrable models. However, it provides a useful didactic system that may be available in most students laboratories to be used in different levels of physics \cite{NelsonOlssonAJP1986}. Besides, this physical model serves as the basis for investigating many different phenomena which exhibit a variety of motions including chaos \cite{BakerBlackburn2005}, and has the basic form that arises in resonance problems \cite{FerrazMello2007}. The case of small oscillations about the stable equilibrium position is customarily studied with linearized dynamics, and was for a long time the basis for implementing traditional timekeeping devices.
\par
The dynamical system is only of one degree of freedom, but the motion may evolve in different regimes, and one must resort to the use of special functions to express its general solution in closed form \cite{Ochs2011}. In fact, this dynamical model is commonly used to introduce the Jacobi elliptic functions \cite{Brizard2009}.
\par
The traditional integration provides the period of the motion as a function of the pendulum's length and the initial angle \cite{LimaAJP2010}. Alternatively, the solution can be computed by Hamiltonian reduction, a case in which action-angle variables play a relevant role \cite{Brizard2013}. In particular, they are customarily accepted as the correct variables for finding approximate solutions of almost integrable problems by perturbation methods \cite{Arnold1989}.
\par
When the closed-form solution of an integrable problem is expressed in terms of standard functions, the transformation to action-angle variables can be made explicitly in closed form, as, for instance, in the case of the harmonic oscillator. However, if the solution relies on special functions, whose evaluation will depend on one or more parameters in addition to the function's argument, the action-angle variables approach may provide the closed-form solution in an implicit form. This fact does not cause trouble when evaluating the solution of the integrable problem, but may deprive this solution of some physical insight. Besides, in usual perturbation methods the disturbing function must be expressed in the action-angle variables of the integrable problem, thus making necessary to expand the (implicit) transformation to action-angle variables as a Fourier series in the argument of the special functions. These kinds of expansions are not trivial at all, and finding them may be regarded as a notable achievement \cite{Sadov1970,SadovRuso1970,Kinoshita1972}.
\par
In the case of elliptic functions, the normal way of proceeding is to replace them by their definitions in terms of Jacobi theta functions, which in turn are replaced by their usual Fourier series expansion in trigonometric functions of the elliptic argument, with coefficients that are powers of the elliptic nome \cite{ByrdFriedman1971,Lawden1989}. This laborious procedure is greatly complicated when the modulus of the elliptic function remains as an implicit function of the action-angle variables, a case that requires an additional expansion and the series reversion of the resulting power series.
\par
On the other hand, the expansion of the transformation to action-angle variables can be constructed directly.
Indeed, the assumption that the transformation equations of the solution are given by Taylor series expansions, with the requirement that this transformation reduces the pendulum Hamiltonian to a function of only the momenta, and the constrain that the transformation be canonical leads naturally to the Lie transforms procedure \cite{Hori1966,Deprit1969}. The procedure is illustrated here for the rotation regime of the simple pendulum \cite{PercivalRichards1982,SussmanWisdom2001}. Indeed, after rearranging the Lie transforms solutions as a Fourier series, we check that both kinds of expansions match term for term. Therefore, there is no need of making any numerical experiment to show the convergence of the Lie transforms solution for the pendulum's rotation regime \citep{SussmanWisdom2001}.
\par
In the case of the pendulum's oscillation regime the series expansion of the closed form solution as a Fourier series in the action-angle variables still can be done. However, these expansions provide a relation between an angle whose oscillations are constrained to a maximum elongation, and an angle that rotates, for this reason not fulfilling the condition required by the Lie transforms method that original and transformed variables are the same when the small parameter vanishes. On the contrary, the Lie transforms solution for the oscillation regime is commonly approached after reformulating the pendulum Hamiltonian as a perturbed harmonic oscillator by making use of a preliminary change of variables to Poincar\'e canonical variables. This case is well documented in the literature \cite{LichtenbergLieberman1992,PercivalRichards1982} and is no tackled here.
\section{Hamiltonian reduction of the simple pendulum}
The Lagrangian of a simple pendulum of mass $m$ and length $l$ under the only action of the gravity acceleration $g$ is written
$\mathcal{L}=T-V$ where, noting $\theta$ the angle with respect to the vertical direction, $V=mgl(1-\cos\theta)$ is the potential energy
and $T=(1/2)I\omega^2$ is the kinetic energy, where $I=ml^2$ is the pendulum's moment of inertia and $\omega=\dot\theta$, where the over
dot means derivation with respect to time.
\par
The conjugate momentum to $\theta$ is given by
\[
\Theta=\frac{\partial\mathcal{L}}{\partial\dot\theta}=ml^2\dot\theta
\]
That is, $\Theta=I\omega$ is the angular momentum.
Hence, the usual construction of the Hamiltonian $\mathcal{H}=\Theta\,\dot\theta(\Theta)-\mathcal{L}(\theta,\dot\theta(\Theta))$ gives
\begin{equation} \label{pendulum:ham}
\mathcal{H}=\frac{1}{2}\frac{\Theta^2}{ml^2}+mgl(1-\cos\theta),
\end{equation}
which represents the total energy for given initial conditions $\mathcal{H}(\theta_0,\Theta_0)=E$. Depending on the energy value the
pendulum may evolve in three different regimes:
\begin{itemize}
\item $0\le{E}<2mgl$, the oscillation regime, with a fixed point of the elliptic type at $E=0$ ($\Theta=0$, $\theta=0$).
\item $E=2mgl$ $\Rightarrow$ $\Theta=\pm2ml^2\sqrt{g/l}\cos(\theta/2)$, the separatrix, with fixed points of the hyperbolic type
$\Theta=0$, $\theta=\pm\pi$.
\item $E>2mgl$, the rotation regime
\end{itemize}
\par
\subsection{Phase space}
For each energy manifold $\mathcal{H}=E$, the Hamiltonian (\ref{pendulum:ham}) has a geometric interpretation as a parabolic cylinder
$\mathcal{P}_E=\{ (x,y,z) \,|\, z^2-2y=h\}$, with
\[
x=l\sin\theta, \qquad y=l\cos\theta, \qquad z=\frac{\Theta}{mlg^{1/2}},
\]
and
\[
h=\frac{2E}{mg}-2l=\mathrm{const.}
\]
\par
Besides, the constraint $x^2+y^2=l^2$ makes that the phase space of the simple pendulum is realized by the intersection of parabolic
cylinders, given by the different energy levels of the Hamiltonian (\ref{pendulum:ham}), with the surface of the cylinder
$\mathcal{C}=\{ (x,y,z) \,|\, x^2+y^2=l^2\}$ of radius $l$. This geometric interpretation of the phase space is illustrated in
Fig.~\ref{f:pendulum:cylinder}.
\par
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.75]{cylinderP.jpg}\quad\includegraphics[scale=0.75]{cylinderL.jpg}
\caption{Geometric construction of the pendulum's phase space. Parabolic cylinders $\mathcal{P}_E=\{ (x,y,z) \,|\, z^2-2y=\mathrm{const.}\}$.}
\label{f:pendulum:cylinder}
\end{figure}
Alternatively, typical trajectories on the cylinder are displayed by means of simple contour plots of the Hamiltonian (\ref{pendulum:ham}),
as illustrated in Fig.~\ref{f:pendulum:phsp} where the trajectories are traveled from left to right for positive heights ($\Theta>0$) and
from right to left for negative heights ($\Theta<0$).
\begin{figure}[htbp]
\centering
\includegraphics[scale=1.2]{PendulumPhaseSpace}
\caption{Phase space of the simple pendulum ($q=\sqrt{2m^2gl^3}$).}
\label{f:pendulum:phsp}
\end{figure}
\par
The traditional solution to the motion of the pendulum is approached by the direct integration of Hamilton equations
\[
\dot\theta=\frac{\partial\mathcal{H}}{\partial\Theta}=\frac{\Theta}{ml^2},
\qquad
\dot\Theta=-\frac{\partial\mathcal{H}}{\partial\theta}=-mlg\sin\theta,
\]
which are usually written as a single, second order differential equation
\begin{equation} \label{pendulum:flow}
\ddot\theta+\frac{g}{l}\sin\theta=0.
\end{equation}
\par
Note that the summand $mgl$ in Eq.~(\ref{pendulum:ham}) is a constant term that does not affect the dynamics derived from Hamilton equations.
Hence, this term is sometimes neglected, which implies a trivial displacement of the energy by a constant level. Besides, the Hamiltonian
can be scaled by the pendulum's moment of inertia to show that it only depends on a relevant parameter. Nevertheless, the dimensional
parameters are maintained in following derivations, because, in addition to the physical insight that they provide, the simple and immediate
test of checking dimensions is very useful in verifying the correctness of the mathematical developments.
\par
\subsection{Hamiltonian reduction}
Alternatively to the classical integration of Eq.~(\ref{pendulum:flow}), the flow can be integrated by Hamiltonian reduction, finding a
transformation of variables
\begin{equation} \label{pendulum:transformation}
\mathcal{T}:(\theta',\Theta')\rightarrow(\theta,\Theta),
\end{equation}
such that the Hamiltonian (\ref{pendulum:ham}) when expressed in the new variables is only a function of the new momentum. Namely,
\begin{equation} \label{pendulum:reduction}
\mathcal{T}:\mathcal{H}\equiv\mathcal{H}(\theta(\theta',\Theta'),\Theta(\theta',\Theta'))=\Phi(-,\Theta').
\end{equation}
The solution of Hamilton equations in the new variables
\[
\dot\theta'=\frac{\partial\Phi}{\partial\Theta'},
\qquad
\dot\Theta'=-\frac{\partial\Phi}{\partial\theta'}=0,
\]
is trivial
\begin{equation} \label{pendulum:sol}
\Theta'=\mathrm{const},
\qquad
\theta'=\theta'_0+\omega't,
\end{equation}
where the frequency $\omega'=\partial\Phi/\partial\Theta'$ only depends on $\Theta'$ and, therefore, is constant.
Plugging Eq.~(\ref{pendulum:sol}) into the transformation $\mathcal{T}$ in Eq.~(\ref{pendulum:transformation}) will give the time solution
in the original phase space $(\theta,\Theta)$.
\par
The Hamiltonian reduction in Eq.~(\ref{pendulum:reduction}) can be achieved by the Hamilton-Jacobi method \cite{GoldsteinPooleSafko2001},
in which the canonical transformation $\mathcal{T}$ is derived from a generating function $\mathcal{S}=\mathcal{S}(\theta,\Theta')$ in
mixed variables, the ``old'' coordinate and the ``new'' momentum, such that the transformation is given by
\begin{equation} \label{pendulum:canonicalpQ}
\theta'=\frac{\partial\mathcal{S}}{\partial\Theta'},\qquad
\Theta=\frac{\partial\mathcal{S}}{\partial\theta}.
\end{equation}
\par
The functional expression of $\Theta$, given by the last of Eq.~(\ref{pendulum:canonicalpQ}) is replaced in the
Hamiltonian (\ref{pendulum:ham}) to give the partial differential equation
\begin{equation} \label{pendulum:PDE}
\frac{1}{2ml^2}\left(\frac{\partial\mathcal{S}}{\partial\theta}\right)^{\!2}+2mgl\sin^2(\theta/2)=\Phi(\Theta'),
\end{equation}
where the trigonometric identity $1-\cos\theta=2\sin^2(\theta/2)$ has been used.
\par
The Hamilton-Jacobi equation (\ref{pendulum:PDE}) is a partial differential equation that can be solved by quadrature
\begin{equation} \label{pendulum:S}
\mathcal{S}=\sqrt{2ml^2}\int\sqrt{\Phi(\Theta')-2mgl\sin^2(\theta/2)}\,\mathrm{d}\theta.
\end{equation}
But, in fact, the generating function $\mathcal{S}$ does not need to be computed because the transformation (\ref{pendulum:canonicalpQ})
only requires the partial derivatives of Eq.~(\ref{pendulum:S}). Thus,
\begin{eqnarray}
\label{pendulum:thetap}
\theta'&=&
\sqrt{2ml^2}\,\frac{\partial\Phi}{\partial\Theta'}\int_{0}^\theta\frac{\mathrm{d}\vartheta}{2\sqrt{\Phi(\Theta')-2mgl\sin^2(\vartheta/2)}},
\\
\label{pendulum:Theta}
\Theta &=& \sqrt{2ml^2}\sqrt{\Phi(\Theta')-2mgl\sin^2(\theta/2)},
\end{eqnarray}
where $\vartheta=0$ corresponds to the upper extreme of the square root in Eq.~(\ref{pendulum:thetap}).
\par
Remarkably, the quadrature in Eq.~(\ref{pendulum:thetap}) can be solved without need of choosing the form of the new Hamiltonian in
advance \cite{SussmanWisdom2001,FerrerLara2010b}.
\par
\subsection{Rotation regime: $\Phi(\Theta')>2mgl$}
Equation (\ref{pendulum:thetap}) is reorganized as
\[
\theta'=\sqrt{\frac{l}{g}}\,k\,\frac{\mathrm{d}\Phi}{\mathrm{d}\Theta'}\,\mathrm{F}(\phi,k^2),
\]
where the partial differentiation has been replaced by the total derivative in view of the single dependency of $\Phi$ on $\Theta'$, and
\[
\mathrm{F}(\phi,k^2)=\int_0^\phi\frac{\mathrm{d}\theta}{\sqrt{1-k^2\sin^2\theta}},
\]
is the incomplete elliptic integral of the first kind of \emph{amplitude}
\begin{equation} \label{pendulum:argument}
\phi=\theta/2,
\end{equation}
and \emph{elliptic modulus}
\begin{equation} \label{pendulum:krotation}
k=\sqrt{\frac{2mgl}{\Phi(\Theta')}}<1.
\end{equation}
\par
Therefore, the rotational motion of the simple pendulum is solved by the transformation in mixed variables
\begin{eqnarray} \label{pendulum:mixed}
\theta' &=& \sqrt{\frac{l}{g}}\,k\,\frac{\mathrm{d}\Phi}{\mathrm{d}\Theta'}\,\mathrm{F}(\phi,k^2),
\\ \label{pendulum:Qc}
\Theta &=& 2m\,l^2\sqrt{\frac{g}{l}}\,\frac{1}{k}\,\sqrt{1-k^2\sin^2\phi},
\end{eqnarray}
which is, in fact, a whole \emph{family} of transformations parameterized by $\Phi$. For instance, a ``simplifying'' option could be
choosing $\Phi\equiv\Theta'^2/(8ml^2)$, which results in
\begin{eqnarray} \label{pendulum:theta0}
\sin\frac{\theta}{2} &=& \mathrm{sn}(\theta',k^2),
\\ \label{pendulum:Theta0}
\frac{\Theta}{2\sqrt{m^2gl^3}} &=& \frac{1}{k}\,\mathrm{dn}(\theta',k^2),
\end{eqnarray}
where $k=4\sqrt{m^2gl^3}/\Theta'$, from Eq.~(\ref{pendulum:krotation}). Note that $\phi=\mathrm{am}(\theta',k^2)$, where the Jacobi
amplitude $\mathrm{am}$ is the inverse function of the elliptic integral of the first kind, and $\mathrm{sn}$ and $\mathrm{dn}$ are the
Jacobi sine amplitude and delta amplitude functions, respectively.
\par
From the definition of $k$ in Eq.~(\ref{pendulum:krotation}), one may express the reduced Hamiltonian in the \emph{standard} form
\begin{equation} \label{pendulum:StandardHam}
\Phi=2mgl\frac{1}{k^2}.
\end{equation}
Then, by simple differentiation,
\begin{equation} \label{pendulum:dPhidThetap}
\frac{\mathrm{d}\Phi}{\mathrm{d}\Theta'}=-4mgl\frac{1}{k^3}\,\frac{\mathrm{d}k}{\mathrm{d}\Theta'}.
\end{equation}
which turns Eq.~(\ref{pendulum:mixed}) into
\begin{equation}
\theta'= -4mgl\sqrt{\frac{l}{g}}\,\frac{1}{k^2}\,\frac{\mathrm{d}k}{\mathrm{d}\Theta'}\,\mathrm{F}(\phi,k^2).
\end{equation}
\section{Action-angle variables}
Note that, in general, the new variables $(\theta',\Theta')$ defined by the mixed transformation in Eqs.~(\ref{pendulum:mixed})
and (\ref{pendulum:Qc}) may be of a different nature than the original variables $(\theta,\Theta)$: an angle and angular momentum,
respectively. Indeed, the new Hamiltonian choice leading to Eqs.~(\ref{pendulum:theta0})--(\ref{pendulum:Theta0}) assigns $\theta'$ the
dimension of time. However, choosing a transformation that preserves the dimensions of the new variables $(\theta',\Theta')$ as angle and
angular momentum ---the so-called transformation to action-angle variables--- may provide a deeper geometrical insight of the solution and,
furthermore, is specially useful in perturbation theory. Alternative derivations to the one given below can be found in the literature \cite{FerrazMello2007}.
\par
Hence, the requirement that $\theta'=\theta'(\theta,\Theta)$ be an angle, given by the condition
$\oint\mathrm{d}\theta'=2\pi$ \cite{Arnold1989}, is imposed to Eq.~(\ref{pendulum:mixed}). That is, when $\theta$ completes a period along
an energy manifold $\mathcal{H}(\theta,\Theta)=E$ then $\theta'$ must vary between $\theta'_0$ and $\theta'_0+2\pi$.
\par
Note in Eq.~(\ref{pendulum:argument}) that when $\theta$ evolves between $0$ and $2\pi$, $\phi=\phi(\theta)$ evolves between $0$ and $\pi$.
Therefore, using Eq.~(\ref{pendulum:mixed}), the angle condition reads
\[
2\pi= \sqrt{\frac{l}{g}}\,k\,\frac{\mathrm{d}\Phi}{\mathrm{d}\Theta'}\left[\mathrm{F}(\pi,k^2)-\mathrm{F}(0,k^2)\right],
\]
where
\[
\mathrm{F}(0,k^2)=0,\qquad \mathrm{F}(\pi,k^2)=2\mathrm{F}(\pi/2,k^2)=2\mathrm{K}(k^2).
\]
Hence,
\begin{equation}\label{pendulum:PhidP}
\frac{\mathrm{d}\Phi}{\mathrm{d}\Theta'}=\sqrt{\frac{g}{l}}\,\frac{\pi}{k\,\mathrm{K}(k^2)},
\end{equation}
which is replaced in Eq.~(\ref{pendulum:mixed}) to give
\begin{equation} \label{ppendulum:thetap}
\theta'=\frac{\pi}{\mathrm{K}(k^2)}\,\mathrm{F}(\phi,k^2).
\end{equation}
\par
On the other hand, by eliminating $\mathrm{d}\Phi/\mathrm{d}\Theta'$ between Eqs.~(\ref{pendulum:dPhidThetap}) and (\ref{pendulum:PhidP})
results in separation of variables. Then, $\Theta'$ is solved by quadrature to give
\begin{equation} \label{ppendulum:pc}
\Theta'=\frac{4}{\pi}\,m\,l^2\,\sqrt{\frac{g}{l}}\,\frac{1}{k}\,\mathrm{E}(k^2),
\end{equation}
where $\mathrm{E}(k^2)=\mathrm{E}(\pi/2,k^2)$ is the \emph{complete} elliptic integral of the second kind, and
\begin{equation} \label{ppendulum:ellipticE}
\mathrm{E}(\phi,k^2)=\int_0^\phi\sqrt{1-k^2\sin^2\theta}\,\mathrm{d}\theta,
\end{equation}
is the \emph{incomplete} elliptic integral of the second kind.
\par
It deserves noting that the elliptic modulus cannot be solved explicitly from Eq.~(\ref{ppendulum:pc}), and hence the reduced
Hamiltonian $\Phi$ must remain as an implicit function of $\Theta'$ in the standard form of Eq.~(\ref{pendulum:StandardHam}). However, the
rotation frequency of $\theta'$ is trivially derived from Hamilton equations by using Eq.~(\ref{pendulum:PhidP}), namely
\begin{equation} \label{pendulum:nr}
n_{\theta'}=\frac{\mathrm{d}\theta'}{\mathrm{d}t}=\frac{\mathrm{d}\Phi}{\mathrm{d}\Theta'}=\sqrt{\frac{g}{l}}\,\frac{\pi}{k\,\mathrm{K}(k^2)}.
\end{equation}
\par
In summary, starting from $(\theta,\Theta)$, corresponding action-angle variables $(\theta',\Theta')$ are computed from the algorithm:
\begin{enumerate}
\item compute $E=\mathcal{H}(\theta,\Theta)$ from Eq.~(\ref{pendulum:ham});
\item compute $k$ from Eq.~(\ref{pendulum:krotation}) with $\Phi=E$;
\item compute $\Theta'$ from Eq.~(\ref{ppendulum:pc}), and $\theta'$ from Eq.~(\ref{ppendulum:thetap}), where $\phi=\theta/2$.
\end{enumerate}
\par
On the other hand, starting from $(\theta',\Theta')$, the inverse transformation is computed as follows.
\begin{enumerate}
\item solve the implicit equation (\ref{ppendulum:pc}) for $k$;
\item invert the elliptic integral of the first kind in Eq.~(\ref{ppendulum:thetap}) to compute
\[
\phi=\mathrm{am}\!\left(\pi^{-1}\mathrm{K}(k^2)\,\theta',k^2\right);
\]
\item $\theta=2\phi$, and $\Theta$ is evaluated from Eq.~(\ref{pendulum:Qc}).
\end{enumerate}
\section{Series expansions}
In spite of the evaluation of elliptic functions is standard these days, using approximate expressions based on trigonometric functions
will provide a higher insight into the nature of the solution and may be accurate enough for different applications.
\par
Thus, Eqs.~(\ref{ppendulum:thetap}) and (\ref{pendulum:Qc}) are written
\begin{eqnarray} \label{pendulum:qk}
\theta &=& 2\,\mathrm{am}(u,k^2),
\\ \label{pendulum:Qk}
\Theta &=& 2ml^2\sqrt{\frac{g}{l}}\,\frac{1}{k}\,\mathrm{dn}(u,k^2),
\end{eqnarray}
which is formally the same transformation as the one in Eqs.~(\ref{pendulum:theta0}) and (\ref{pendulum:Theta0}), except for the argument
\[
u\equiv\mathrm{F}(\phi,k^2)=\frac{\mathrm{K}(k^2)}{\pi}\theta',
\]
which now fulfils the requirement that $\theta'$ be an angle.
\par
The elliptic functions in Eqs.~(\ref{pendulum:qk}) and (\ref{pendulum:Qk}) are expanded using standard relations,\footnote{See \texttt{http://dlmf.nist.gov/22.11E3} and \texttt{http://dlmf.nist.gov/22.16.E9} for Eqs.~(\ref{pendulum:dns}) and (\ref{pendulum:ams}), respectively.}
\begin{eqnarray} \label{pendulum:dns}
\mathrm{dn}(u,k^2) &=&
\frac{\pi}{\mathrm{K}(k^2)}\left[\frac{1}{2}+2\sum_{n=1}^{\infty}\frac{q^{n}}{1+q^{2n}}\cos(n\theta')\right]\!,\quad
\\ \label{pendulum:ams}
\mathrm{am}(u,k^2) &=&
\frac{1}{2}\theta'+2\sum_{n=1}^{\infty}\frac{q^{n}}{n(1+q^{2n})}\sin(n\theta'),
\end{eqnarray}
where the \emph{elliptic nome} $q$ is defined as
\begin{equation} \label{pendulum:nome}
q= \exp\left[-\pi\,\mathrm{K}(1-k^2)/\mathrm{K}(k^2)\right],
\end{equation}
and $k$ must be written explicitly in terms of $\Theta'$. This is done by expanding $\Theta'$ as given by Eq.~(\ref{ppendulum:pc}) in power
series of $k$, viz.
\begin{eqnarray*}
\Theta' &=& \frac{2\sqrt{g l^3 m^2}}{k}\left[
1-\frac{1}{4}k^2-\frac{3}{64}k^4-\frac{5}{256}k^6-\frac{175}{16384}k^8 \right. \\ && \left.
-\frac{441}{65536}k^{10}-\frac{4851}{1048576}k^{12}+O(k^{13})\right].
\end{eqnarray*}
Then, by series reversion,
\begin{equation} \label{pendulum:kThetap}
k=2\sqrt{\epsilon}\left[1-\epsilon
+\frac{5}{4}\epsilon^2
-\frac{7}{4}\epsilon^3
+\frac{161}{64}\epsilon^4
-\frac{239}{64}\epsilon^5 +O(\epsilon^6)
\right],
\end{equation}
where the notation
\begin{equation} \label{pendulum:epsilonThetap}
\epsilon=\frac{m^2gl^3}{\Theta'^2},
\end{equation}
has been introduced. Needles to say that $\epsilon$ must be small in order to the expansions make sense.
\par
Once the required operations have been carried out, the transformation to action-angle variables given by Eqs.~(\ref{pendulum:qk})
and (\ref{pendulum:Qk}) is written as the expansion
\begin{eqnarray} \nonumber
\theta &=& \theta'
+\left(1+\frac{11}{16}\epsilon^2+\frac{247}{256}\epsilon^4\right)\epsilon\sin\theta' \\ \nonumber
&& +\left(\frac{1}{8}+\frac{3}{16}\epsilon^2\right)\epsilon^2\sin2\theta'
+\left(\frac{1}{48}+\frac{3}{64}\epsilon^2\right)\epsilon^3\sin3\theta' \\ \label{pendulum:qseries}
&& +\frac{1}{256} \epsilon^4 \sin4\theta'
+\frac{1}{1280}\epsilon^5 \sin5\theta' +O(\epsilon^6),
\\ [1ex] \nonumber
\Theta &=& \Theta'\left[
1-\frac{1}{2}\epsilon^2-\frac{15}{32}\epsilon^4
+\left(1+\frac{3}{16}\epsilon^2+\frac{39}{256}\epsilon^4\right)\epsilon\cos\theta' \right. \\ \nonumber
&&
+\left(\frac{1}{4}+\frac{1}{4}\epsilon^2\right)\epsilon^2\cos2\theta'
+\left(\frac{1}{16}+\frac{7}{64}\epsilon^2\right)\epsilon^3\cos3\theta' \\ \label{pendulum:Qseries}
&& \left.
+\frac{1}{64}\epsilon^4\cos4\theta'
+\frac{1}{256}\epsilon^5\cos5\theta'+O(\epsilon^6)
\right].
\end{eqnarray}
\par
Proceeding analogously, the standard Hamiltonian in Eq.~(\ref{pendulum:StandardHam}) is expanded like
\begin{equation} \label{pendulum:hamSeries}
\Phi=\frac{\Theta'^2}{2l^2m}\left(1+\frac{1}{2}\epsilon^2+\frac{5}{32}\epsilon^4+\frac{9}{64}\epsilon^6+\dots\right).
\end{equation}
\par
\section{Action-angle variables expansions by Lie transforms}
In spite of modern algebraic manipulators provide definite help in handling series expansion and reversion, the procedure described in the previous section can be quite awkward. Then, it emerges the question if the expanded solution in Eqs.~(\ref{pendulum:qseries}), (\ref{pendulum:Qseries}) and (\ref{pendulum:hamSeries}) could be constructed directly. The answer is in the affirmative, and dealing with elliptic functions and series reversion can be totally avoided by setting the rotation regime of the pendulum as a perturbed rotor. Indeed, since previous expansions are based in the convergence of Eq.~(\ref{pendulum:kThetap}), the function $\epsilon\equiv\epsilon(\Theta')$ must be small enough.
Hence, neglecting the constant summand $mgl$, Eq.~(\ref{pendulum:ham}) can be written as
\begin{equation} \label{pendulum:hame}
\mathcal{H}=\frac{\Theta^2}{2ml^2}\left(1-\frac{2m^2gl^3}{\Theta^2}\cos\theta\right),
\end{equation}
and in those cases in which $2m^2gl^3/\Theta^2\ll1$, that is
\[
mgl\ll\frac{\Theta^2}{2ml^2}
\]
the reduction of the Hamiltonian (\ref{pendulum:hame}) can be achieved as follows.
\par
\subsection{Deprit's triangle}
First, Eq.~(\ref{pendulum:hame}) is written
\[
\mathcal{H}=\frac{\Theta^2}{2ml^2}-\kappa\cos\theta,
\]
where the physical parameter $\kappa=mgl$ is small.\footnote{In fact, one can always choose the units so that $\Theta^2/(2ml^2)$ is of order 1 and $\kappa\ll1$.} Then, the pendulum Hamiltonian can be written as the Taylor series
\begin{equation} \label{pendulum:HTS}
\mathcal{H}=\sum_{n\ge0}\frac{\kappa^n}{n!}H_{n,0}(\theta,\Theta),
\end{equation}
with the coefficients
\begin{eqnarray}
H_{0,0} &=& \frac{\Theta^2}{2ml^2}, \\
H_{1,0} &=& -\cos\theta, \\ \label{Hn0}
H_{n,0} &=& 0 \quad (n>1).
\end{eqnarray}
The reasons for using a double subindex notation will be apparent soon.
\par
The desired transformation $T$ is formally written
\begin{eqnarray*}
\theta &=& \theta(\theta',\Theta';\kappa), \\
\Theta &=& \Theta(\theta',\Theta';\kappa),
\end{eqnarray*}
which is assumed to be analytic.
Then, when this transformation is applied to Eq.~(\ref{pendulum:hame}), one gets
\[
\Psi\equiv\mathcal{H}(\theta(\theta',\Theta';\kappa),\Theta(\theta',\Theta';\kappa);\kappa),
\]
which is expanded as a Taylor series in the new variables
\begin{equation} \label{psi}
\Psi=\sum_{n\ge0}\frac{\kappa^n}{n!}H_{0,n}, \qquad
H_{0,n}=\left.\frac{\mathrm{d}^n\Psi}{\mathrm{d}\kappa^n}\right|_{\kappa=0},
\end{equation}
where the required derivatives with respect to the small parameter $\kappa$ are computed by the chain rule
\begin{equation} \label{chainrule}
\frac{\mathrm{d}\Psi}{\mathrm{d}\kappa}=
\frac{\partial\mathcal{H}}{\partial\theta}\frac{\mathrm{d}\theta}{\mathrm{d}\kappa}
+\frac{\partial\mathcal{H}}{\partial\Theta}\frac{\mathrm{d}\Theta}{\mathrm{d}\kappa}
+\frac{\partial\mathcal{H}}{\partial\kappa}
\end{equation}
We want the transformation to be canonical and explicit, and hence impose that $T$ is derived from some generating function $\mathcal{W}\equiv\mathcal{W}(\theta,\Theta;\kappa)$. In particular, it is imposed that $T$ be a Lie transform, so that it is defined by the solution of the differential system \cite{LichtenbergLieberman1992,MeyerHallOffin2009}
\begin{equation} \label{lds}
\frac{\mathrm{d}\theta}{\mathrm{d}\kappa}=\frac{\partial\mathcal{W}}{\partial\Theta},
\qquad
\frac{\mathrm{d}\Theta}{\mathrm{d}\kappa}=-\frac{\partial\mathcal{W}}{\partial\theta},
\end{equation}
for the initial conditions $\kappa=0$,
\begin{eqnarray*}
\theta_0 &=& \theta(\theta',\Theta';0)=\theta', \\
\Theta_0 &=& \Theta(\theta',\Theta';0)=\Theta'.
\end{eqnarray*}
The existence of the solution of Eq.~(\ref{lds}) in a neighborhood of the origin, is guaranteed by the theorem of existence and uniqueness
of the ordinary differential equations. Besides, since the transformation is computed as the solution of a Hamiltonian system, its
canonicity is also guaranteed.
\par
Then, taking into account Eq.~(\ref{lds}), Eq.~(\ref{chainrule}) is rewritten as
\begin{equation} \label{chainpoiss}
\frac{\mathrm{d}\Psi}{\mathrm{d}\kappa}=
\frac{\partial\mathcal{H}}{\partial\theta}\frac{\partial\mathcal{W}}{\partial\Theta}
-\frac{\partial\mathcal{H}}{\partial\Theta}\frac{\partial\mathcal{W}}{\partial\theta}
+\frac{\partial\mathcal{H}}{\partial\kappa}
=\{\mathcal{H},\mathcal{W}\}+\frac{\partial\mathcal{H}}{\partial\kappa},
\end{equation}
where $\{\mathcal{H},\mathcal{W}\}$ is the Poisson bracket of $\mathcal{H}$ and $\mathcal{W}$.
Replacing $\mathcal{W}$ in Eq. (\ref{chainpoiss}) by its Taylor series expansion
\begin{equation}\label{ltW}
\mathcal{W}=\sum_{n\ge0}\frac{\kappa^n}{n!}\,{W}_{n+1}(\theta,\Theta),
\end{equation}
where the subindex $n+1$ is now used for convenience in subsequents derivations of the procedure, after straightforward manipulations, one arrives to the recurrence
\begin{equation} \label{ltdeprittriangle}
H_{n,q+1}=H_{n+1,q}+
\sum_{0\le{i}\le{n}}{n\choose{i}}\,\{H_{n-i,q};{W}_{i+1}\}.
\end{equation}
Finally, since the terms $H_{0,n}$ needed in Eq.~(\ref{psi}) are evaluated at $\kappa=0$, the original variables are replaced by corresponding prime variables. Full details on the derivation of Eq.~(\ref{ltdeprittriangle}), may be found in Deprit's original reference \cite{Deprit1969} or in modern textbooks in celestial mechanics.
\par
The recurrence in Eq.~(\ref{ltdeprittriangle}) is customarily known as Deprit's triangle. Indeed, the computation of the term $H_{0,n}$ requires the computation of all intermediate terms in a ``triangle'' of ``vertices'' $H_{0,0}$, $H_{0,n}$, and $H_{n,0}$. For instance, for $n=4$ the recurrence in Eq.~(\ref{ltdeprittriangle}) results in a triangle made of:
\begin{center}
\begin{tabular}{lllllllllll}
& & & & ${H}_{0,0}$ \\[1.5ex]
& & & ${H}_{1,0}$ & & ${H}_{0,1}$ \\[1.5ex]
& & ${H}_{2,0}$ & & ${H}_{1,1}$ & & ${H}_{0,2}$ \\[1.5ex]
& ${H}_{3,0}$ & & ${H}_{2,1}$ & & ${H}_{1,2}$ & & ${H}_{0,3}$ \\[1.5ex]
${H}_{4,0}$ & & ${H}_{3,1}$ & & ${H}_{2,2}$ & & ${H}_{1,3}$ & & ${H}_{0,4}$ \\
\end{tabular}
\end{center}
\par
Deprit's triangle is not constrained to the case of one degree of freedom Hamiltonians and generally applies for the transformation of any function $F\equiv{F}(x,X;\kappa)$ where $x\in\mathrm{I\!\!R}^n$, $X\in\mathrm{I\!\!R}^n$ are coordinates and their conjugate momenta, respectively. In particular, when the generation function is given by its Taylor series expansion, it is easily checked that the solution of the differential system (\ref{lds}) is equivalent to applying Deprit's triangle to the functions
\begin{equation} \label{teqs}
\theta=\sum_{n\ge0}\frac{\kappa^n}{n!}\theta_{n,0}(\theta,\Theta), \qquad
\Theta=\sum_{n\ge0}\frac{\kappa^n}{n!}\Theta_{n,0}(\theta,\Theta),
\end{equation}
where $\theta_{0,0}=\theta$, $\Theta_{0,0}=\Theta$, and $\theta_{n,0}=\Theta_{n,0}=0$ for $n>0$.
\par
\subsection{Perturbations by Lie transforms}
Deprit's triangle provides an efficient and systematic way of computing higher orders of a transformation when the generating function is known. However, it is precisely the generating function that is needed to be computed with the only condition that the transformation reduces the pendulum Hamiltonian to a function of only the (new) momentum. Therefore, a new Hamiltonian fulfilling this requirement must be also constructed. The whole procedure of constructing the action-angle Hamiltonian and computing the corresponding generating function can be done in an iterative way as follows.
The construction of $\Phi$ is an iterative procedure based on Deprit's triangle. At each order $n$ Eq.~(\ref{ltdeprittriangle}) is
rearranged in the form
\begin{equation}\label{lthomological}
\{{W}_n,H_{0,0}\}+H_{0,n}=\widetilde{H}_n,
\end{equation}
which is dubbed as the \textit{homological} equation, where
\begin{itemize}
\item $\widetilde{H}_n$ is known from previous computations
\item $H_{0,n}$ is chosen in agreement with some simplification criterion such that
\item ${W}_n$ is obtained as a solution of the partial differential equation (\ref{lthomological})
\end{itemize}
Commonly, the simplification criterion for choosing $H_{0,n}$ is eliminating cyclic variables in order to reduce the number of degrees of freedom of the new, transformed Hamiltonian after truncation to $O(\kappa^n)$. This is exactly the case required for the reduction of the pendulum Hamiltonian.
\par
In practice, the procedure starts by reformulating all the functions involved in the homological equations in the prime variables, so that the new Hamiltonian is directly obtained in the correct set of variables.
\par
\subsection{The pendulum as a perturbed rotor}
We start from Eqs.~(\ref{pendulum:HTS})--(\ref{Hn0}). For any term $n$ of the generating function,
\[
\{{W}_n,\mathcal{H}_{0,0}\}=-\frac{\Theta'}{ml^2}\,\frac{\partial{W}_n}{\partial\theta'}.
\]
\par
The first order of the homological equation is
\begin{equation} \label{homo1}
\frac{\Theta'}{ml^2}\,\frac{\partial{W}_1}{\partial\theta'}=H_{0,1}-\widetilde{H}_1,
\end{equation}
with $\widetilde{H}_1=H_{1,0}$. Then, $H_{0,1}$ is chosen by averaging terms depending on the cyclic variable $\theta'$
\[
H_{0,1}=\frac{1}{2\pi}\int_0^{2\pi}\widetilde{H}_1\,\mathrm{d}\theta'=0,
\]
and ${W}_1$ is solved from Eq.~(\ref{homo1}) by quadrature
\[
{W}_1=\frac{ml^2}{\Theta'}\int(H_{0,1}-\widetilde{H}_1)\,\mathrm{d}\theta'=-\frac{ml^2}{\Theta'}\sin\theta'.
\]
\par
Once ${W}_1$ is obtained, the first order of the transformation is calculated by application of Deprit's triangle to Eq.~(\ref{teqs})
with $x\equiv\theta$ and $X\equiv\Theta$. It is obtained
\begin{eqnarray*}
\theta_{0,1} &=& \{\theta',W_1\}=\frac{\epsilon}{\kappa}\sin\theta', \\
\Theta_{0,1} &=& \{\Theta',W_1\}=\Theta'\frac{\epsilon}{\kappa}\cos\theta',
\end{eqnarray*}
where the abbreviation $\epsilon=m^2gl^3/\Theta'^2$ is used, in agreement with Eq.~(\ref{pendulum:epsilonThetap}).
\par
The second order of the homological equation is
\begin{equation} \label{homo2}
\frac{\Theta'}{ml^2}\,\frac{\partial{W}_2}{\partial\theta}=H_{0,2}-\widetilde{H}_2,
\end{equation}
where $\widetilde{H}_2$ is computed from Deprit's triangle
\[
\widetilde{H}_2=\{H_{1,0},{W}_1\}=\frac{ml^2}{\Theta'^2}\sin^2\theta',
\]
$H_{0,2}$ is chosen by averaging
\[
H_{0,2}=\frac{1}{2\pi}\int_0^{2\pi}\widetilde{H}_2\,\mathrm{d}\theta'=\frac{ml^2}{2\Theta'^2}=\frac{\Theta'^2}{2ml^2}\frac{\epsilon^2}{\kappa^2},
\]
and ${W}_2$ is solved from Eq.~(\ref{homo2}) by quadrature
\[
{W}_2=\frac{ml^2}{\Theta'}\int(H_{0,2}-\widetilde{H}_2)\,\mathrm{d}\theta'=-\frac{m^2l^4}{4\Theta'^3}\sin2\theta'.
\]
Then, a new application of Deprit's triangle to the variables provides the second order of the transformation
\begin{eqnarray*}
\theta _{0,2} &=& \{\theta_{0,1},W_1\}+\{\theta',W_2\}=\frac{1}{4}\frac{\epsilon^2}{\kappa^2}\sin2\theta', \\
\Theta _{0,2} &=& \{\Theta_{0,1},W_1\}+\{\Theta',W_2\}=\Theta'\frac{\epsilon^2}{\kappa^2}\left(\frac{1}{2}\cos2\theta'-1\right),
\end{eqnarray*}
\par
New iterations of the procedure lead to the third order
\begin{eqnarray*}
H_{0,3} &=& 0, \\
\theta _{0,3} &=& \frac{\epsilon^3}{\kappa^3}\left(\frac{33}{8}\sin\theta'+\frac{1}{8}\sin3\theta'\right), \\
\Theta _{0,3} &=& \Theta'\frac{\epsilon^3}{\kappa^3}\left(\frac{9}{8}\cos\theta'+\frac{3}{8}\cos3\theta'\right),
\end{eqnarray*}
the fourth order,
\begin{eqnarray*}
H_{0,4} &=& \frac{\Theta'^2}{2ml^2}\,\frac{15}{4}\frac{\epsilon^4}{\kappa^4}, \\
\theta _{0,4} &=& \frac{\epsilon^4}{\kappa^4} \left(\frac{9}{2}\sin2\theta'+\frac{3}{32}\sin4\theta'\right), \\
\Theta _{0,4} &=& \Theta'\frac{\epsilon^4}{\kappa^4}\left(6\cos2\theta'+\frac{3}{8}\cos4\theta'-\frac{45}{4}\right),
\end{eqnarray*}
the fifth order,
\begin{eqnarray*}
H_{0,5} &=& 0, \\
\theta _{0,5} &=& \frac{\epsilon^5}{\kappa^5}\left(\frac{3705}{32}\sin\theta'+\frac{45}{8}\sin3\theta'+\frac{3}{32}\sin5\theta'\right), \\
\Theta _{0,5} &=& \Theta'\frac{\epsilon^5}{\kappa^5}\left(\frac{585}{32}\cos\theta'+\frac{105}{8}\cos3\theta'+\frac{15}{32}\cos5\theta'\right),
\end{eqnarray*}
and so on.
Now, it is simple to check that the new Hamiltonian $\mathcal{K}=\sum_{n\ge0}(\kappa^n/n!)H_{0,n}$ matches Eq.~(\ref{pendulum:hamSeries}). Besides, rearranging the perturbation series as Fourier series
\begin{eqnarray*}
\theta &=& \sum_{n\ge0}\frac{\kappa^n}{n!}\theta_{0,n}(\theta',\Theta')=\sum_{n\ge0}s_{0,n}(\Theta')\sin(n\theta'), \\
\Theta &=& \sum_{n\ge0}\frac{\kappa^n}{n!}\Theta_{0,n}(\theta',\Theta')=\sum_{n\ge0}c_{0,n}(\Theta')\cos(n\theta'),
\end{eqnarray*}
it is simple to check that they match Eqs.~(\ref{pendulum:qseries}) and (\ref{pendulum:Qseries}), respectively.
\section{Conclusion}
Expanding the transformation to action-angle variables of the simple pendulum solution as a Fourier series, is a cumbersome task because of the elliptic integrals and functions on which the solution relies upon, on the one hand, and the implicit character of the transformation, on the other. Alternatively, these Fourier series expansions can be computed by the Lie transforms method. The latter is straightforward, avoids dealing with special functions, provides the transformation series explicitly without need of series reversion, and is amenable of automatic programming by machine, thus allowing one to compute higher orders of the solution.
\par
|
\section{The minimax risk of a noisy coin}
In this appendix we show that the minimax risk of estimating the bias of a \emph{noisy} coin is $O(1/\sqrt N)$ (in contrast to the $O(1/N)$ minimax risk for a noiselessly observed coin), and derive a simple lower bound on it.
Now, suppose a coin with bias $p=\mathrm{Pr}(\mathrm{``heads''})$ is flipped $N$ times and a sequence of binary outcomes $\vec{n} = \{n_k\}$ are recorded. But these observations are unreliable; each outcome is recorded incorrectly with trial-dependent probabilities $\vec{\alpha} = \{\alpha_k\}$ (all taken from the interval $\left[0,\frac12\right)$).
The distribution of the outcomes is
\begin{equation}\label{likelihood function}
\Pr(\vec{n}|p,\vec{\alpha})=\prod_{k=1}^N \Pr(n_k|p,\alpha_k) = \prod_{k=1}^N q_k^{n_k}(1-q_k)^{1-n_k},
\end{equation}
where the probability of observing ``heads'' on trial $k$ is not $p$, but
\begin{equation}
q_k(p) = \alpha_k + p(1-2\alpha_k) = p + \alpha_k\left(1-2p\right).
\end{equation}
We recover a standard noiseless coin when $\alpha_k=0$ for all $k$.
For each prior distribution $\mu(p)$, the estimator with the smallest risk (expected cost) is the \emph{Bayes estimator} for $\mu(p)$, and its risk is the \emph{Bayes risk} of $\mu(p)$. Bayes estimators need not be simple, but because relative entropy is a \emph{Bregman divergence}, the Bayes estimator is always the mean of the posterior distribution (obtained via Bayes' Rule) \cite{blume2010optimal}. The prior with the highest Bayes risk is the \emph{least favorable} prior, and its risk is the minimax risk. Thus, the Bayes risk of any prior is a lower bound for the minimax risk, which suggests an obvious variational approach to bounding the minimax risk by choosing a prior whose risk is high but easy to calculate. Obviously, some priors have very low risk [e.g., $\mu(p) = \delta(p-p_0)$], and provide useless lower bounds. A common approach is to use the uniform (Lebesgue) prior, but for the noisy coin this prior actually has rather low [$O(1/N)$] risk. So instead, we consider the set of \emph{bimodal priors},
\begin{equation}
\pi(p) = \frac{\delta(p-p_0) + \delta(p-p_1)}{2}.
\end{equation}
(Varying the weights yields a slightly less favorable prior, but doesn't change the asymptotic scaling). We will choose $p_0=0$ and $p_1 \approx 1/\sqrt{N}$.
The risk of $\pi$ is given by $\overline{d}(\pi) = [\overline{d}(0) + \overline{d}(p_1)]/2$, and by observing that $\overline{d}(p_1)\geq0$ we obtain the lower bound
\begin{equation}\label{nc_bimodal}
\overline{d}(\pi)\geq \frac 12\overline{d}(0)= \frac12 \mathbb E_{\vec{n}|p=0}[D(0\|\hat p(\vec{n}))],
\end{equation}
where the Bayes estimator is the posterior mean, given by
\begin{equation}
\hat p (\vec{n}) = \frac{p_1\Pr(\vec{n}|p_1)}{\Pr(\vec{n}|p_1)+\Pr(\vec{n}|0)} = \frac{p_1}{1+\Lambda(\vec{n})},
\end{equation}
in terms of the \emph{likelihood ratio}
\begin{equation}
\Lambda(\vec{n}) = \frac{\Pr(\vec{n}|0)}{\Pr(\vec{n}|p_1)}.
\end{equation}
We can lower-bound the relative entropy term by $D(0\|\hat p)=-\log(1-\hat{p}) \geq \hat{p}$, so
\begin{equation}
\overline{d}(0)\geq \mathbb E_{\vec{n}|p=0}[\hat p(\vec{n})] = p_1\mathbb E_{\vec{n}|p=0}\left[\frac{1}{1+\Lambda(\vec n)}\right].
\end{equation}
If we define $\lambda(\vec{n}) = -2\log\Lambda(\vec n)$ and apply Jensen's inequality, we obtain
\begin{equation}
\overline{d}(0) \geq p_1 e^{-\frac12\mathbb E_{\vec{n}|p=0}\left[\lambda(\vec n)\right]}.
\end{equation}
Next, we perform a Taylor expansion of the expectation $E_{\vec{n}|p=0}\left[\lambda(\vec n)\right]$ around $p_1=0$. The derivatives of the likelihood function \eqref{likelihood function} are
\begin{align}
\frac{\partial}{\partial p_1} \log \Pr(\vec{n}|p_1) & = \sum_k \left(\frac{n_k(1-2\alpha_k)}{q_k}-\frac{(1-n_k)(1-2\alpha_k)}{1-q_k}\right),\\
\frac{\partial^2}{\partial p_1^2} \log \Pr(\vec{n}|p_1) & = \sum_k \left(-\frac{n_k(1-2\alpha_k)^2}{q_k^2}-\frac{(1-n_k)(1-2\alpha_k)^2}{(1-q_k)^2}\right).
\end{align}
Evaluating these at $p_1=0$ and taking the expectation $\mathbb E_{\vec{n}|p=o}[n_k] = \alpha_k$, we have
\begin{align}
\mathbb E_{\vec{n}|p=o}\left[\frac{\partial}{\partial p_1} \log \Pr(\vec{n}|p_1)\right]_{p_1=0} & = 0,\\
\mathbb E_{\vec{n}|p=o}\left[\frac{\partial^2}{\partial p_1^2} \log \Pr(\vec{n}|p_1)\right]_{p_1=0} & = \sum_k \frac{(1-2\alpha_k)^2}{\alpha_k(1-\alpha_k)}.
\end{align}
Putting everything together in the Taylor series, we obtain
\begin{equation}\label{Eseries}
\mathbb E_{\vec{n}|p=0}\left[\lambda(\vec n)\right] = \sum_{k=1}^N\frac{(1-2\alpha_k)^2}{\alpha_k(1-\alpha_k)}p_1^2 + O(p_1^3),
\end{equation}
where the $O(p_1^3)$ term does not scale with $N$ w/r.t. the leading order term.
To simplify this quantity, we define the per-sample ``resolution'' $\beta_k$,
$$
\beta_k = \frac{(1-2\alpha_k)^2}{\alpha_k(1-\alpha_k)},
$$
(which, at least for small $\alpha_k$, is approximately $1/\alpha_k$). The expectation value in Eq. \ref{Eseries} can be written concisely in terms of the average $\beta$, $\overline{\beta} = N^{-1}\sum_{k=1}^N{\beta_k}$, as
$$
\mathbb E_{\vec{n}|p=0}\left[\lambda(\vec n)\right]\sim N \overline{\beta} p_1^2.
$$
Finally, we set
\begin{equation}\label{pdef}
p_1 = \frac{1}{\sqrt{\overline{\beta}}}\frac{1}{\sqrt N},
\end{equation}
which ensures that $p_1\to 0$ as $N\to\infty$ and justifies truncating the series expansion Eq. \ref{Eseries} above at leading order. This yields a lower bound on the minimax risk of
\begin{equation}
\overline{d}_{\rm max} \geq \overline{d}(\pi) \geq \frac{1}{2\sqrt{e\overline{\beta}}}\frac1{\sqrt N}.
\end{equation}
It is worth noting that the risk is \emph{not} determined by the average value of $\alpha$ (the per-sample noise probability), but by the average of $\beta$, which behaves roughly like $1/\alpha$. In particular, if any constant fraction of the samples are observed noiselessly, then those samples have $\beta=\infty$, and they dominate the minimax risk -- $\overline{\beta}\to\infty$, and the minimax risk collapses to $O(1/N)$, as is appropriate for a noiseless coin.
\section{Minimax risk for quantum tomography}
In this section, we derive a lower bound for the minimax risk of qubit state estimation using the same framework that we used for the noisy coin. The difficulty in doing this is that a qubit's state space (the Bloch sphere) is more complex than that of a coin -- instead of a single parameter ($p$) there are three ($x,y,z$). However, the minimax risk is dominated by (i) states $\rho$ that are very close to pure, and (ii) errors in estimating the \emph{spectrum} of $\rho$ (rather than errors in its eigenvectors, which contribute much less to the risk). This observation allows us to simplify the analysis greatly by choosing a bimodal prior for the qubit, supported on two states that differ only in their eigenvalues. In this circumstance, each measurement provides information equivalent (in its effect on the final estimate) to a noisy coin flip whose noisyness depends on what was measured (or, most generally, on which outcome was observed). Because we have chosen a very simple prior that is not least favorable, our analysis only guarantees a lower bound. However, it captures the dominant component of the minimax risk, and (for such a simple model) turns out to be surprisingly close to tight.
Suppose we are given $N$ samples of a qubit state $\rho$. The state is drawn from a bimodal prior supported on (i) a pure state $\rho_0 = \proj{\psi}$, and (ii) a slightly more mixed state $\rho_1 = (1-p_1)\proj{\psi} + p_1(1\!\mathrm{l}-\proj{\psi})$:
\begin{equation}\label{qubit_bimodal}
\pi(\rho) = \frac12\left(\delta(\rho-\rho_0) + \delta(\rho-\rho_1)\right).
\end{equation}
We will specify $\ket\psi$ and $p_1\approx1/\sqrt{N}$ later. Each sample is measured in some basis; on the $k$th sample we perform the [orthogonal basis] POVM $\{\ketbra{\phi_k}{\phi_k},\mathbbm 1-\ketbra{\phi_k}{\phi_k}\}$, and list the outcomes as a binary vector $\vec n:=\{n_k\}$.
The likelihood function for a single observation is\begin{align}
\Pr(n_k|\rho_0) &= |\braket{\phi_k}{\psi}|^2=:\alpha_k\\
\Pr(n_k|\rho_1) &= (1-2p_1)|\braket{\phi_k}{\psi}|^2+p_1=(1-2\alpha_k)p_1 +\alpha_k.
\end{align}
These are identical to the likelihoods for the noisy coin. The Bayes estimator is
\begin{equation}\label{qubit_bimodal_bayes}
\hat \rho (\vec{n}) =\frac{[\Pr(\vec{n}|\rho_0)+(1-2p_1)\Pr(\vec{n}|\rho_1)]\ketbra\psi\psi +p_1)\Pr(\vec{n}|\rho_1)1\!\mathrm{l}}{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)}.
\end{equation}
Now, to compute the expected risk, we observe that the Bayes estimate is always of the form $\hat\rho = \alpha \ketbra\psi\psi + \beta \mathbbm 1$, with
\begin{equation}
\alpha = \frac{[\Pr(\vec{n}|\rho_0)+(1-2p_1)\Pr(\vec{n}|\rho_1)]}{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)},\ \beta = \frac{p_1\Pr(\vec{n}|\rho_1)}{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)}.
\end{equation}
and for any such mixture $\sigma = \alpha \ketbra\psi\psi + \beta \mathbbm 1$, the relative entropy can be computed as
\begin{align}
D\left(\ketbra\psi\psi\|\sigma\right) &= - \braopket{\psi}{\log\sigma}{\psi} \\
&= -\braopket{\psi}{\log(\alpha+\beta)\ketbra\psi\psi + \log\beta (\mathbbm 1 - \ketbra\psi\psi)}{\psi}\\
&= -\log(\alpha+\beta).
\end{align}
Thus, in the limit of $p_1\to0$ and $N\to\infty$, the risk \emph{given} that $\rho=\rho_0$ is given by
\begin{align}
D(\rho_0|\hat\rho(\vec{n})) &= -\log\left[\frac{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)\left(1-p_1\right)}{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)}\right]\\
&= -\log\left[1-\frac{p_1\Pr(\vec{n}|\rho_1)}{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)}\right]\\
&= p_1 \frac{\Pr(\vec{n}|\rho_1)}{\Pr(\vec{n}|\rho_0)+\Pr(\vec{n}|\rho_1)} + O(p_1^2).
\end{align}
This is identical to the risk of the noisy coin.
As in Eq. \ref{pdef}, we choose
\begin{equation}
p_1 = \frac{1}{\sqrt{\bar\beta}}\frac{1}{\sqrt N},
\end{equation}
where $\bar\beta$ is defined in Equation \ref{alpha_def} as
\begin{equation*}
\bar \beta = \frac{1}{N}\sum_{k=1}^N{\beta_k} = \frac{1}{N}\sum_{k=1}^N\frac{(1-2\alpha_k)^2}{\alpha_k(1-\alpha_k)}.
\end{equation*}
This yields a near-final lower bound of
\begin{equation}
\overline{d}_{\rm max} \geq r(\pi) \geq \frac{e^{-\frac12}}{2\sqrt{\bar\beta}}\frac1{\sqrt N}.
\end{equation}
To obtain a concrete lower bound on the risk, we must choose $\proj{\psi}$. To ensure that the average resolution $\bar\beta$ is as small as possible near $\ket\psi$, we want $\alpha_k$ to be uniformly large. For the case of a qubit or a rebit, the solution is to pick the state ``furthest away'' from all the measurement axes, which yields $\alpha_k = \frac12\left(1-\frac1{\sqrt D}\right)$ where $D=2$ for a rebit and $D=3$ for a qubit. This yields the simple result $\bar{\beta} = 4/(D-1)$, and therefore
\begin{equation}
\overline{d}_{\rm max} \geq r(\pi) \geq \frac{e^{-\frac12}}{4}\frac{\sqrt{D-1}}{\sqrt{N}}.
\end{equation}
This argument can be extended to any discrete POVM, by choosing $\ket\psi$ so that it is not orthogonal to any effect of the POVM. Then $\alpha_k$ for each $k$ will be lower-bounded by the minimum overlap of $\ket\psi$ with any effect, and $\bar\beta$ will be finite, and so the minimax risk will scale as $1/\sqrt{N}$.
But, by exactly the same argument, the \emph{best} nonadaptive tomographic measurement must be unitarily symmetric. And since it must also be rank-1, it is the Haar-uniform POVM whose effects include all pure states $\proj\phi$ with the unitarily invariant measure. The analysis given so far breaks down for this uniform POVM, because no matter what $\ket\psi$ we choose, the measurement has effects that diagonalize it. The effective ``noise''
\begin{equation}
\Pr(n_k|\rho_0) = |\braket{\phi_k}{\psi}|^2=:\alpha_k
\end{equation}
is distributed uniformly over $[0,1]$. If we attempt to replace the sum
\begin{equation}
\sum_{k=1}^{N} \frac{(1-2\alpha_k)^2}{\alpha_k(1-\alpha_k)}= : N \overline{\beta},
\end{equation}
with its average, then the integral diverges and our lower bound collapses to $\overline{d}\geq0$.
Instead, we observe that the Haar-uniform POVM can be described as a two-step process: (1) choose a Haar uniform-basis, and (2) measure in that basis. So, a tomography experiment involving $N$ samples can be described by a sequence $[\alpha_1,\alpha_2,\ldots \alpha_N]$, in which each $\alpha_k$ is drawn from the uniform distribution over $[0,1]$. The minimax risk is the [probability-weighted] average over all such sequences. We divide them into two subsets: those in which all the $\{\alpha_k\}$ lie in the interval
$$\alpha_k \in \left[\frac{1}{2N},1-\frac{1}{2N}\right]$$
and those in which at least one does not.
The probability that any given $\alpha_k$ lies outside the interval is exactly $1/N$, so \emph{all} of them lie within it with probability
$$p = \left(1-\frac{1}{N}\right)^N \geq \frac{1}{e}.$$
Conditional on all the $\{\alpha_k\}$ lying within the interval, the minimax risk can be lower bounded by integrating $\beta$ over the interval, which yields
\begin{equation}
\overline{\beta} = 2\log N + O(1).
\end{equation}
This happens with probability at least $1/e$, so a lower bound on the minimax risk for the Haar-uniform POVM (as $N\to\infty$) is
\begin{equation}
\overline{d} \geq \frac{1}{(2e)^{3/2}}\frac{1}{\sqrt{N\log N}}.
\end{equation}
\section{Least favorable priors}
The ``Optimization Toolbox'' in MATLAB 2011a, for example, contains a method \texttt{fminimax} which directly solves the optimization problem we are interested in.
However, finding minimax estimators by brute force seems impossibly difficult. There are uncountably many estimators; each one is a density-matrix valued function on the set of all possible datasets. Each estimator's performance is quantified by maximizing its risk profile $\overline{d}(\rho)$ over all density matrices $\rho$. Even computing the maximum risk of a single specified estimator is nontrivial; finding its minimum over the uncountable set of \emph{all} estimators seems intractable.
Fortunately, we have some useful mathematical tools that simplify matters greatly (see, for example, \cite{lehmann1998theory}):
\begin{enumerate}
\item The minimax estimator is \emph{also} the Bayes estimator for some measure. This fact is called \emph{Minimax-Bayes duality}, and the measure in question is called a \emph{least favorable prior} (LFP).
\item Relative entropy is a Bregman divergence (a.k.a. strictly proper scoring rule), and therefore the Bayes estimator for any given measure $\mu$ is Bayesian mean estimation (BME).
\item The least favorable priors for this problem are (empirically) always \emph{discrete}, with a finite number of support points. This is not proven, but it is often the case in similar problems, and is easy to verify numerically for this problem.
\end{enumerate}
Minimax-Bayes duality is enormously helpful, both as a technical tool and as an aid to problem-solving. The reasoning behind this duality is fairly straightforward:
\begin{enumerate}
\item Any estimator involves trade-offs in accuracy, which are quantified by its risk profile $\overline{d}(\rho)$. For example, the constant estimator $\hat{\rho}(D) = \rho_0$ is exceptionally accurate \emph{if} the true state happens to be $\rho_0$! That is, $\overline{d}(\rho_0) = 0$. No other estimator can match its accuracy at $\rho_0$. But there is a price to be paid; $\overline{d}(\rho)$ is dreadfully high for any state $\rho$ that is far from $\rho_0$.
\item Averaging $\overline{d}(\rho)$ over a measure $\mu$ quantifies these tradeoffs. In order to minimize that average, the Bayes estimator for $\mu$ must achieve fairly low expected risk in regions where $\mu$ is concentrated, but can tolerate high risk where $\mu$ is sparse.
\item If we consider the Bayes estimator for a specific measure $\mu_0$, its risk profile $\overline{d}(\rho)$ will typically be non-constant -- so it will have at least one maximum, which we denote $\rho_0$. Now suppose that we modify $\mu_0$ (to $\mu'$) by slightly increasing the probability density around $\rho_0$. The new measure $\mu'$ will have a higher Bayes risk (since $\rho_0$ has higher-than-average risk, and is now slightly more probable). But the Bayes estimator for $\mu'$ will be slightly different as well; it will achieve a \emph{lower} value of $\overline{d}(\rho_0)$ because by increasing the probability of $\rho_0$ we have increased the value of achieving low risk at $\rho_0$.
\item Iterating this process defines a flow -- probability flows towards high-risk states (decreasing their expected risk) and away from low-risk states (increasing their expected risk). Every step in this flow defines a new prior (and its associated Bayes estimator) with higher Bayes risk and lower maximum risk.
\item If $\mu$ is a stationary point of this flow, then the risk profile of its Bayes estimator $\hat{\rho}_\mu(D)$ is: (i) equal to a constant $C$ on the support of $\mu$, and (ii) no greater than $C$ at every point \emph{not} in the support of $\mu$. This estimator is necessarily minimax, because:
\begin{itemize}
\item No estimator can achieve lower \emph{average} risk on $\mu$ (by the definition of Bayes estimator),
\item So no estimator can achieve lower \emph{maximum} risk on the support of $\mu$ (since $\hat{\rho}_\mu(D)$'s risk is constant),
\item And therefore no estimator can achieve lower maximum risk over all states (since ``all states'' is a superset of $\mu$'s support).
\end{itemize}
\item Such a stationary measure can occur in one of two ways. Either $\hat{\rho}_\mu(D)$ has constant risk on \emph{all} states, or $\mu$ is supported on a discrete set. (Because $\overline{d}(\rho)$ is analytic, it cannot be constant over a limited range, which means either it is constant everywhere or it has discrete maxima).
\end{enumerate}
\section{Numerical recipes}
The above argument is the basis for the numerical algorithm we used to compute LFPs, and hence the minimax estimators. Our results were generated using an implementation of Algorithm \ref{alg:kempthorne}, a variant of the one given by Kempthorne \cite{Kempthorne1987Numerical}. Although this algorithm is drastically more efficient than the brute force approach, it is still insufficient to extract the asymptotic form of the scaling for the risk of minimax estimator. In particular, it takes $\sim$week to compute the LFP shown in Figure \ref{fig:2LFPs} using an implementation of Algorithm \ref{alg:kempthorne} in MATLAB 2011a on four 2.2GHz processors.
To remedy this, we devised an efficient Monte Carlo algorithm (Algorithm \ref{alg:mc}). The core insight is that varying only the weights of the prior renders maximizing the Bayes risk a convex optimization problem. The algorithm proceeds by randomly choosing $n$ states according to the Hilbert-Schmidt prior. Then, the Bayes risk is maximized keeping the location of the states fixed. Both upper and lower bounds on the minimax risk can be obtained. If these are not close, then we resample near those points whose weights have not be set to zero and repeat the process.
\begin{figure}[ht]
\centering
\includegraphics[width=.45\textwidth]{LFPprior_rebitN8}
\includegraphics[width=.45\textwidth]{crown.pdf}
\caption{\label{fig:2LFPs}
Here we show the support points of numerical approximations to least favorable priors (LFPs) for $N=16$ ($M=8$) Pauli measurements on a \emph{rebit} (a qubit with the constraint $\expect{\sigma_y}=0$. The weights on these points are not uniform, but we shown a Gaussian kernel density estimate of them on the right. The LFP found using the highly accurate Algorithm \ref{alg:kempthorne} is supported on the large black dots, while the one found using the much faster Algorithm \ref{alg:mc} is supported on the smaller gray dots. Note that in this case (and all others where we could use Algorithm \ref{alg:kempthorne}), while the LFPs are evidently different, the resulting minimax risks are indistinguishable. We conclude that the maximum risk is insensitive to certain visible variations in the prior.}
\end{figure}
Least favorable priors produced by Algorithm \ref{alg:mc} are noticeably different from the (more) exact solutions obtained by Algorithm \ref{alg:kempthorne} (Fig. \ref{fig:2LFPs}). However, the corresponding Bayes estimators are nearly identical, and these LFPs yield very tight upper and lower bounds on $\dbar_{\mathrm{max}}$ (see Figure \ref{fig:compare1}). We conclude that the minimax risk is very insensitive to certain variations in the prior. This explains the discrepancies in the LFPs obtained via Algorithms 1-2, and also justifies our use of the estimators and risks obtained via Algorithm \ref{alg:mc}. Using this algorithm, we were able to find good approximations to the minimax risk up to $N=192$, but this is still insufficient to clearly show the asymptotic behavior of $\dbar_{\mathrm{max}}$. For that purpose, we developed the ``noisy coin'' model.
\begin{figure}[ht]
\centering
\includegraphics[width=.45\textwidth]{scaling-qubit_rebit_fit-1}
\caption{\label{fig:compare1}
Comparison of the minimax risk computed using Algorithms \ref{alg:kempthorne} and \ref{alg:mc}. The minimax risk of qubit and rebit tomography with $N$ Pauli measurements ($N=3\ldots192$ for qubits and $N=2\ldots512$ for rebits) was computed by finding least favorable priors, using Algorithm \ref{alg:kempthorne} (large black dots) and Algorithm \ref{alg:mc} (small gray dots). In all cases where both algorithms could be applied, results agreed to high precision.}
\end{figure}
\begin{algorithm}
\caption{Kempthorne (deterministic) algorithm for finding the least favorable prior \cite{Kempthorne1987Numerical}.}
\label{alg:kempthorne}
\begin{algorithmic}
\Require Number of measurements $N>0$.
\Require Support points of initial guess prior $x_i$, $i \in \{1, \dots, n\}$.
\Require Probability weights of the support points $w_i$, $i \in \{1, \dots, n\}$ such that $\sum_i w_i = 1$.
\Require Tolerance $\mathtt{tol}>0$.
\Require Mixing parameter $\alpha$.
\Ensure Least favorable prior $\{\vec{x},\vec{w}\}$ with $m>n$ support points.
\Ensure Lower bound on the minimax risk $\mathtt{av\_risk}$.
\Ensure Upper bound on the minimax risk $\mathtt{max\_risk}$.
\Function{DeterministicLFP}{$N$, $\{\vec{x},\vec{w}\}$,$\mathtt{tol}$}
\State $\mathtt{diff} \gets \mathtt{tol}+1$
\While{$\mathtt{diff} > \mathtt{tol}$}
\State $\{\vec{x},\vec{w}\} \gets$ prior with same number of support points which maximizes the Bayes risk
\State $\mathtt{av\_risk} \gets$ the maximum value of the Bayes risk for the prior found above
\State $\mathtt{max\_risk} \gets$ global maximum of risk using the Bayes estimator of the $(\{\vec{x},\vec{w}\})$
\State $\mathtt{diff}\gets |\mathtt{av\_risk}-\mathtt{max\_risk}|/\mathtt{av\_risk}$
\If{$\mathtt{diff} > \mathtt{tol}$}
\State Add a new support where the maximum risk is attained
\State $w_{\mathtt{length}(\vec{x})}\gets\alpha$
\State for each $i\leq\mathtt{length}(\vec{x})-1$, $w_i\gets w_i-\alpha/(\mathtt{length}(\vec{x})-1)$
\EndIf
\EndWhile
\State \Return $\{\vec{x},\vec{w}\},\mathtt{av\_risk},\mathtt{max\_risk}$
\EndFunction
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{Monte Carlo algorithm for finding the least favorable prior.}
\label{alg:mc}
\begin{algorithmic}
\Require Number of measurements $N>0$.
\Require Number of support points $n>0$.
\Require Tolerance on accuracy $\mathtt{tol}>0$.
\Require Tolerance on the weights to remove supports $\mathtt{weight\_tol}>0$.
\Require Number of support points to add at each iteration $m>0$ for each current support point.
\Require Variance of normal distribution to sample new points from $\sigma$.
\Ensure Least favorable prior $\{\vec{x},\vec{w}\}$ with $m>n$ support points.
\Ensure Lower bound on the minimax risk $\mathtt{av\_risk}$.
\Ensure Upper bound on the minimax risk $\mathtt{max\_risk}$.
\Function{MCLFP}{$N$, $n$,$\mathtt{tol}$,$\mathtt{weight\_tol}$}
\State $\mathtt{diff} \gets \mathtt{tol}+1$
\State $\{\vec{x},\vec{w}\} \gets$ uniform distribution ($w_i=1/n$) sampled according to uniform distribution over $\vec{x}$
\While{$\mathtt{diff} > \mathtt{tol}$}
\State $\vec{w} \gets$ weights which maximize the Bayes risk keeping the support points $\vec{x}$ fixed
\State $\mathtt{av\_risk} \gets$ the maximum value of the Bayes risk for the prior found above
\State $\mathtt{max\_risk} \gets$ global maximum of risk using the Bayes estimator of the $(\{\vec{x},\vec{w}\})$
\State $\mathtt{diff}\gets |\mathtt{av\_risk}-\mathtt{max\_risk}|/\mathtt{av\_risk}$
\State Remove all $x_i$ such that $w_i<\mathtt{weight\_tol}$
\If{$\mathtt{diff} > \mathtt{tol}$}
\For{each $x_i$ left}
\State Add $m$ new support sampled randomly from $\mathcal N (x_i,\sigma)$
\EndFor
\State each $w_i\gets 1/\mathtt{length}(\vec{x})$
\EndIf
\EndWhile
\State \Return $\{\vec{x},\vec{w}\},\mathtt{av\_risk},\mathtt{max\_risk}$
\EndFunction
\end{algorithmic}
\end{algorithm}
\end{document}
|
\section{}
\begin{abstract}
We develop a variational method of deriving stochastic partial differential equations whose solutions follow the flow of a stochastic vector field. As an example in one spatial dimension we numerically simulate singular solutions (peakons) of the stochastically perturbed Camassa-Holm (CH) equation derived using this method. These numerical simulations show that peakon soliton solutions of the stochastically perturbed CH equation persist and provide an interesting laboratory for investigating the sensitivity and accuracy of adding stochasticity to finite dimensional solutions of stochastic partial differential equations (SPDE). In particular, some choices of stochastic perturbations of the peakon dynamics by Wiener noise (canonical Hamiltonian stochastic deformations, or CH-SD) allow peakons to interpenetrate and exchange order on the real line in overtaking collisions, although this behaviour does not occur for other choices of stochastic perturbations which preserve the Euler-Poincar\'e structure of the CH equation (parametric stochastic deformations, or P-SD), and it also does not occur for peakon solutions of the unperturbed deterministic CH equation. The discussion raises issues about the science of stochastic deformations of finite-dimensional approximations of evolutionary PDE and the sensitivity of the resulting solutions to the choices made in stochastic modelling.
\end{abstract}
\newpage
\section{Introduction}\label{intro}
Two main approaches have arisen recently for implementing variational principles in stochastic geometric mechanics. In one of them, sometimes called \emph{stochastic deformation} \cite{ArChCr2014}, the Lagrangian in Hamilton's principle is the classical one, but it is evaluated on underlying stochastic processes and their mean derivatives. This perspective was initially motivated by the quantization of classical systems \cite{ChZa2003,Ya1980,Ya1983} and a probabilistic version of Feynman's path integral approach to quantum mechanics. The stochastic deformation approach brings to bear the full theory of stochastic partial differential equations (SPDE). For more details about the history of the applications of this approach, in particular for fluid dynamics, see \cite{ArChCr2014,Ho2015}.
Here we will advocate a simpler, and more restricted approach. Our approach is based on a generalisation in \cite{Ho2015} of earlier work by Bismut \cite{Bi1981}, L\'azaro-Cam\'i and Ortega \cite{LaCa-Or2008}, and Bou-Rabee and Owhadi \cite{BR-O2009} for stochastic ordinary differential equations (SDE), which unifies their Hamiltonian and Lagrangian approaches to temporal stochastic dynamics, and extends them to stochastic partial differential equations (SPDE) in the case of cylindrical noise in which the spatial dependence is \emph{parametric}, while temporal dependence is stochastic. The advantage of this approach is that the parametric spatial dependence of the stochastic dynamical variables in the resulting SPDEs allows essentially finite-dimensional stochastic methods to be applied at each point of space. This feature allows us to safely assume from the onset that all the objects we introduce in this context are semimartingales. To distinguish between the two approaches, we will call the approach taken here \emph{parametric stochastic deformation} (P-SD).
Parametric stochastic partial differential equations (P-SPDE) result in the present approach by applying P-SD to a deterministic variational principle. These P-SPDE contain a type of multiplicative, cylindrical, Stratonovich noise that depends on both the solution variables and their \emph{spatial gradients}. This unfamiliar feature does not interfere with the passage to the It\^o representation, though, since the space variable is treated merely as a parameter when dealing with cylindrical noise. That is, one may regard the cylindrical noise process as a finite dimensional stochastic process parametrized by $\MM x$ (the space variable). Then, the Stratonovich equation makes analytical sense pointwise, for each fixed $\MM x$. Once this is agreed, then the transformation to It\^o by the standard method also makes sense pointwise in space. For more details about P-SPDE explained in a fluid dynamics context, see \cite{Ho2015}.
In this paper we develop an approach for inserting parametric stochastic deformation with cylindrical noise into systems of evolutionary partial differential equations which derive from deterministic variational principles that are invariant under a Lie group action. The corresponding deterministic dynamical systems are called Euler-Poincar\'e equations. The set of Euler-Poincar\'e equations includes the equations of ideal fluid dynamics, which follow from variational principles whose Lagrangians satisfy certain invariance properties under smooth invertible maps (diffeomorphisms) \cite{MaRa1994,HoMaRa1998}.
\paragraph{Objectives.} This paper has two main objectives. The first objective is the inclusion of parametric stochastic deformation (P-SD) in the variational principle for the EPDiff partial differential equation
\footnote{EPDiff is the PDE which arises when the Lagrangian in Hamilton's principle is a functional of continuous Eulerian vector fields, whose flows are smooth invertible maps (diffeomorphisms). EPDiff is the Euler-Poincar\'e equation arising for Lagrangians which are invariant under the diffeomorphism group. When the Lagrangian is chosen to be the $H^1$ norm of the vector fields, the EPDiff equation becomes the Camassa-Holm equation \cite{CaHo1993}, which is a completely integrable Hamiltonian system and possesses singular soliton solutions known for their peaked shape as \emph{peakons}.} The second objective is the numerical study of the statistical effects of parametric and canonically Hamiltonian stochastic deformations (CH-SD) on the soliton-like solutions of EPDiff which arise in the deterministic case in one spatial dimension, when the Lagrangian in Hamilton's principle is a Sobolev norm on the continuous vector fields. When the $H^1$ norm is chosen this results in the CH equation \cite{CaHo1993}. We study the stochastic generalised Camassa-Holm equation because its singular momentum map \cite{HoMa2004} persists under the stochastic deformations we introduce here and thereby allows its solutions to be investigated on a finite dimensional invariant manifold.
\subsection{Stochastic EPDiff variational principle}
The action integral for the stochastic variational principle we shall study for EPDiff is a \emph{stochastically constrained} variational principle $\delta S = 0$, with action integral, $S$, given by
\begin{align}
S(u,p,q) &=
\int \bigg( \ell(u) dt
+ \left\langle p\,,\,{dq} + \widetilde{\pounds}_{\widetilde u} q\,\right\rangle_V \bigg)
\,,\label{SVP1}
\end{align}
where $\ell(u)$ is the unperturbed deterministic Lagrangian, written as a functional of velocity vector field $u\in\mathfrak{X}(\mathbb{R}^3)$. The angle brackets
\begin{equation}
\langle \,p\,,\,q\,\rangle_V:=\int <p(x,t),q(x,t)>dx
\label{L2pairing}
\end{equation}
denote the spatial $L^2$ integral over the domain of flow of the pairing $<p\,,\,q>$ between stochastic dynamical variables $q$, which take values in a tensor space $V$, and their dual elements $p$ taking values in $V^*$. In \eqref{SVP1}, the quantity $p\in V^*$ is a Lagrange multiplier and $\widetilde{\pounds}_{\widetilde u}q$ is the \emph{stochastic Lie differential} of the dynamical variable $q\in V$ with respect to a stochastic vector field $\widetilde u(x,t)$ which is defined by the following sum of a drift velocity $u(x,t)$ and \emph{Stratonovich} stochastic process with \emph{cylindrical noise} parameterised by spatial position $x$, \cite{Pa2007,Sc1988}
\begin{equation}
\widetilde u(x,t) = u(x,t)\,dt - \sum_i \xi_i(x)\circ dW_i(t)
\,.\label{vel-vdt}
\end{equation}
We give a precise definition of the stochastic Lie differential $\widetilde{\pounds}_{\widetilde u} q$ in formula \eqref{eq: Stochastic Lie differential definition} in Section~\ref{sec: Stratonovich stochastic EPDiff equation}. One may interpret equation \eqref{vel-vdt} as the decomposition of a vector field $\widetilde u(x,t)$ defined at spatial position $x$ and time $t$ into a time-dependent drift velocity $u(x,t)$ and a stochastic vector field with cylindrical noise (note that for notational convenience we define $\widetilde u$ as a stochastic differential with respect to the time variable; we will use the tilde to denote objects defined as stochastic differentials throughout the rest of this work). The time-independent vector fields $\xi_i(x)$ with $i=1,2,\dots,M$ in the cylindrical stochastic process are usually interpreted as ``diffusivities" of the stochastic vector field, and the choice of these $M$ quantities must somehow be specified from the physics of the problem to be considered. In the present considerations, a natural choice will arise from the singular momentum map admitted by the deterministic EPDiff equation \cite{HoMa2004}.
The $L^2$ pairing $\left\langle\,\cdot\,,\,\cdot\,\right\rangle_V$ in the Stochastic Variational Principle (SVP) $\delta S=0$ for action functional $S$ in \eqref{SVP1} with Lagrange multiplier $p\in T^*V$ enforces the \emph{advection condition} that the quantity $q\in V$ is preserved along the Stratonovich stochastic integral curves of the vector field \eqref{vel-vdt} for any tensor space $V$. Namely, $q$ satisfies the advection condition
\begin{equation}
{dq} + \widetilde{\pounds}_{\widetilde u} q = 0
\,.\label{advec-cond}
\end{equation}
The advection condition \eqref{advec-cond} for the quantities $q\in V$ may be regarded as a \emph{stochastic constraint} imposed on the variational principle \eqref{SVP1} via the Lagrange multiplier $p$.
At this point, we have introduced parametric Stratonovich stochasticity into the variational principle \eqref{SVP1} with dynamical variables $(u,q,p)$ through the constraint that the advected quantities $q\in V$ should evolve by following the Stratonovich stochastic vector field $\widetilde u(x,t)$ in equation \eqref{vel-vdt}. This advection law is formulated as a Lie differential with respect to the Stratonovich stochastic vector field acting on a tensor space.
We use the Stratonovich formulation, so the normal rules of calculus apply.
For mathematical discussions of Lie derivatives with respect to stochastic vector fields, see, e.g., \cite{IkWa1981, KaSh1994}.
\paragraph{Plan of the paper.}
After setting out the general theory of Euler-Poincar\'e evolutionary P-SPDE in $(x,t)\in \mathbb{R}^3\times \mathbb{R}$ in the remainder of this introductory section, we will specialise to one spatial dimension for $(x,t)\in \mathbb{R}\times \mathbb{R}$ and study the corresponding soliton-like solution behaviour for the case when the Lagrangian $\ell(u)$ in the variational principle \eqref{SVP1} is chosen as a Sobolev norm of the vector field $u$.
The objective of the remainder of the paper is to use the Stratonovich stochastic EPDiff Theorem proved below to study the effects of introducing this type of stochasticity on the interactions of the peakon solutions of the CH equation with parametric stochastic deformation (P-SD). The P-SD of the equations of motion for the singular solutions of stochastic EPDiff in one spatial dimension will be introduced via Hamilton's principle in Section \ref{sec: ParametricStochasticDeformations} and their effects on the numerical solutions will be studied thereafter in comparison with a more general canonical Hamiltonian stochastic deformation (CH-SD) in the sense of \cite{Bi1981,LaCa-Or2008}, which includes P-SD but can be more general.
The Fokker-Planck equations for the probability density evolution associated with P-SD and CH-SD of the EPDiff equation will be discussed in Section \ref{sec: The Fokker-Planck equation}. In Section~\ref{sec: Stochastic variational integrator} we discuss the numerical algorithm we use, and in Section~\ref{sec: Numerical experiments} we present the results of our numerical studies, including sample paths and mean solutions, the probability distribution for crossing of singular solutions on the real line, statistics of the \emph{first crossing time}, effects of noise screening, comparison with other types of noise such as additive noise in the canonical momentum equation and the results of convergence tests for our stochastic variational integrator. Section \ref{sec: Summary and future work} is devoted to a brief summary of results and discussion of some open problems for possible future work.
\subsection{Stratonovich stochastic EPDiff equation}
\label{sec: Stratonovich stochastic EPDiff equation}
If the drift velocity vector field $u(x,t)$ and the diffusivity vector fields $\xi_i(x)$ satisfy some standard measurability and regularity conditions, then the stochastic vector field \eqref{vel-vdt} possesses a pathwise unique stochastic flow $F_{t,s}(x)$. By definition, this flow almost surely satisfies $F_{s,s}(x)=x$ and the integral equation
\begin{equation}
\label{eq: Integral definition of the stochastic flow}
F_{t_2,s}(x)-F_{t_1,s}(x) = \int_{t_1}^{t_2} u\big( F_{t,s}(x), t \big)\,dt - \sum_i \int_{t_1}^{t_2} \xi_i\big( F_{t,s}(x) \big)\circ dW_i(t),
\end{equation}
\noindent
or shortly, in the differential form,
\begin{equation}
\label{eq: Stochastic differential of the stochastic flow}
dF_{t,s}(x) = u\big( F_{t,s}(x), t \big)\,dt - \sum_i \xi_i\big( F_{t,s}(x) \big)\circ dW_i(t).
\end{equation}
\noindent
It can be proved that for fixed $t,s$ this flow is mean-square differentiable with respect to the $x$ argument, and also almost surely is a diffeomorphism (see \cite{ArnoldSDE}, \cite{IkWa1981}, \cite{KloedenPlatenSDE}, \cite{Kunita}). These properties allow us to generalise the differential-geometric notion of the Lie derivative, which we do in the following definition and theorem.
\begin{definition}[Stochastic Lie differential]
\label{thm: Stochastic Lie differential}
Let $q$ be a smooth tensor field. The stochastic Lie differential $\widetilde{\pounds}_{\widetilde u} q$ is the almost surely unique stochastic differential satisfying
\begin{equation}
\label{eq: Stochastic Lie differential definition}
d(F_{t,s}^* q) := F_{t,s}^* \widetilde{\pounds}_{\widetilde u} q.
\end{equation}
\end{definition}
\begin{theorem}
\label{thm: Stochastic Lie differential formula}
The stochastic Lie differential $\widetilde{\pounds}_{\widetilde u} q$ is almost surely unique and given by
\begin{equation}
\label{eq: Stochastic Lie differential formula}
\widetilde{\pounds}_{\widetilde u} q = \pounds_u q \, dt - \sum_i \pounds_{\xi_i}q \circ dW_i(t),
\end{equation}
\noindent
where $\pounds_u$ and $\pounds_{\xi_i}$ are the standard Lie derivatives.
\begin{proof}
The proof is a straightforward generalisation of the standard differential-geometric construction of the Lie derivative of tensor fields (see \cite{AbrahamMarsdenRatiuMTA}). Whenever necessary, we replace time-differentiation with stochastic differentials and use the weak property \eqref{eq: Stochastic differential of the stochastic flow} of the flow. We first prove \eqref{eq: Stochastic Lie differential formula} when $q$ is a smooth, real-valued function. Then $(F_{t,s}^* q)(x) =q\circ F_{t,s}(x)$. Formula \eqref{eq: Stochastic Lie differential formula} is proved by calculating the stochastic differential $d(F_{t,s}^* q)$ using the rules of Stratonovich calculus. Next we consider the case when $q$ is a smooth vector field. Let $G_\lambda$ be the smooth flow of $q$. Then for fixed $t,s$ the flow $H_\lambda$ of the vector field $F_{t,s}^* q$ satisfies $G_\lambda \circ F_{t,s} = F_{t,s}\circ H_\lambda$. From the mean-square differentiability of $F_{t,s}$ we have mean-square differentiability of both sides with respect to $\lambda$. Differentiating both sides with respect to $\lambda$, evaluating at $\lambda=0$, calculating the stochastic differential with respect to $t$ and comparing terms, we obtain formula \eqref{eq: Stochastic Lie differential formula} for vector fields. For the case when $q$ is a differential one-form we use the property $F_{t,s}^*\langle q, v \rangle = \langle F_{t,s}^*q, F_{t,s}^*v \rangle$, where $v$ is an arbitrary smooth vector field and $\langle \cdot,\cdot\rangle$ is the dual pairing between one-forms and vector fields. Calculating the stochastic differential of both sides and using our already established results for functions and vector fields, we prove \eqref{eq: Stochastic Lie differential formula} for differential one-forms. It is now straightforward to complete the proof for a general tensor field. Almost sure uniqueness of $\widetilde{\pounds}_{\widetilde u} q$ follows from our construction and pathwise uniqueness of the flow $F_{t,s}$.
\end{proof}
\end{theorem}
In order to choose the form of the spatial correlations, or \emph{diffusivities}, $\xi_i(x)$ of the cylindrical Stratonovich stochasticity in \eqref{vel-vdt}, we notice that the action integral for the variational principle in \eqref{SVP1} may be rearranged into the equivalent form
\begin{align}
S(u,p,q) =
\int \bigg( \ell(u) dt
+ \left\langle p\,,\,{dq} + \widetilde{\pounds}_{\widetilde u} q\,\right\rangle_V \bigg)
=
\int \bigg( \ell(u) dt
+ \left\langle p\,,\,{dq}\right\rangle_V
- \left\langle p\diamond q\,,\,{\widetilde u} \,\right\rangle_{\mathfrak{X}} \bigg)
\,,
\label{SVP2}
\end{align}
where we define the diamond operation $(\diamond)$ in the expression $p\diamond q\in \mathfrak{X}^*$ via the real-valued nondegenerate pairing $\langle\,\cdot\,,\,\cdot\,\rangle_{\mathfrak{X}}: \mathfrak{X}^*\times \mathfrak{X}\to\mathbb{R}$ between a vector field $\eta\in \mathfrak{X}$ and its dual under $L^2$ pairing $\mu\in \mathfrak{X}$ as
\begin{align}
\left\langle p\diamond q\,,\,\eta \,\right\rangle_{\mathfrak{X}}
:=
\left\langle p\,,\,-\pounds_\eta q\right\rangle_V
\,.
\label{diamond-def}
\end{align}
The diamond operation $(\diamond)$ will be instrumental in deriving the Stratonovich form of the stochastic EPDiff equation from the stochastic variational principle for the action integral in \eqref{SVP1}, as stated in the following theorem.
\begin{theorem}[Stratonovich Stochastic EPDiff equation]\label{SEP-thm}\rm$\,$
The parametric Stratonovich stochastic deformation in the action $S(u,p,q)$ for the stochastic variational principle $\delta S = 0$ for EPDiff given by
\begin{align}
S(u,p,q) &=
\underbrace{\
\int \bigg( \ell(u) - \left\langle p\diamond q\,,\,u(x,t) \,\right\rangle_{\mathfrak{X}} \bigg)dt
}_{\hbox{Lebesgue integral}}
+
\underbrace{\
\int \left\langle p(x,t)\,,\,dq(x,t) \,\right\rangle_V
}_{\hbox{Stratonovich integral wrt $q$}}
+
\underbrace{
\int \sum_{i} \left\langle p\diamond q\,,\,\xi_i(x) \,\right\rangle_{\mathfrak{X}} \circ dW_i(t)
}_{\hbox{Stratonovich integral}}
\,,\label{SVP2-redux-FIRST}
\end{align}
yields the following Stratonovich form of the \emph{stochastic} EPDiff equation
\begin{align}
dm + \widetilde{\pounds}_{\widetilde u}m
=
0\,.
\label{SEP-eqns-thm-FIRST}
\end{align}
The momentum density $m(x,t)$ and velocity vector field $u(x,t)$ in \eqref{SEP-eqns-thm-FIRST} are related by $m=\frac{\delta \ell}{\delta u}$ and the stochastic vector field $\widetilde u(x,t)$ is given by
\begin{align}
\widetilde u(x,t) = u(x,t)\,dt - \sum_{i} \xi_i(x) \circ dW_i(t)
\,.
\label{path-thm-FIRST}
\end{align}
\end{theorem}
\begin{proof}
The first step is to take the elementary variations of the action integral \eqref{SVP2}, to find
\begin{align}
\delta u:\quad
\frac{\delta \ell}{\delta u} - p\diamond q = 0
\,,\quad
\delta p:\quad
dq + \widetilde{\pounds}_{\widetilde u} q = 0
\,,\quad
\delta q:\quad
- dp
+ \widetilde{\pounds}_{\widetilde u}^Tp = 0
\,.
\label{var-eqns-thm-FIRST}
\end{align}
The first equation in \eqref{var-eqns-thm-FIRST} follows from the definition of the diamond operation $(\diamond)$ in \eqref{diamond-def}.
The second and third equations immediately follow from variations of the equivalent form of the action $S(u,p,q)$ in equation \eqref{SVP2} and integrations by parts with vanishing endpoint and boundary conditions.
The governing equation for $m$ will be recovered by using the result of the following Lemma.
\begin{lemma}\rm\label{Lemma-m-eqn}
Together, the three equations in \eqref{var-eqns-thm-FIRST} imply \eqref{SEP-eqns-thm-FIRST}.
\begin{proof}
For an arbitrary $\eta\in \mathfrak{X}$, one computes the pairing
\begin{align}
\begin{split}
\left\langle
dm \,,\, \eta
\right\rangle_{\mathfrak{X}}
&=
\left\langle
dp\diamond q + p\diamond dq\,,\, \eta
\right\rangle_{\mathfrak{X}}
\\
\hbox{By equation \eqref{var-eqns-thm-FIRST} } &=
\left\langle
(\widetilde{\pounds}_{\widetilde u}^Tp) \diamond q
- p\diamond \widetilde{\pounds}_{\widetilde u} q\,,\, \eta
\right\rangle_{\mathfrak{X}}
\\&=
\left\langle
p\,,\, (-\widetilde{\pounds}_{\widetilde u} \pounds_{\eta} + \pounds_{\eta} \widetilde{\pounds}_{\widetilde u} )q\,
\right\rangle_{V}
\\&=
\left\langle
p\,,\, {\rm \widetilde{ad}}_{\widetilde u}{\eta}\,q\,
\right\rangle_{V}
=
-\left\langle
p\diamond q\,,\, {\rm \widetilde{ad}}_{\widetilde u}{\eta}\,
\right\rangle_{\mathfrak{X}}
\\&=
-\left\langle
{\rm \widetilde{ad}}^*_{\widetilde u}(p\diamond q)\,,\,{\eta}\,
\right\rangle_{\mathfrak{X}}
=
-\,\Big\langle
\widetilde{\pounds}_{\widetilde u}m\,,\,{\eta}\,
\Big\rangle_{\mathfrak{X}}\,,
\end{split}
\label{calc-lem}
\end{align}
where ${\rm \widetilde{ad}}_{\widetilde u} \eta = -[\widetilde{\pounds}_{\widetilde u},\pounds_\eta] = {\rm ad}_{u} \eta \,dt - \sum_i {\rm ad}_{\xi_i} \eta \circ dW_i(t)$ is the stochastic adjoint action differential. Since $\eta\in \mathfrak{X}$ was arbitrary, the last line completes the proof of the Lemma. In the last step we have used the fact that coadjoint action $({\rm ad}^*)$ is identical to Lie-derivative action $(\pounds)$ for vector fields acting on 1-form densities.
\end{proof}
\end{lemma}
The result of Lemma \ref{Lemma-m-eqn} now produces the $m$-equation in \eqref{SEP-eqns-thm-FIRST} of Theorem \ref{SEP-thm}. This completes the proof.
\end{proof}
\begin{remark}[Multiplicative noise in the 3D vector stochastic EPDiff equation]\rm$\,$\\
In 3D vector notation, the 1-form density $m$ is expressed as $m=\MM{m}\cdot {\rm d}\MM{x}\otimes {\rm d}^3x$ and equation \eqref{SEP-eqns-thm-FIRST} becomes
\begin{align}
d\MM{m} + \Big( \partial_j(\MM{m}u^j) + m_j\MM{\nabla} u^j \Big)dt
+
\sum_{i}\Big(\partial_j\big(\MM{m} \xi^j_i(x)\big)
+ m_j\MM{\nabla} \xi^j_i(x) \Big)\circ dW_i(t)
= 0
\,,
\label{SEP-1D}
\end{align}
with $m={\delta \ell}/{\delta u}$.
Importantly, the noise terms in \eqref{SEP-1D} multiply both the solution and its gradient. The latter is not a common form for stochastic PDEs. In addition, both the spatial correlations $\xi_i(x)$ and their derivatives $\nabla\xi_i(x)$ are involved. The effects of these noise terms on the singular solutions of stochastic EPDiff in one spatial dimension will be treated in Section \ref{sec: ParametricStochasticDeformations} and its numerical solutions will be studied thereafter.
\end{remark}
\subsection{It\^o version of the stochastic EPDiff equation}\label{Strat-Ito-subsec}
In the It\^o version of the stochastic EPDiff equation, noise terms have zero mean, but \textit{additional drift terms arise}. These drift terms are double Lie derivatives, which are diffusive, as shown in \cite{Ho2015} for stochastic fluid dynamics.
The corresponding It\^o forms of the stochastic EPDiff equation in \eqref{SEP-eqns-thm-FIRST} and the second and third equations in \eqref{var-eqns-thm-FIRST} are found by using It\^o's formula to identify the quadratic covariation terms \cite{Sh1996} as
\begin{align}
\begin{split}
dm + \widehat {\pounds}_{\widetilde u}m
&=
\frac12
\sum_{j} \pounds_{\xi_j(x)}\left(\pounds_{\xi_j(x)}m\right) \,dt
\,,\\
dq + \widehat {\pounds}_{\widetilde u}q
&=
\frac12
\sum_{j} \pounds_{\xi_j(x)}\left(\pounds_{\xi_j(x)}q\right) \,dt
\,,\\
dp - \widehat {\pounds}_{\widetilde u}^Tp
&=
- \frac12
\sum_{j} \pounds_{\xi_j(x)}^T\left(\pounds_{\xi_j(x)}^Tp\right) \, dt
\,,
\end{split}
\label{ItoEP-eqns}
\end{align}
\noindent
where $\widehat {\pounds}_{\widetilde u}q = \pounds_u q \, dt - \sum_i \pounds_{\xi_i}q \, dW_i(t)$ is an It\^o stochastic differential related to the stochastic Lie differential \eqref{eq: Stochastic Lie differential formula} (it should be noted that $\widehat {\pounds}_{\widetilde u}$ is \emph{not} a Lie differential) and we have used $[dW_i(t),dW_j(t)]=\delta_{ij}dt$ for Brownian motion to identify the quadratic covariation terms as drift terms. For more details about this sort of calculation in the geometric mechanics context, see \cite{Ho2015}.
\subsection{Legendre transform to Stratonovich stochastic Lie-Poisson Hamilton equations}\label{LPform-sec}
\begin{theorem}[Lie-Poisson representation of Stratonovich stochastic EP equations]\label{LPform-thm}\rm$\,$\\
The Stratonovich stochastic EP system in \eqref{SEP-eqns-thm-FIRST} may be written equivalently in terms of a standard semidirect product Lie-Poisson Hamiltonian structure \cite{MaRa1994} with a canonical Poisson bracket $\{q,p\}$, as
\begin{align}
\begin{bmatrix}
dm \\ dq \\ dp
\end{bmatrix}
&=
\begin{bmatrix}
-{\rm \widetilde{ad}}^*_{(\,\cdot\,)} m & -(\,\cdot\,)\diamond q & p\diamond(\,\cdot\,)
\\
-\,\widetilde{\pounds}_{(\,\cdot\,)}q & 0 & 1
\\
\widetilde{\pounds}_{(\,\cdot\,)}^Tp & -1 & 0
\end{bmatrix}
\begin{bmatrix}
\delta \widetilde{h} / \delta m \\ \delta \widetilde{h} / \delta q \\ \delta \widetilde{h} / \delta p
\end{bmatrix}
=:
\left\{
\begin{bmatrix}
m \\ q \\ p
\end{bmatrix}
,
\widetilde{h}
\right\}_{LP}
\,,
\label{SLPstructure-thm}
\end{align}
where $\widetilde{h}(m,q,p)$ is a stochastic differential representing the Legendre transform of the stochastic Lagrangian and $\{\,\cdot\,,\,\cdot\,\}_{LP}$ denotes the Lie--Poisson bracket.
\end{theorem}
\begin{proof}
As usual, the Legendre transform of the stochastic Lagrangian determines the stochastic Hamiltonian and its variational derivatives. In a slight abuse of notation, we may write this Legendre transform as
\begin{align}
\begin{split}
\widetilde{h}(m,q,p) &= \left\langle m\,,\, u \right\rangle dt - \ell(u,q)dt
- \sum_{i} \left\langle m \,,\,\xi_i(x)
\right\rangle_{\mathfrak{X}}\circ dW_i(t)
\\
&=: H(m,q,p)dt
- \sum_{i} \left\langle m \,,\,\xi_i(x)
\right\rangle_{\mathfrak{X}}\circ dW_i(t)
\,,\end{split}
\label{Stochastic-Ham}
\end{align}
where now we allow $q$-dependence in the Lagrangian $\ell(u,q)$. Varying the stochastic Hamiltonian in \eqref{Stochastic-Ham} gives
\begin{align}
\begin{split}
\delta \widetilde{h}(m,q,p)
&=
\left\langle \delta m\,,\,
\frac{\delta \widetilde{h}}{\delta m} \right\rangle
+
\left\langle \delta u\,,\,
m - \frac{\delta \ell}{\delta u} \right\rangle dt
+
\left\langle \frac{\delta \widetilde{h}}{\delta q} \,,\, \delta q \right\rangle
+
\left\langle \frac{\delta \widetilde{h}}{\delta p} \,,\, \delta p \right\rangle
\\ &=
\left\langle \delta m\,,\, udt - \sum_i \xi_i(x)\circ dW_i(t) \right\rangle
+
\left\langle \delta u\,,\,
m - \frac{\delta \ell}{\delta u} \right\rangle dt
+ \left\langle -\, \frac{\delta \ell}{\delta q}\,,\, \delta q \right\rangle dt
+ \left\langle 0 \,,\, \delta p \right\rangle
\,.\end{split}
.\label{Stochastic-Ham-var-FIRST}
\end{align}
Consequently, the corresponding variational derivatives of the stochastic Hamiltonian are
\begin{align}
\frac{\delta \widetilde{h}}{\delta m}
= udt - \sum_i \xi_i(x)\circ dW_i(t)
= \widetilde u
\,,\quad
\frac{\delta \widetilde{h}}{\delta q} = - \frac{\delta \ell}{\delta q}dt
\quad\hbox{and}\quad
\frac{\delta \widetilde{h}}{\delta p} = 0
\,.\label{Stochastic-Ham-var}
\end{align}
The resulting Lie-Poisson Hamiltonian form of the system of Stratonovich stochastic variational equations in \eqref{SEP-eqns-thm} is then given by
\begin{align}
\begin{bmatrix}
dm \\ dq \\ dp
\end{bmatrix}
&=
\begin{bmatrix}
-{\rm \widetilde{ad}}^*_{(\,\cdot\,)} m & -\,(\,\cdot\,)\diamond q & p\diamond(\,\cdot\,)
\\
-\widetilde{\pounds}_{(\,\cdot\,)}q & 0 & 1
\\
\widetilde{\pounds}_{(\,\cdot\,)}^Tp & -1 & 0
\end{bmatrix}
\begin{bmatrix}
\delta \widetilde{h} / \delta m \\ \delta \widetilde{h} / \delta q \\ \delta \widetilde{h} / \delta p
\end{bmatrix}
\label{SLPstructure}
=
\begin{bmatrix}
-{\rm \widetilde{ad}}^*_{\widetilde u} m - \frac{\delta \widetilde{h}}{\delta q}\diamond q
\\
-\widetilde{\pounds}_{\widetilde u}q
\\
\widetilde{\pounds}_{\widetilde u}^Tp - \frac{\delta \widetilde{h}}{\delta q}
\end{bmatrix}.
\end{align}
Of course, the terms involving ${\delta \widetilde{h}}/{\delta q}$ vanish, when ${\delta \ell}/{\delta q}=0$, as for EPDiff.
\end{proof}
\begin{remark}\rm
The matrix operator in \eqref{SLPstructure} is the Hamiltonian operator for a standard semidirect product Lie-Poisson structure \cite{HoMaRa1998} with the canonical Poisson bracket (two-cocycle) $\{q,p\}$ between $q$ and $p$. Similar Lie-Poisson structures also appear in the Hamiltonian formulation of the dynamics of complex fluids \cite{GBRa2009}.
Stochastic Lie-Poisson Hamiltonian dynamics has also been previously studied and developed in applications in spin dynamics in \cite{BrElHo2008, BrElHo2009} and in image registration in \cite{CoCoVi2012, HoRaTrYo2004, TrVi2012}.
\emph{Canonical} stochastic Hamilton equations were introduced in Bismut \cite{Bi1981} and were recently developed further in the context of geometric mechanics in L\'azaro-Cam\'i and Ortega \cite{LaCa-Or2008}.
\end{remark}
\section{Stochastic variational perturbations in one spatial dimension}
\label{sec: ParametricStochasticDeformations}
\subsection{Singular peakon solutions of the EPDiff equations}
The EPDiff$(H^1)$ equation in the one-dimensional case when $\ell(u)=\frac12\|u\|^2_{H^1}=\frac12\int u^2 + \alpha^2u_x^2\, {\rm d} x$ is called the Camassa-Holm (CH) equation for $m=\delta \ell/\delta u = u - \alpha^2u_{xx}$ with positive constant $\alpha^2$; namely \cite{CaHo1993},
\begin{align}
m_t + (um)_x + mu_x = 0
\quad\hbox{with}\quad
m = u - \alpha^2u_{xx}
\,.
\label{CH-eqn}
\end{align}
This equation has singular \emph{peakon} solutions, given by
\begin{align}
m(x,t):=\frac{\delta \ell}{\delta u} = u - \alpha^2u_{xx} = \sum_{a=1}^N p_a(t) \delta (x - q_a(t))
\,,\quad\hbox{so that}\quad
u(x,t): = \sum_{b=1}^N p_b(t) K (x - q_b(t)),
\label{m-def-peakon}
\end{align}
where $K (x - y)=\exp (-|x-y|/\alpha)$ is the Green's function for the Helmholtz operator $1-\alpha^2\partial_x^2$. The peaked shape of the velocity profile for each individual peakon solution of the CH equation $u(x,t): = p(t) \exp (-|x - q(t)|/\alpha)$ gives them their name.
Peakons are emergent singular solutions which dominate the initial value problem, as shown in Figure \ref{CH-peakons}. An initially confined smooth velocity distribution will decompose into peakon solutions and, in fact, \emph{only} peakon solutions. Substituting the (weak) solution Ansatz \eqref{m-def-peakon} into the CH equation \eqref{CH-eqn} and integrating against a smooth test function yields the following dynamical equations for the $2N$ solution parameters $q_a(t)$ and $p_a(t)$
\begin{equation}
\frac{dq_a}{dt} = u(q_a(t),t)
\quad\hbox{and}\quad
\frac{dp_a}{dt} = - \,p_a(t) \frac{\partial u(q_a(t),t)}{\partial q_a}
\,.
\label{pkn-pq-eqns}
\end{equation}
The system of equations for the peakon parameters comprises a completely integrable canonical Hamiltonian system, whose solutions determine the positions $q_a(t)$ and amplitudes $p_a(t)$, for all $N$ solitons, $a=1,\dots,N$, and also describe the dynamics of their multi body interactions, as shown, for example, in Figure \ref{CH-peakons}.
\begin{figure}
\begin{center}
\includegraphics[width=.5\textwidth,angle=0]{waterb2.pdf}
\end{center}
\caption{\label{CH-peakons} Singular peakon solutions emerge from smooth initial conditions and form a finite dimensional solution set for the CH equation, EPDiff($H^1$). The velocity profile for each individual peakon has the peaked shape given by $u(x,t): = p(t) \exp (-|x - q(t)|/\alpha)$, which is the Green's function for the Helmholtz operator. The main point to notice is that the distance between any two peaks never passes through zero. That is, the peakons keep their order, even after any number of overtaking collisions. (The taller peakons travel faster.)
}
\end{figure}
As mentioned earlier, the objective of the remainder of the paper is to use the Stratonovich stochastic EPDiff Theorem \ref{SEP-thm} to study the effects of introducing this type of stochasticity on the interactions of the peakon solutions of the CH equation with Stratonovich parametric stochastic deformation (P-SD) and canonical Hamiltonian stochastic deformation (CH-SD). The first step is to adapt Theorem \ref{SEP-thm} to accommodate the peakon solutions. For this adaptation, the advection condition \eqref{advec-cond} used previously will be replaced by the definition of peakon velocity as the time derivative of peakon position, as in the first equation in \eqref{pkn-pq-eqns}.
\subsection{Singular momentum map version of the Stratonovich stochastic EPDiff equations}
\begin{theorem}[Canonical Hamiltonian Stochastic Deformation (CH-SD) of EPDiff ]\label{SEP-thm-peakons}\rm$\,$
The action $S(u,p,q)$ for the stochastic variational principle $\delta S = 0$ given by
\begin{align}
S(u,p,q) &=
\int \bigg( \ell(u)\,dt + \sum_{a} \left\langle p_a\,,\,dq_a - {u}( q_a, t)\,dt\,\right\rangle \bigg)
-
\int \sum_{i} h_i(q,p)\circ dW_i(t)
\,,\label{SVP2-redux}
\end{align}
leads to the following Stratonovich form of the \emph{stochastic} EPDiff equation
\begin{align}
\begin{split}
&
dm = - \pounds_u m \, dt + \sum_{i} \big\{m \,,\, h_i(q,p)\big\} \circ dW_i(t)
\,,\\
&dq_a = u(q_a,t) \,dt + \sum_{i} \big\{q_a \,,\, h_i(q,p)\big\} \circ dW_i(t)
\,,\\
&dp_a = - p_a(t) \frac{\partial u}{\partial x}(q_a,t)\,dt
+ \sum_{i} \big\{p_a \,,\, h_i(q,p)\big\} \circ dW_i(t)
\,,
\end{split}
\label{SEP-eqns-thm}
\end{align}
where the momentum density $m$ and velocity $u$ are given by
\begin{align}
m(x,t):=\frac{\delta \ell}{\delta u} = \sum_{a=1}^N p_a \delta (x - q_a(t))
\,,\quad\hbox{and}\quad
u(x,t): = \sum_{b=1}^N p_b K (x - q_b(t)).
\label{m-def-thm}
\end{align}
\end{theorem}
\begin{proof}
As in the proof of Theorem \ref{SEP-thm}, the first step is to take the variations of the action integral \eqref{SVP2-redux}, to find
\begin{align}
\begin{split}
\delta u:\quad&
\frac{\delta \ell}{\delta u} - \sum_{a=1}^N p_a \delta (x - q_a(t)) = 0
\,,\\
\delta p:\quad&
dq_a - u(q_a,t) \,dt - \sum_{i} \frac{\partial h_i}{\partial p_a}(q,p) \circ dW_i(t) = 0
\,,\\
\delta q:\quad&
- dp_a - p_a(t) \frac{\partial u}{\partial x}(q_a,t)\,dt
- \sum_i \frac{\partial h_i}{\partial q_a}(q,p)\circ dW_i(t)
= 0
\,,\end{split}
\label{var-eqns-thm}
\end{align}
after integrations by parts with vanishing endpoint and boundary conditions.
The first variational equation captures the relation \eqref{m-def-thm}, and latter two equations in \eqref{var-eqns-thm} produce the corresponding equations in \eqref{SEP-eqns-thm}. Substituting the latter two equations in \eqref{var-eqns-thm} into the time derivative of the first one yields the first equation in \eqref{SEP-eqns-thm}.
\end{proof}
The particular choice of the functions $h_i(q,p) = \sum_{a=1}^N p_a \xi_i(q_a)$ reproduces the results of Theorem~\ref{SEP-thm} for parameterised stochastic deformation (P-SD) of the peakon solutions. We summarise this observation in the following Corollary.
\begin{corollary}\rm[P-SD is a special case of CH-SD for EPDiff]
\label{Lemma-m-eqn-peakon}
Given the set of diffusivities $\xi_i(x)$, $i=1,\ldots,M$, let $h_i(q,p) = \sum_{a=1}^N p_a \xi_i(q_a)$. Then the momentum density $m(x,t)$ satisfies the equation
\begin{equation}
\label{m-eqn-peakons}
dm + \widetilde{\pounds}_{\widetilde u}m =0\,,
\end{equation}
where the stochastic vector field $\widetilde u(x,t)$ is given by the P-SD formula,
\begin{align}
\widetilde u(x,t) = u(x,t) \,dt + \sum_{i} \xi_i(x) \circ dW_i(t)\,.
\label{path-thm}
\end{align}
\begin{proof}
Specialise to $h_i(q,p) = \sum_{a=1}^N p_a \xi_i(q_a)$ in the first line of equation \eqref{SEP-eqns-thm} in Theorem \ref{SEP-thm-peakons}.
\end{proof}
\end{corollary}
\begin{remark}[Outlook: Comparing results for P-SD and CH-SD]
\label{sec: Remark on P-SD and CH-SD}\rm
In Section~\ref{sec: The Fokker-Planck equation} and Section~\ref{sec: Numerical experiments} we will investigate the effects of choosing between two slightly different types of stochastic potentials on the interaction of two peakons, $N=2$, corresponding to P-SD and CH-SD. The two options are $h^{(1)}_i(q,p) = \sum_{a=1}^N p_a \xi_i(q_a)$ and $h^{(2)}_i(q,p) = \sum_{a=1}^N p_a \varphi_{ia}(q)$, respectively, for $i=1,\ldots,M$. These are both linear in the peakon momenta and in the simplest case they have constant coefficients. We will consider numerical simulations for two cases: $h^{(1)}_1(q,p)= c(p_1+p_2)$ (P-SD for $M=1$) and $h^{(2)}_1(q,p)= \beta_1 p_1$, $h^{(2)}_2(q,p)= \beta_2 p_2$ (CH-SD for $M=2$) with constants $c,\beta_1,\beta_2$. Although these choices for $h^{(1)}_1$ and $h^{(2)}_i$ are very similar, they will produce quite different solution behaviour in our numerical simulations of peakon-peakon overtaking collisions in Section~\ref{sec: Numerical experiments}.
\end{remark}
\begin{remark}[Stratonovich stochastic EPDiff equations in one dimension]\rm$\,$
\begin{enumerate}
\item
In one spatial dimension, equation \eqref{m-eqn-peakons} becomes
\begin{align}
dm + \big( um_x + 2mu_x\big)dt
+
m_x\sum_{i} \xi_i(x) \circ dW_i(t) + 2m\sum_{i} \xi'_i(x) \circ dW_i(t)
= 0
\,.
\label{SEP-1D-1}
\end{align}
Importantly, the multiplicative noise multiplies both the solution and its gradient. The latter is not a common form for stochastic PDEs. In addition, both the spatial correlations $\xi_i(x)$ and their derivatives $\xi'_i(x)$ are involved.
\item
The equations for $dq_a$ and $dp_a$ in \eqref{SEP-eqns-thm} are stochastic canonical Hamiltonian equations (SCHEs) in the sense of Bismut \cite{Bi1981,LaCa-Or2008}. These equations for $dq_a$ and $dp_a$ may be rewritten as
\begin{align}
\begin{split}
&
dq_a = \frac{\partial H}{\partial p_a}(q,p) \,dt + \sum_{i} \frac{\partial h_i}{\partial p_a}(q,p) \circ dW_i(t)
\,,\\&
dp_a = - \frac{\partial H}{\partial q_a}(q,p)\,dt
- \sum_i \frac{\partial h_i}{\partial q_a}(q,p) \circ dW_i(t)
\,,
\end{split}
\label{SEP-eqns-thm-qp}
\end{align}
where the deterministic Hamiltonian is given by
\begin{align}
H(q,p) = \frac12\sum_{a,b} p_ap_bK(q_a-q_b)
\,.
\label{qp-Ham}
\end{align}
The stochastic canonical Hamilton equations \eqref{SEP-eqns-thm-qp} can also be obtained by extremising the \emph{phase-space action functional}
\begin{equation}
\label{eq: Stochastic phase-space action functional}
S\big[q(t),p(t)\big] = \int_0^T \Big( \sum_{a=1}^N p_a \circ dq_a - H(q,p)\,dt \Big)
- \int_0^T \sum_{i=1}^M h_i(q,p) \circ dW_i(t)\,.
\end{equation}
This is the restriction of \eqref{SVP2-redux} to the submanifold defined by the Ansatz \eqref{m-def-thm}.
\item
In the It\^o version of stochastic canonical Hamiltonian equations, the noise terms have zero mean, but \textit{additional drift terms arise}. These drift terms are double canonical Poisson brackets, which are diffusive \cite{LaCa-Or2008}:
\begin{align*}
&\delta p :\quad dq_a = u(q_a,t)dt +
\underbrace{
\sum_i \{q_a,h_i\} dW_i(t) \
}_{\hbox{It\^o Noise for \it q}}
+
\underbrace{
\frac12\sum_i \{\{q_a,h_i\},h_i\}dt\
}_{\hbox{It\^o Drift for \it q}},
\\
&\delta q :\quad dp_a = -p_a \frac{\partial u}{\partial x}(q_a,t) \,dt
+
\underbrace{
\sum_i \{p_a,h_i\} dW_i(t)\
}_{\hbox{It\^o Noise for \it p}}
+
\underbrace{
\frac12\sum_i \{\{p_a,h_i\},h_i\}dt\
}_{\hbox{It\^o Drift for \it p}}.
\end{align*}
The It\^o stochastic dynamics of landmark points in the image registration problem discussed in Trouv\'e and Vialard \cite{TrVi2012} is recovered when we choose $h_i(p,q)=-\sigma q_i$, with $i=1,2,3,$ for $q\in\mathbb{R}^3$ and constant $\sigma$. In that particular case, the double bracket terms vanish. In the present study, we will take $h_i(p,q)$ as the two cases mentioned in Remark~\ref{sec: Remark on P-SD and CH-SD} and compare their effects on the dynamics of \emph{peakons} with $K (x - y)=\exp (-|x-y|/\alpha)$ and \emph{pulsons} with $K (x - y)=\exp (-(x-y)^2/\alpha^2)$ in one spatial dimension. We have introduced the latter Gaussian shaped pulsons, in order to determine how sensitively the numerical results we shall discuss below depend on the jump in derivative of the velocity profile for peakons.
\end{enumerate}
\end{remark}
\section{The Fokker-Planck equation}
\label{sec: The Fokker-Planck equation}
The stochastic process in \eqref{SEP-eqns-thm-qp} for $(q(t),p(t))$ can be described with the help of a transition density function $\rho(t,q,p; \bar q, \bar p)$ which represents the probability density that the process, initially in the state $(\bar q, \bar p)$, will reach the state $(q,p)$ at time $t$. The transition density function satisfies the Fokker-Planck equation corresponding to \eqref{SEP-eqns-thm-qp} (see \cite{GardinerStochastic}, \cite{KloedenPlatenSDE}).
Let us examine the form of this equation in the case $h^{(2)}_1(q,p)= \beta_1 p_1$, $h^{(2)}_2(q,p)= \beta_2 p_2$. In that case the noise in \eqref{SEP-eqns-thm-qp} is additive, and the Stratonovich and It\^o calculus yield the same equations of motion.
\subsection{Single-pulson dynamics}
\label{sec: Single-pulson dynamics}
Consider a single pulson ($N=1$) subject to one-dimensional (i.e., $M=1$) Wiener process, with the stochastic potential $h(q,p)=\beta p$, where $\beta$ is a nonnegative real parameter. The stochastic Hamiltonian equations \eqref{SEP-eqns-thm-qp} take the form $dq = p \, dt + \beta \circ dW(t)$, $dp =0,$ which are easily solved by
\begin{equation}
\label{eq: Solution of the stochastic Hamiltonian equations for one pulson}
q_\beta(t) = \bar q + \bar p t + \beta W(t), \qquad \qquad p_\beta(t) = \bar p,
\end{equation}
where $(\bar q, \bar p)$ are the initial conditions. Note that the pulson/peakon retains its initial momentum/height $\bar p$. We will use this solution as a reference for the convergence test in Section~\ref{sec: Convergence test}. The corresponding Fokker-Planck equation takes the form
\begin{equation}
\label{eq: F-P equation for one pulson}
\frac{\partial \rho}{\partial t} + p \frac{\partial \rho}{\partial q} - \frac{1}{2} \beta^2 \frac{\partial^2 \rho}{\partial q^2}=0
\end{equation}
with the initial condition $\rho(0,q,p; \bar q, \bar p) = \delta(q-\bar q)\delta(p-\bar p)$. This advection-diffusion equation is easily solved with the help of the fundamental solution for the heat equation, and the solution yields
\begin{equation}
\label{eq: Solution of F-P equation for one pulson}
\rho_\beta(t,q,p; \bar q, \bar p) = \frac{1}{\beta \sqrt{2 \pi t}} e^{-\frac{(q-\bar q-pt)^2}{2 \beta^2 t}} \delta(p-\bar p).
\end{equation}
This solution means that the initial momentum $\bar p$ is preserved, which is consistent with \eqref{eq: Solution of the stochastic Hamiltonian equations for one pulson}. The position has a Gaussian distribution which widens with time, and whose maximum is advected with velocity $\bar p$.
\subsection{Two-pulson dynamics}
\label{sec: Two-pulson dynamics}
The dynamics of two interacting pulsons has been thoroughly studied and possesses interesting features (see \cite{FringerHolmGeodesic}, \cite{HolmGMS}).
It is therefore intriguing to see how this dynamics is affected by the presence of noise. Consider $N=2$ pulsons subject to a two-dimensional (i.e., $M=2$) Wiener process, with the stochastic potentials $h_1(q,p) = \beta_1 p_1$ and $h_2(q,p) = \beta_2 p_2$, where $q=(q_1,q_2)$, $p=(p_1,p_2)$, and $\beta_1,\beta_2 \geq 0$. The corresponding Fokker-Planck equation takes the form
\begin{align}
\label{eq: F-P equation for two pulsons}
\frac{\partial \rho}{\partial t} + \frac{\partial}{\partial q_1} \big[a_1(q,p) \rho \big] + \frac{\partial}{\partial q_2} \big[a_2(q,p) \rho \big] + \frac{\partial}{\partial p_1} \big[a_3(q,p) \rho \big] + \frac{\partial}{\partial p_2} \big[a_4(q,p) \rho \big] - \frac{1}{2} \beta_1^2 \frac{\partial^2 \rho}{\partial q_1^2} - \frac{1}{2} \beta_2^2 \frac{\partial^2 \rho}{\partial q_2^2}=0
\end{align}
with the initial condition $\rho\big(0,q,p; \bar q, \bar p \big) = \delta(q_1-\bar q_1) \delta(p_1-\bar p_1) + \delta(q_2-\bar q_2) \delta(p_2-\bar p_2)$, where
\begin{align}
\label{eq: Drift functions}
\begin{split}
a_1(q,p) &= p_1 + p_2 K(q_1-q_2), \qquad\qquad a_3(q,p) = -p_1 p_2 K'(q_1-q_2), \\
a_2(q,p) &= p_2 + p_1 K(q_1-q_2), \qquad\qquad a_4(q,p) = p_1 p_2 K'(q_1-q_2).
\end{split}
\end{align}
Despite its relatively simple structure, it does not appear to be possible to solve this equation analytically. It is nevertheless an elementary exercise to verify that the function
\begin{equation}
\label{eq: Asymptotic solution to the F-P equation for two pulsons}
\rho(t,q_1,q_2,p_1,p_2; \bar q_1, \bar q_2, \bar p_1, \bar p_2) = \rho_{\beta_1}(t,q_1,p_1; \bar q_1, \bar p_1) + \rho_{\beta_2}(t,q_2,p_2; \bar q_2, \bar p_2),
\end{equation}
where $\rho_{\beta_i}$ is given by \eqref{eq: Solution of F-P equation for one pulson}, satisfies \eqref{eq: F-P equation for two pulsons} asymptotically as $q_1-q_2 \longrightarrow \pm \infty$, assuming the Green's function and its derivative decay in that limit. This simple observation gives us an intuition that stochastic pulsons should behave like individual particles when they are far from each other, just like in the deterministic case. In order to study the stochastic dynamics of the collision of pulsons, we need to resort to Monte Carlo simulations.
In Section~\ref{sec: Stochastic variational integrator} we discuss our numerical algorithm, and in Section~\ref{sec: Numerical experiments} we present the results of our numerical studies.
\subsection{Two-pulson dynamics with P-SD}
\label{sec: Two-pulson dynamics with P-SD}
The stochastic interaction of two (or more) pulsons can be analysed explicitly when the stochastic potential has the form $h(q,p)=\beta (p_1+p_2)$ for $\beta\geq 0$ (P-SD for $M=1$; see Remark~\ref{sec: Remark on P-SD and CH-SD}). It is an elementary exercise to show that in that case the stochastic Hamiltonian equations \eqref{SEP-eqns-thm-qp} are solved by
\begin{equation}
\label{eq: P-SD exact solution}
q_a(t) = q^D_a(t) + \beta W(t), \qquad\qquad p_a(t) = p^D_a(t), \qquad\qquad a=1,2,
\end{equation}
\noindent
where $q^D_a(t)$ and $p^D_a(t)$ are the solutions of the deterministic system \eqref{pkn-pq-eqns}. We verify this numerically in Section~\ref{eq: Additive noise in the momentum equation}.
\section{Stochastic variational integrator}
\label{sec: Stochastic variational integrator}
Given the variational structure of the problem we have formulated in Theorem \ref{SEP-thm-peakons}, it is natural to employ variational integrators for numerical simulations. For an extensive review of variational integrators we refer the reader to Marsden \& West \cite{MarsdenWestVarInt} and the references therein. Stochastic variational integrators were first introduced in Bou-Rabee \& Owhadi \cite{BR-O2009}. These integrators were derived for Lagrangian systems using the Hamilton-Pontryagin variational principle. In our case, however, we find it more convenient to stay on the Hamiltonian side and use the discrete variational Hamiltonian mechanics introduced in Lall \& West \cite{LallWestHamiltonian}. We combine the ideas of \cite{BR-O2009} and \cite{LallWestHamiltonian}, and propose the following discretization of the phase-space action functional \eqref{eq: Stochastic phase-space action functional}:
\begin{equation}
\label{eq: Discrete action}
S_d = \sum_{k=0}^{K-1} \sum_{i=1}^N \bigg(p_i^k (q_i^{k+1}-q_i^k) - H(q^{k+1},p^k)\Delta t \bigg) - \sum_{k=0}^{K-1} \sum_{m=1}^M \frac{h_m(q^k,p^k) + h_m(q^{k+1},p^{k+1})}{2} \Delta W_k^m,
\end{equation}
where $\Delta t = T/K$ is the time step, $(q^k, p^k)$ denote the position and momentum at time $t_k = k \Delta t$, and $\Delta W^m_k \sim N(0,\Delta t)$ are independent normally distributed random variables for $m=1,\ldots,M$ and $k=0,\ldots, K-1$. Let $L(q, \dot q)$ denote the Lagrangian related to $H(q,p)$ via the standard Legendre transform $\dot q_i = \partial H / \partial p_i$. Then one can easily show that \eqref{eq: Discrete action} is equivalent to the discretization
\begin{equation}
\label{eq: Discrete H-P action}
S_d = \sum_{k=0}^{K-1} \bigg( L(q^k,v^k) + \sum_{i=1}^N p_i^{k+1} \Big( \frac{q_i^{k+1}-q_i^k}{\Delta t} - v^{k+1} \Big) \bigg) \Delta t - \sum_{k=0}^{K-1} \sum_{m=1}^M \frac{h_m(q^k,p^k) + h_m(q^{k+1},p^{k+1})}{2} \Delta W_k^m
\end{equation}
of the Hamilton-Pontryagin principle used in \cite{BR-O2009}. Omitting the details, \eqref{eq: Discrete action} is obtained by computing the left discrete Hamiltonian $H^-(q^{k+1},p^k)$ corresponding to the discrete Lagrangian $L_d(q^k, q^{k+1}) = \Delta t L(q^{k+1},(q^{k+1}-q^k)/\Delta t)$. The interested reader is referred to \cite{LallWestHamiltonian} for more details. Extremizing \eqref{eq: Discrete action} with respect to $q^k$ and $p^k$ yields the following implicit stochastic variational integrator:
\begin{align}
\label{eq: Stochastic Variational Integrator}
\frac{q_i^{k+1} - q_i^k}{\Delta t} &= \frac{\partial H}{\partial p_i}\big(q^{k+1},p^k\big) +\sum_{m=1}^M \frac{\partial h_m}{\partial p_i}\big(q^k,p^k\big) \frac{\Delta W_{k-1}^m+\Delta W_k^m}{2 \Delta t}, \nonumber \\
\frac{p_i^{k+1} - p_i^k}{\Delta t} &= -\frac{\partial H}{\partial q_i}\big(q^{k+1},p^k\big) -\sum_{m=1}^M \frac{\partial h_m}{\partial q_i}\big(q^{k+1},p^{k+1}\big) \frac{\Delta W_k^m+\Delta W_{k+1}^m}{2 \Delta t},
\end{align}
for $i=1,\ldots, N$. Knowing $(q^k,p^k)$ at time $t_k$, the system above allows to solve for the position $q^{k+1}$ and momentum $p^{k+1}$ at the next time step. For increased computational efficiency, it is advisable to solve the first (nonlinear) equation for $q^{k+1}$ first, and then the second equation for $p^{k+1}$.
Note that in \eqref{eq: Discrete action} we used $p_i^k (q_i^{k+1}-q_i^k)$ to approximate the Stratonovich integral $\int_{t_k}^{t_{k+1}}p_i\circ dq_i$ in \eqref{eq: Stochastic phase-space action functional}, which means the numerical scheme \eqref{eq: Stochastic Variational Integrator} will not be convergent for general nonlinear stochastic potentials $h_i(q,p)$ (i.e., multiplicative noise). If we used $(p_i^k+p_i^{k+1})(q_i^{k+1}-q_i^k)/2$, the resulting integrator would be a two-step method, i.e., it would not be self-starting, and its geometric/symplectic properties would be in question. Nevertheless, for additive noise, i.e., when the stochastic potentials $h_i(q,p)$ are linear in their arguments, the integrator \eqref{eq: Stochastic Variational Integrator} is a simple modification of the Euler-Maruyama method and its convergence can be established using similar techniques (see \cite{Milstein1995}). The integrator \eqref{eq: Stochastic Variational Integrator} has strong order of convergence 0.5, and weak order of convergence 1. We further verify this fact numerically in Section~\ref{sec: Convergence test}.
The integrator \eqref{eq: Stochastic Variational Integrator} is symplectic, and preserves momentum maps corresponding to (discrete) symmetries of the discrete Hamiltonian---for instance, if $H(q,p)$ and all $h_i(q,p)$ are translationally invariant, as in our simulations in Section~\ref{sec: Numerical experiments}, then the total momentum $\sum_{i=1}^N p_i$ is numerically preserved. The proof of these facts trivially follows from \cite{BR-O2009}, keeping in mind that the momenta $p_i$ and velocities $\dot q_i$ are related via the Legendre transform.
\section{Numerical experiments}
\label{sec: Numerical experiments}
We performed numerical simulations of the rear-end collision of two pulsons for two different Green's functions, namely $K(q_1-q_2) = e^{-(q_1-q_2)^2}$ and $K(q_1-q_2) = e^{-2|q_1-q_2|}$. In the latter case, the corresponding pulsons are commonly called \textquoteleft peakons'. We investigated the initial conditions $\bar q_1 = 0$, $\bar q_2 = 10$, $\bar p_2 =1$ together with the following four initial values: $\bar p_1 = 8$, $\bar p_1 = 4$, $\bar p_1 = 2$, $\bar p_1 = 1.$ That is, we varied the initial momentum of the faster pulson. We perturbed the slower pulson by introducing a one-dimensional Wiener process with the stochastic potential $h(q,p)=\beta p_2$ (this corresponds to $\beta_1=0$, $\beta_2=\beta$ in Section~\ref{sec: Two-pulson dynamics}). The pulsons were initially well-separated, so their initial evolution was described by \eqref{eq: Asymptotic solution to the F-P equation for two pulsons}. The parameter $\beta$ was varied in the range $[0, 6.5]$. We used the time step $\Delta t = 0.02$, and for each choice of the parameters 50000 sample solutions were computed until the time $T=100$.
\subsection{Sample paths and mean solutions}
\label{sec: Sample paths and mean solutions}
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{Sample_Paths_p40_beta40-eps-converted-to.pdf}
\caption{Example numerical sample paths for Gaussian pulsons for the simulations with $\bar p_1=4$ and $\beta=4$. The positions are depicted in the plots in the upper row, and the corresponding momenta are shown in the plots in the lower row.}
\label{fig: Sample paths 1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.8\textwidth]{Mean_Paths_p40_b15_b25_b45-eps-converted-to.pdf}
\caption{Numerical mean paths for Gaussian pulsons for the simulations with $\bar p_1=4$. Results for three example choices of the parameter $\beta$ are presented: $\beta=1.5$ ({\it left}), $\beta=2.5$ ({\it middle}), and $\beta=4.5$ ({\it right}). The positions are depicted in the plots in the upper row, and the corresponding momenta are shown in the plots in the lower row.}
\label{fig: Mean paths 1}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth]{Sample_Paths_p10_beta50-eps-converted-to.pdf}
\caption{Example numerical sample paths for Gaussian pulsons for the simulations with $\bar p_1=1$ and $\beta=5$. The positions are depicted in the plots in the upper row, and the corresponding momenta are shown in the plots in the lower row.}
\label{fig: Sample paths 2}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth]{Mean_Paths_p10_b05_b10_b20-eps-converted-to.pdf}
\caption{Numerical mean paths for Gaussian pulsons for the simulations with $\bar p_1=1$. Results for three example choices of the parameter $\beta$ are presented: $\beta=0.5$ ({\it left}), $\beta=1$ ({\it middle}), and $\beta=2$ ({\it right}). The positions are depicted in the plots in the upper row, and the corresponding momenta are shown in the plots in the lower row.}
\label{fig: Mean paths 2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{MeanHamiltonian-eps-converted-to.pdf}
\caption{ Mean Hamiltonian for the simulations with $\bar p_1 = 4$ for Gaussian pulsons (\emph{left)} and peakons (\emph{right)}. }
\label{fig: Hamiltonian}
\end{figure}
Figure~\ref{fig: Sample paths 1} shows a few sample paths from the simulations of the interaction of Gaussian pulsons for the case with $\bar p_1=4$ and $\beta=4$. The simulations for $\bar p_1=8$ and $\bar p_1=2$, as well as the simulations for peakons, gave qualitatively similar results. The most striking feature is that the faster pulson/peakon may in fact cross the slower one. In the deterministic case one can show that the faster pulson can never pass the slower one---they just exchange their momenta. The proof relies on the fact that both the Hamiltonian and total momentum are preserved (see \cite{FringerHolmGeodesic}, \cite{HolmGMS}). In our case, however, the Hamiltonian \eqref{qp-Ham} is not preserved due to the presence of the time-dependent noise (see Figure~\ref{fig: Hamiltonian}), which allows much richer dynamics of the interactions. This may find interesting applications in landmark matching---see the discussion in Section~\ref{sec: Summary and future work}.
Looking at Figure~\ref{fig: Sample paths 1} we also note that our variational integrator exactly preserves the total momentum, as expected. Figure~\ref{fig: Mean paths 1} depicts the mean solution for Gaussian pulsons with the initial condition $\bar p_1=4$ for different values of the noise intensity $\beta$. We see that for small noise the mean solution resembles the deterministic one, but as the parameter $\beta$ is increased, the mean solution represents two pulsons passing through each other with increasingly less interaction. We study the probability of crossing in more detail in Section~\ref{sec: Probability of crossing}.
We observed that pulsons may cross even when they have the same initial momentum (see Figure~\ref{fig: Sample paths 2}). In the deterministic case they would just propagate in the same direction, retaining their relative distance. Nevertheless, the mean solution (see Figure~\ref{fig: Mean paths 2}) does not show any crossing.
\subsection{Probability of crossing}
\label{sec: Probability of crossing}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{DistributionDensity_p40_b15_b25_b45-eps-converted-to.pdf}
\caption{Numerical probability density $\rho$ of the distance $\Delta q(t)=q_2(t)-q_1(t)$ at time $t=100$ for Gaussian pulsons for the simulations with $\bar p_1=4$ (cf. Figure~\ref{fig: Mean paths 1}). Results for three example choices of the parameter $\beta$ are presented: $\beta=1.5$ ({\it left}), $\beta=2.5$ ({\it middle}), and $\beta=4.5$ ({\it right}). }
\label{fig: Distribution Density 1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{DistributionDensity_p10_b05_b10_b20-eps-converted-to.pdf}
\caption{Numerical probability density $\rho$ of the distance $\Delta q(t)=q_2(t)-q_1(t)$ at time $t=100$ for Gaussian pulsons for the simulations with $\bar p_1=1$ (cf. Figure~\ref{fig: Mean paths 2}). Results for three example choices of the parameter $\beta$ are presented: $\beta=0.5$ ({\it left}), $\beta=1$ ({\it middle}), and $\beta=2$ ({\it right}). }
\label{fig: Distribution Density 2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{CrossingProbability-eps-converted-to.pdf}
\caption{The probability of crossing, that is, the probability that $q_2(t)<q_1(t)$ at time $t=100$, as a function of the parameter $\beta$ for Gaussian pulsons ({\it top}) and peakons ({\it bottom}). }
\label{fig: Crossing Probability}
\end{figure}
We studied in more detail the distance between the pulsons $\Delta q(t) = q_2(t)-q_1(t)$ at the end of the simulation, that is, at time $t=100$. Figure~\ref{fig: Distribution Density 1} presents the experimental probability density function of $\Delta q$ computed for Gaussian pulsons with $\bar p_1=4$. The density appears to have two local maxima, with the global maximum shifting from positive to negative values of $\Delta q$ as the noise intensity $\beta$ is increased. The simulations for $\bar p_1=8$ and $\bar p_1=2$, as well as the simulations for peakons, gave qualitatively similar results. Figure~\ref{fig: Distribution Density 2} depicts analogous results for the case of Gaussian pulsons with $\bar p_1=1$. In this case the density function also has two local maxima, but the global maximum never shifts to negative values of $\Delta q$. The probability of crossing as a function of the noise intensity $\beta$ is depicted in Figure~\ref{fig: Crossing Probability}. We see that this probability seems to approach unity for the simulations with $\bar p_1 >1$, and 0.5 for $\bar p_1 = 1$.
\subsection{First crossing time}
\label{sec: First crossing time}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth]{FirstCrossingTimeDistributionDensity_p40_b25-eps-converted-to.pdf}
\caption{ The blue histogram presents the experimental probability density of the first crossing time $T_c$ for Gaussian pulsons for the simulations with $\bar p_1=4$ and $\beta=2.5$. More precisely, this is the conditional probability density given that $T_c<\infty$, i.e., assuming the pulsons do cross, the integral $\int_a^b \rho(\tau)\, d\tau$ yields the probability that the first crossing occurs at time $T_c \in [a,b]$. The black line depicts the inverse Gaussian distribution \eqref{eq: Inverse Gaussian distribution} with the parameters \eqref{eq: Mean and shape parameter}. }
\label{fig: Distribution density for the first crossing time}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{MeanFirstCrossingTime-eps-converted-to.pdf}
\caption{The mean first crossing time $E(T_c)$ as a function of the parameter $\beta$ for Gaussian pulsons ({\it top}) and peakons ({\it bottom}). More precisely, this is the conditional expectation $E(T_c | T_c<\infty)$ given that the pulsons do cross (i.e., $T_c<\infty$). }
\label{fig: Mean first crossing time}
\end{figure}
It might be of interest to investigate the earliest time when the pulsons cross $T_c=\inf \{t>0: \Delta q(t)=0 \}$---let us call it the first crossing time (also known as the \emph{first exit time} or the \emph{hitting time}; see \cite{GardinerStochastic}, \cite{KloedenPlatenSDE}). Assume $T_c=+\infty$ if no crossing occurs. We can approximate the probability density of $T_c$ using the single-pulson solution \eqref{eq: Solution of the stochastic Hamiltonian equations for one pulson} and the asymptotic property \eqref{eq: Asymptotic solution to the F-P equation for two pulsons}. The two pulsons are initially far from each other, so we have
\begin{equation}
\label{eq: Asymptotic Delta q}
\Delta q(t) \approx \Delta \bar q - (\bar p_1 -\bar p_2) t + \beta W(t),
\end{equation}
\noindent
that is, $\Delta q(t)$ is approximated by a Brownian motion starting at $\Delta \bar q = \Delta q(0)>0$ with the drift $\bar p_1 -\bar p_2$. Assuming that $\bar p_1>\bar p_2$, the probability density of $T_c$ is given by the inverse Gaussian distribution $T_c \sim IG(\mu,\lambda)$ with the mean $\mu$ and shape parameter $\lambda$, respectively,
\begin{equation}
\label{eq: Mean and shape parameter}
\mu = \frac{\Delta \bar q}{\bar p_1 -\bar p_2}, \qquad\qquad \lambda=\frac{\Delta \bar q^2}{\beta^2},
\end{equation}
\noindent
where the density function is
\begin{equation}
\label{eq: Inverse Gaussian distribution}
\rho_{\text{IG}}(\tau) = \sqrt{\frac{\lambda}{2 \pi \tau^3}} \exp{\frac{-\lambda(\tau-\mu)^2}{2 \mu^2 \tau} }.
\end{equation}
\noindent
An example of a conditional probability density function of $T_c$ is depicted in Figure~\ref{fig: Distribution density for the first crossing time}. It shows very good agreement between the experimental and approximate theoretical distributions. The corresponding density functions for all other simulations look qualitatively similar, differing in the mean, variance, etc. When $\bar p_1 \longrightarrow \bar p_2^+$, then $\mu\longrightarrow +\infty$ and the inverse Gaussian distribution \eqref{eq: Inverse Gaussian distribution} tends to the L\'evy distribution, whose mean is infinite. However, we observed that for the simulations with $\bar p_1=\bar p_2=1$ the first crossing time is still well-approximated by the inverse Gaussian distribution with the shape parameter $\lambda$ as in \eqref{eq: Mean and shape parameter} and mean $\mu$ which seems to asymptotically decrease with the noise intensity $\beta$. The conditional mean crossing time $E(T_c | T_c<\infty)$ (given that a crossing occurs) as a function of the noise intensity $\beta$ is depicted in Figure~\ref{fig: Mean first crossing time}. The mean crossing time shows minor variations for the simulations with $\bar p_1 > 1$ and agrees well with \eqref{eq: Mean and shape parameter}, while it appears to asymptotically decrease for the simulations with $\bar p_1 = 1$.
\subsection{Noise screening}
\label{sec: Noise screening}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth]{NoiseScreening_p4_b4_b4-eps-converted-to.pdf}
\caption{ Example numerical sample paths for Gaussian pulsons for the simulations with the initial conditions $\bar q_1 = 0$, $\bar p_1 = 4$, $\bar q_2 = 10$, and $\bar p_2 =1$, and the stochastic potential \eqref{eq: Noise screening stochastic potential} with the parameters $\beta=4$ and $\gamma=4$. The positions are depicted in the plots in the upper row, and the corresponding momenta are shown in the plots in the lower row. }
\label{fig: Noise screening sample paths}
\end{figure}
In the numerical experiments described above we observed that the presence of noise causes pulsons to cross with a non-zero probability. The functions $q_1(t)$, $p_1(t)$, $q_2(t)$ and $p_2(t)$ define a transformation of the real line through \eqref{m-def-thm}. In the deterministic case this transformation is a diffeomorphism, but not when noise is added, since the crossing of pulsons introduces topological changes in the image of the real line under this transformation. This may be of interest in image matching, as in \cite{TrVi2012}, when one would like to construct a deformation between two images which are not exactly diffeomorphic. However, with that application in mind, one may want to restrict the stochastic effects only to the situation when two pulsons get close to each other. This can be obtained by applying the stochastic potential
\begin{equation}
\label{eq: Noise screening stochastic potential}
h(q,p) = \beta p_2 e^{-\frac{(q_2-q_1)^2}{\gamma}}.
\end{equation}
\noindent
The parameter $\beta \geq 0$ adjusts the noise intensity, just as before, while the parameter $\gamma>0$ controls the range over which the stochastic effects are non-negligible. We performed a few simulations with this stochastic potential. Since this potential is nonlinear, the integrator \eqref{eq: Stochastic Variational Integrator} is not applicable here. Instead, we used the stochastic symplectic midpoint rule (see \cite{MilsteinRepin}). A few sample paths are depicted in Figure~\ref{fig: Noise screening sample paths}. Note that this stochastic potential is translation-invariant, so the total momentum is preserved.
\subsection{Restriction to parametric noise and additive noise in the momentum equation}
\label{eq: Additive noise in the momentum equation}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{ParametricNoise_p4_b4.pdf}
\caption{ Example numerical sample paths for Gaussian pulsons for the simulations with the initial conditions $\bar q_1 = 0$, $\bar p_1 = 4$, $\bar q_2 = 10$, and $\bar p_2 =1$, and the stochastic potential $h(q,p)=\beta (p_1+p_2)$ with the parameter $\beta=4$.}
\label{fig: Parametric noise sample paths}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{AdditiveNoiseQ_p4_b4-eps-converted-to.pdf}
\caption{ Example numerical sample paths for Gaussian pulsons for the simulations with the initial conditions $\bar q_1 = 0$, $\bar p_1 = 4$, $\bar q_2 = 10$, and $\bar p_2 =1$, and the stochastic potential $h(q,p)=\beta q_2$ with the parameter $\beta=4$. The positions are depicted in the plots in the upper row, and the corresponding momenta are shown in the plots in the lower row. }
\label{fig: Additive noise in q sample paths}
\end{figure}
Interestingly, crossing of pulsons does not seem to occur for the case of parametric stochastic deformation with the restriction $\varphi_{ia}(q) = \xi_i(q_a)$ as in Corollary~\ref{Lemma-m-eqn-peakon}. We ran numerical experiments for the potential $h(q,p)= \beta (p_1+p_2)$, which has the form as in Corollary~\ref{Lemma-m-eqn-peakon} with $\xi(x) = \beta$, but observed no interpenetration (see Figure~\ref{fig: Parametric noise sample paths}). This is consistent with our observation in Section~\ref{sec: Two-pulson dynamics with P-SD} and the fact that pulsons never cross in the deterministic case. We also did not observe crossing when the stochastic potential is independent of $p$. For instance, we performed simulations with the potential $h(q,p) = \beta q_2$. Such a potential results in additive noise in the momentum equation in \eqref{SEP-eqns-thm-qp} only, as in \cite{TrVi2012}. A few sample paths are depicted in Figure~\ref{fig: Additive noise in q sample paths}. Note that in this case the total momentum is not preserved, since $h(q,p)$ is not translationally invariant. In many cases the pulsons would asymptotically approach each other, but never pass. We observed similar behavior for the (translationally invariant) potential $h(q,p)=\beta \exp (-(q_1-q_2)^2 / \gamma)$ with $\beta, \gamma >0$.
\subsection{Convergence tests}
\label{sec: Convergence test}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{ConvergencePlot-eps-converted-to.pdf}
\caption{ Dependence of the absolute and mean errors on the time step for the single Gaussian pulson (\emph{left}), two point vortices (\emph{center}) and Kubo oscillator (\emph{right}). }
\label{fig: Convergence plot}
\end{figure}
\subsubsection{Single pulson}
In order to test the convergence of the numerical algorithm \eqref{eq: Stochastic Variational Integrator} we performed computations for $N=1$ Gaussian pulson subject to one-dimensional (i.e., $M=1$) Wiener process with the stochastic potential $h(q,p)=\beta p$ (cf. Section~\ref{sec: Single-pulson dynamics}). Simulations with the initial conditions $\bar q=0$, $\bar p =4$ and the noise intensity $\beta=4$ were carried until the time $T=2$ for a number of decreasing time steps $\Delta t$. In each case 50,000 sample paths were generated. Let $z_{\Delta t}(t) = (q_{\Delta t}(t), p_{\Delta t}(t) )$ denote the numerical solution. We used the exact solution \eqref{eq: Solution of the stochastic Hamiltonian equations for one pulson} as a reference for computing the absolute error $E(|z_{\Delta t}(T)-z_\beta(T)|)$ and the mean error $| E(z_{\Delta t}(T))-E(z_\beta(T)) |$, where $z_\beta(t) = (q_\beta(t), p_\beta(t) )$. The dependence of these errors on the time step $\Delta t$ is depicted in Figure~\ref{fig: Convergence plot}. We verified that our algorithm has strong order of convergence $0.5$, and weak order of convergence 1.
\subsubsection{Two planar point vortices}
We performed a similar test for $N=2$ planar point vortices subject to a one-dimensional (i.e., $M=1$) Wiener process. The system is described by
\begin{equation}
H(q,p)=-\frac{1}{4 \pi} \Gamma_1 \Gamma_2 \log \bigg[ \Big( \frac{q_1}{\sigma_1}-\frac{q_2}{\sigma_2} \Big)^2 + \Big( \frac{p_1}{\lambda_1}-\frac{p_2}{\lambda_2} \Big)^2 \bigg], \qquad\quad h(q,p) = \beta \lambda_1 q_1 + \gamma \sigma_1 p_1 + \beta \lambda_2 q_2 + \gamma \sigma_2 p_2,
\end{equation}
\noindent
where $\Gamma_1, \Gamma_2$ are the circulations of the vortices, $\sigma_i = \sqrt{|\Gamma_i|}\,\text{sgn}\,\Gamma_i$, $\lambda_i = \sqrt{|\Gamma_i|}$ are scaling factors, $\beta, \gamma$ are the noise intensities, and $q_i, p_i$ denote the $x$- and $y$-coordinate of the $i$-th vortex, respectively (see \cite{Flandoli}, \cite{Newton}). Simulations for $\Gamma_1=2$, $\Gamma_2=1$ with the initial conditions $\bar q_1 = \sigma_1 R_1$ , $\bar q_2 = \sigma_2 R_2$, $\bar p_1=\bar p_2=0$, where $R_1=\Gamma_2/(\Gamma_1+\Gamma_2)$, $R_2=-\Gamma_1/(\Gamma_1+\Gamma_2)$, and the noise intensities $\beta=\gamma=0.5$ were carried out until the time $T=6.4$ for a number of decreasing time steps $\Delta t$. In each case 50000 sample paths were generated. We used the exact solution (see \cite{Flandoli}, \cite{Newton})
\begin{align}
q_1(t) &= \sigma_1 R_1 \cos \omega t + \gamma \sigma_1 W(t), \qquad p_1(t) = \lambda_1 R_1 \sin \omega t - \beta \lambda_1 W(t), \\
q_2(t) &= \sigma_2 R_2 \cos \omega t + \gamma \sigma_2 W(t), \qquad p_2(t) = \lambda_2 R_2 \sin \omega t - \beta \lambda_2 W(t),
\end{align}
\noindent
where $\omega = (\Gamma_1+\Gamma_2)/(2 \pi)$, as a reference for computing the absolute and mean errors (see Figure~\ref{fig: Convergence plot}). We verified that our algorithm has strong order of convergence $0.5$, and weak order of convergence 1.
\subsubsection{Kubo oscillator}
To demonstrate that the integrator \eqref{eq: Stochastic Variational Integrator} fails to converge for multiplicative noise, we performed computations for the Kubo oscillator, which is defined by $H(q,p)=p^2/2+q^2/2$ and $h(q,p)=\beta(p^2/2+q^2/2)$, where $\beta$ is the noise intensity (see \cite{MilsteinRepin}). The exact solution is given by
\begin{equation}
q(t)=\bar p \sin(t+\beta W(t)) + \bar q \cos(t+\beta W(t)), \qquad\quad p(t)=\bar p \cos(t+\beta W(t)) - \bar q \sin(t+\beta W(t)).
\end{equation}
\noindent
A similar convergence test with $\bar q=0$, $\bar p=4$, $\beta=1$, and $T=6.4$ revealed that the integrator \eqref{eq: Stochastic Variational Integrator} failed to converge, although the errors remained bounded (see Figure~\ref{fig: Convergence plot}).
\section{Summary}
\label{sec: Summary and future work}
We have seen in Section \ref{sec: ParametricStochasticDeformations} that the finite-dimensional peakon solutions for the EPDiff partial differential equation in one spatial dimension persist under both parametric stochastic deformation (P-SD) and canonical Hamiltonian stochastic deformations (CH-SD) of the EPDiff variational principle. Being both finite-dimensional and canonically Hamiltonian, the dynamics of the peakon solution set for EPDiff admits the entire range of CH-SD in the sense of \cite{Bi1981,LaCa-Or2008}, which includes P-SD but can be more general. Therefore, the peakon solution set offers a finite-dimensional laboratory for comparing the effects of P-SD and CH-SD on the stochastically deformed EPDiff SPDE solution behaviour. In fact, as it turns out, the peakon solution set for EPDiff offers a particularly sensitive assessment of the effects of stochasticity on finite-dimensional solutions of SPDE. In Section \ref{sec: ParametricStochasticDeformations}, we took advantage of the flexibility of CH-SD to study stochastic peakon-peakon collisions in which noise was introduced into \emph{only one} of the peakon position equations (rather than symmetrically into both of the canonical position equations, as occurs with P-SD), while at the same time not introducing any noise into either of the corresponding canonical momentum equations. The precision and flexibility of the CH-SD approach to stochastic peakon-peakon collision dynamics revealed that its asymmetric case with noise in only one canonical position equation allows the soliton-like singular peakon and pulson solutions of EPDiff to interpenetrate and change order on the real line, although this is not possible for the diffeomorphic flow represented by the solutions of the unperturbed deterministic EPDiff equation. This crossing of peakon paths was observed and its statistics were studied in detail for CH-SD in numerical experiments in Section~\ref{sec: Numerical experiments}. In contrast, crossing of peakon paths was \emph{not} observed for the corresponding P-SD simulations in which the noise enters symmetrically in both position equations. Crossing of peakon paths was also not observed when stochasticity was added only in the canonical momentum equations, as studied in \cite{TrVi2012}.
Thus, for the deterministic EPDiff, adding stochasticity of \emph{constant amplitude} with either CH-SD of P-SD to a finite dimensional invariant solution set has been found to produce different SDE solution behaviour. Here, the difference has introduced the possibility of a topological change in the order of points moving on a line in the CH-SD approach, while no such change in topology seems to be available via the P-SD subclass. One can also imagine that changing the level of noise in the P-SD EPDiff SPDE could change the number of peaks or pulsons; a feature which would not have been available if the level of noise were changed after the reduction to a fixed $N$-peakon solution sector. The intriguing idea of creation of singular EPDiff solutions by P-SD noise in the SPDE is under current investigation.
The investigation of stochastic EPDiff in this paper has raised and illustrated a potentially important issue. The need for assessing the validity of approximating the stochastic solution behaviour of nonlinear SPDE by SDE obtained from adding noise to finite-dimensional projections (or discretisations) of the solutions is likely to be encountered quite often in many other circumstances and can be expected to be of frequent future concern. In particular, this issue is likely to occur in considerations of model error in stochastic data assimilation. For example, the endeavours of computational anatomy must face this issue in the use of the singular solutions of EPDiff known as \emph{landmarks} in the task of registration of noisy images \cite{TrVi2012}. This issue of the validity of stochastic deformations of finite-dimensional approximations of evolutionary PDE is a challenge for continuing research in P-SD of EPDiff, as well as in stochastic deformations of more general continuum equations, such as Euler's equations for an ideal fluid, or the Navier-Stokes equations for a viscous fluid. The present work has shown that the introduction of even constant stochasticity into the equations of motion for \emph{exact solutions} (peakons and pulsons, or landmarks, for EPDiff) can produce unexpected changes in topology of the solution in one dimension. The corresponding introduction of stochasticity into the equations of motion for finite-dimensional \emph{approximations} such as discretisation, or projections of the solutions of nonlinear evolutionary PDE may result in other surprises.
\subsection{Acknowledgements}
We are very grateful for the encouragement of the many people who took the time to discuss these matters with us, or comment on drafts, especially our friends and colleagues N. Bou-Rabee, A. Castro, M. Chekroun, C. J. Cotter, D. Crisan, M. O. Hongler, J. P. Ortega and H. Owhadi. However, as usual, any mistakes belong to the authors. This work was partially supported by the European Research Council Advanced Grant 267382 FCCA.
|
\section*{Introduction}
The statistical properties of Twitter and, more generally, of human activity, are characterized by a strong heterogeneity in different dimensions. First, human behavior is known to generate bursty temporal patterns, significantly deviating from independent Poisson processes, as a majority of events take place over short time scales while a few events take place over very large times. This property translates into fat-tailed distributions for the timings $\Delta\tau$ between occurrences of a certain type of events, e.g. between two phone calls or two emails emitted by an individual. For instance, the inter-event time distribution $P(\Delta\tau)$ for the timings between two tweets of a user, or the use of a hashtag is well fitted by a power law such as $P(\Delta\tau)\approx\Delta\tau^\alpha$~\cite{scientificrumor}. The deviation from an exponential (uncorrelated) distribution may be either driven by complex decision-making and cascading mechanisms~\cite{BarbarasiOriginofBursts, Competition_memes, cooperationandcompetition} or by the time dependency of the underlying process, partly because of its intrinsic circadian and weekly rhythms~\cite{Amaral, plane}, as described in Fig. 1, or by a combination of these factors~\cite{Circadianpatternburstiness_mobilephone, burstydynamicsTwitter, 2014arXiv1411.0722F, 2015arXiv150203224M}. Importantly, the nonstationarity of the signal is known to broaden $P(\Delta\tau)$ and therefore to artificially increase the value of standard metrics, such the variance or the Fano factor, originally defined for stationary processes. Recently, a stochastic model for a stationary process also suggests a broad distribution in online user activity level on long time scales, longer than $\Delta\tau$~\cite{2015arXiv150203224M}.
\begin{figure}[]
\includegraphics[width= 10cm]{Fig1_dailytweetComp.pdf}
\caption{Circadian pattern of tweeting activity. Increasing amount of tweets from midday (12:00) to midnight (00:00) is shown in the yellow shaded regions. Significant decays of activity are observed during nights. Activity increases during mornings as shown in purple shaded rectangles. In the inset, we show the temporal evolution at a finer scale, where fluctuations are visible. The data exhibits two peaks of activity in the evening of a political debate, on May 2 2012 (first peak) and on election day, May 6 2012 (second peak).}
\end{figure}
In addition to temporal heterogeneity in $\Delta\tau$, online human activity often generates a heterogeneity in popularity~\cite{onlinepopularityheterogeneity}. In the following, we focus on the popularity of hashtags in Twitter. Twitter is a micro-blogging service allowing users to post short messages, and to follow those published by other users. Messages often incorporate hashtags, keywords identified by the symbol $\#$, which users can track and respond to the message content and makes the platform interactive. Hashtags play a significant role in information diffusion by enhancing information and rumor spreading and consequently increase the impact of news. Discussions on protests~\cite{spanishprotestPLoSONE, spanishprotestSciRep} and political elections, advertisement of new products in marketing, announcements of scientific innovations~\cite{scientificrumor}, panic events such as earthquakes~\cite{earthquakeTwittercollectiveattention}, and comments on TV shows are some examples where hashtags are widely used. Additionally, hashtags can be even used to track and locate crisis~\cite{determinecrisisviahash} and can spread under the influences of both endogeneous factors, that is the propagation between Twitter users following each others, and exogeneous sources such as TV and newspapers~\cite{Dynamicsofbooksale}.
The popularity $p$ of a hashtag is measured by the number of times that it appears in an observation time window. While a majority of hashtags attracts no attention only very few of them propagate heavily~\cite{Viralityprediction_memes, Competition_memes}. Understanding the mechanisms by which certain hashtags or messages gain attention is a central topic of research in the study of online social media~\cite{leskovec}. Potential mechanisms for the emergence of this heterogeneity include forms of preferential attachment and competition-induced forces~\cite{Competition_memes_limitedattention, Competition_inducedcriticality, Competition_advertisement, 2015arXiv150105956G} driven by the limited amount of attention of users.
Our main purpose is to explore connections between temporal and popularity heterogeneity. As a first contribution, we introduce a temporal measure for online human dynamics, suited for the analysis of nonstationary time series to quantify bursts, regularity, and temporal correlations. Originally defined for the study of inter-spike intervals of neurons~\cite{Lv1, Lv2, Lv3, Lv4, Lv5}, the so-called local variation $L_V$ is shown to identify and characterize deviations from Poisson (uncorrelated) processes, and to help predict successful hashtags.
\section*{Data mining and basic analysis}
\subsection*{Data collection and basic overview}
The data set has been collected via the publicly open Twitter streaming API between April 30, 2012, 10 pm and May 10, 2012, 10 pm. Only the geographical constraint has been applied as follows: The actions of all Twitter users located in France have been considered in order to avoid the existence of time differences between countries and regions, and no language filtering has been applied. The time resolution is 1 second and multiple activity can be recorded in the same second. During this time period, two major public events took place: An important political debate held on May 2 and the French presidential election-2012 held on May 6. These events are not the topic of this work, but they are clearly visible in the time series, as shown in Fig. 1.
The total number of tweets, including retweets, captured during the data collection is 9,747,351. The total number of tweets including at least one hashtag is 2,942,239. Around $30\%$ of the tweets therefore contain a hashtag. The fact that hashtags are used in regular tweets or in retweets is not specified. Moreover, any message (identical or not) considering at least one hashtag is recorded. Due to the debate and the election taking place during the data collection, the most popular hashtags are related to politics, as seen in Table 1. The time series of the hashtag study in this paper are provided in Supporting
Information S1. A total number of 473,243 individual users has been identified. Among those, 228,525 users published at least one hashtag, e.g. almost half of the social network is associated with hashtag diffusion.
In order to further characterize the importance of hashtags in Twitter activity, we compare the total number of seconds when any action is performed in the data set, 763,262 s $\approx$ 8.8 days and thus $88\%$ of the total duration, to the number of seconds when at least one hashtag is published, 667,996 s $\approx$ 7.7 days, that is $ 77\%$ of the total duration. In any case, the hashtag data cover a majority of the time window, even during off-peak hours. These numbers confirm the importance of hashtags in the Twitter ecosystem, and their prevalence in a variety of contexts.
\begin{table}[ht]
\caption{{\bf Ranking of popular hashtags.} The first 40 most used hashtags are listed with the corresponding popularity $p$. The hashtags related to the debate and the presidential election such as ledebat, hollande, sarkozy, votehollande, france2012, and présidentielle are recognized.}
\centering
\begin{tabular}{c c c c c c}
\hline\hline
rank & hashtag & popularity $p$ & rank & hashtag & popularity $p$ \\ [0.5ex]
\hline
1 & ledebat & 180946 & 21 & ns & 18715 \\
2 & hollande & 143636 & 22 & ps & 18492 \\
3 & sarkozy & 116906 & 23 & teamfollowback & 18476 \\
4 & votehollande & 99908 & 24 & ggi & 17734 \\
5 & radiolondres & 97622 & 25 & bastille & 16056 \\
6 & bahrain & 71571 & 26 & présidentielle & 13799 \\
7 & fh2012 & 67759 & 27 & afp & 13710 \\
8 & avecsarkozy & 67549 & 28 & france2 & 12906 \\
9 & ledébat & 66668 & 29 & syria & 11594 \\
10 & ff & 49499 & 30 & psg & 10566 \\
11 & ns2012 & 40337 & 31 & sarko & 10503 \\
12 & ump & 25125 & 32 & tf1 & 10201\\
13 & thevoice & 24696 & 33 & mutualite & 10093 \\
14 & fr & 24249 & 34 & egypt & 9970 \\
15 & bayrou & 23029 & 35 & lavictoire & 9949 \\
16 & fh & 22369 & 36 & fn & 9763 \\
17 & rt & 21598 & 37 & franceforte & 9626 \\
18 & france2012 & 20635 & 38 & placeaupeuple & 9211 \\
19 & reseaufdg & 19488 & 39 & jemesouviens & 9098 \\
20 & france & 19268 & 40 & bfmtv & 9010 \\ [1ex]
\hline
\end{tabular}
\label{table:nonlin}
\end{table}
Any type of human activity is influenced by circadian and weekly cycles. This observation has been verified in recent years in a variety of social data sets, going from mobile phone ~\cite{Circadianpatternburstiness_mobilephone} to online social media~\cite{burstydynamicsTwitter,2014arXiv1411.0722F, 2015arXiv150203224M}. In addition, deviations from these cycles can help at detecting atypical events such as responses to catastrophes \cite{scientificrumor, earthquakeTwittercollectiveattention, determinecrisisviahash}. Fig. 1 in the introduction shows the total number of tweets per minute over a sub-period of 6 days and confirms these findings, with clear circadian patterns and two peaks during major public events related to the French presidential election-2012. Besides this smooth periodic behavior, the data also exhibits a noisy signal at a finer time scale, as shown in the inset of Fig. 1. In the following, we will analyze the properties of this complex time series, by decomposing it into groups of hashtags depending on their popularity, and uncover temporal statistical differences between these groups.
\subsection*{Heterogeneity in popularity of hashtags}
The success of a hashtag can be measured by its popularity $p$, defined as its number of occurrences, and equivalent to its frequency. Fig. 2 presents the Zipf-plot and the probability density function (PDF) of $p$, for the 295,697 unique hashtags observed in the data set. The Zipf-plot [Fig. 2(a)] indicates that more than half of the hashtags ($\approx60\%$) appear just once in the data set, with $p$ = 1. Moreover, around $83\%$ of the hashtags have $p<5$, in the pink-colored region in the last (right) rectangle of Fig. 2(a). For moderate values of $p$, if we set a threshold of $p$ to 1000 with an upper-bound to 25000, only $0.15\%$ of the hashtags fit in the yellow-colored rectangle. Finally, top hashtags with $p>25000$, in the red-colored rectangle, are very rare ($\approx0.0001\%$), but more frequent than would be expected for values so large as compared to the median. These observations are confirmed in Fig. 2(b), where we show the probability distribution of $p$, $P(p)$ in a log-log plot. $P(p)$ is a clear example of a fat-tailed distribution associated with a strong heterogeneity in the system.
\begin{figure}[]
\includegraphics[width= 15.5cm]{Fig2_heterogeneitypopularityComp.pdf}
\caption{Heterogeneity in the hashtag popularity $p$ is shown in (a) Zipf-plot and (b) probability density function (PDF), $P(p)$. (a) Diversity in $p$ (frequency) is visible in a power-law scaling in the log-log plot. We rank hashtag from high $p$ (left) to low $p$ (right). Different colored shaded rectangles highlight the value of $p$ from red and orange (high $p$) to purple and pink (low $p$). The percentages describe the overall contributions of the corresponding rectangles. (b) Similarly, $P(p)$ obeys a slowly decaying function and presents a power-law distribution with a fat tail. The same colored schema in (a) is applied to visualize the contributions of different values of $p$.}
\end{figure}
The heterogeneity in $p$ has been already observed~\cite{plane, Viralityprediction_memes, Competition_memes, onlinepopularityheterogeneity}. A mechanism proposed for its emergence is the competition between information overload and the limited capacity of each user~\cite{Competition_memes_limitedattention, Competition_inducedcriticality, Competition_advertisement, 2015arXiv150105956G}, sometimes coupled with cooperative effects~\cite{Competition_memes, cooperationandcompetition}. It has been also shown that hashtags having unique textual features become more popular than hashtags presenting common textual features~\cite{AverageBoring_MemeSuccess}. In this paper, we are not interested in the origin of the heterogeneity, but in its relation with temporal characteristics of hashtags.
\section*{Hashtag spike trains}
\subsection*{Temporal heterogeneity}
We will draw an analogy between hashtag dynamics and neuron spike trains. To this end, we introduce standard methods from spike train analysis into the field of hashtag dynamics. Hashtags are keywords associated to different topics, which can be created, tracked and reused by users. Their popularity and unambiguity make them an essential mechanism for information diffusion in Twitter. The statistical description of neuron spike sequences is essential for extracting underlying information about the brain~\cite{train1}. It was originally believed that in vivo cortical neurons behave as time-dependent Poisson random spike generators, where successive inter-spike intervals are independently chosen from an exponential distribution with a time-dependent firing rate~\cite{train2}. However, more recent observations have shown that the inter-spike interval distribution exhibits significant deviations from the exponential distribution, which has led to the construction of appropriate tools to describe neuron signals~\cite{Lv1, Lv2, Lv3, Lv4, Lv5}.
Similarly, a hashtag spike train is defined as the sequence of timings at which a hashtag is observed in Twitter. In this framework, we do not specify the type of dynamics of hashtags, endogeneous or exogeneous~\cite{Dynamicsofbooksale}, i.e. endogeneous, hashtag diffusion among members of the social network, or exogeneous, the diffusion driven by external factors such as TV and newspapers, but only in the timings. Each hashtag thus generates a unique hashtag spike train with a characteristic popularity $p$. As a first basic indicator, in Figs. 3(a,b) we show the inter-hashtag spike interval cumulative and probability distributions, $CDF(\Delta\tau)$ and $P(\Delta\tau)$, respectively. In order to avoid artificially deforming the distributions because of heterogeneity in $p$, we classify $CDF(\Delta\tau)$ and $P(\Delta\tau)$ in classes depending on $p$, illustrated by different colors in Fig. 2. We observe similar behavior across the classes, as
$P(\Delta\tau)$ deviates strongly from an exponential distribution (Poisson), $P(\Delta\tau)$ = $\xi e^{-\xi\Delta\tau}$, where $\xi$ is a firing rate (frequency and so $p$ in our concept) at which hashtags appear. Instead, we observe fat-tailed distributions~\cite{scientificrumor, onlinepopularityheterogeneity, BarbarasiOriginofBursts, TaroNaokiVotermodel_IEI, Circadianpatternburstiness_mobilephone, ABarratHeterogeneityIEI, Altmann_extreme_events} as shown in Fig. 3(b) for high and moderate $p$. As mentioned in the introduction, this deviation may either originate from temporal correlations or non-stationary patterns, making the system different from a stationary, uncorrelated random signal.
\begin{figure}[]
\includegraphics[width= 8cm]{Fig3_CDFPDF.pdf}
\caption{The cumulative (a), $CDF(\Delta\tau)$, and probability (b), $P(\Delta\tau)$, distributions of the inter-hashtag spike intervals. We observe that $P(\Delta\tau)$ exhibits, for different classes of hashtags distinguished by their popularity, non-exponential features. The different colors correspond to those in Fig. 2. The legend provides the average popularity $\langle p\rangle$ in each hashtag class. The dash lines indicate the positions of 1 day, 2 days, and 3 days, where $P(\Delta\tau)$ gives peaks for low $p$ (pink symbols). The binning is varied from 8 minutes to 2 hours depending on $p$, e.g. 8 min. for high $p$ (red-orange), 1.5 hour for moderate $p$ (yellow-green-blue-purple), and 2 hours for low $p$ (pink). All $P(\Delta\tau)$ present maxima at 1 second, which is not shown to describe tails in a larger window.}
\end{figure}
\subsection*{Real and randomized data sets}
We will analyze two sets of data, which we now describe: The empirical data set, directly coming from the data, and a randomized data set, serving as a null model in our analysis.
The \textit{real data set} contains one spike train per hashtag, as illustrated in Fig. 4(a). The time resolution of the spikes is the same as that of the data set, that is 1 second. In situations when multiple spikes of the same hashtag take place at the same time only one event is considered. The statistics of such events are provided at the end of this subsection. In each spike train, the appearance time of the spikes is ordered from the earliest time to the latest time.
\begin{figure}[]
\includegraphics[width= 16.5cm]{Fig4_hashspiketrainComp.pdf}
\caption{Real and artificial hashtag spike trains. (a) As an illustration of different hashtag spike trains representing different types of hashtag propagation of the data set. (b) Merging hashtag spike trains from the real data. The black spikes describe that only one activity is counted if multiple activities occur at the same time. (c) Randomization procedure by randperm (Matlab). $T$ contains full hashtag activity of the data set. The randperm gives a matrix $p$, unique independent numbers out of $T$, and constructing random time series $\ldots$, $\tau^r_{i-1}$, $\tau^r_i$, $\tau^r_{i+1}$, $\ldots$ from full hashtag activity matrix $T$. (d) The resultant artificial hashtag spike train.}
\end{figure}
The \textit{random data set} is randomized version of the real data set, where each spike train of size $p$ generates a spike train of the same size with random times. In practice, we first combine all hashtag spike trains and obtain one merged hashtag spike train as illustrated in Fig. 4(b). This train carries the full history of all hashtags and, importantly, reproduces the nonstationary features of the original data in the presence of temporal correlations, burstiness, and the cyclic rhythm. As before, if two or more spikes generated in the same time, only one spike is shown in that time in the merged spike train, e.g. see the black spikes in Fig. 4(b).
Randomization is performed by permuting elements, as shown in Fig. 4(c), for instance by using randperm($T$, $p$) in Matlab. Here, $T$ represents the full matrix of times in the merged spike train and $p$ is the desired popularity, number of total spikes in a train. The permutation procedure generates $p$ times uniformly distributed unique numbers out of $T$ and these numbers define the artificial spike train, e.g. $\ldots$, $\tau^r_{i-1}$, $\tau^r_i$, $\tau^r_{i+1}$, $\ldots$, as shown in Fig. 4(c). In our data set, $p\ll T$ is always verified, as the maximum $p$ is 180,900 and the length of $T$ is 667,996. This procedure is applied to each spike train of size $p$ [Fig. 4(d)]. Generating independent, yet time-dependent events, the procedure is expected to create time-dependent Poisson random processes, $P(\Delta\tau,t)$ = $\xi(t)e^{-\xi(t)\Delta\tau}$, where the firing rate $\xi(t)$ in this case explicitly depends on the time of the day and of the week.
\textit{Statistics of multiple tweets in 1 second.} We detect multiple occurrences in 1 second for 6661 hashtags. Fig. 5 presents the probability distribution $P(c_h)$ of observing $c_h$, occurrences of an hashtag during one second, for different hashtag popularity class. Even though $c_h>1$ occurs rarely, we observe that this possibility is more probable for popular hashtags (red open circles), as expected. For the most popular hashtag, ledebat, one finds $max(c_h)$ = 40.
\begin{figure}[]
\includegraphics[width= 8cm]{Fig5_PDFch.pdf}
\caption{The probability distribution of count of hashtag activity per second $P(c_h)$. We show that, except for the top most popular hashtags listed in Table 1 with ranking 1-11 and presented here in red symbols, multiple activity in 1 second is very rare. The different colors correspond to those in Figs. 2 and 3. The legend provides the average popularity $\langle p\rangle$ in each hashtag class.}
\end{figure}
\section*{Local variation}
The time series of spike trains are inherently nonstationary, as shown in Fig. 1. For this reason, metrics defined for stationary processes are inadequate and might lead to incorrect conclusions. For instance, the non-exponential shapes of the inter-event time distribution $P(\Delta\tau)$ in Fig. 3 might originate either from correlated and collective dynamics, or from the nonstationarity of the hashtag propagation. Similarly, statistical indicators based on this distribution, such as its variance or Fano factor, might be affected in a similar way. For this reason, we consider here the so-called local variation $L_V$, originally defined to determine intrinsic temporal dynamics of neuron spike trains~\cite{Lv1, Lv2, Lv3, Lv4, Lv5}.
Unlike quantities such as $P(\Delta\tau)$, $L_V$ compares temporal variations with their local rates and is specifically defined for nonstationary processes~\cite{Lv5}
\begin{equation}
L_V=\frac{3}{N-2}\sum\limits_{i=2}^{N-1} \left(\frac{(\tau_{i+1}-\tau_i)-(\tau_{i}-\tau_{i-1})}{(\tau_{i+1}-\tau_i)+(\tau_{i}-\tau_{i-1})}\right)^2
\label{Eq:Lv_tau}
\end{equation}
Here, $N$ is the total number of spikes and $\ldots$, $\tau_{i-1}$, $\tau_i$, $\tau_{i+1}$, $\ldots$ represents successive time sequence of a single hashtag spike train. Eq.~\ref{Eq:Lv_tau} also takes the form~\cite{Lv5}
\begin{equation}
L_V=\frac{3}{N-2}\sum\limits_{i=2}^{N-1} \left(\frac{\Delta\tau_{i+1}-\Delta\tau_{i}}{\Delta\tau_{i+1}+\Delta\tau_{i}}\right)^2
\label{Eq:Lv_Deltatau}
\end{equation}
where $\Delta\tau_{i+1}$ = $\tau_{i+1}-\tau_i$ and $\Delta\tau_{i}$ = $\tau_{i}-\tau_{i-1}$. $\Delta\tau_{i+1}$ quantifies forward delay and $\Delta\tau_{i}$ represents backward waiting time for an event at $\tau_{i}$. Importantly, the denominator normalizes the quantity such as to account for local variations of the rate at which events take place. By definition, $L_V$ takes values in the interval [0:3].
The local variation $L_V$ presents properties making it an interesting candidate for the analysis of hashtag spike trains~\cite{Lv1, Lv2, Lv3, Lv4, Lv5}. In particular, $L_V$ is on average equal to 1 when the random process is either a stationary or a non-stationary Poisson process~\cite{Lv1}, with the only condition that the time scale over which the firing rate $\xi(t)$ fluctuates is slower than the typical time between spikes.
Deviations from 1 originate from local correlations in the underlying signal, either under the form of pairwise correlations between successive inter-event time intervals, e.g. $\Delta\tau_{i+1}$ and $\Delta\tau_{i}$ which tend to decrease $L_V$, or because the inter-event time distribution is non-exponential. An interesting case is given by Gamma processes~\cite{Lv1, Lv3}
\begin{equation}
P(\Delta\tau, t; \xi, \kappa) = (\xi \kappa)^\kappa\Delta\tau^{(\kappa-1)} e^{-\xi \kappa\Delta\tau}/\Gamma(\kappa)
\end{equation}
where
$\kappa$ is called a shape parameter and determines the shape of the distribution and $\Gamma$ is the Gamma function. Here, $\xi$ and $\kappa$ are the two parameters of the Gamma process. While $\xi$ determines the speed of the dynamics, $\kappa$ controls for the burstiness (irregularity) of the spike trains. Assuming that events are independently drawn, the shape factor is related to $L_V$ as follows~\cite{Lv1, Lv3}
\begin{equation}
\langle L_V\rangle = \frac{3}{2 \kappa +1}
\end{equation}
Here, the brackets describe the average taken over the given distribution~\cite{Lv1}.
When $\kappa$ = 1, an exponential is recovered, and one finds $\langle L_V\rangle$ = 1 as expected. Smaller values of $\kappa$ increase the variance in $\Delta\tau$ and therefore its burstiness, making $L_V$ larger than 1. On the other hand, larger values of $\kappa$ decrease the variance of $\Delta\tau$ and the burstiness of the process, making $\langle L_V\rangle\approx0$ smaller than 1.
We measure $L_V$ of hashtag spike trains and group the values depending on the popularity $p$ of their hashtag as was done in Figs. 2 and 3. Fig. 6 shows scatter plots of $L_V$ for the real data set (a), the empirical sequence $\ldots$, $\tau_{i-1}$, $\tau_i$, $\tau_{i+1}$, $\ldots$, and the random data set (b), the random sequence $\ldots$, $\tau^r_{i-1}$, $\tau^r_i$, $\tau^r_{i+1}$, $\ldots$, on linear-log plots. Different colors are used to distinguish the different groups and the inset legend provides the average popularity $\langle p\rangle$ in the groups.
\begin{figure*}[]
\includegraphics[width= 16cm]{Fig6_LvHashSpikeTrainsComp.pdf}
\caption{The local variation $L_V$ of hashtag spike trains versus popularity $p$ on a linear-log plot. Each color and symbol summarized in the legend present different range of $p$: Low $p$, pink and purple colors, and moderate $p$, blue, green, and yellow colors, and then high $p$, orange and red colors. In addition, the average $p$, $\langle p\rangle$, indicated in the legend ranks colors and symbols quantitatively. (a) Hashtag spike trains of the data set. (b) Artificial hashtag spike trains.}
\end{figure*}
A more readable representation is provided in Fig. 7, where we show histograms $P(L_V)$ of the values of $L_V$, for the two data sets and for the distinguished hashtag groups in $p$. The results clearly show that $L_V$ fluctuates around 1 in the random data set, as expected for a time-dependent Poisson process. On the other hand, $L_V$ systematically deviates from 1 in the original data set, where temporal correlations are present.
\begin{figure}[]
\includegraphics[width= 8cm]{Fig7_PDFLV.pdf}
\caption{Probability density function (PDF) of the local variation $L_V$ of real hashtag propagation (a) and random hashtag time sequence (b). Two distinct shapes are visible: (a) From high $p$ to low $p$, the peak position of $P(L_V)$ shifts from low values of $L_V$ to higher values of $L_V$. (b) $P(L_V)$ always peaks around 1 for the random sequences generated by artificial hashtag spike trains. The same color coding is applied as used in Fig. 6.}
\end{figure}
The observation is confirmed in Fig. 8(a), where we plot the mean $\mu(L_V)$ of $L_V$, with error bars, as a function of $\langle p\rangle$. Furthermore, $L_V$ of the original data indicates that high impact hashtags (high $p$) are associated with lower values of $L_V$ suggesting more homogeneous (regular) time distributions. These results confirm the potential use of $L_V$ as a metric to capture deviations from Poisson (temporarily uncorrelated) processes, but also to identify distinct statistical properties generated specifically in high $p$. Moreover, Fig. 8(b) presents the statistical differences between the real and the random spike trains in detail. The deviations from Poisson processes where $\mu_{0}(L_V)$ = 1 are calculated by $z$ = $\mu(L_V)-\mu_{0}(L_V)/\sigma(L_V)/\sqrt{n}$ with the standard deviations of $L_V$, $\sigma(L_V)$, and the number of the data points given in the distributions in Fig. 7, $n$. We observe that $z-$values for the random spikes are almost equal to 0, excluding in high $p$, indicating the agreement between Poisson signals and our random spike trains, which is not the case for the real trains giving $z\ncong0$ in any of $\langle p\rangle$.
To conclude, we perform an analysis to test the persistence of the temporal characteristics of hashtags, as measured by $L_V$, through time. To do so, we divide each hashtag time series into two time series. The resulting values of local variations are $L_V(t_1)$ for the first half of a spike train and $L_V(t_2)$ for the second half of the train, and then we calculate the Pearson correlation coefficient $r(L_V(t_1), L_V(t_2))$ between these values~\cite{LVcorrelation}. In Fig. 9(a) we show the linear relation between $L_V(t_1)$ and $L_V(t_2)$. Fig. 9(b) shows $r(L_V(t_1), L_V(t_2))$ as a function of the average popularity $\langle p\rangle$. Both indicate that values of $L_V$ for the same hashtag at different times is significantly and temporarily correlated. We also observe that bursty (low $p$) and regular (high $p$) signals give small $r$, while the spike trains with moderate $p$ provide the largest values of $r$ where $L_V$ suggests more uniform temporal behavior through the individual trains.
\begin{figure}[]
\includegraphics[width= 9cm]{Fig9_corrLV.pdf}
\caption{Linear correlation of $L_V$ through real hashtag spike trains. (a) The linear relation of the first and the second halves of the empirical spike trains, $L_V(t_1)$ and $L_V(t_1)$, respectively, are investigated. The legend ranks $\langle p\rangle$ in different colors and symbols. (b) The Pearson correlation coefficient $r(L_V(t_1), L_V(t_2))$ between these quantities show that while the temporal correlation through moderately popular hashtag is maximum, $r$ reaches the minimum values for both bursty (high $L_V$ and low $p$) and regular (low $L_V$ and high $p$) spike trains.}
\end{figure}
\section*{Discussion}
The main purpose of this paper is to introduce a statistical measure suitable for the analysis of nonstationary time series, as they often take place in online social media and communications in social systems. As a test case, we have focused on the dynamics of hashtags in Twitter. However, the same methodology could be also applied to the other types of correlated, bursty, and nonstationary signals, for instance the dynamics of cascades in Twitter and Facebook or phone call activity.
Instead of measuring standard statistical properties of noisy hashtag signals such as the inter-event time distribution, its variance or the Fano factor, conventionally applied to characterize the burstiness of a signal, we have focused on the local variation $L_V$, a metric capturing the fluctuations of the signal as compared to a local characteristic time. This measure, previously defined for neuron spike train analysis, nicely uncovers the regularity and the firing rate of the trains~\cite{Lv1, Lv2, Lv3, Lv4, Lv5} and so helps to identify local temporal correlations. It is important to stress that the current analysis exclusively focuses on properties of time series, and does considers neither the mechanisms leading to the observed statistical dynamic properties nor the effect of the underlying topology, e.g. through following-follower relations. An interesting line of research would study the relation between $L_V$, the underlying topology~\cite{Uncoveringtemporaldynamics} and diffusive models, for instance Hawkes process~\cite{Burstingpointprocess, Hawkespointprocesses}. In addition, both neurons~\cite{train2} and hashtags can be driven by multiple firing rates and $L_V$ analysis associated to Gamma distributions would provide more concrete results on hashtag spike trains, as done for neuron spikes~\cite{Lv3}.
We should also note that the finite temporal resolution of the data (1 sec), and the fact that multiple events per time window are neglected, tends to make $L_V$ artificially decrease for popular hashtags. In an extreme case, the time series is indeed regular, with events taking place every second. In this work, we have therefore carefully verified that fluctuations in $L_V$ are not artificially driven by these limitations. To do so, we have compared the values of $L_V$ in the empirical data with those of a null model. We observe a small decay of $L_V$ for popular hashtags in the null model (see Fig. 8), but this decay is much more limited than the one observed in the empirical data, e.g. $L_V$ = 0.89 for $\langle p\rangle$ = $10^5$ in the null model while it is equal to $L_V$ = 0.54 for the real data. In addition, a decay of $L_V$ in real hashtag data is also present in moderately popular hashtags, where multiple events per second are very rare. An interesting research direction would be to generalize the definition of local variation in order to allow for the analysis of multiple events per time window, thereby evaluating the deviations of dense time series to non-stationary Poisson processes. Finally, in a finite time window, as observed in empirical data, the statistics of high frequency hashtags is much better than that of low frequency hashtags, simply because the former occur many more times than the latter. For this reason, measurements of $L_V$ for low popularity hashtags are more subject to noise.
The empirical analysis also reveals an interesting pattern observed in the data, as more popular hashtags tend to present a more regular temporal behavior. This lack of burstiness ensures that hashtags do not disappear from the social network for very long periods of time, thereby allowing for a regular activation of the interest of Twitter users. These findings are reminiscent of recent observation in numerical simulations showing that
burstiness hinders the size of cascades~\cite{PhysRevE.89.062815}, and
should be incorporated into the modeling of theoretical information diffusion models, in particular threshold~\cite{Karimi20133476} and stochastic~\cite{Kawamoto20133470} models, on temporal networks but also into the ranking models capturing online heterogeneity in the empirical data~\cite{onlinepopularityheterogeneity}.
\section*{Supporting Information}
{\bf Supporting Information S1 Supporting data files.}\\
\noindent (Compressed zip folder of .dat files - will be publicly available online)
\begin{acknowledgments}
We thank Takaaki Aoki and Taro Takaguchi for their useful comments and Lionel Tabourier for providing data set. C. Sanl{\i} acknowledges supports from the European Union 7th Framework OptimizR Project, FNRS (le Fonds de la Recherche Scientifique, Wallonie, Belgium), and National Institute of Informatics, Tokyo, Japan. This paper presents research results of the Belgian Network DYSCO (Dynamical Systems, Control, and Optimization), funded by the Interuniversity Attraction Poles Programme, initiated by the Belgian State, Science Policy Office.
\end{acknowledgments}
|
\section{Introduction}\label{introduction}
This paper is devoted to {\em Bratteli diagrams}, the object that is widely used for constructions of transformation models in various dynamics.
It is difficult to overestimate the significance of Bratteli diagrams for the study of dynamical systems. A class of graduated infinite graphs, later called Bratteli diagrams, was originally introduced by Bratteli \cite{Br72} in his breakthrough article on the classification of approximately finite (AF) $C^*$-algebras.
It turned out that the close ideas developed by Vershik in the study of sequences of measurable partitions led to a realization of any ergodic automorphism of a standard measure as a transformation acting on a path space of a graph (afterwards called a Vershik map) \cite{vershik:1981}, \cite{vershik:1982}. The dynamical systems obtained in this way are called Bratteli-Vershik dynamical systems. During the last two decades, Bratteli diagrams turned out to be a very powerful and productive tool for the study of dynamical systems not only on a measure space but also on Cantor and Borel spaces \cite{herman_putnam_skau:1992}, \cite{BDK06}. By a Cantor dynamical system we mean a pair $(X,T)$ consisting of a Cantor set $X$ and a homeomorphism $T : X \to X$. The results proved in \cite{herman_putnam_skau:1992} build a bridge between Cantor dynamics and Bratteli diagrams. It was proved that any minimal Cantor dynamical system $(X, T)$ is realized as a Bratteli-Vershik homeomorphism defined on the path space $X_B$ of a Bratteli diagram $B$. The diagrams arising in this way have a nice property: they are {\em simple}.
Our goal is to show that a large part of results proved in the context of Cantor minimal dynamical systems remains true for a much wider class of aperiodic homeomorphisms of a Cantor set. First of all, every aperiodic homeomorphism admits its realization as a Vershik map on a {\em non-simple} Bratteli diagram \cite{medynets:2006}. Moreover, the non-simple stationary and finite rank Bratteli diagrams correspond to substitution dynamical systems and expansive homeomorphisms (the same as in minimal dynamics) (see \cite{durand_host_skau}, \cite{BKM09}, \cite{DM}). On the other hand, the problem of the existence of a continuous dynamics on a non-simple Bratteli diagram is much harder. It turns out that there are Bratteli diagrams that cannot serve as Bratteli-Vershik dynamical systems. But nevertheless one can consider the {\em tail equivalence relation} on such diagrams. This relation determines a kind of dynamical system on the path space. Our discussions of this issue are mostly based on \cite{medynets:2006}, \cite{BKY14}, \cite{BY}, \cite{JQY14}. Also there are no general results about classification of aperiodic homeomorphisms with respect to the orbit equivalence relation. In contrast to minimal case, where a number of nice theorems were proved (see \cite{giordano_putnam_skau:1995}, \cite{glasner_weiss:1995-1}, \cite{GPS99},
\cite{giordano_putnam_skau:2004}, \cite{GMPS08}, \cite{GMPS10}, \cite{HKY12}), we are aware of only some sort of negative results which show that the invariants used in minimal case do not work, in general, for aperiodic homeomorphisms \cite{S.B.O.K.} (the only exclusion is full groups \cite{Med}).
The main reason why Bratteli diagrams are convenient to use for the study of homeomorphisms $T: X\to X$ is the fact that various properties of $T$ become more transparent when one deals with corresponding Bratteli-Vershik dynamical systems. This observation is related to $T$-invariant measures and their supports, to minimal components of $T$, structure of $T$-orbit, etc.
In particular, the problem of finding all ergodic $T$-invariant measures (the extreme points of the Choquet simplex $M(X,T)$) and their supports for a given $(X, T)$ is traditionally a central one in the theory of dynamical systems, especially for specific interesting examples of homeomorphisms $T$. But being considered in general settings, this problem looks rather vague, and there are very few universal results that can be applied to a given homeomorphism $T$. But for Bratteli-Vershik realization of $T$, we are going to discuss some natural methods for the study of $M(X,T)$ based on the structure of the underlying diagram. Moreover, these methods work even for Bratteli diagrams that do not support any Vershik map. To emphasize the difference between simple and non-simple Bratteli diagrams, we remark that for an aperiodic homeomorphism $T$ the simplex $M(X,T)$ may contain infinite measures. This is impossible for minimal dynamical systems.
There are important applications of Bratteli diagrams. One of them is the theory of countable dimension groups. For every such a group $G$ there exists a Bratteli diagram whose $K^0$ group is order isomorphic to $G$. In case of simple Bratteli diagrams this group can be realized as the quotient group of $C(X, \Z)$ by $T$-boundaries. Then $T$-invariant measures are in one-to-one correspondence with traces of the dimension group. Dimension groups have a large repertoire of results and examples, and sometimes it is convenient to translate the original problems of measures to this context. For instance, we consider the notion of good measures (defined by Akin \cite{Akin2}) for stationary Bratteli diagrams. It is interesting to see how this notion can be extended to traces of dimension groups \cite{BH}. Handelman studied this notion and other amazing properties of traces in his recent works. Another application is considered in \cite{BJ} where some representations of the Cuntz-Krieger algebras are constructed by stationary Bratteli diagrams.
We should say that the list of papers devoted to different aspects of Bratteli diagrams is very long. It is impossible to cover (or even mention) most of them. We have to restrict our choice of the material for this work to several topics that are clearly mentioned in the title. Unfortunately, we were unable to discuss relations between Bratteli diagrams and operator algebras otherwise it would doubled the size of the paper. Also the notion of dimension groups and impressive results establishing links to Bratteli diagrams are not included in this survey.
There are several recent surveys that are focused on other important directions of the study of Bratteli diagrams (see, for example, \cite{durand:2010}, \cite{P10}, \cite{S00}). Our survey has minor intersections with them.
\medskip
\textbf{Notation}
\\
\begin{tabular}{ll}
$(X,T)$ & Cantor dynamical system \\
$B = (V, E)$ & Bratteli diagram \\
$B = (V, E, \omega)$ & ordered Bratteli diagram \\
$X_B$ & path space of a Bratteli diagram\\
$\varphi_\om, \varphi$ & Vershik map on ordered diagram $B$\\
$\mathcal{R}$ & tail equivalence relation\\
$Orb_T(x)$ & $T$-orbit of $x$\\
\end{tabular}
\section{Fundamentals of Cantor dynamics and Bratteli diagrams}\label{Preliminaries}
This section contains the basic definitions and facts about Cantor dynamical systems and Bratteli diagrams. They are the main objects of our consideration in this paper.
\subsection{Cantor dynamics} By definition, a {\em Cantor set (space)} $X$ is a zero-dimensional compact metric space without isolated points. The topology on $X$ is generated by a countable family of clopen subsets. It is called the clopen topology. All such Cantor sets are homeomorphic.
A {\em homeomorphism} $T : X \to X$ is a continuous bijection. Denote
$Orb_T(x) := \{T^n (x) \; |\; n \in \mathbb{Z}\}$; the set $Orb_T(x)$ is called the {\em orbit} of $x\in X$ under action of $T$ (or simply $T$-orbit). We consider here only {\em aperiodic} homeomorphisms $T$, i.e., for every $x$ the set $Orb_T(x)$ is countably infinite. In fact, some definitions and facts make sense for arbitrary homeomorphisms. We focus on the case of aperiodic homeomorphisms for convenience mostly.
A homeomorphism $T : X \to X$ is called {\em minimal} if for every $x \in X$ the set $Orb_T(x)$ is dense. A {\em minimal Cantor system} is a pair $(X,T)$ where $X$ is a Cantor space, and $T \colon X \rightarrow X$ is a minimal homeomorphism. Any (aperiodic) homeomorphism $T$ of a Cantor set has a {\em minimal component} $Y$: this is a $T$-invariant closed non-empty subset $Y$ of $X$ such that $T|_Y$ is minimal on $Y$.
Given a minimal Cantor system $(X,T)$ and a clopen $A\subset X$, the first return function $r_A(x) = \min\{n \geq 1 : T^n(x) \in A\}$ is a well defined continuous integer-valued function with domain $A$. Then $T_A(x) = T^{r_A}(x)$ is a homeomorphism of $ A$, and $(A, T_A)$ is called the \textit{induced} Cantor minimal system.
There are several notions of equivalence for Cantor dynamical systems. We give these definitions for single homeomorphisms of Cantor sets.
\begin{definition}
Let $(X,T )$ and $(Y, S)$ be two aperiodic Cantor systems. Then
(1) $(X,T)$ and $(Y,S)$ are \textit{conjugate} (or \textit{isomorphic}) if there exists a homeomorphism $h \colon X \rightarrow Y$ such that $h \circ T = S \circ h$.
(2) $(X,T)$ and $(Y,S)$ are \textit{orbit equivalent} if there exists a homeomorphism $h \colon X \rightarrow Y$ such that $h(Orb_T(x)) = Orb_S(h(x))$ for every $x \in X$. In other words, there exist functions $n, m \colon X \rightarrow \mathbb{Z}$ such that for all $x \in X$, $h \circ T(x) = S^{n(x)} \circ h(x)$ and $h \circ T^{m(x)} = S \circ h(x)$. The functions $n,m$ are called the orbit cocycles associated to $h$.
(3) $(X,T)$ and $(Y,S)$ are \textit{strong orbit equivalent} if they are orbit equivalent and each of the corresponding orbit cocycles $n, m$ has at most one point of discontinuity.
(4) $(X,T)$ and $(Y,S)$ are \textit{Kakutani equivalent} if they both have clopen subsets such that the corresponding induced systems are conjugate.
(5) $(X,T)$ and $(Y,S)$ are \textit{Kakutani orbit equivalent} if they both have clopen subsets such that the corresponding induced systems are orbit equivalent.
\end{definition}
Given a Cantor dynamical system $(X,T)$, a Borel measure $\mu$ on $X$ is called $T$-\textit{invariant} if $\mu(TA) = \mu(A)$ for any Borel set $A$. Let $M(X,T)$ be the set of all invariant measures. It is well known that $M(X,T)$ is a Choquet simplex whose extreme points are $T$-ergodic measures. This simplex includes probability measures (when $\mu(X) =1$) and infinite measures (when $\mu(X) = \infty$). We observe that infinite invariant measures may exist for aperiodic homeomorphisms; in minimal dynamics this is impossible. If $M(X,T) =\{\mu\}$, then $T$ is called {\em uniquely ergodic}.
It is not hard to see that every Borel measure $\mu$ on a Cantor set is completely determined by its values on clopen sets. This means that there is a one-to-one correspondence between $\mu$ and the collection of numbers $S(\mu) = \{\mu(A) : A\ \mbox{clopen}\} \subset [0,1]$.
It follows from \cite{herman_putnam_skau:1992}, \cite{giordano_putnam_skau:1995}, and \cite{medynets:2006} that any minimal (and even aperiodic) Cantor dynamical system $(X,T)$ admits a realization as a Bratteli-Vershik dynamical system $(X_B, \varphi_B)$ acting on a path space $X_B$ of a Bratteli diagram (see Section~\ref{BV}). Thus, the study of $T$-invariant measures is reduced to the case of measures defined on the path space of a Bratteli diagram. The advantage of this approach is based on the facts that (i) any such a measure is completely determined by its values on cylinder sets of $X_B$, and (ii) there are simple and explicit formulas for measures of cylinder sets. Especially transparent this method works for stationary and finite rank Bratteli diagrams, simple and non-simple ones \cite{BKMS_2010}, \cite{BKMS_2013}.
It is worth pointing out that the study of measures on a Bratteli diagram is a more general problem than that in Cantor dynamics. This observation follows from the existence of Bratteli diagrams that do not support any continuous dynamics on their path spaces which is compatible with the tail equivalence relation. The first example of such a Bratteli diagram was given in \cite{medynets:2006}; a more comprehensive coverage of this subject can be found in \cite{BKY14} and \cite{BY} (we discuss this stuff below in Section~\ref{VM}). If a Bratteli diagram does not admit a Bratteli-Vershik homeomorphism, then we have to work with the {\em tail equivalence relation} $\mathcal R$ on $X_B$ and study measures invariant with respect to $\mathcal R$.
\subsection{Bratteli diagrams}
A {\it Bratteli diagram} is an infinite graph $B=(V,E)$ such that the vertex
set $V =\bigcup_{i\geq 0}V_i$ and the edge set $E=\bigcup_{i\geq 0}E_i$
are partitioned into disjoint subsets $V_i$ and $E_i$ where
(i) $V_0=\{v_0\}$ is a single point;
(ii) $V_i$ and $E_i$ are finite sets, $\forall i \geq 0$;
(iii) there exist $r : V \to E$ (range map $r$) and $s : V \to E$ (source map $s$), both from $E$ to $V$, such that $r(E_i)= V_{i+1}$, $s(E_i)= V_{i}$, and
$s^{-1}(v)\neq\emptyset$, $r^{-1}(v')\neq\emptyset$ for all $v\in V$
and $v'\in V\setminus V_0$.
The set of vertices $V_i$ is called the $i$-th level of the
diagram $B$. A finite or infinite sequence of edges $(e_i : e_i\in E_i)$
such that $r(e_{i})=s(e_{i+1})$ is called a {\it finite} or {\it infinite
path}, respectively. For $m<n$, $v\, \in V_{m}$ and $w\,\in V_{n}$, let $E(v,w)$ denote the set of all paths $\overline{e} = (e_{1},\ldots, e_{p})$
with $s(\ol e) = s(e_{1})=v$ and $r(\ol e) = r(e_{p})=w$. If $m < n$ let $E(m,n)$ denote all paths whose source belongs to $V_m$ and whose range belongs to $V_n$.
For a Bratteli diagram $B$,
let $X_B$ be the set of infinite paths starting at the top vertex $v_0$.
We endow $X_B$ with the topology generated by cylinder sets
$[\ol e]$ where $\ol e= (e_0, ... , e_n)$, $n \in \mathbb N$, and
$[\ol e]:=\{x\in X_B : x_i=e_i,\; i = 0, \ldots, n\}$.
With this topology, $X_B$ is a 0-dimensional compact metric space.
By assumption, we will consider only such Bratteli diagrams $B$ for which $X_B$ is a {\em Cantor set}, that is $X_B$ has no isolated points. Letting $|A|$ denote the cardinality of the set $A$, this means that for every $(x_0,x_1,\ldots)\in X_B$
and every $n\geq 1$ there exists $m>n$ such that $|s^{-1}(r(x_m))|>1$.
Given a Bratteli diagram $B$, the $n$-th {\em incidence matrix}
$F_{n}=(f^{(n)}_{v,w}),\ n\geq 0,$ is a $|V_{n+1}|\times |V_n|$
matrix such that $f^{(n)}_{v,w} = |\{e\in E_{n+1} : r(e) = v, s(e) = w\}|$
for $v\in V_{n+1}$ and $w\in V_{n}$. Every vertex $v \in V$ is connected with $v_0$ by a finite path, and the set of $E(v_0, v) $ of all such paths is finite. If $h^{(n)}_v = |E(v_0, v)|$, then for all $n \geq 1$
\begin{equation}
h_v^{(n+1)}=\sum_{w\in V_{n}}f_{v,w}^{(n)}h^{(n)}_w \ \mbox{or} \ \
h^{(n+1)}=F_{n}h^{(n)}
\end{equation}
where $h^{(n)}=(h_w^{(n)})_{w\in V_n}$. The numbers $h_w^{(n)}$ are usually called heights (see Section~\ref{BV}).
We define the following important classes of Bratteli diagrams that we work with in this article.
\begin{definition}\label{rank_d_definition} Let $B$ be a Bratteli diagram.
\begin{enumerate}
\item
We say that $B$ has \textit{finite rank} if for some $k$, $|V_n| \leq k$ for all $n\geq 1$.
\item
Let $B$ have finite rank. We say that $B$ has \textit{rank $d$} if $d$ is the smallest integer such that $|V_n|=d$ infinitely often.
\item
We say that $B$ is {\em simple} if for any level
$n$ there is $m>n$ such that $E(v,w) \neq \emptyset$ for all $v\in
V_n$ and $w\in V_m$. Otherwise, $B$ is called {\em non-simple}.
\item
We say that $B$ is \textit{stationary} if $F_n = F_1$ for all $n\geq 2$.
\end{enumerate}
\end{definition}
Let $x =(x_n)$ and $y =(y_n)$ be two paths from $X_B$. It is said that $x$ and $y$ are {\em tail equivalent} (in symbols, $(x,y) \in \mathcal R$) if there exists some $n$ such that $x_i = y_i$ for all $i\geq n$. Since $X_B$ has no isolated points, the $\mathcal R$-orbit of any point $x\in X_B$ is infinitely countable. The diagrams with infinite $\mathcal R$-orbits are called {\em aperiodic}. Note that a Bratteli diagram is simple if the tail equivalence relation $\mathcal R$ is minimal.
In order to illustrate the above definitions, we give an example of a Bratteli diagram.
\unitlength = 0.5cm
\begin{center}
\begin{graph}(17,15)
\graphnodesize{0.4}
\roundnode{V0}(8.5,14)
\roundnode{V11}(0.0,9) \roundnode{V12}(4.25,9)
\roundnode{V13}(8.5,9) \roundnode{V14}(12.75,9)
\roundnode{V15}(17.0,9)
\roundnode{V21}(0.0,5) \roundnode{V22}(4.25,5)
\roundnode{V23}(8.5,5) \roundnode{V24}(12.75,5)
\roundnode{V25}(17.0,5)
\roundnode{V31}(0.0,1) \roundnode{V32}(4.25,1)
\roundnode{V33}(8.5,1) \roundnode{V34}(12.75,1)
\roundnode{V35}(17.0,1)
\edge{V11}{V0} \edge{V12}{V0} \edge{V13}{V0} \edge{V14}{V0}
\edge{V15}{V0}
\edge{V21}{V11}\bow{V21}{V11}{0.06} \edge{V21}{V12} \edge{V22}{V11} \edge{V22}{V12}
\edge{V23}{V13} \edge{V23}{V14} \edge{V24}{V13} \edge{V24}{V14}
\edge{V25}{V11}\bow{V25}{V11}{0.06} \edge{V25}{V13} \bow{V25}{V15}{0.06}
\bow{V25}{V15}{-0.06} \edge{V25}{V15}
\edge{V31}{V21} \edge{V31}{V22} \bow{V31}{V22}{0.06}\edge{V32}{V21} \edge{V32}{V22}
\edge{V33}{V23}\bow{V33}{V23}{0.06} \edge{V33}{V24} \edge{V34}{V23} \edge{V34}{V24}
\edge{V35}{V21} \edge{V35}{V23} \bow{V35}{V25}{0.06}
\bow{V35}{V25}{-0.06} \edge{V35}{V25}
\end{graph}
\centerline{.\ .\ .\ .\ .\ .\ .\ .\ .\ .\ .\ .\ .\ .\ .\ .\ }
\\
Fig. 1. Example of a Bratteli diagram
\end{center}
\medskip
This diagram is a non-simple finite rank Bratteli diagram that has exactly two minimal components (they are clearly seen).
We will constantly use the \textit{telescoping} procedure for a Bratteli diagram:
\begin{definition} \label{telescoping_definition}
Let $B$ be a Bratteli diagram, and $n_0 = 0 <n_1<n_2 < \ldots$ be a strictly increasing sequence of integers. The {\em telescoping of $B$ to $(n_k)$} is the Bratteli diagram $B'$, whose $k$-level vertex set $V_k'$ is $V_{n_k}$ and whose incidence matrices $(F_k')$ are defined by
\[F_k'= F_{n_{k+1}-1} \circ \ldots \circ F_{n_k},\]
where $(F_n)$ are the incidence matrices for $B$.
\end{definition}
Roughly speaking, in order to telescope a Bratteli diagram, one takes a subsequence of levels $\{n_k\}$ and considers the set $E(n_k, n_{k+1})$ of all finite paths between the levels $\{n_k\}$ and $\{n_{k+1}\}$ as edges of the new diagram.
In particular, a Bratteli diagram $B$ has rank $d$ if and only if there is
a telescoping $B'$ of $B$ such that $B'$ has exactly $d$ vertices
at each level. When telescoping diagrams, we often do not specify to which levels $(n_k)$ we telescope, because it suffices to know that such a sequence of levels exists.
Two Bratteli diagrams are {\em isomorphic} if they are isomorphic as graded graphs.
Define an equivalence relation $\sim$ on the set of Bratteli diagrams generated by isomorphism and telescoping. One can show that $B_1 \sim B_2$ if there exists a Bratteli diagram $B$ such that telescoping of $B$ to odd levels yields, say, $B_1$ and telescoping to even levels yields $B_2$.
In order to avoid consideration of some trivial cases, we will assume that the following \textit{convention} always holds: {\em our Bratteli diagrams are not unions of two or more disjoint subdiagrams.}
\subsection{Ordered Bratteli diagrams and Vershik maps} The concept of an ordered Bratteli diagram is crucial for the existence of dynamics on the path space of a Bratteli diagram.
\begin{definition}\label{order_definition} A Bratteli diagram $B=(V,E) $ is called {\it ordered} if a linear order `$>$' is defined on every set $r^{-1}(v)$, $v\in
\bigcup_{n\ge 1} V_n$. We use $\om$ to denote the corresponding partial
order on $E$ and write $(B,\om)$ when we consider $B$ with the ordering $\om$. Denote by $\mathcal O_{B}$ the set of all orderings on $B$.
\end{definition}
Every $\omega \in \mathcal O_{B}$ defines the \textit{lexicographic}
ordering on the set $E(k,l)$ of finite paths between
vertices of levels $V_k$ and $V_l$: $(e_{k+1},...,e_l) > (f_{k+1},...,f_l)$
if and only if there is $i$ with $k+1\le i\le l$, such that $e_j=f_j$ for $i<j\le l$
and $e_i> f_i$.
It follows that, given $\om \in \mathcal O_{B}$, any two paths from $E(v_0, v)$
are comparable with respect to the lexicographic ordering generated by $\om$. If two infinite paths are tail equivalent, and agree from the vertex $v$ onwards, then we can compare them by comparing their initial segments in $E(v_0,v)$. Thus $\om$ defines a partial order on $X_B$, where two infinite paths are comparable if and only if they are tail equivalent.
\begin{definition}
We call a finite or infinite path $e=(e_i)$ \textit{ maximal (minimal)} if every
$e_i$ is maximal (minimal) amongst the edges
from $r^{-1}(r(e_i))$.
\end{definition}
Notice that, for $v\in V_i,\ i\ge 1$, the
minimal and maximal (finite) paths in $E(v_0,v)$ are unique. Denote
by $X_{\max}(\om)$ and $X_{\min}(\om)$ the sets of all maximal and
minimal infinite paths in $X_B$, respectively. It is not hard to see that
$X_{\max}(\om)$ and $X_{\min}(\om)$ are \textit{non-empty closed subsets} of
$X_B$; in general, $X_{\max}(\om)$ and $X_{\min}(\om)$ may have
interior points. For a finite rank Bratteli diagram $B$, the sets $X_{\max}(\om)$
and $X_{\min}(\om)$ are always finite for any $\om$, and if $B$ has rank $d$,
then each of them have at most $d$ elements
(\cite{BKM09}). For an aperiodic Bratteli diagram $B$, we see that $X_{\max}(\om)\cap X_{\min}(\om) =\emptyset$.
We say that an ordered Bratteli diagram $(B, \om)$ is \textit{properly ordered} if the sets $X_{\max}(\om)$ and $X_{\min}(\om)$ are singletons.
If we denote by $\mathcal O_B(j)$ the set of all orders on $B$ which have $j$ maximal and $j$ minimal paths, then, in this notation, $\mathcal O_B(1)$ is the set of proper orders on $B$.
Let $(B,\omega)$ be an ordered Bratteli diagram, and suppose that $B'=(V',E')$ is the telescoping of $B$ to levels $(n_k)$. Let $v' \in V'$ and suppose that the two edges $e_1',$ $e_2'$, both with range $v'$, correspond to the finite paths $e_1$, $e_2$ in $B$, both with range $v$. Define the order $\om'$ on $B'$ by $e_1'<e_2'$ if and only if $e_1<e_2$. Then $\om' $ is called the {\em lexicographic order generated by $\om$} and is denoted by $\om'=L(\om)$.
It is not hard to see that if $\om' = L(\om)$, then
$$
|X_{\max}(\om)| = |X_{\max}(\om')|,\ \ |X_{\min}(\om)| = |X_{\min}(\om')|.
$$
A Bratteli diagram $B $ is called \textit{ regular} if for any ordering
$\omega \in \mathcal O_{B}$ the sets $X_{\max}(\omega)$ and
$X_{\min}(\omega)$ have empty interior.
In particular, finite rank Bratteli diagrams are automatically regular, and if all incidence matrix entries of $B$ are at least 2, then $B$ is regular. We consider here only regular Bratteli diagrams.
We will need the notion of the language associated to an ordered Bratteli diagram. If $V$ is a finite alphabet, let $V^{+}$ denote the set of nonempty
words over $V.$ We use the notation $W' \subseteq W$ to indicate that $W'$ is a subword of $W$.
If $W_1, W_2, \ldots ,W_n$, are words, then we let $\prod_{i=1}^{n}W_i$ refer to their concatenation.
Let $\om$ be an order on a Bratteli diagram $B$. Fix a vertex $v\in
V_n$ and some level $m<n$, consider the set $E(V_m,v) = \bigcup_{v'\in V_{m}}E(v',v)$ of all finite paths between vertices of level $m$ and $v$. This set can be ordered by $\om$: $E(V_m, v) = \{e_{1},\ldots e_{p}\}$ where $e_{i}<e_{i+1}$ for
$1\leq i\leq p-1$. Define the word $w(v,m,n):=s(e_{1})s(e_{2})\ldots
s(e_{p})$ over the alphabet $V_{m}$. If $W=v_{1}\ldots v_{r} \in
V_{n}^{+}$, let $w(W,n-1, n):= \prod_{i=1}^{r}w(v_{i},n-1,n).$
\begin{definition}\label{language_definition}
The {\em level-$n$ language $\mathcal
L(B,\om, n)$ of $(B,\om)$ } is
$$\mathcal L(B,\om,n):=
\{W : \,\, W\subseteq w(v,n,N), \,\, \mbox{ for
some } v\in V_{N}, N>n\}\, .
$$
\end{definition}
If $B$ has strict rank $d$, then each of the level-$n$ languages can be defined on a common alphabet $V$, and in this case we have the {\em language} of $B$
$$ \mathcal L(B, \om):=\limsup_{n}\mathcal L(B,\om, n)\, .
$$
The idea to use an order on a Bratteli diagram to define a transformation acting
on the path space $X_B$ was firstly developed by Vershik \cite{vershik:1981}, and then it was applied in many papers (see, e.g. \cite{durand:2010}, \cite{giordano_putnam_skau:1995},
\cite{herman_putnam_skau:1992})
\begin{definition}\label{VershikMap}
Let $(B, \omega)$ be an ordered, regular Bratteli
diagram. We say that $\varphi = \varphi_\omega : X_B\rightarrow X_B$
is a {\it (continuous) Vershik map} if it satisfies the following conditions:
(i) $\varphi$ is a homeomorphism of the Cantor set $X_B$;
(ii) $\varphi(X_{\max}(\omega))=X_{\min}(\omega)$;
(iii) if an infinite path $x=(x_0,x_1,\ldots)$ is not in $X_{\max}(\omega)$,
then $\varphi(x_0,x_1,\ldots)=(x_0^0,\ldots,x_{k-1}^0,\overline
{x_k},x_{k+1},x_{k+2},\ldots)$, where $k=\min\{n\geq 1 : x_n\mbox{
is not maximal}\}$, $\overline{x_k}$ is the successor of $x_k$ in
$r^{-1}(r(x_k))$, and $(x_0^0,\ldots,x_{k-1}^0)$ is the minimal path
in $E(v_0,s(\overline{x_k}))$.
\end{definition}
If $\om$ is an ordering on $B$, then one can always define the map
$\varphi_0$ that maps $X_B \setminus X_{\max}(\om)$ onto $X_B
\setminus X_{\min}(\om)$ according to (iii) of Definition
\ref{VershikMap}. The question about the existence of the Vershik map is
equivalent to that of an extension of $\varphi_0 : X_B \setminus
X_{\max}(\om) \to X_B \setminus X_{\min}(\om)$ to a homeomorphism
of the entire set $X_B$. If $\om$ is a proper ordering, then
$\varphi_\om$ is a homeomorphism. In particular any simple Bratteli diagram has a Vershik map. For a finite rank Bratteli diagram $B$, the
situation is simpler than for a general Bratteli diagram because the sets
$X_{\max}(\om)$ and $X_{\min}(\om)$ are finite.
\begin{definition}\label{Good_and_bad} Let $B$ be a Bratteli diagram
$B$. We say that an ordering $\om\in \mathcal O_{B}$ is \textit{perfect}
if $\om$ admits a Vershik map $\varphi_{\om}$ on $X_B$.
Denote by $\mathcal P_B $ the set of
all perfect orderings on $B$.
\end{definition}
We observe that for a regular Bratteli diagram with an ordering $\om$, the Vershik map $\varphi_\om$, if it exists, is
defined in a unique way. Also, a necessary condition for $\om\in
\mathcal P_{B}$ is that $|X_{\max}(\om)|=|X_{\min}(\om)|$. Given $(B,\om)$ with $\om \in \mathcal P_B$, the uniquely defined system $(X_B, \varphi_\om)$ is called a {\em Bratteli-Vershik} or {\em adic} system.
Thus, we can summarize the above definitions and results in the following statement.
\begin{theorem} Let $B= (V,E, \om)$ be an ordered Bratteli diagram with perfect order
$\om \in \mathcal P_B$. Then there exists an aperiodic homeomorphism (Vershik map) $\varphi_{\om}$ acting on the path space $X_B$ according to Definition \ref{VershikMap}. The homeomorphism $\varphi_\om$ is minimal if and only if $B$ is simple.
\end{theorem}
The pair $(X_B, \varphi_\om)$ is called the {\em Bratteli-Vershik dynamical system}.
The simplest example of a Bratteli diagram is an {\em odometer}. Any odometer can be realized as a Bratteli diagram $B$ with $|V_n| =1$ for all $n$. Then any order on $B$ is proper and defines the Vershik map.
It is worth noticing that a general Bratteli diagram may have a rather complicated structure. In particular, the tail equivalence relation may have uncountably many minimal components or, in other words, uncountably many simple subdiagrams that do not have connecting edges.
\section{Bratteli-Vershik representations of Cantor dynamical systems
and orbit equivalence}\label{BV}
\subsection{Bratteli-Vershik model of a Cantor dynamical system}
Let $(X,T)$ be a Cantor aperiodic dynamical system. Is it possible to represent $(X,T)$ as a Bratteli-Vershik system? In other words, we want to associate a refining sequence of clopen partitions to the homeomorphism $T$ whose elements generate the clopen topology on $X$. If $T$ is minimal, the answer is well known. Take any clopen subset $A$ of $X$ and consider the forward $T$-orbit that starts at a point $x \in A$. By minimality of $T$, the orbit returns to $A$ in a finitely many steps. Since the function of the first return $r_A(x)$ has finite values and $A$ is compact, we get a finite partition of $X$ into clopen $T$-towers $\xi_i = (A_i, TA_i, ... , T^{i-1}A)$ with base $A = \bigcup_i A_i$ where $A_i = \{x\in A : r_A(x) = i\}$. For aperiodic $T$, this partition can be obtained only when $A$ has some additional properties.
We say that $\xi$ is a {\em Kakutani-Rokhlin partition} of a Cantor set $X$ if $\xi$ is a finite union of disjoint $T$-towers $\xi_i = (C_i, ..., T^{h_i -1}C_i)$. Then $h_i$ is called the height of $\xi$ and $C_i$ is the base of $\xi_i$.
One says that a partition $\xi_1$ refines a partition $\xi$ if every element (atom) of $\xi$ is a union of elements of $\xi_1$.
A closed subset $Y$ of $X$ is called {\em basic} if (1) $Y \cap T^iY = \emptyset, i\neq 0$, and (2) every clopen neighborhood $A$ of $Y$ is a complete $T$-section, i.e., $A$ meets every $T$-orbit at least once. This means that every point from $A$ is recurrent. It is clear that if $T$ is minimal then every point of $X$ is a basic set.
The following result was well known for a minimal homeomorphism $T$ after the paper by Putnam \cite{putnam:1989}. The case of aperiodic Cantor system is much subtler and was considered in \cite{BDM05} and \cite{medynets:2006}. Based on the results of \cite{BDM05} Medynets proved in \cite{medynets:2006} that {\em every Cantor aperiodic system $(X, T)$ has a basic set}. Then one can prove
\begin{theorem} [\cite{medynets:2006}] \label{K-R part} Let $(X,T)$ be an aperiodic Cantor system, and let $Y \subset X$ be a basic set for $T$. Then there exists a sequence of Kakutani-Rokhlin clopen partitions $\xi(n)$ such that for all $n \in \N$:
(i) $\xi(n+1)$ refines $\xi(n)$ and atoms of these partitions generate the clopen topology on $X$;
(ii) $B(\xi(n)) \subset B(\xi(n+1)) $ and $Y = \bigcap_n B(\xi(n))$;
(iii) $h_n \to \infty$ (as $n\to \infty$) where $h_n$ is the minimal height of the towers that form $\xi(n)$.
\end{theorem}
The ideas developed in the papers by Vershik \cite{vershik:1981}, \cite{vershik:1982}, where sequences of refining measurable partitions of a measure space were used to construct a realization of an ergodic automorphisms of a measure space, turned out to be very fruitful for finding a model of any minimal homeomorphism $T$ of a Cantor set $X$. In \cite{herman_putnam_skau:1992}, Herman, Putnam, and Skau found an explicit construction that allows one to define an ordered simple Bratteli diagram $B = (V, E, \om)$ such that $T$ is conjugate to the corresponding Vershik map $\varphi_\omega$. Since this construction is described in many papers (not only in \cite{herman_putnam_skau:1992}), we will not give the details here referring to the original paper \cite{herman_putnam_skau:1992} (and \cite{durand:2010}) for detailed explanation. We discuss here the main idea of constructing such a diagram $B$ in a non-rigorous way.
Let $T$ be an aperiodic (minimal) homeomorphism of a Cantor set $X$. First, find a sequence of Kakutani-Rokhlin partitions $(\xi(n))$ satisfying Theorem \ref{K-R part}: $\xi(0) = X$,
\begin{equation*}\xi(n)=\{T^iA_v(n)\; :\; v=1,\ldots, m_v(n); \ \ i=0,\ldots,h_v(n)-1\},\;n\geq 1,
\end{equation*}
that generates the clopen topology on $X$. Simultaneously, we will define an ordered Bratteli diagram $B = (V,E, \om)$ as follows.
(i) Let
$$
\xi_v(n)=\{A_v(n),\ldots,T^{h_v(n)-1}A_v(n)\} \mbox{ for } v=1,\ldots, m_v(n).
$$
The vertex set $V = \coprod_{n \geq 0} V_n$ where $V_0$ is a singleton, and $V_n = \{1, ... , m_v(n)\}$.
(ii) Define the set of edges $E_n$ between the consecutive
levels $V_{n}$ and $V_{n+1}$ by the incidence matrix
$F_n=\{m_{vw}(n) : v\in V_{n+1},\;w\in V_{n}\}$, where
$$
f_{vw}^{(n)}= |\{0\leq i<h_v(n) : T^iA_v(n+1)\subset A_w(n)\}|.
$$
In other words, we fix a vertex $v\in V_{n+1}$ and define $V(v,n)$
as the set of all vertices from $V_{n}$ such that $\xi_v(n +1)$
intersects $\xi_w(n)$, and a vertex $w$ appears in $V(v,n)$ as
many times as $\xi_v(n+1)$ intersects $\xi_w(n)$. Then we connect
$v$ to each vertex $w\in V(v,n)$ taking into account the
multiplicity of appearance of $w$ in $V(v,n)$.
(iii) To define the ordering $\om$ on $E$ we take the clopen set
$A_v(n+1)$. Then tracing the orbit of $A_v(n+1)$ within the $T$-tower
$\xi_v(n+1)$, we see that $T^iA_v(n+1)$ consecutively meets the sets
$A_{w_1}(n)$, ..., $A_{w_{k_v}}(n)$ (some of them can
occur several times). This defines the set of edges $r^{-1}(v)$.
Enumerate the edges from $r^{-1}(v)$ as follows:
$e(w_1,v) < e(w_2,v) <\ldots<e(w_{k_v},v)$.
Since the partitions $\{\xi_n\}$ generate the topology of
$X$, for each point $x\in X$ there is a unique sequence
$i(x)=\{(v_n,i_n)\}_{v\in V_n;\,0\leq i_n<h_{v_n}(n)}$ such that
$$
\{x\}=\bigcap_{n=1}^\infty T^{i_n}A_{v_n}(n).
$$
Define the map $\theta : X\rightarrow X_B$ by
$$\theta(x)=\bigcap_{n\geq 1}U(y_1,\ldots,y_n)
$$ where
$(y_1,\ldots,y_n)$ is the $i_n$-th finite path in $E(v_0,v_n)$
with respect to the lexicographical ordering on $E(v_0,v_n)$. It can be easily checked that $\theta$ is a homeomorphism.
Define $\varphi_\om=\theta \circ T\circ\theta^{-1}$. Then $\varphi_\om$ is the
Vershik map defined by order $\om$. Thus, we
obtain that $(X,T)$ is conjugate to the Bratteli-Vershik system
$(X_B,\varphi_\om)$.
This gives the theorem proved in \cite{medynets:2006} for an aperiodic homeomorphism (the case of a minimal homeomorphism was considered in \cite{herman_putnam_skau:1992}).
\begin{theorem} \label{existenceBD}
Let $(X, T)$ be a Cantor aperiodic system with a basic set $Y$. There exists an ordered Bratteli diagram $B = (V,E,\om)$ such that $(X, T)$ is conjugate to a Bratteli-Vershik dynamical system $(X_B, \varphi_\om)$. The homeomorphism implementing the conjugacy between $T$ and $\varphi_\om$ maps the basic set $Y$ onto the set $X_{\min}(\om)$ of all minimal paths of $X_B$. The equivalence class of $B$ does not depend on a choice of $\{\xi(n)\}$ with the property $\bigcap_n B(\xi(n)) =Y$.
\end{theorem}
Thus, for every aperiodic homeomorphism $T$ of a Cantor set $X$, there exists an ordered Bratteli diagram $(B, \om)$ such that $T$ is conjugate to the Vershik map $\varphi_\om$. Is the converse true? In case of a simple Bratteli diagram, the answer is obviously affirmative: there exists a proper order $\om$ on any simple Bratteli diagram $B$ so that $(X_B, \varphi_\om)$ is a minimal Cantor system.
The main difficulty for the study of non-simple Bratteli diagrams is illustrated by the following amazing observation made by Medynets in \cite{medynets:2006}.
He proved that there exists an (even stationary) non-simple Bratteli diagram such the {\em Vershik map $\varphi_\om$ is not continuous for any order $\om$.} Here is an example of such a diagram:
\begin{center}
\unitlength=0.6cm
\begin{graph}(5,7)
\graphnodesize{0.2}
\roundnode{V0}(3,6
\roundnode{V11}(1,5)
\roundnode{V12}(3,5)
\roundnode{V13}(5,5)
\roundnode{V21}(1,3)
\roundnode{V22}(3,3)
\roundnode{V23}(5,3)
\roundnode{V31}(1,1)
\roundnode{V32}(3,1)
\roundnode{V33}(5,1)
\graphlinewidth{0.025}
\edge{V0}{V11
\edge{V0}{V12}
\edge{V0}{V13}
\bow{V21}{V11}{0.06}
\bow{V21}{V11}{-0.06}
\edge{V22}{V11}
\bow{V22}{V12}{0.06
\bow{V22}{V12}{-0.06}
\edge{V22}{V13}
\bow{V23}{V13}{0.06}
\bow{V23}{V13}{-0.06}
\bow{V31}{V21}{0.06}
\bow{V31}{V21}{-0.06}
\edge{V32}{V21}
\bow{V32}{V22}{0.06
\bow{V32}{V22}{-0.06}
\edge{V32}{V23}
\bow{V33}{V23}{0.06}
\bow{V33}{V23}{-0.06}
\freetext(3,0.1){.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.\,.}
\end{graph}
\end{center}
\subsection{Orbit equivalence and full groups}
The circle of problems related to the classification of transformations with respect to orbit equivalence was originated in the famous papers by Dye \cite{Dye1}, \cite{Dye2}. It was proved that any two ergodic finite measure preserving automorphisms of a standard measure space are orbit equivalent. Also he defined a new invariant of orbit equivalence, the so called {\em full group} of automorphisms. We give its definition for homeomorphisms of a Cantor set, but it is obvious that this definition can be reformulated for dynamical systems acting on other underlying spaces as a Borel space or standard measure space.
\begin{definition}\label{Full GR def}
Let $\Gamma$ be a group of homeomorphisms of a Cantor set $X$. Then the full group $[G]$ (generated by $\Gamma$) is formed by homeomorphisms of $X$ that keep the partition of $X$ into $\Gamma$-orbits fixed. That is
$$
[\Gamma] := \{R \in Homeo (X) : Rx \in Orb_{\Gamma}(x), \forall x \in X\}.
$$
\end{definition}
The most interesting is the case when $\Gamma$ is an action of an amenable countable group $G$. If $\Gamma$ is generated by a single homeomorphism $T$ we write $[T]$ for the corresponding full group. It is well known that in the case of measurable dynamics any ergodic finite measure preserving $G$-action is orbit equivalent to an action of $\Z$ \cite{OW80}, \cite{CFW81}. One of the central problems of Cantor dynamics consists of an extension of this result to the group of homeomorphisms of a Cantor set.
In 1995, Giordano, Putnam, and Skau \cite{giordano_putnam_skau:1995} found complete invariants of orbit equivalence for minimal homeomorphisms. We need to give more definitions to formulate their results. Giving a Cantor minimal system $(X, T)$, denote by $C(X, \Z)$ the countable group of integer-valued continuous functions, and by $B(X,\mathbb Z)$ the subgroup of $C(X,\Z)$ generated by $T$-coboundaries ($f \in B(X,\mathbb Z)$ if $f = g - g\circ T$ for some $g\in C(X,\mathbb Z)$).
Then
$$
K^0(X, T) = C(X, \mathbb Z)/B(X,\mathbb Z)
$$
is considered as an ordered abelian group whose cone of positive elements $K^0(X, T)^+$ is determined by $\{ \widehat f \in K^0(X, T) : f \geq 0, f \in C(X, \mathbb Z)\}$ with distinguished order unit $\widehat 1$. The group $C(X, \Z)$ contains also the subgroup of all infinitesimal functions with respect to $T$: by definition, $f \in Inf(X, T)$ if
$$
\int_X f d\mu = 0 \ \mbox{ for\ all} \ \mu\in M(X,T).
$$
Since $Inf(X, T) \supset B(X,T)$, one can define the ordered abelian group
$$
\wt K^0(X, T) = K^0(X, \mathbb Z)/Inf(X,\mathbb Z).
$$
\begin{theorem}[\cite{giordano_putnam_skau:1995}, \cite{GPS99}] \label{OE GPS}
Let $(X,T)$ and $(Y,S)$ be minimal Cantor systems. The following are equivalent:
(1) $(X,T)$ and $(Y,S)$ are orbit equivalent;
(2) the groups $\wt K^0(X, T)$ and $\wt K^0(Y, S)$ are order isomorphic by a map preserving the distinguished order units;
(3) there exists a homeomorphism $F : X \to Y$ carrying the $T$-invariant measures onto the $S$-invariant measures;
(4) the full groups $[T]$ and $[S]$ are isomorphic as abstract groups.
\end{theorem}
There are two impressive corollaries of this theorem related to uniquely ergodic homeomorphisms. Let $S(\mu) := \{\mu (A) : A \ \mbox{clopen}\}$ where $\mu$ is a probability measure on a Cantor set $X$.
\begin{corollary} [\cite{giordano_putnam_skau:1995}] \label{OE corollary}
(1) The uniquely ergodic Cantor minimal systems $(X,T)$ and $(Y,S)$ are orbit equivalent if and only if $S(\mu) = S(\nu)$ where $ \mu \circ T = \mu$ and $\nu\circ T = \nu$.
(2) Any uniquely ergodic minimal Cantor system is orbit equivalent either an odometer (when $S(\mu) \subset \mathbb Q$) or to a Denjoy homeomorphism.
\end{corollary}
An action $\Gamma$ of a group $G$ is called {\em hyperfinite or affable (approximately finite)} if it is orbit equivalent to a $\Z$-action. It is an intriguing question which countable groups admit hyperfinite actions by homeomorphisms of a Cantor set. So far, the best known result is
\begin{theorem} [\cite{GMPS10}] \label{Z^n} Any minimal action of $\Z^n$ on a Cantor set is hyperfinite.
\end{theorem}
Hyperfinite equivalence relations have been studied in a series of papers that contain very deep results on their structure like the Absorbtion Theorem (see \cite{GMPS08}). We refer to a recent paper by Putnam \cite{P10} devoted to this subject.
The paper \cite{HKY12} is devoted to the interplay between the notions of topological and measurable orbit equivalence. Let $(X,T)$ and $(Y,S)$ be two Cantor minimal systems. A homeomorphism $\rho\:\ X\rightarrow Y$ is called universally measure preserving if $\rho$ gives rise to an affine bijection between $T$-invariant probability measures and $S$-invariant probability measures. The main result of the paper is a dynamical proof of the fact that two Cantor minimal systems $(X,T)$ and $(Y,S)$ are orbit equivalent if and only if there exists a universally measure preserving homeomorphism between $X$ and $Y$. The proof is based on a scrupulous Kakutani-Rokhlin tower analysis (in its Bratteli-Vershik interpretation). Another important ingredient of the proof is a topological version of the famous copying lemma which is developed in \cite{HKY12}.
\medskip
It is unclear to what extend this result can be generalized. This is an open problem whether any minimal action of a countable amenable group is hyperfinite. Moreover very few facts are known about orbit equivalence of aperiodic homeomorphisms of a Cantor set. We will see below that the Theorem \ref{OE GPS} is not true for aperiodic homeomorphisms. It is obvious that aperiodic actions have more invariants of orbit equivalence than minimal actions have. For example, the cardinality of the set of minimal components must be the same for orbit equivalent aperiodic homeomorphisms.
To finish this subsection, we mention the following result proved by Medynets in \cite{Med}. It is proved that (under some very mild conditions) the Cantor dynamical systems $(X, G_1)$ and $(X, G_2)$ are orbit equivalent if and only if the full groups $[G_1]$ and $[G_2]$ are isomorphic as abstract groups. This theorem, in particular, works for aperiodic actions of countable groups.
\section{Bratteli diagrams that admit a continuous Vershik map}\label{VM}
In this section we are interested in the following problem. How can we recognize whether a given Bratteli diagram admit an order generating a continuous Vershik map? In other words, we are interested in the problem of finding those Bratteli diagrams that can serve as Bratteli-Vershik models of aperiodic homeomorphisms.
We can describe the set of all orderings $\mathcal O_{B}$ on a Bratteli diagram $B$ in the following way. Let $P_v$ denote the set of all orders on $r^{-1}(v)$; an element in $P_v$ is denoted by $\om_v$. Then $\mathcal O_{B}$ can be
represented as
\begin{equation}\label{orderings_set}
\mathcal O_{B} = \prod_{v\in V}P_v .
\end{equation}
Giving each set $P_v$ the discrete topology, we see from (\ref{orderings_set}) that $\mathcal O_{B}$ is a Cantor set with respect to the product topology. In other words, two orderings $\omega= (\om_v)$ and $\omega' = (\om'_v)$ from $\mathcal O_{B}$ are close
if and only if they agree on a sufficiently long initial segment: $\om_v = \om'_v,
v \in \bigcup_{i=0}^k V_i$.
It is worth noticing that the order space $\mathcal O_B$ is sensitive with
respect to a telescoping. Indeed, let $B$ be a Bratteli diagram and $B'$
denote the diagram obtained
by telescoping of $B$ with respect to a subsequence $(n_k)$ of levels.
We see that any ordering $\omega$ on $B$ can be extended to the
(lexicographic) ordering $\om'$ on $B'$. Hence the map $L : \om \to \om'=L(\om)$
defines a closed proper subset $L(\mathcal O_B)$ of $\mathcal O_{B'}$.
The set of all orderings $\mathcal O_{B} $ on a Bratteli diagram $B $
can be considered also as a \textit{measure space} whose Borel structure is
generated by cylinder sets. On the set
$\mathcal O_B $
we take the product measure $\mu = \prod_{v\in V} \mu_v$ where $\mu_v$
is a measure on the set $P_v$. The case where each $\mu_v$
is the uniformly distributed measure on $P_v$ is of particular interest:
$\mu_v(\{i\}) = (|r^{-1}(v)|!)^{-1}$ for every $i \in P_v$ and $v\in V^* \backslash V_0$. Unless $|V_n|=1$ for almost all $n$, if $B'$ is a telescoping of $B$, then in $\mathcal O_{B'}$, $L(\mathcal O_{B})$ is a set of zero measure.
We recall that $\mathcal P_B$ is the subset of $\mathcal O_B$ consisting of \textit{perfect} orderings that produce Bratteli-Vershik topological dynamical systems (Vershik maps).
One can show that the map $\Phi : \omega \to \varphi_\om $ is continuous.
Also the sets $\mathcal P_B$ and ${\mathcal O_B} \setminus {\mathcal P}_B$ are both dense in $\mathcal O_B$ (see \cite{BKY14}).
In \cite{BKY14}, a class of Bratteli diagrams that do not have perfect orders was described.
\begin{definition}\label{aperiodic_definition}
We define the family $\mathcal A$ of Bratteli diagrams, all of whose incidence matrices are of the form
\[F_{n}:=
\left(
\begin{array}{ccccc}
A_{n}^{(1)} & 0 & \ldots & 0 & 0 \\
0 & A_{n}^{(2)} & \ldots & 0 & 0 \\
\vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & \ldots & A_{n}^{(k)} & 0 \\
B_{n}^{(1)}& B_{n}^{(2)} & \ldots & B_{n}^{(k)} & C_{n} \\
\end{array}
\right), \ \ n\geq 1,
\]
where
\begin{enumerate}
\item
for $1\leq i \leq k$ there is some $d_{i}$ such that
$A_{n}^{(i)}$ is a $d_{i}\times d_{i}$ matrix for each $n\geq 1$,
\item all matrices $A_{n}^{(i)}$, $B_{n}^{(i)}$ and $C_{n}$ are strictly
positive,
\item $C_{n}$ is a $d\times d$ matrix,
\item there exists $j\in \{\sum_{i=1}^{k}d_{i} +1, \ldots \sum_{i=1}^{k}d_{i} +d\}$ such that for each $n\geq 1$, the $j$-th row of $F_{n}$ is strictly positive.
\end{enumerate}
If a Bratteli diagram's incidence matrices are of the form above, we shall say that it has $k$ {\em minimal components}.
\end{definition}
It follows from this definition that every Bratteli diagram has exactly $k$ minimal components.
The next proposition describes how for some aperiodic diagrams $B$ that
belong to the special class $\mathcal A$ (see Definition \ref{aperiodic_definition}), there are structural obstacles to the existence of perfect orders on $B$.
\begin{proposition}[\cite{BKY14}]\label{corollary_1}
Let $B\in \mathcal A$ have $k$ minimal components, and such that for each $n\geq 1$, $C_{n}$ is an $s\times s$ matrix where $1\leq s\leq k-1$. If $k>2$, then there is no perfect ordering on $B$. If $k=2$, there are perfect orderings on $B$ only if $C_{n}=(1)$ for all but finitely many $n$.
\end{proposition}
The technique used in \cite{BKY14} and \cite{BY} is based on study new notions related to any ordered Bratteli diagram. They are skeletons and associated graphs.
These notions are especially useful in the case of finite rank diagrams.
Suppose that $B$ has strict rank $d$, i.e., $|V_n| = d$ for all $n\geq 1$.
If a maximal (minimal) path $M$ ($m$) goes through the same vertex $v_{M}$
($v_{m}$) at each level of $B$, we will call this path {\em
vertical}. The following proposition characterizes when $\om$ is
a perfect order on a finite rank Bratteli diagram.
\begin{proposition} [\cite{BKY14}] \label{existence_vershik_map}
Let $(B, \om)$ be an ordered Bratteli diagram.
\begin{enumerate}
\item Suppose that $B$ has strict rank $d$ and that the
$\om$-maximal and $\om$-minimal paths $M_1,...,M_k$ and
$m_1,...,m_{k'}$
are vertical passing through the vertices $v_{M_1},\ldots , v_{M_k}$ and $v_{m_1},\ldots , v_{m_{k'}}$ respectively. Then $\om$ is perfect if and only if
\begin{enumerate}\item $ k = k'$,
\item there is a permutation $\sigma$ of $\{1,\ldots k\}$ such that for
each $i\in \{1,...,k\}$, $v_{M_i}v_{m_{j}} \, \in \mathcal L(B,\om)$ if and only if $j=\sigma(i)$.
\end{enumerate}
\item
Let $B'$ be a telescoping of
$B$. Then $\omega \in \mathcal P_{B}$ if and only if $\om' = L(\om)\in \mathcal P_{B'}$.
\end{enumerate}
\end{proposition}
Let $\om$ be an order on a Bratteli diagram $B$. If $v\in V\backslash V_0 $, we denote the minimal edge with range $v$ by $\ol e_v$ , and we denote the maximal edge with range $v$ by $\wt e_v$.
\begin{definition} \label{well_telescoped_definition}
Let $(B,\om)$ be an ordered rank $d$ diagram. We say that $(B,\om)$ is {\em well telescoped} if
\begin{enumerate}
\item $B$ has strict rank $d$,
\item all $\om$-extremal paths are vertical, with $\wt V$, $\ol V$ denoting the sets of vertices through which maximal and minimal paths run respectively,
\item $s(\wt e_v)\in \wt V$ and $s(\ol e_v) \in \ol V$ for each $v\in V\backslash (V_0 \cup V_1 )$, and this is independent of $n$.
\end{enumerate}
If $(B,\om)$ is perfectly ordered, for it to be considered well telescoped, it will also have to satisfy
\begin{enumerate}
\setcounter{enumi}{3}
\item if $\wt v\ol v $ appears as a subword of some $w(v,m,n)$ with $m\geq 1$, then , then $\sigma(\wt v) = \ol v$ defines a one-to-one correspondence between the sets $\wt V$ and $\ol V$.
\end{enumerate}
\end{definition}
Given an ordered finite rank $(B,\om)$, it can always be telescoped so that it is well telescoped.
For details of how this can be done, see Lemma 3.11 in \cite{BKY14}. Thus, when we talk about a (finite rank) ordered diagram, we assume without loss of generality that it is well telescoped.
For well telescoped ordered diagrams $(B,\om)$, we have $s(\wt e_v) \in \wt V_n$ and $s(\ol e_v) \in \ol V_n$ for any $v\in V_{n+1},\ n \geq 1$.
\begin{definition}
Given a well telescoped $(B,\om)$, we call the set
$$
\mathcal F_{\omega} = (\wt V, \ol V, \{\wt e_{v}, \ol e_{v} : v\in V_{n}, \,\, n\geq 2\})
$$
the {\em skeleton associated to $\om$}. If $\om$ is a perfect order on $B$, it follows that $|\ol V |= |\wt V|$, and
if $\sigma:\wt V\rightarrow \ol V$ is the
permutation given by Proposition \ref{existence_vershik_map}, we call
$\sigma$ the {\em accompanying permutation}.
\end{definition}
The notion of a skeleton of an ordered diagram can be extended to an unordered diagram. Namely, given a strict rank $d$ diagram $B$, we select, two subsets $\wt V$ and $\ol V$ of $V$, of the same cardinality, and, for each $v\in V\backslash V_0 \cup V_1$, we select two edges $\wt e_v$ and $\ol e_v$, both with range $v$, and such that $s(\wt e_v) \in \wt V$,
$s(\ol e_v) \in \ol V$. In this way we can extend the definition of a skeleton
$\mathcal F = (\wt V, \ol V, \{\wt e_{v}, \ol e_{v}: v\in V_{n}, \,\, n\geq 2\})$ to an unordered strict rank Bratteli diagram.
A more detailed discussion can be found in \cite{BKY14}). Arbitrarily choosing a
bijection $\sigma:\wt V \rightarrow \ol V$, we can consider the set of orders on $B$ which have $\mathcal F$ as skeleton and $\sigma$ as accompanying permutation.
\begin{definition} Given a skeleton $\mathcal F$ on a strict finite rank diagram $B$,
for any vertices $\wt v \in
\wt V$ and $\ol v \in \ol V$, we set
\begin{equation}\label{W} W_{\wt v} = \{w \in V : s(\wt e_w) =
\wt v\} \mbox{ and } W'_{\ol v} = \{w \in V : s(\ol e_w) =
\ol v\}.
\end{equation}
Then $W = \{W_{\wt v} : \wt v \in \wt V\}$ and $W' = \{W_{\wt v}' : \ol v \in \ol V\}$ are both partitions of $V$. We call $W$ and $W'$ the {\em partitions generated by $\mathcal F$.}
\end{definition}
Let $[\ol v, \wt v]:= W'_{\ol v} \cap W_{\wt v}$, and
define the partition \[W \cap W' := \{[\ol v,\wt v] : \ol v\in \ol V, \wt v \in \wt V, [\ol v,\wt v]\neq \emptyset\} .\]
\begin{definition}\label{associated_directed_graph}
Let $\mathcal F$ be a skeleton on
the strict finite rank $B$ with
accompanying permutation $\sigma$.
Let $\mathcal H=(T,P)$ be the directed graph where
the set $T$ of vertices of $\mathcal H$ consists of partition elements $[\ol v, \wt v]$ of $W'\cap W$, and where there is an edge in $P$ from $[\ol v, \wt v]$ to $[\ol v', \wt v']$
if and only if $\ol v' \, =\sigma(\wt v)$.
We call $\mathcal H$ the {\em directed graph associated to $(B, \mathcal F, \sigma)$}.
\end{definition}
In order to see how these notion work, we can formulate the following illustrating results. Recall that a directed graph is {\em strongly connected} if for any two
vertices $v$, $v'$, there is a path from $v$ to $v'$, and also a path from $v'$
to $v$. If at least one of these paths exist, then $G$ is {\em weakly
connected}. We notice that, given $(B, \mathcal F, \sigma)$, an associated graph $\mathcal H =(T,P)$ is not connected, in general.
\begin{proposition}[\cite{BKY14}]\label{connectedness and goodness}
Let $(B,\om)$ be a finite rank, perfectly ordered and well telescoped Bratteli diagram, and suppose $\om$ has skeleton $\mathcal F_\om$ and permutation $\sigma$.
\begin{enumerate}
\item If $B$ is simple, then the associated graph $\mathcal H$ is
strongly connected.
\item If $B \in \mathcal A$, then the associated graph $\mathcal H$ is
weakly connected.
\end{enumerate}
\end{proposition}
If we wanted to define similar notions for a not finite rank Bratteli diagram $B$, then we would come across the obvious difficulties because $|V_n|$ is not bounded. Nevertheless, one can overcome these problems and define a skeleton of $B$ and a sequence of associated graphs. The details are rather cumbersome, so that we refer to \cite{BY} where these definitions and more facts were given.
Suppose a Bratteli diagram $B$ is given. The structure of the diagram is completely defined by the sequence of incidence matrices $(F_n)$. The following question seems to be challenging: {\em is it possible to determine using the entries of $F_n$ whether the Bratteli diagram $B$ admits a perfect order?} We give a criterion that answers the above question based on \cite{BKY14} (for a finite rank diagram) and \cite{BY} (for an arbitrary Bratteli diagram). Here we discuss the case of a finite rank diagram since it is more transparent, and the general result from \cite{BY} extends the method used in \cite{BKY14}.
Let $(B,\om)$ be a perfectly ordered simple and well telescoped Bratteli diagram of finite rank diagram (that is $|V_n| = d$). Let $\mathcal F = \mathcal F_\om$ be the skeleton generated by $\om$, $\sigma:\wt V \rightarrow \ol V$ the permutation, and $\varphi = \varphi_\om$ be the corresponding Vershik map such that $\varphi_{\om} (M_v) = m_{\sigma(v)}$ for $ v \in \wt V$.
Also we define the two partitions $W = \{W_{\wt v} : v \in \wt V\}$ and $W' = \{W'_{\ol v} : v \in \ol V\}$ of $V$ generated by $\mathcal F$.
Let $E(V_n, u)$ be the set of all finite paths
between vertices of level $n$ and a vertex $u\in V_m$ where $m >
n$. The symbols $\wt e(V_n, u)$ and $\ol e(V_n, u)$ are used to denote
the maximal and minimal finite paths in $E(V_n, u)$, respectively.
Fix maximal and minimal vertices $\wt v$ and $\ol v$ in $\wt V_{n-1}$ and $\ol V_{n-1}$ respectively. Denote $E(W_{\wt
v}, u) = \{e \in E(V_n, u): s(e) \in W_{\wt v}, r(e) = u\}$ and
$\wt E(W_{\wt v}, u) = E(W_{\wt v}, u) \setminus \{\wt e(V_n,
u)\}$. Similarly, $\ol E(W'_{\ol v}, u) = E(W'_{\ol v}, u)
\setminus \{\ol e(V_n, u)\}$. Clearly, the sets $\{E(W_{\wt v}, u)
: \wt v \in \wt V\}$ form a partition of $E(V_n, u)$.
Let $e$ be a non-maximal finite path, with $r(e)=v$ and $s(e)\in V_m$, which determines the cylinder set $U(e)$. It is clear that for any finite path $e \in \wt E(W_{\wt v}, u)$ we have $\varphi_\om(e) \in \ol E(W'_{\sigma(\wt v)}, u)$.
Therefore,
$$
|\wt E(W_{\wt v}, u)| = |\ol E(W'_{\sigma(\wt v)}, u)|.
$$
Define two sequences of matrices $\wt F_n = (\wt
f_{w,v}^{(n)})$ and $\ol F_n = (\ol f_{w,v}^{(n)})$ by the following
rule (here $w\in V_{n+1}$, $ v\in V_n$ and $ \ n\geq 1$):
\begin{equation}\label{wt f}
\wt f_{w,v}^{(n)} = \left\{
\begin{array}{ll}
f_{w,v}^{(n)} - 1, & \mbox{ if }\wt e_w \in E(v, w); \\
f_{w,v}^{(n)}, & \mbox{otherwise},
\end{array}
\right.
\end{equation}
\begin{equation}\label{ol f}
\ol f_{w,v}^{(n)} = \left\{
\begin{array}{ll}
f_{w,v}^{(n)} - 1, & \mbox{ if }\ol e_w \in E(v, w); \\
f_{w,v}^{(n)}, & \mbox{otherwise}.
\end{array}
\right.
\end{equation}
Then for any $u\in V_{n+1}$ and $ \wt v\in \wt V_{n-1}$, we obtain that the entries of incidence matrices have the property:
\begin{equation}\label{sum condition1}
\sum_{w\in W_{\wt v}} \wt f_{u,w}^{(n)} \ \ = \sum_{w'\in W'_{\sigma(\wt v)}} \ol f_{u,w'}^{(n)},\ \ n \geq 2.
\end{equation}
We call relations (\ref{sum condition1}) the {\em balance relations}.
\begin{theorem} [\cite{BKY14}] \label{existence of good order}
Let $B$ be a simple strict rank $d$ Bratteli diagram, let $\mathcal F = \{M_{\wt v}, m_{\ol v}, \wt e_w, \ol e_w: w \in V \backslash V_0,\,\, \wt v \in \wt V\mbox{ and } \ol v \in \ol V \}$ be a skeleton on $B$, and let $\sigma : \wt V \to \ol V$ be a bijection. Suppose that eventually all associated graphs $\mathcal H_n$ are positively strongly connected, and suppose that the entries of incidence matrices $(F_n)$ eventually satisfy the balance relations (\ref{sum condition1}).
Then there is a perfect ordering $\om$ on $B$ such that $\mathcal F = \mathcal F_\om$ and the Vershik map $\varphi_\om$ satisfies the relation $\varphi_\om(M_{\wt v}) = m_{\sigma(\wt v)}$.
\end{theorem}
The next results is related to finite rank Bratteli diagrams. Recall that on the set of all orders $\mathcal O_B$ we can consider the product measure $P$ such that the probability to pick up an order for $r^{-1} (v) $ is uniform. The following theorem is somewhat surprising because it states that the diagram $B$ ``knows a priori'' how many maximal and minimal paths it must have.
\begin{theorem} [\cite{BKY14}]\label{goodbaddichotomy}
Let $B$ be a finite rank $d$ aperiodic Bratteli diagram. Then there
exists $j \in \{1,..., d\}$ such that $P$-almost all orders
have $j$ maximal and $j$ minimal elements.
\end{theorem}
In the recent work \cite{JQY14}, random orders have been studied on simple 0-1 Bratteli diagrams, i.e., all entries of incidence matrices are either zeros or ones.
It was proved that a random order has uncountably many infinite paths if and only if the growth rate of the level-n vertex sets is superlinear.
\begin{theorem}[\cite{JQY14}] \label{JQY} Let $B$ be a Bratteli diagram with incidence matrices $F_n$ whose entries are 1 and $|V_n| \geq 1$. The following dichotomy holds:
(1) If $\sum_n 1/|V_n| = \infty$, then there exists $P$-almost sure a unique maximal path.
(2) If $\sum_n 1/|V_n| < \infty$, then there exists $P$-almost sure
uncountably many maximal paths.
\end{theorem}
Thus, a random order on a slowly growing Bratteli diagram satisfying \ref{JQY}
admits a homeomorphism, while a random order on a quickly growing Bratteli diagram does not.
\section{Stationary Bratteli diagrams}\label{stationary}
In this section, stationary Bratteli diagrams are considered. We explicitly describe the class of homeomorphisms represented by stationary simple and non-simple diagrams. We also consider their invariant measures and orbit equivalence classes.
\subsection{Substitution dynamical systems}
Papers~\cite{forrest, durand_host_skau} explore minimal Cantor systems and their relation to stationary Bratteli diagrams and the corresponding dimension groups (for example, see~\cite{Effros} for details on dimension groups). It turns out that the class of minimal homeomorphisms which can be represented by stationary Bratteli diagrams is constituted by minimal substitution dynamical systems and odometers. In \cite{BKM09}, the systematical study of non-primitive substitutions was initiated. In particular, it is shown that one can prove the analogue of the above mentioned result for aperiodic homeomorphisms, i.e., the assumption of minimality is not essential for finding the corresponding Bratteli-Vershik model.
Let $A$ denote a finite alphabet and $A^+$ the set of all non-empty words over $A$. For a word $w=w_0\ldots w_{n-1}$
with $w_i\in A$ let $|w|=n$ stand for its length. By a {\it substitution}, we mean any map $\sigma : A\rightarrow A^+$. A map $\sigma : A\rightarrow A^+$ can be extended to the map $\sigma :A^+\rightarrow A^+$ by concatenation. We define the {\it language} $L(\sigma)$ of a substitution $\sigma$ as the set of all words which appear as factors of $\sigma^n(a)$, $a\in A$, $n\geq 1$. We set also $\sigma^0(a)=a$ for all $a\in A$. Let $T : A^\mathbb {Z} \to A^\mathbb {Z}$ be the shift, that is $T(x_i) = (y_i)$ where $y_i = x_{i+1}$, $i\in \Z$. By a {\it substitution dynamical system associated to a substitution $\sigma$}, we mean a pair $(X_\sigma,T_\sigma)$, where
$$
X_\sigma=\{x\in A^\mathbb Z : x{[-n,n]}\in L(\sigma)\mbox{ for any }n\}
$$
is a closed $T$-invariant subset of $A^\mathbb Z$, and $T_\sigma$ is the shift $T$ restricted to the set $X_\sigma$. We will denote the $k$th coordinate of $x \in X_\sigma$ by $x[k]$.
The following theorem is one of the main results in~\cite{durand_host_skau}. A part of this theorem was proved in~\cite{forrest}.
\begin{theorem}[\cite{durand_host_skau}]\label{DHSmain}
The family $\mathfrak{B}$ of Bratteli-Vershik systems associated with stationary,
properly ordered Bratteli diagrams is (up to isomorphism) the disjoint union of
the family of substitution minimal systems and the family of stationary odometer
systems. Furthermore, the correspondence in question is given by an explicit
and algorithmic effective construction. The same is true of the computation of the
(stationary) dimension group associated with a substitution minimal system.
\end{theorem}
In the proof of Durand, Host and Skau, the explicit construction of homeomorphism is given, while the proof of Forrest has more existential nature. For instance, let $B$ be a stationary properly ordered Bratteli diagram with a simple hat
(a Bratteli diagram has a simple hat whenever it has only simple edges between the top vertex and the first level).
The substitution $\sigma$ \textit{read on} $B$ is defined as follows. Since $B$ is
stationary, all information about it is given by the first level. Let $\mathcal{A} =
\{a_1,\ldots,a_K\}$ be an alphabet, where $K = |V_1| = |V_n|$. With a vertex
number $i$ we associate a letter $a_i$, this operation does not depend on the level.
Consider a letter (vertex) $a \in V_2$ and the ordered list $(e_1, \ldots, e_r)$ of
the edges in $E_1$ with $r(e_j) = a$. Let $(a_1,\ldots, a_r)$ be the ordered list of
the sources of these edges. Then define $\sigma(a) = a_1 \ldots a_r$. If $\sigma$
is periodic then $(X_B, \varphi_B)$ is isomorphic to an odometer with stationary
base, otherwise $(X_B, \varphi_B)$ is isomorphic to $(X_\sigma, T_\sigma)$.
On the other hand, any substitution $\sigma$ defines a stationary Bratteli diagram: one has to find a stationary ordered Bratteli diagram $B$ with simple hat such that the substitution read on $B$ is exactly $\sigma$.
A substitution $\sigma$ on an alphabet $\mathcal{A}$ is called \textit{proper} if there exists an integer $n > 0$ and two letters $a,b \in \mathcal{A}$ such that for every $c \in \mathcal{A}$, $a$ is the first letter and $b$ is the last letter of $\sigma^n(c)$.
In the case when $\sigma$ is primitive, aperiodic and proper, the systems $(X_\sigma, T_\sigma)$ and $(X_B, \varphi_B)$ are isomorphic. In the case of a non-proper substitution one has to build Bratteli diagram using return words~\cite{durand_host_skau}.
The following theorem extends the result from the minimal case to the aperiodic one.
\begin{theorem}[\cite{BKM09}]\label{TheoremVershikMapIsSubstitution}
(i) Suppose that $(X_B,\varphi_B)$ is an aperiodic Bratteli-Vershik system with $B$ a stationary ordered Bratteli diagram and $X_B$ is perfect. Then the system $(X_B,\varphi_B)$ is conjugate to an aperiodic substitution dynamical system (with substitution read on $B$) if and only if no restriction of $\varphi_B$ to a minimal component is isomorphic to an odometer.
(ii) Let $\sigma : A\rightarrow A^+$ be an aperiodic substitution such that $|\sigma^n(a)|\rightarrow \infty$ as $n \rightarrow \infty$ for all $a \in A$. Then the substitution dynamical system $(X_\sigma,T_\sigma)$ is conjugate to the Vershik map of a stationary Bratteli diagram.
\end{theorem}
The technique presented in the proof is applicable for a wide class of substitutions
including various Chacon-like substitutions.
A substitution is {\it recognizable} if for each $x\in X_\sigma$ there exist a unique $y\in X_\sigma$ and unique $i\in \{0,\ldots,|\sigma(y[0])|-1\}$ such that $x = T^i_\sigma\sigma(y)$. Every aperiodic primitive substitution is recognizable~\cite{mosse:1992, mosse:1996}. In~\cite{BKM09}, it is shown that an arbitrary aperiodic substitution is recognizable:
\begin{theorem}[\cite{BKM09}]\label{TheoremRecognizability}
Each aperiodic substitution $\sigma : A\rightarrow A^+$ is recognizable.
\end{theorem}
The proof involves the usage of Downarowicz-Maass' techniques~\cite{DM} (see also Section~\ref{finrank}).
\subsection{Invariant measures on stationary diagrams and their supports}
In this subsection, we give an explicit description of all ergodic probability measures on stationary Bratteli diagrams invariant with respect to the tail equivalence relation (or the Vershik map).
Let $B = (V,E)$ be a Bratteli diagram with incidence matrices ${F_n}$.
For $w \in V_{n}$, the set $E(v_{0}, w)$ defines the clopen subset $X_{w}^{(n)} := \{x = (x_{i}) \in X_{B} : r(x_{n}) = w \}$ of $X_B$. Then $\{X_{w}^{(n)} : w \in V_{n}\}$ is a clopen partition of $X_{B}$. Analogously, the sets $X_{w}^{(n)}(\overline{e}) := \{x = (x_{i}) \in X_{B} : x_{i} = e_{i}, i = 1,...,n\}$ determine a clopen partition of $X_{w}^{(n)}$ where $\overline{e} = (e_{1}, \ldots ,e_{n}) \in E(v_{0}, w)$, $n \geq 1$. Note that any two paths $x$, $y$ from $X_{B}$ are $\mathcal{R}$-equivalent if and only if there exists $w \in V$ such that $x \in X_{w}^{(n)}(\overline{e})$ and $y \in X_{w}^{(n)}(\overline{e'})$ for some $\overline{e}, \overline{e'} \in E(v_{0}, w)$. Recall that a measure $\mu$ on $X_B$ is called $\mathcal{R}$-\textit{invariant} if for any two paths $\overline{e}$ and $\overline{e'}$ from $E(v_{0}, w)$ and any vertex $w$, one has $\mu(X_{w}^{(n)}(\overline{e})) = \mu(X_{w}^{(n)}(\overline{e'}))$. Then
$$
\mu(X_{w}^{(n)}(\overline{e})) = \frac{1}{h_{w}^{(n)}}\mu(X_{w}^{(n)}), \ \ \overline{e} \in E(v_{0}, w).
$$
For $x = (x_1, \ldots, x_N) \in \mathbb{R}^{N}$, we wil write $x \geq 0$ if $x_i \geq 0$ for all $i$. Let $\mathbb{R}^N_+ = \{x \in \mathbb{R}^N : x \geq 0\}$. Let
$$
C_k^{(n)} = F_k^T \ldots F_n^T(\mathbb{R}^{|V_{n+1}|}_+) \mbox{ for } 1 \leq k \leq n.
$$
Clearly, $\mathbb{R}^{|V_{k}|}_+ \supset C_k^{(n)}\supset C_k^{(n+1)}$ for all $n \geq 1$. Let
$$
C_k^{(\infty)} = \bigcap_{n \geq k}C_k^{(n)} \mbox{ for } k \geq 1.
$$
The following theorem describes a construction of invariant measures on Bratteli diagram of general form.
\begin{theorem}[\cite{BKMS_2010}]\label{Theorem_measures_general_case} Let $B = (V,E)$ be a Bratteli diagram such that the tail equivalence relation
$\mathcal R$ on $X_B$ is aperiodic.
If $\mu\in M(\Rk)$,
then the vectors $p^{(n)}=(\mu(X_w^{(n)}(\ov{e})))_{w\in
V_n}$, $\ov{e}\in E(v_0,w)$, satisfy the following conditions for $n\geq 1$:
(i) $p^{(n)} \in C_n^\infty$,
(ii) $F_n^T p^{(n+1)} = p^{(n)}$.
Conversely, if a sequence of vectors $\{ p^{(n)}\}$ from
$\mathbb R^{|V_n|}_+$ satisfies condition (ii), then there
exists a non-atomic finite Borel $\mathcal R$-invariant measure $\mu$ on
$X_B$ with $p_w^{(n)}=\mu(X_w^{(n)}(\overline e))$ for all $n\geq 1$
and $w\in V_n$.
The $\Rk$-invariant measure $\mu$ is a probability measure if and only if
(iii) $\sum_{w\in V_n}h_w^{(n)}p_w^{(n)}=1$ for $n=1$,
\noindent in which case this equality holds for all $n\ge 1$.
\end{theorem}
In \cite{BKMS_2010}, all invariant ergodic measures on a stationary Bratteli diagram were described as follows. It was first shown that the study of any stationary Bratteli diagram $B = (V,E)$ (with $|V| = N$) can be reduced to the case when the incidence matrix $F$ of size $N\times N$ has the Frobenius Normal Form:
\begin{equation}\label{Frobenius Form}
F =\left(
\begin{array}{ccccccc}
F_1 & 0 & \cdots & 0 & 0 & \cdots & 0 \\
0 & F_2 & \cdots & 0 & 0 & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots & \vdots & \cdots& \vdots \\
0 & 0 & \cdots & F_s & 0 & \cdots & 0 \\
X_{s+1,1} & X_{s+1,2} & \cdots & X_{s+1,s} & F_{s+1} & \cdots & 0 \\
\vdots & \vdots & \cdots & \vdots & \vdots & \ddots & \vdots \\
X_{m,1} & X_{m,2} & \cdots & X_{m,s} & X_{m,s+1} & \cdots & F_m \\
\end{array}
\right)
\end{equation}
We can telescope the diagram $B$ and regroup the vertices in such a way that the matrix $F$ will have the following property:
\begin{eqnarray} \label{prop-I}
& F & \mbox{has the form (\ref{Frobenius Form}) where every nonzero matrix $F_i$}\\[-1ex]
& & \mbox{on the main diagonal is primitive.} \nonumber
\end{eqnarray}
We can telescope the diagram $B$ further to make sure that
\begin{eqnarray} \label{prop-I'}
& F & \mbox{has the form (\ref{Frobenius Form}) where every nonzero matrix $F_i$}\\[-1ex]
& & \mbox{on the main diagonal is strictly positive.} \nonumber
\end{eqnarray}
The matrices $F_i$ determine the partition of the vertex set $V$ into subsets (classes) $V_i$ of vertices. In their turn, these subsets generate subdiagrams $B_i$. The non-zero matrices $X_{j,k}$ indicate which subdiagrams are linked by some edges (or finite paths). Notice that each subdiagram $B_i$, $i=1,...,s$, corresponds to a minimal component of the cofinal equivalence relation $\mathcal R$.
We denote by $F_\alpha$, $\alpha \in \Lambda$, the \textit{non-zero} matrices on the main diagonal in (\ref{Frobenius Form}).
Let $\alpha \geq \beta$. It is said that the class of vertices $\alpha$ {\it has access} to a class $\beta$, in symbols $\alpha \succeq \beta$, if and only if either $\alpha = \beta$ or there is a finite path in the diagram from a vertex which belongs to $\beta$ to a vertex from $\alpha$. In other words, the matrix $X_{\alpha, \beta}$ is non-zero. A class $\alpha$ is called {\it final (initial)} if there is no class $\beta$ such that $ \alpha \succ \beta$ ($\beta \succ \alpha$).
Let $\rho_\alpha$ be the spectral radius of $F_\alpha$. A class $\alpha\in \{1,...,m\}$ is called {\it distinguished} if $\rho_\alpha > \rho_\beta$ whenever $ \alpha \succ \beta$. Notice that all classes $\alpha = 1,\ldots,s$ are necessarily distinguished. A real number $\lambda$ is called a {\it
distinguished eigenvalue} if there exists a non-negative eigenvector $x$ with $Fx=\lambda x$.
A real number $\lambda$ is a distinguished eigenvalue if and only if there exists a distinguished class $\alpha$ such that $\rho_\alpha = \lambda$. In this case we denote $\lambda_\alpha = \lambda$.
If $x = (x_1,...,x_N)^T$ is an eigenvector corresponding to a distinguished eigenvalue $\lambda_\alpha$, then $x_i >0$ if and only if $i \in \beta$ and $\alpha \succeq \beta$.
The main result of \cite{BKMS_2010} completely describes the simplex of $\mathcal R$-invariant probability measures
of a stationary Bratteli diagram. Denote
$$
core (A) = \bigcap_{k \geq 1} A^{k}(\mathbb{R}_+^N).
$$
Let $\{\xi_1,\ldots,\xi_k\}$ be the extreme vectors of the cone $core(A)$. Normalize each vector so that $\sum_{w \in V_1} h_w^{(1)} (\xi_i)_w = 1$. Set $D = \{x \in core(A) : \sum_{w \in V_1} h_w^{(1)} x_w = 1 \}$.
\begin{theorem}[\cite{BKMS_2010}]\label{Theorem_MeasuresErgodicMeasures}
Suppose that $B$ is a stationary Bratteli diagram such that the
tail equivalence relation $\Rk$ is aperiodic and
the incidence matrix $F$ satisfies (\ref{prop-I}).
Then there is a one-to-one correspondence between vectors $p^{(1)}\in D$ and $\mathcal R$-invariant probability
measures on $X_B$. This correspondence is given by the rule $\mu \leftrightarrow
p^{(1)}= (\mu(X_w^{(1)})/h_w^{(1)})_{w\in V_1}$. Furthermore, ergodic measures correspond to the extreme vectors $\{\xi_1,\ldots,\xi_k\}$. In particular, there exist exactly $k$ ergodic measures.
\end{theorem}
More precisely, fix a distinguished eigenvalue $\lambda$ and let $x = (x_1,...,x_N)^T$ be the probability non-negative eigenvector corresponding to $\lambda$. Then the ergodic probability measure $\mu$ defined by $\lambda$ and $x$ satisfies the relation:
\begin{equation}\label{mucount}
\mu(X_{i}^{(n)}(\overline{e})) = \frac{x_i}{\lambda^{n - 1}},
\end{equation}
where $i\in V_n$ and $\overline e$ is a finite path with $s(\overline e) =i$. Therefore, the clopen values set for $\mu$ has has the form:
\begin{equation}\label{formulaSmu}
S(\mu) = \left\{\sum_{i = 1}^{N} k^{(n)}_{i}\frac{x_i}{\lambda^{n - 1}} : 0 \leq k^{(n)}_{i} \leq h^{(n)}_{i}; \; n = 1, 2, \ldots \right\}.
\end{equation}
If $\lambda$ is a non-distinguished Perron-Frobenius eigenvalue for $A$, then the corresponding class of vertices is non-distinguished and $\mathcal R$-invariant measure on $X_B$ is infinite. The following result describes infinite $\Rk$-invariant measures for stationary Bratteli diagrams.
\begin{theorem} [\cite{BKMS_2010}]\label{th-infmeas}
Suppose that $B$ is a stationary Bratteli diagram such that the
tail equivalence relation $\Rk$ is aperiodic and
the incidence matrix $F$ satisfies (\ref{prop-I}).
Then the set of ergodic infinite ($\sig$-finite) invariant measures, which are positive
and finite on at least one open set (depending on the measure), modulo a
constant multiple, is in 1-to-1 correspondence with the set of non-distinguished eigenvalues of $A=F^T$.
\end{theorem}
The following theorem gives necessary and sufficient conditions under which a measure on $X_B$ will be $\mathcal{R}$-invariant.
\begin{theorem}[\cite{BKMS_2010}]\label{TheoremStationaryMeasures}
Suppose that $B$ is a stationary Bratteli diagram such that the
tail equivalence relation $\Rk$ is aperiodic and
the incidence matrix $F$ satisfies (\ref{prop-I}).
Let $\mu$ be a finite Borel $\mathcal R$-invariant measure on $X_B$.
Set $p^{(n)}=(\mu(X_w^{(n)}(\overline e)))_{w\in V_n}$ where $\overline e\in E(v_0,w)$. Then for the matrix $A = F^T$ associated to $B$ the following properties hold:
(i) $p^{(n)}=A p^{(n+1)}$ for every $n\geq 1$;
(ii) $ p^{(n)}\in core(A),\ n\geq 1$.
Conversely, if a sequence of vectors $\{ p^{(n)}\}$ from
$\mathbb R^{N}_+$ satisfies condition (ii), then there
exists a finite Borel $\mathcal R$-invariant measure $\mu$ on
$X_B$ with $p_w^{(n)}=\mu(X_w^{(n)}(\overline e))$ for all $n\geq 1$
and $w\in V_n$.
The $\Rk$-invariant measure $\mu$ is a probability measure if and only if
(iii) $\sum_{w\in V_n}h_w^{(n)}p_w^{(n)}=1$ for $n=1$,
\noindent in which case this equality holds for all $n\ge 1$.
\end{theorem}
Let $\sigma : \mathcal A\to \mathcal A^+$ be an aperiodic substitution such that $|\sigma^n(a)| \to \infty$ as $n\to \infty$ for any $a\in \mathcal A$. Denote by $(X_\sigma, T_\sigma)$ the corresponding substitution dynamical system. Let $M_\sigma = (m_{ab})$ be the matrix of substitution $\sigma$, i.e. $m_{ab} = L_a(\sigma(b))$ where $L_a(\sigma(b))$ is the number of $a$ occurring in $\sigma(b)$. Denote by $B_\sigma$ the stationary Bratteli diagram ``read on substitution''. This means that $M_\sigma$ is the matrix transpose to the incidence matrix of $B_\sigma$. It follows from Theorem \ref{TheoremVershikMapIsSubstitution} that there exists a stationary ordered Bratteli diagram $B(X_\sigma, T_\sigma) = B$ whose Vershik map $\varphi_B$ is conjugate to $T_\sigma$. Thus, we have two Bratteli diagrams associated to $(X_\sigma, T_\sigma)$. Note that the diagram $B$ may have considerably more vertices than the diagram $B_\sigma$.
\begin{theorem}[\cite{BKMS_2010}]\label{measures_and_substitutions}
There is a one-to-one correspondence $\Phi$ between the set of ergodic $T_\sigma$-invariant probability measures on the space
$X_\sigma$ and the set of ergodic $\mathcal R$-invariant probability measures on the path space $X_{B_\sigma}$ of the stationary diagram $B_\sigma$ defined by substitution $\sigma$. The same statement holds for non-atomic infinite invariant measures.
\end{theorem}
\subsection{Good measures on stationary Bratteli diagrams}
Two probability measures $\mu$ and $\nu$ defined on Borel subsets of a topological space $X$ are called \textit{homeomorphic} or \textit{topologically equivalent} if there exists a self-homeomorphism $h$ of $X$ such that $\mu = \nu\circ h$, i.e. $\mu(E) = \nu(h(E))$ for every Borel subset $E$ of $X$. In such a way, the set of all Borel probability measures on $X$ is partitioned into equivalence classes. The classification of Borel measures with respect to homeomorphisms started in the paper by Oxtoby and Ulam~\cite{Oxt-Ul}, where they found a criterion for a measure to be homeomorphic to Lebesgue measure on the interval $[0,1]$. The classification of measures on Cantor sets started with classification of Bernoulli measures (see \cite{Navarro, Navarro-Oxtoby} and later papers \cite{Austin, D-M-Y, Akin3, Yingst}). Akin began a systematic study of homeomorphic measures on a Cantor space \cite{Akin1, Akin2}. It was noted in \cite{Akin1} that there exist continuum classes of equivalent full non-atomic probability measures on a Cantor set. This fact is based on the existence of a countable base of clopen subsets of a Cantor set. Recall that for a measure $\mu$ on a Cantor space $X$, the \textit{clopen values set} $S(\mu) = \{\mu(U):\,U\mbox{ is clopen in } X\}$.
The set $S(\mu)$ is a countable dense subset of the unit interval, and this set provides an invariant for topologically equivalent measures, although it is not a complete invariant, in general. But for the class of the so called good measures, $S(\mu)$ \textit{is} a complete invariant.
\begin{definition}
A full non-atomic probability measure $\mu$ is \textit{good} if whenever $U$, $V$ are clopen sets with $\mu(U) < \mu(V)$, there exists a clopen subset $W$ of $V$ such that $\mu(W) = \mu(U)$.
\end{definition}
It turns out that such measures are exactly invariant measures of uniquely ergodic
minimal homeomorphisms of Cantor sets (see \cite{Akin2},
\cite{glasner_weiss:1995-1}).
It is obvious that all above mentioned definitions (clopen values set, good measures, etc.) are applicable to the measures on Bratteli diagrams. In particular, we note that the ergodic measures corresponding to minimal components are automatically good: on a simple stationary Bratteli diagram any Vershik map is minimal and uniquely ergodic.
The following results deal with the classification of ergodic probability Borel measures on stationary non-simple Bratteli diagrams which are invariant with respect to the tail equivalence relation $\mathcal{R}$.
Let $B = (V, E)$ be a stationary non-simple Bratteli diagram, $F$ be its incidence $K \times K$ matrix and $A = F^T$. Let $\mu$ be the measure defined by a distinguished class of vertices $\alpha$ and $\lambda$ the corresponding distinguished eigenvalue of $A$. Denote by $(y_1,...,y_K)^T$ the probability eigenvector of the matrix $A$ corresponding to $\lambda$. Notice that the vector $(y_1,...,y_K)^T$ may have zero entries. These zero entries are assigned to the vertices from $B$ that are not accessible from the class $\alpha$. Denote by $(x_1, \ldots, x_n)^T$ the positive vector obtained from $(y_1,...,y_K)^T$ by crossing out zero entries. We call $(x_1,...,x_n)^T$ the \textit{reduced vector} corresponding to the measure $\mu$.
The following theorem gives a criterion for $\mu$ to be good.
\begin{theorem}[\cite{S.B.O.K.}]\label{KritGood}
Let $\mu$ be an ergodic $\mathcal R$-invariant measure on a stationary diagram $B$ defined by a distinguished eigenvalue $\lambda$ of the matrix $A = F^T$. Denote by $x = (x_1,...,x_n)^T$ the corresponding reduced vector. Let the vertices $m+1, \ldots, n$ belong to the distinguished class $\alpha$ corresponding to $\mu$. Then $\mu$ is good if and only if there exists $R \in \mathbb{N}$ such that $\lambda^R x_1,...,\lambda^R x_m$ belong to the additive group generated by $\{x_j\}_{j=m+1}^n$.
If the clopen values set of $\mu$ is rational and $(\frac{p_1}{q}, \ldots, \frac{p_n}{q})^T$ is the corresponding reduced vector, then $\mu$ is good if and only if $\gcd(p_{m+1}, ...,p_n) | \; \lambda^R$ for some $R \in \mathbb N$.
\end{theorem}
The idea of the proof is as follows. The support of measure $\mu$ is a stationary simple subdiagram $B_\alpha$ which corresponds to the distinguished class $\alpha$. The measure $\mu\mid_{B_\alpha}$ is good since it is a unique ergodic invariant measure for a Vershik homeomorphism. The property of goodness can be destroyed when $\mu\mid_{B_\alpha}$ is extended to the measure $\mu$ on the whole diagram $B$. The measures of cylinder sets that end in the vertices of $B$ that do not belong to class $\alpha$ are obtained as infinite sums of measures of cylinder sets that end in the vertices of class $\alpha$. Thus, new values in $S(\mu)$ might appear which do not belong to $S(\mu\mid_{B_\alpha})$, and this will be the reason for $\mu$ to be not good.
For a good measure $\mu$, there always exists a subgroup $G \subset \mathbb R$ such that $S(\mu) = G\cap [0,1]$, i.e. $S(\mu)$ is \textit{group-like} (see~\cite{Akin2}). Indeed, let $U$ and $V$ be any two clopen sets and $\mu$ be a good measure. Suppose $\mu(U) = \alpha$, $\mu(V) = \beta$ and without loss of generality $\alpha < \beta$. Thus, $\alpha, \beta \in S(\mu)$. Since $\mu$ is good, there is a clopen set $W \subset V$ such that $\mu(W) = \alpha$. Hence $V \setminus W$ is a clopen set of measure $\beta - \alpha$, and $\beta - \alpha \in S(\mu)$. It easily follows that $S(\mu)$ is group-like.
It turns out that for any ergodic invariant measure $\mu$ on a stationary diagram $B$ the set $S(\mu)$ is group-like~\cite{S.B.O.K.}. Moreover, it was proved that for an ergodic invariant measure $\mu$ defined by a distinguished eigenvalue $\lambda$ and distinguished class of vertices $\alpha$, the following equation holds:
$$
S(\mu) = \left(\bigcup_{N=0}^\infty \frac{1}{\lambda^N} H\right) \cap [0,1],
$$
where $(x_1, \ldots, x_n)^T$ is the corresponding reduced vector and $H$ is the additive subgroup of $\mathbb{R}$ generated by $\{x_1, \ldots , x_n\}$.
Two cases were considered in the proof: (i) $\lambda$ is rational and hence $S(\mu)\subset\mathbb Q$; (ii) $\lambda$ is an irrational algebraic integer of degree $k$, then $S(\mu)\subset \mathbb{Q}[\lambda]$, where $\mathbb{Q}[\lambda]$ is the least ring that contains both $\mathbb{Q}$ and $\lambda$. The case (i) is relatively simple. The following asymptotics mentioned in \cite{BKMS_2010} was used:
$$
(A^n)_{ij} \sim \lambda^n,\ \ n\to \infty,\ \ \mbox{for}\ i\in \beta,\ j\in \alpha,\ \mbox{with}\
\alpha \succeq \beta.
$$
Here $\sim$ means that the ratio tends to a positive constant. On the other hand,
$$
(A^n)_{ij} = o(\lambda^n),\ \ n\to \infty,\ \ \mbox{for}\ j\in
\beta \prec \alpha.
$$
In the case (ii), the eigenvector entries and eigenvalues are represented as vectors with rational entries (the dimension of vectors is equal to the algebraic degree of $\lambda$). Then the operation of dividing by $\lambda$ is a linear transformation in the vector space $\mathbb{Q}^{k}$. The methods of linear algebra and matrix theory were used to obtain the proof.
Theorem~\ref{KritGood} was used to prove the following result. We showed that one can build infinitely many homeomorphic ergodic invariant measures on stationary diagrams such that the corresponding tail equivalence relations are non-orbit equivalent.
\begin{theorem}[\cite{S.B.O.K.}]\label{goodhomeo}
Let $\mu$ be a good ergodic $\mathcal{R}$-invariant probability measure on a stationary (non-simple) Bratteli diagram $B$. Then there exist stationary Bratteli diagrams $\{B_i\}_{i=0}^\infty$ and good ergodic $\mathcal{R}_i$-invariant probability measures $\mu_i$ on $B_i$ such that each measure $\mu_i$ is homeomorphic to $\mu$ and the dynamical systems $(B_i, \mathcal R_i)$, $(B_j, \mathcal R_j)$ are topologically orbit equivalent if and only if $i = j$. Moreover, the diagram $B_i$ has exactly $i$ minimal components for the tail equivalence relation $\mathcal R_i, i\in \mathbb N$.
\end{theorem}
In~\cite{K12}, the notion of good measure is extended to the case of infinite measures on Cantor sets.
Borel infinite measures arise as ergodic invariant measures for aperiodic
homeomorphisms of a Cantor set. The study of homeomorphic infinite measures is
of crucial importance for the classification of Cantor aperiodic systems up to orbit equivalence.
The direct analogues of Theorems~\ref{KritGood}, \ref{goodhomeo} for infinite measures
on stationary Bratteli diagrams can be found in~\cite{K12}. Measures on stationary
Bratteli diagrams can be also considered as extensions of good measures on
non-compact locally compact Cantor sets (see~\cite{K122} for details).
In the paper \cite{BH} the authors initiated the study of properties of traces of dimension groups which are motivated by Cantor dynamics. In particular, it was shown that there are example of minimal Cantor systems whose ergodic measures have completely different (in some sense even opposite) properties.
\subsection{Complexity and orbit equivalence}
Let $\mu$ be an ergodic invariant measure for a proper primitive substitution dynamical system. By Theorem~\ref{DHSmain}, we can find easily the clopen values set $S(\mu)$ in terms of the matrix of substitution. By Corollary~\ref{OE corollary}, to construct a minimal substitution dynamical system which is orbit equivalent to a given one, it suffices to find another stationary simple Bratteli diagram such that the clopen values set is kept unchanged. In the paper~\cite{BK14}, two constructions are used to build countably many non-isomorphic orbit equivalent minimal substitution dynamical systems. In both constructions, the complexity function $n\mapsto p_\sigma(n)$ is used to distinguish non-isomorphic systems. Recall that the function $p_\sigma(n)$ counts the number of words of length $n$ in the infinite sequence invariant with respect to $\sigma$. If minimal substitution dynamical systems $(X_\sigma, T_\sigma)$ and $(X_\zeta, T_\zeta)$ are topologically conjugate, then there exists a constant $c$ such that, for all $n > c$ one has $p_\sigma(n - c) \leq p_\zeta(n) \leq p_\sigma(n + c)$ (see~\cite{F}). In the proof of the following theorem, the complexity of the systems $(X_{\zeta_n}, T_{\zeta_n})$ is forced to grow by increasing the number of letters in the alphabets of $\zeta_n$.
\begin{theorem}[\cite{BK14}]\label{diff_alphabets}
Let $\sigma$ be a proper substitution.
Then there exist countably many proper substitutions $\{\zeta_n\}_{n=1}^\infty$ such that $(X_\sigma, T_\sigma)$ is orbit equivalent to $(X_{\zeta_n}, T_{\zeta_n})$, but the systems $\{(X_{\zeta_n}, T_{\zeta_n})\}_{n=1}^\infty$ are pairwise non-isomorphic.
\end{theorem}
Moreover, if one wants to have a substitution dynamical system which is strongly orbit equivalent to a given $(X_\sigma, T_\sigma)$, then additionally the dimension group of the diagram $B_\sigma$ must be unchanged. If $A$ is the incidence matrix for $\sigma$ and $A^N$ is the incidence matrix for $\zeta$ for some $N \in \mathbb{N}$, then the dimension groups associated to minimal Cantor systems $(X_\sigma, T_\sigma)$ and $(X_\zeta, T_\zeta)$ are order isomorphic (see~\cite{giordano_putnam_skau:1995}). In the following theorem, incidence matrices of built substitution systems are the powers of the incidence matrix of the initial substitution system $(X_\sigma, T_\sigma)$. Hence substitutions $\sigma$ and $\zeta_n$ have the same fixed alphabet, and the complexity function is made increasing by enlarging the length of substitution $\zeta_n$ and by an appropriate permutation of letters. Thus, we produce a countable family of pairwise non-isomorphic strong orbit equivalent substitution systems.
\begin{theorem}[\cite{BK14}]
Let $\sigma$ be a primitive proper substitution. Let $(B, \leq)$ be the corresponding stationary properly ordered simple Bratteli diagram.
Then there exist countably many telescopings $B_n$ of $B$ with proper orders $\leq_n$ and corresponding substitutions $\zeta_n$ read on $B_n$ such that the substitution dynamical
systems $\{(X_{\zeta_n}, T_{\zeta_n})\}_{n=1}^\infty$ are pairwise non-isomorphic and strong orbit equivalent to $(X_\sigma, T_\sigma)$.
\end{theorem}
In the following theorem, given a Bratteli-Vershik system $(X_B, \varphi_B)$ on a stationary simple diagram, an orbit equivalent Bratteli-Vershik system is found with the least possible number of vertices. The vector technique developed in~\cite{S.B.O.K.} is used to prove the result. Let $\lambda$ be a Perron-Frobenius eigenvalue for the incidence matrix of $B$. Since the algebraic degree $\deg \lambda$ is equal to $k$, the dimension of the vector space corresponding to any other orbit equivalent Bratteli-Vershik system is at least $k$. Hence there is no stationary Bratteli-Vershik system with less than $k$ vertices which is orbit equivalent to $(X_B, \varphi_B)$.
\begin{theorem}[\cite{BK14}]\label{minBD}
Let $\sigma$ be a primitive substitution whose incidence matrix has a Perron-Frobenius eigenvalue $\lambda$ and $k = \deg \lambda$. Then $(X_\sigma, T_\sigma)$ is orbit equivalent to a Bratteli-Vershik system defined on a stationary Bratteli diagram with $k$ vertices on each level. Moreover, there is no stationary Bratteli-Vershik system with less than $k$ vertices which is orbit equivalent to $(X_\sigma, T_\sigma)$.
\end{theorem}
\section{Finite rank Bratteli diagrams}\label{finrank}
This section is devoted to the study of aperiodic Cantor dynamical systems which can be represented by Bratteli diagrams with uniformly bounded number of vertices on each level. It is an open question which classes of Cantor dynamical systems admit such a representation.
\begin{definition}
A Cantor dynamical system $(X,S)$ has the {\it topological rank} $K>0$ if it admits a Bratteli-Vershik model $(X_B,\varphi_B)$ such that the number of vertices of the diagram $B$ at each level is not greater than $K$ and $K$ is the least possible number of vertices for any Bratteli-Vershik realization.
\end{definition}
If a system $(X,S)$ has the rank $K$, then, by an appropriate telescoping, we can assume that the diagram $B$ has exactly $K$ vertices at each level. It is said that a homeomorphism $S: X\rightarrow X$ is {\it expansive} if there exists $\delta>0$ such that for any distinct $x,y\in X$ there is $m\in\mathbb Z$ with $d(S^mx,S^my)>\delta$. The number $\delta$ is called an {\it expansive constant}. Note that the notion of expansiveness does not depend on the choice of the metric $d$, see \cite{walters:book}. The series of papers is devoted to the characterization of Cantor minimal systems through expansive Bratteli diagrams, namely, expansive diagrams with constant number of incoming edges characterize Toeplitz systems~\cite{gjerde_johansen:2000}, expansive diagrams with a finite set of incidence matrices characterize linearly recurrent subshifts~\cite{CDHM03}. The main result in~\cite{DM} states that
\begin{theorem}[\cite{DM}]\label{minfinranCSexpansive}
Every Cantor minimal system of finite rank $d > 1$ is expansive.
\end{theorem}
In~\cite{DM}, the authors suggested a method of coding the
dynamics on Bratteli diagrams by means of the so-called $j$-symbols.
A new proof of Theorem \ref{minfinranCSexpansive} is given in \cite{H14}.
In~\cite{BKM09}, the ideas and results from~\cite{DM} are
generalized, and the abstract definitions of $j$-symbols, $j$-sequences
etc. are given. They can be used to study dynamics of different nature,
for instance, Bratteli-Vershik systems and substitution dynamical
systems. The main advantage of this approach is that it allows one to
use the machinery of symbolic dynamics for solving some problems of
Cantor dynamics.
The following result is an extension
of Theorem~\ref{minfinranCSexpansive} and was proved by using the
technique of $j$-sequences.
\begin{theorem}[\cite{BKM09}]\label{TheoremFiniteRank}
Let $(X,S)$ be an aperiodic Cantor dynamical system of finite rank. If the restriction of $(X,S)$ to every minimal component is not conjugate to an odometer, then $(X,S)$ is expansive.
\end{theorem}
In~\cite{BKMS_2013}, the structure of invariant measures on finite rank Bratteli diagrams is considered. In particular, it is shown that every ergodic invariant measure (finite or ``regular'' infinite) can be obtained as an extension from a simple vertex subdiagram.
\begin{definition}
Let $B$ be a Bratteli diagram. By a {\it vertex subdiagram} of $B$, we mean a Bratteli diagram $\ov B = (\ov V,\ov E)$ constructed by taking some vertices at each level $n$ of the diagram $B$ and then considering {\it all} the edges of $B$ that connect these vertices.
\end{definition}
Let $\ov B = (\ov V,\ov E)$ be a subdiagram of $B$. Consider the set $X_{\ov B}$ of all infinite paths of the subdiagram $\ov B$. Then the set $X_{\ov B}$ is naturally seen as a subset of $X_B$. Let $\ov \mu$ be a finite invariant (with respect to the tail equivalence relation $\mathcal R$) measure on $X_{\ov B}$. Let $\wh X_{\ov B}$ be the saturation of $X_{\ov B}$ with respect to $\mathcal R$. In other words, a path $x\in X_B$ belongs to $\wh X_{\ov B}$ if it is $\mathcal R$-equivalent to a path $y\in X_{\ov B}$. Then $\wh X_{\ov B}$ is $\mathcal R$-invariant and $X_B$ is a complete section for $\mathcal R$ on $\wh X_{\ov B}$. By the {\it extension of measure $\ov \mu$ to $\wh X_{\ov B}$} we mean the $\mathcal R$-invariant measure $\widehat {\ov \mu}$ on $\wh X_{\ov B}$ (finite or infinite) such that $\widehat {\ov\mu}$ induced on $X_{\ov B}$ coincides with $\ov \mu$. To extend the measure $\widehat {\ov\mu}$ to the $\mathcal R$-invariant measure on the whole space $X_B$, we set $\widehat {\ov\mu}( X_B \setminus \wh X_{\ov B}) = 0$.
In the next theorem, we describe the structure of the supports of ergodic invariant measures. Everywhere below the term ``measure'' stands for an $\mathcal R$-invariant measure. By an infinite measure we mean any $\sigma$-finite non-atomic measure which is finite (non-zero) on some clopen set. The support of each ergodic measure turns out to be the set of all paths that stabilize in some subdiagram, which geometrically can be seen as ``vertical'', i.e., they will eventually stay in the subdiagram. Furthermore, these subdiagrams are pairwise disjoint for different ergodic measures. It is shown in~\cite{BKMS_2013}, that for any finite rank diagram $B$ one can find finitely many vertex subdiagrams $B_\alpha$ such that each finite ergodic measure on $X_{B_\alpha}$ extends to a (finite or infinite) ergodic measure on $X_B$. It is also proved that each ergodic measure (both finite and infinite) on $X_B$ is obtained as an extension of a finite ergodic measure from some $X_{B_\alpha}$. Moreover, the following theorem holds:
\begin{theorem}[\cite{BKMS_2013}]\label{TheoremGeneralStructureOfMeasures}
Let $B$ be a Bratteli diagram of finite rank $d$.
The diagram $B$ can be telescoped in such a way that for every
probability ergodic measure $\mu$ there exists a subset $W_\mu$ of
vertices from $\{1,\ldots,d\}$ such that the support of $\mu$ consists of
all infinite paths that eventually go along the vertices of $W_\mu$
only. Furthermore,
(i) $W_\mu\cap W_\nu = \emptyset$ for different ergodic measures
$\mu$ and $\nu$;
(ii) given a probability ergodic measure $\mu$, there exists a constant $\delta>0$
such that for any $v\in W_\mu$ and any level $n$
$$\mu(X_v^{(n)})\geq \delta
$$
where $X_v^{(n)}$ is the set of all paths that go through the vertex $v$ at level $n$;
(iii) the subdiagram generated by $W_\mu$ is simple and uniquely
ergodic. The only ergodic measure on the path space of the subdiagram
is the restriction of measure $\mu$.
If a probability ergodic measure $\mu$ is the extension of a measure
from the vertical subdiagram determined by a proper subset $W\subset\{1,\ldots,d\}$, then
$$\lim_{n\to\infty}\mu(X_v^{(n)}) = 0\mbox{ for all }v\notin W.$$
\end{theorem}
We observe that condition (ii) can be used in practice to determine the support
of an ergodic measure $\mu$.
A number of sufficient conditions for unique ergodicity of a finite rank Bratteli-Vershik system are obtained in~\cite{BKMS_2013}. Here we present one of them.
\begin{definition}
(i) For two positive vectors $x,y\in\mathbb R^d$, the {\it projective metric (Hilbert metric)} is
$$
D(x,y) = \ln\max_{i,j}\frac{x_iy_j}{x_jy_i} = \ln \frac{\max_i\frac{x_i}{y_i}}{\min_j \frac{x_j}{y_j}},
$$
where $(x_i)$ and $(y_i)$ are entries of the vectors $x$ and $y$.
(ii) For a non-negative matrix $A$, the {\it Birkhoff contraction
coefficient} is $$\tau(A) = \sup_{x,y>0}\frac{D(Ax,Ay)}{D(x,y)}.$$
\end{definition}
\begin{theorem}[\cite{BKMS_2013}]\label{theoremUniqueErgodicityInTermsTau}
Let $B$ be a simple Bratteli diagram of finite rank
with incidence matrices $\{F_n\}_{n\geq 1}$. Let $A_n = F_n^T$.
Then the diagram $B$ is uniquely ergodic if and only if
$$\lim_{n\to
\infty}\tau(A_m\ldots A_n) = 0\mbox{ for every }m.
$$
\end{theorem}
For a positive matrix $A = (a_{i,j})$, set
$$\phi(A) = \min_{i,j,r,s} \frac{a_{i,j}a_{r,s}}{a_{r,j}a_{i,s}}.
$$
If $A$ has a zero entry, then, by definition, we put $\phi(A)=0$.
As noticed in \cite{hartfiel:book:2002},
$$\tau(A) = \frac{1-\sqrt{\phi(A)}}{1+\sqrt{\phi(A)}}$$
when $A$ has a nonzero entry in each row.
The following results give computable sufficient conditions for measure
uniqueness.
\begin{proposition}\label{PropositionSufficientConditionsUniqueErgodicity}
Let $\{A_n\}_{n\geq 1} = F^T_n$ be primitive incidence matrices of a finite rank diagram $B$.
(1) If $$\sum_{n = 1}^\infty \sqrt{\phi(A_n)} = \infty,$$
then $B$ admits a unique invariant probability measure.
(2) If
$$
\sum_{n = 1}^\infty \left(\frac{m_n}{M_n} \right)=\infty,
$$
where $m_n$ and $M_n$ are the smallest and the largest entry of
$A_n$ respectively, then $B$ admits a unique invariant probability
measure.
(3) If $||F_n||_1\leq Cn$ for some $C>0$ and all sufficiently
large $n$, then the diagram admits a unique invariant probability
measure\footnote{Here $||A||_1 = \sum_{i,j}|a_{i,j}|$.}. In particular, this result holds if the diagram has only finitely
many different incidence matrices.
\end{proposition}
\begin{example} Let $B$ be a simple Bratteli diagram with incidence matrices
$$
F_n = \left(
\begin{array}{cccc}
f_1^{(n)} & 1 & \cdots & 1 \\
1 & f_2^{(n)} & \cdots & 1 \\
\vdots & \vdots & \ddots & \vdots \\
1 & 1 & \cdots & f_d^{(n)} \\
\end{array}
\right).
$$
Let $q_n = \mbox{max}\{f_i^{(n)}f_j^{(n)} : i\neq j \}$. If for $A_n = F_n^T$
$$
\sum_{n=1}^\infty \sqrt{\phi(A_n)} = \sum_{n=1}^\infty
\frac{1}{\sqrt q_n} =\infty,
$$
then there is a unique invariant probability measure on $B$.
\end{example}
We say that a Bratteli diagram of a finite rank is of {\it exact finite rank} if there is a finite invariant measure $\mu$ and a constant $\delta >0$ such that after a telescoping $\mu(X_v^{(n)})\geq \delta$ for all levels $n$ and vertices $v$.
The following result shows that the Vershik map on the path space of an exact finite rank diagram cannot be strongly mixing independently of the ordering.
\begin{theorem}[\cite{BKMS_2013}]\label{TheoremAbsenceStrongMixing} Let $B = (V,E,\omega)$
be an ordered simple Bratteli diagram of exact finite
rank. Let $\varphi_{\omega}:X_B\rightarrow X_B$ be the Vershik map defined by the
order $\omega$ on $B$ ($\varphi_{\omega}$ is not necessarily continuous everywhere).
Then the dynamical system $(X_B,\mu,\varphi_{\omega})$ is not strongly mixing
with respect to the unique invariant measure $\mu$.
\end{theorem}
On the other hand, it is proved in the same paper that for the so-called ``consecutive'' ordering, the Vershik map is not strongly mixing on all finite rank diagrams.
\medskip
The last part of this section is devoted to the following problem. Let $B$ be a Bratteli diagram of finite rank $k$. It is known (see, e.g., \ref{TheoremGeneralStructureOfMeasures} that $B$ can support at most $k$ ergodic (finite and infinite) measures. Is it possible to determine under what conditions on the incidence matrices of $B$ there exist exactly $k$ ergodic measures? We give a criterion for the existence of $k$ measures when the incidence matrices satisfy the equal row sum property.
\begin{theorem}[\cite{ABKK}]
Let $B = (V,E)$ be a Bratteli diagram of rank $k \geq 2$; identify $V_n$ with $\{1, ... ,k\}$ for any $n \geq 1$. Let $F_n = (f_{i,j}^{(n)})$ form a sequence of incidence matrices of $B$ such that
$\sum_{j \in V_n} f_{i,j}^{(n)} = r_n \geq 2$ for every $i \in V_{n+1}$. Suppose that $\rank\ F_n = k$ for all $n$. Denote
$$
z^{(n)} =
\det
\begin{pmatrix}
\dfrac{f_{1,1}^{(n)}}{r_n} & \ldots & \dfrac{f_{1,{k-1}}^{(n)}}{r_n} & 1\\
\vdots & \ddots & \vdots & \vdots\\
\dfrac{f_{{k},1}^{(n)}}{r_n} & \ldots & \dfrac{f_{{k},{k-1}}^{(n)}}{r_n} & 1
\end{pmatrix}.
$$
Then there exist exactly $k$ ergodic invariant measures on $B$ if and only if
$$
\prod_{n = 1}^{\infty} |z^{(n)}| > 0,
$$
or, equivalently,
$$
\sum_{n = 1}^{\infty}(1 - |z^{(n)}|) < \infty.
$$
\end{theorem}
In the paper \cite{BDM10} the authors addressed the following question: how can combinatorial properties of Bratteli diagrams be used to describe continuous and measurable eigenvalues? They used the sequence of incidence matrices to produce necessary (algebraic) conditions for a number $\lambda$ to be an eigenvalue that is both continuous and measurable. It is remarkable that the necessary conditions do not depend on an order of the diagram (that specifies a Vershik map). These results were applied to Toeplitz minimal systems.
\section{Measures and subdiagrams}\label{sbd}
In this section, we consider ergodic $\mathcal R$-invariant measures on arbitrary Bratteli diagrams related to subdiagrams. Suppose that we have a Bratteli diagram $B$ and an ergodic measure $\mu$. It is still an open question whether one can explicitly describe the support of $\mu$ on $X_B$ in terms of the diagram $B$. It would be nice to have a statement similar to \ref{TheoremGeneralStructureOfMeasures}. We give answers to the following questions:
(A) Given a subdiagram $B'$ of $B$ and an ergodic measure $\mu$ on $X_B$, under what conditions on $B'$ the subset $X_{B'}$ has positive measure $\mu$ in $X_B$?
(B) Let $\nu$ be a measure supported by the path space $X_{B'}$ of a subdiagram $B' \subset B$. Then $\nu$ is extended to the subset $\mathcal R(X_{B'})$ by invariance with respect to the tail equivalence relation $\mathcal R$. Under what conditions $\nu(\mathcal R(X_{B'}))$ is finite (or infinite)?
By a Bratteli subdiagram, we mean a Bratteli diagram $B'$ that can be obtained from $B$ by removing some vertices and edges from each level of $B$. Then $X_{B'}\subset X_B$. We will consider two extreme cases of Bratteli subdiagrams: vertex subdiagram (when we fix a subset of vertices at each level and take all edges between them) and edge subdiagram (some edges are removed from the initial Bratteli diagram but the vertices are not changed). It is clear that an arbitrary subdiagram can be obtained as a combination of these cases.
We keep the following notation:
$\overline{X}_v^{(n)}$ stands for the tower in a subdiagram $\overline{B}$ that is determined by a vertex $v$ of $\ol B$. Thus, we consider the paths in $\overline{X}_v^{(n)}$ that contain edges from $\overline{B}$ only. Let $\overline{h}_v^{(n)}$ be the height of the tower $\ol X_v^{(n)}$. As a rule, objects related to a subdiagram $\ol B$ are denoted by barred symbols.
The following theorem gives criteria for finiteness of the measure extension.
\begin{theorem}[\cite{BKK_14}]\label{thm from BKK}
Let $B$ be a Bratteli diagram with the sequence of incidence matrices $\{F_n\}_{n=0}^\infty$, and let $\ov B$ be a vertex subdiagram of $B$ defined by the sequence of subsets $\{W_n\}_{n = 0}^\infty, \ W_n \subset V_n$. Suppose that $\ov \mu$ is a probability $\ol{\mathcal R}$-invariant measure on $X_{\ov B}$. Then the following properties are equivalent:
\begin{eqnarray*}
\widehat{\overline{\mu}}(\wh X_{\ov B}) < \infty & \Longleftrightarrow & \sum_{n=1}^\infty \sum_{v \in W_{n+1}} \sum_{w \in W'_n} f_{v,w}^{(n)} h_w^{(n)} \ov p_v^{(n+1)} < \infty\\
& \Longleftrightarrow & \sum_{n=1}^{\infty} \sum_{w \in W_{n+1}} \wh{\ov\mu}(X_{w}^{(n+1)})\sum_{v \in W^{'}_{n}} q_{w,v}^{(n)}
< \infty\\
& \Longleftrightarrow & \sum_{i=1}^{\infty} \left(\sum_{w\in W_{i+1}} h_w^{(i+1)} \ov p_w^{(i+1)} - \sum_{w\in W_{i}} h_w^{(i)} \ov p_w^{(i)}\right)< \infty.
\end{eqnarray*}
\end{theorem}
If $(F_n)$ is a sequence of incidence matrices of a Bratteli diagram $B$, then we can also define the sequence of stochastic matrices $(Q_n)$ with entries
$$
q^{(n)}_{v,w} = f^{(n)}_{v,w}\frac{h_w^{(n)}}{h_v^{(n+1)}},\ \ v\in V_{n+1},\ w\in V_{n}.
$$
Paper \cite{BKK_14} contains also some necessary and sufficient conditions for finiteness of the measure extension. For instance, it is shown that if
\begin{equation*}\label{suffic}
\sum_{n=1}^{\infty} \sum_{v \in W_{n+1}} \sum_{w \in W^{'}_{n}} q_{v,w}^{(n)} < \infty,
\end{equation*}
then any probability measure $\ov \mu$ defined on the path space $X_{\ol B}$ of the vertex subdiagram $\ol B$ extends to a finite measure $\widehat{\ov \mu}$ on $\wh X_{\ol B}$.
The analogue of Theorem~\ref{thm from BKK} can be proved also for edge subdiagrams:
\begin{theorem}[\cite{ABKK}]\label{edge_fin_krit}
Let $\overline{B}$ be an edge subdiagram of a Bratteli diagram $B$. For a probability invariant measure $\overline{\mu}$ on $X_{\overline{B}}$, the extension $\widehat{\overline{\mu}} (\widehat{X}_{\overline{B}})$ is finite if and only if
$$
\sum_{n=1}^{\infty} \sum_{v \in V_{n+1}} \sum_{w \in V_{n}} \widetilde{f}_{v,w}^{(n)} h_w^{(n)}\overline{p}_v^{(n+1)} < \infty
$$
where $\widetilde{f}_{v,w} = {f}_{v,w} - \ol{f}_{v,w}$.
\end{theorem}
The following theorem gives a necessary and sufficient condition for a subdiagram $\ol B$ of $B$ to have a path space of zero measure in $X_B$. Though the theorem is formulated for a vertex subdiagram, the statement remains true also for any edge subdiagram $\ol B$.
\begin{theorem}[\cite{ABKK}]\label{main}
Let $B$ be a simple Bratteli diagram, and let $\mu$ be any probability ergodic measure on $X_B$. Suppose that $\ov B$ is a vertex subdiagram of $B$ defined by a sequence $(W_n)$ of subsets of $V_n$. Then $\mu(X_{\ov B}) = 0$ if and only if
\begin{equation}\label{thin set condition}
\forall \varepsilon > 0 \ \exists n = n(\varepsilon) \ \mathrm{such\ that} \ \forall w \in W_n \ \mathrm{one\ has} \ \frac{\ov h_w^{(n)}}{h_w^{(n)}} < \varepsilon.
\end{equation}
\end{theorem}
In fact, Theorem \ref{main} states that if a subdiagram $\ol B$ satisfies (\ref{thin set condition}), then $X_{\ol B}$ has measure zero with respect to every ergodic invariant measure, that is the set $X_{\ol B}$ is {\em thin} according to the definition from \cite{giordano_putnam_skau:2004}. The following result is a corollary of Theorem \ref{main}:
\begin{theorem}[\cite{ABKK}]\label{corol_infty}
Let $\ol B$ be a subdiagram of $B$ such that $X_{\ol B}$ is a thin subset of $X_B$. Then for any probability invariant measure $\ov \mu$ on $\ov B$ we have $\widehat{\overline{\mu}}(\wh X_{\ov B}) = \infty$.
\end{theorem}
\begin{remark}
There are a lot of papers, where invariant measures for various Bratteli diagrams are studied. For instance, in~\cite{Frick_Petersen, PeVa} the authors consider ergodic invariant probability measures on a Bratteli diagram of a special form, called an Euler graph; the combinatorial properties of the Euler graph are connected to those of Eulerian numbers. The authors of~\cite{FrO} study spaces of invariant measures for a class of dynamical systems which is called polynomial odometers. These are adic maps on regularly structured Bratteli diagrams and include the Pascal and Stirling adic maps as examples. K. Petersen~\cite{Petersen} considers ergodic invariant measures on a Bratteli-Vershik dynamical system, which is based on a diagram whose path counts from the root are the Delannoy numbers.
We would like to mention also the interesting paper by Fisher \cite{F09} where various properties of Bratteli diagrams and measures are discussed.
\end{remark}
\medbreak
{\bf Acknowledgement.} We are thankful to our colleagues and co-authors for numerous useful discussions of the concept of Bratteli diagrams. S.B. would like to thank the University of Iowa for the warm hospitality where the paper was finished.
\bibliographystyle{amsalpha}
|
\section{Introduction}
The principle of maximum entropy (MaxEnt for short) is considered a fundamental tool in the
application of information theory to the construction of probabilistic models~\cite{Presse2013}.
First proposed in the context of Statistical Mechanics by J. W. Gibbs~\cite{Gibbs1902} as a derivation
of the canonical ensemble from the requirement of fixed average internal energy, it was recast by
Jaynes~\cite{Jaynes1957} as a \emph{general principle of reasoning} following
the insights of Shannon~\cite{Shannon1948}.
This principle of reasoning under uncertainty, that we call MaxEnt, postulates that the most unbiased
probability assignment $\rho^*(\bm u)$ among those probability densities $\rho(\bm u)$ consistent
with given information $E$, is the one that maximizes the universal \textbf{entropy} functional
\begin{equation}
\mathcal{S}[\rho; \rho_0] = -\int_U d^n u\; \rho(\bm u)\ln \frac{\rho(\bm u)}{\rho_0(\bm u)},
\label{eq_shannon}
\end{equation}
under the constraints imposed by $E$. Here $U$ is an $n$-dimensional state space and $\bm{u} \in U$
a possible state of the system. The density $\rho_0(\bm{u})$ is the initial probability density
representing complete ignorance about the value of $\bm{u}$.
The maximization problem is solved using the method of Lagrange multipliers, and for the case of information
$E$ expressed as the given expectation $\Big<f(\bm{u})\Big>=F$ the solution is well-known~\cite{Jaynes1957,Jaynes2003},
and corresponds to the standard \emph{exponential family} of models,
\begin{equation}
\rho^*(\bm{u}) = \frac{1}{Z(\lambda)}\rho_0(\bm{u})\exp(-\lambda f(\bm{u}))
\label{eq_canonical}
\end{equation}
where $\lambda$ is the Lagrange multiplier which can be determined from the implicit equation
\begin{equation}
-\frac{\partial}{\partial \lambda}\ln Z(\lambda)=F.
\label{eq_lambda}
\end{equation}
The uniqueness of the MaxEnt procedure relies on the uniqueness of the entropy functional $\mathcal{S}$
of Eq. \ref{eq_shannon}, and this has been established from consistency requirements~\cite{Shore1980, Skilling1988,
Garrett2001, Caticha2006}. These axiomatic approaches have allowed a useful separation between the operational
formalism of MaxEnt updating (in the sense of formulating an optimal inference rule) and the meaning of the
entropy $\mathcal{S}$ being maximized.
In the purely operational vision, entropy is just a convenient device which allows the ranking
of the different candidate distributions from best to worst, and we do not need to assign any meaning to it.
However, this ranking must, at some point, be understood in terms of some quality of the candidate models in
terms of which they are rejected or accepted. According to the standard interpretation of the entropy $\mathcal{S}$ by
Shannon (later adopted by Jaynes), the quantity $\mathcal{S}$ is a measure of the \emph{missing information}
needed to completely determine the state $\bm{u}$. In maximizing $\mathcal{S}$ we are choosing the least informative
model that agrees with all our constraints.
The original entropy axioms by Shannon~\cite{Shannon1948} as well as the Shore and Johnson axioms~\cite{Shore1980}
have been recenly put into question~\cite{Tsallis2015} because they rule out the so called generalized entropies
(such as the Tsallis or Renyi entropies) for use in inference~\cite{Presse2013b}. In this ongoing development it
would be highly desirable to have alternative methods for updating probabilities, and an obvious candidate is Bayes'
theorem itself. Both MaxEnt and Bayes' theorem solve essentially the same problem, namely updating a prior to a
posterior incorporating some new information or evidence, regardless of the nature of that evidence (measured data
or model constraints). Looking closely, we can in fact recognize the exponential form given in Eq.~\ref{eq_canonical}
as a Bayesian updating from a \emph{prior} distribution $\rho_0(\bm{u})$ to a posterior distribution
$\rho^*(\bm{u})=M(\bm{u}) \rho_0(\bm{u})$ with an ``updating factor''
\begin{equation}
M(\bm{u}) \propto \exp\big(-\sum_i \lambda_i f_i(\bm{u})\big).
\label{eq_M_factor}
\end{equation}
If we can arrive at this form of $M$ without invoking a maximization of some functional, then we can effectively
bypass the issue of adequacy of $\mathcal{S}$. In a frequentist context (taking expectations as the limit of statistical
averages) Campenhout and Cover~\cite{Campenhout1981} have produced this factor $M$ in the limit of infinite samples.
In this work we provide a general derivation of the update rule in Eq. \ref{eq_canonical} from Bayesian inference
subjected to expectation constraints. Unlike the result by Campenhout and Cover, the proof does not depend on frequentist
assumptions such as the identification of expectations with averages over samples, or even the assumption of data samples being processed.
It is based simply on imposing consistency conditions (in similar spirit to Cox~\cite{Cox1961} and Shore and Johnson) on the Bayesian updating
rules themselves to constrain the form of $M$.
The work is organized as follows. First, in section \ref{sec_notation}, we present the notation to be used in
the rest of the paper and give an outline of the assumptions needed for the central claim of the paper.
In sections \ref{sec_repre}, \ref{sec_cvt} and \ref{sec_indep}, the exponential form is obtained from the assumptions.
\section{Mathematical notation and statement of the derivation}
\label{sec_notation}
We will consider in the following a system described by states $\bm{u}$ in an
$n$-dimensional state space $U$. We start from an arbitrary state of knowledge
$I_0$ with probability density $P(\bm{u}|I_0)$ and our goal is to perform a
Bayesian update to a posterior density $P(\bm{u}|I)$ where $I=I_0\wedge E$ is the new
state of knowledge that includes the evidence $E$. We will denote the expectation of a
quantity $A(\bm{u})$ under the state of knowledge $J$ by $\big<A\big>_J$,
defined as\footnotemark
\begin{equation}
\Big<A\Big>_J = \int_U d^nu\; P(\bm{u}|J) A(\bm{u}).
\end{equation}
\footnotetext{
Note that, if $A$ is a scalar (i.e. invariant under a change of coordinate
system) then $\big<A\big>_J$ is also a scalar, because $P(\bm{u}|J)$ is a scalar
density, which transforms just as the invariant measure $\sqrt{g(\bm{u})}$.
Therefore we do not include $\sqrt{g}$ explicitly in the integral.}
For the problem of inference with prior $P(\bm{u}|I_0)$ and evidence $E$ given
by $\big<f(\bm{u})\big>_I=F$, we define the following \textbf{consistency conditions}:
\vspace{10pt}
\begin{enumerate}
\item [(a)] The probability ratio $P(E|\bm{u},I_0)/P(E|I_0)$ is a unique functional of the constraining
function $f$, evaluated using the state $\bm{u}$ and the constraining value $F$. That is, we can write
\begin{equation}
\frac{P(E|\bm{u}, I_0)}{P(E|I_0)} = M[f](\bm{u}, F).
\end{equation}
The functional $M$ encodes the method of inference, and we are looking for a unique method. The only
information we have in using this method is the evidence $E$, which consists of the function $f$ and
its expectation $F$, so $M$ cannot depend on any other piece of information, such as additional parameters.
\vspace{10pt}
\item [(b)] The inference considering $f$ as a function of $\bm{u}$ is
consistent with the inference considering $f$ as a fundamental variable. In
other words, it should be possible to ignore the degrees of freedom $\bm{u}$ and perform an
inference over $f$ itself, using the same functional $M$. This inference has to
be valid and consistent with the full inference using $\bm{u}$.
\vspace{10pt}
\item [(c)] Logically independent subsystems $U_1$ and $U_2$ can be analyzed separately or
jointly as $U=U_1\otimes U_2$, producing the same result. This \textbf{does not} restrict the method
to separable systems only, it only ensures that \textbf{if} we decide to apply it to a pair of independent
systems, the method should preserve their independence. The same functional $M$ must be valid for systems
of arbitrary correlation.
\end{enumerate}
In the following we will show that these conditions imply a posterior
\begin{equation}
P(\bm{u}|E,I_0) = \frac{1}{Z(\lambda)}\exp(-\lambda f(\bm{u}))P(\bm{u}|I_0).
\label{eq_expform}
\end{equation}
Here $\lambda$ is a parameter to be fixed by the value of $F$, not a Lagrange multiplier, as there
is no variational procedure in the derivation.
The proof is divided in two parts. First, we prove that conditions (a) and (b)
imply that the functional $M[f](\bm{u}, F)$ is actually a function of two arguments, $m(f(\bm{u}), F)$.
Finally, we show that condition (c) implies $m(f, F) \propto \exp(-\lambda(F)f)$, which immediately
leads to Eq. \ref{eq_expform} after normalization.
\section{Consistency between different representations}
\label{sec_repre}
The evidence $E$, consisting of the given expectation $\big<f\big>_I=F$, can be
used in two different ways. If we regard the quantity $f$ as a variable in
itself, it can be used to update the prior $P(f|I_0)$ to a posterior $P(f|I)$, given by
\begin{equation}
P(f|I) = P(f|I_0)M[\mathbb{I}](f, F).
\label{eq_prob_f_1}
\end{equation}
Here the functional depends on the identity function $f \rightarrow f$, denoted by $\mathbb{I}$.
On the other hand, we can take $f$ as a function of the degrees of freedom $\bm{u}$, and use the
evidence $E$ to update the prior $P(\bm{u}|I_0)$ to a posterior $P(\bm{u}|I)$,
\begin{equation}
P(\bm{u}|I) = P(\bm{u}|I_0)M[f](\bm{u}, F).
\label{eq_prob_f_2}
\end{equation}
For both problems the same functional $M$ should be used and yield consistent
posteriors. Through the laws of probability, the probability density of $f$ is
always connected to the probability density of $\bm{u}$ by
\begin{equation}
P(f|J) = \Big<\delta(f(\bm{u})-f)\Big>_J,
\end{equation}
for every state of knowledge $J$, which in our case produces two independent relations,
for $J=I$ and $J=I_0$, namely
\begin{eqnarray}
P(f|I) = \int d^nu\; P(\bm{u}|I)\delta(f(\bm{u})-f), \\
P(f|I_0) = \int d^nu\; P(\bm{u}|I_0)\delta(f(\bm{u})-f).
\label{eq_bar}
\end{eqnarray}
By replacing Eqs. \ref{eq_prob_f_1} and \ref{eq_prob_f_2} we find a constraint for $M$,
namely
\begin{equation}
\int d^nu\; P(\bm{u}|I_0)\delta(f(\bm{u})-f)\Big[M[f](\bm{u}, F)-M[\mathbb{I}](f, F)\Big] = 0
\end{equation}
for any prior $P(\bm{u}|I_0)$. The delta function allows us to replace $f$ by
$f(\bm{u})$ in the arguments to $M[\mathbb{I}]$, so we have
\begin{equation}
\int d^nu\; P(\bm{u}|I_0)\delta(f(\bm{u})-f)\Big[M[f](\bm{u}, F)-M[\mathbb{I}](f(\bm{u}), F)\Big] = 0.
\end{equation}
Then, taking the functional derivative on both sides with respect to the prior, we have
\begin{equation}
M[f](\bm{u}, F) = M[\mathbb{I}](f(\bm{u}), F) = m(f(\bm{u}), F).
\end{equation}
This means that for any constraining function $f$ the functional $M$ reduces to a function of
$f(\bm{u})$ (the constraining function evaluated at the state $\bm{u}$) and $F$ (the constraining value);
it does not, for instance, depend on derivatives of $f$. We have then constrained the form of the posterior
distribution to be
\begin{equation}
P(\bm{u}|I) = P(\bm{u}|I_0)m(f(\bm{u}), F)
\label{eq_posterior_m}
\end{equation}
with $m$ a unique function, to be determined.
\section{A theorem for posterior expectations}
\label{sec_cvt}
In order to find this universal function $m(f, F)$ we first present a useful
identity between expectations. Applying Stoke's theorem to an arbitrary probability
density $P(\bm{u}|J)$ we obtain (see the Appendix) the following expectation identity,
\begin{equation}
\Big<\nabla \cdot \bm{v}\Big>_J + \Big<\bm{v}\cdot\nabla \ln P(\bm{u}|J)\Big>_J = 0,
\end{equation}
valid for any differentiable vector function $\bm{v}(\bm{u})$. This is a generalization of the result in Ref.~\cite{Davis2012}.
Replacing the form of the posterior found in Eq. \ref{eq_posterior_m}, we have,
for the state of knowledge $I$,
\begin{equation}
\Big<\nabla \cdot \bm{v}\Big>_I + \Big<\bm{v}\cdot\nabla \ln
P(\bm{u}|I_0)\Big>_I = -\Big<\bm{v}\cdot \nabla \ln m(f(\bm{u}), F)\Big>_I,
\end{equation}
or, more compactly,
\begin{equation}
\Big<\mathcal{D}_0 \bm{v}\Big>_I = \Big<\hat{\lambda}(f(\bm{u}), F)\bm{v}\cdot\nabla f\Big>_I,
\end{equation}
where the operator $\mathcal{D}_0$ is defined as $\mathcal{D}_0 \bm{v}=\nabla \cdot \bm{v}+\bm{v}\cdot\nabla \ln P(\bm{u}|I_0)$
and
\begin{equation}
\hat{\lambda}(f, F)= -\frac{\partial}{\partial f}\ln m(f, F).
\label{eq_hat_lambda}
\end{equation}
In the next section we will constrain the form of this function $\hat{\lambda}$
using the condition (c) on independent systems. We will show that $\hat{\lambda}$ only
depends on its second argument $F$.
\section{Consistency between separate and joint treatment of independent subsystems}
\label{sec_indep}
Now let us consider the following situation: two logically independent systems, $U_1$ and $U_2$
with priors $P(\bm{u}_1|I_0)$ and $P(\bm{u}_2|I_0)$ respectively. We can decide
to update these priors separately using the evidence $E$ given by
\begin{eqnarray}
\Big<f_1(\bm{u}_1)\Big>_I = F_1 \label{eq_constraint1}, \\
\Big<f_2(\bm{u}_2)\Big>_I = F_2.
\label{eq_constraint2}
\end{eqnarray}
We could also decide to update the joint system prior $P(\bm{u}|I_0)=P(\bm{u}_1|I_0)P(\bm{u}_2|I_0)$,
using the evidence
\begin{equation}
\Big<f_1(\bm{u}_1)+f_2(\bm{u}_2)\Big>_I = F_1+F_2.
\label{eq_constraint_combined}
\end{equation}
In both cases we are forced, by condition (a), to use Eq. \ref{eq_posterior_m}
with the same function $m$ on each case. From Eqs. \ref{eq_constraint1} and \ref{eq_constraint2}
applied separately on $\bm{u}_1$ and $\bm{u}_2$, we obtain
\begin{eqnarray}
\Big<\mathcal{D}_0 \bm{v}_1(\bm{u}_1)\Big>_I = \Big<\hat{\lambda}(f_1(\bm{u}_1), F_1)\bm{v}_1\cdot\nabla f_1\Big>_I, \\
\Big<\mathcal{D}_0 \bm{v}_2(\bm{u}_2)\Big>_I = \Big<\hat{\lambda}(f_2(\bm{u}_2), F_2)\bm{v}_2\cdot\nabla f_2\Big>_I,
\end{eqnarray}
and from Eq. \ref{eq_constraint_combined},
\begin{equation}
\Big<\mathcal{D}_0 \bm{v}(\bm{u})\Big>_I = \Big<\hat{\lambda}(f_1(\bm{u}_1)+f_2(\bm{u}_2), F_1+F_2)\bm{v}\cdot\nabla f\Big>_I.
\end{equation}
Choosing first $\bm{v}(\bm{u}_1)=\bm{v}_1$ and then $\bm{v}(\bm{u}_2)=\bm{v}_2$ we find (calling $f_{12}=f_1+f_2$),
\begin{eqnarray}
\Big<\hat{\lambda}(f_{12}, F_1+F_2)\bm{v}_1\cdot\nabla f_1\Big>_I
= \Big<\hat{\lambda}(f_1, F_1)\bm{v}_1\cdot\nabla f_1\Big>_I, \\
\Big<\hat{\lambda}(f_{12}, F_1+F_2)\bm{v}_2\cdot\nabla f_2\Big>_I
= \Big<\hat{\lambda}(f_2, F_2)\bm{v}_2\cdot\nabla f_2\Big>_I,
\end{eqnarray}
for every choice of $\bm{v}_1$ and $\bm{v}_2$. This implies
\begin{equation}
\hat{\lambda}(f_{12}, F_1+F_2) = \hat{\lambda}(f_1, F_1) = \hat{\lambda}(f_2, F_2).
\end{equation}
As $f_1$ only depends on $\bm{u}_1$, and $f_2$ only depends on $\bm{u}_2$, the last
equality means that, in general, the function $\hat{\lambda}(f, F)$ does not
depend on $f$; it is only a function of the second argument $F$, so that we can
replace $\hat{\lambda(f(\bm{u}), F)}$ by $\lambda(F)$ such that
\begin{equation}
\lambda(F_1+F_2) = \lambda(F_1) = \lambda(F_2).
\end{equation}
Replacing $\hat{\lambda}(f(\bm{u}), F)$ by $\lambda(F)$ into the definition of $\hat{\lambda}$
(Eq. \ref{eq_hat_lambda}), we have
\begin{equation}
-\frac{\partial}{\partial f}\ln m(f, F) = \lambda(F),
\end{equation}
and therefore,
\begin{equation}
m(f, F) = m_0(F)\exp(-\lambda(F)f).
\end{equation}
The posterior distribution for $\bm{u}$ in Eq. \ref{eq_posterior_m} then reads,
\begin{equation}
P(\bm{u}|I) = m_0(F)\exp(-\lambda(F)f(\bm{u}))P(\bm{u}|I_0),
\end{equation}
with $m_0(F)$ fixed by normalization to be
\begin{equation}
\frac{1}{m_0(F)} = \int d^nu\; \exp(-\lambda(F)f(\bm{u}))P(\bm{u}|I_0) = Z(\lambda(F)).
\end{equation}
As all the dependence on $F$ is through $\lambda(F)$, we can finally write the
distribution entirely as a function of $\lambda$, so
\begin{equation}
P(\bm{u}|I) = \frac{1}{Z(\lambda)}\exp(-\lambda f(\bm{u}))P(\bm{u}|I_0),
\end{equation}
with $\lambda=\lambda(F)$ a number to be determined to agree with the constraint
$\big<f(\bm{u})\big>_I = F$. Given this functional form of $P(\bm{u}|I)$ we can
write the expectation of $f$ as a derivative of $\ln Z(\lambda)$,
\begin{equation}
\Big<f(\bm{u})\Big>_I = -\frac{\partial}{\partial \lambda}\ln Z(\lambda)
\end{equation}
and then it follows that the constraint fixes $\lambda$ through Eq. \ref{eq_lambda}, as expected.
\section{Conclusion}
We have proved that Bayesian updating given an expectation constraint can be established as a uniquely
defined procedure, leading to an exponential family posterior density, which is the same result produced
by the application of the principle of maximum entropy (MaxEnt) under the same constraints.
We only require that a unique ``updating factor'' is used in all cases, and that its use is consistent
through different definitions of the state space. That the same answer is revealed using an alternative method
of inference shows that the essence of MaxEnt is already contained in Bayes' theorem under our additional
consistency requirements. This allows a conceptual unification of the fields of Bayesian inference and MaxEnt
inference under a common framework for continuous degrees of freedom. Our derivation also shows that the
existence of an entropy functional is not central to the core of inference; although it certainly provides
a more than convenient device for practical calculations, it is conceptually not required for a consistent
formulation of a theory of inference.
|
\section{Introduction}
After the pioneering $\text{C}_{60}${}+$\text{C}_{60}${} collision experiment of E.E.B. Campbell et.al. in 1993~\cite{Campbell93},
cluster-cluster collisions became a versatile new field of research
(for reviews see~\cite{Schmidt1995,Jellinek99,Campbell00_coll,Campbell2003}) with lasting interest~\cite{Schmidt91,Seifert91,Knospe1993,Schmidt1993,Schmidt92,Schmidt94,Schulte95,Zhang1994,Shen95,Farizon1999,Braeuning03,Rogan2004,Kamalou2006,Alamanova2007,Jakowski2012}.
In particular, fusion between fullerenes has been studied in great detail, both experimentally and theoretically~\cite{Campbell93,Yeretzian92,Strout93,Zhang93,Kim1994,Robertson95,Xia96,Rohmund96,Knospe96,Rohmund96_jpb,Xia97,Glotov01,Campbell2002,Zhao2002,Han2004,Kaur2008,Jakowski2010,Zettergren2014}.
Fusion is a universal phenomenon in collisions between complex particles
covering many orders in size and energy from heavy ion collisions in nuclear physics~\cite{Bock1980} to macroscopic liquid droplets~\cite{Brenn1989,Menchahca1997} or even colliding galaxies~\cite{Struck1999}.
It is a great challenge of ongoing interest to reveal universal similarities and basic differences of these mechanisms.
Usually, the gross features of fusion can be understood with macroscopic arguments~\cite{Schmidt92,Bock1980,Menchahca1997,Hasse1988} leading to the general expression for the fusion cross section $\sigma$ as function of the center mass energy $E_\text{c.m.} \equiv E$ of
\begin{equation}
\sigma(E)= \pi R_{\scriptscriptstyle 12}^2 \left( 1- \frac{V_B}{E} \right) \label{eq:sigma}
\end{equation}
with $R_{\scriptscriptstyle 12} = R_{\scriptscriptstyle 1}+R_{\scriptscriptstyle 2}$ the sum of the radii of the colliding partners and $V_B$ the fusion barrier (for a derivation see e.g.~\cite{Jellinek99,Bock1980,Levine1987}).
This formula (known as ``critical distance model'' in nuclear physics~\cite{Bock1980} or ``absorbing sphere model'' in chemistry~\cite{Levine1987}) describes \emph{quantitatively} the experimental fusion cross section for atomic nuclei (with $V_B > 0$) and liquid droplets (with $V_B = 0$)~\cite{Schmidt92,Brenn1989,Menchahca1997} in the low energy range with $E \gtrsim V_B$~\footnote{We do not consider here the high energy range, where the cross sections generally decreases with $E_\text{c.m.}$, see~\cite{Schmidt92}.}.
It is expected to hold also for collisions between metallic clusters~\cite{Schmidt92} (even with
$V_B< 0$); see last paper in the series~\cite{Schmidt91,Seifert91,Knospe1993,Schmidt1993}.
The physics behind formula~(\ref{eq:sigma}) is indeed simple: fusion takes place, if the colliding partners touch, owing to the larger binding energy of the fused compound~\cite{Hasse1988}.
For colliding fullerenes one naturally expects a fusion barrier $V_B$ of at least the $sp^2$ bond
breaking energy, or more pertinently the Stone-Wales transformation energy~\cite{Zhao2002,Han2004} of a few~eV~\cite{Dresselhaus} and, according to~(\ref{eq:sigma}),
a fusion cross section of the order of the geometrical one, $\pi R_{\scriptscriptstyle 12}^2 \sim 150$~\AA{}$^2$~\cite{Rohmund96_jpb}.
Experimentally, however, the fusion barriers are about one order of magnitude larger (around 80~eV)~\cite{Rohmund96_jpb} and the cross sections are even two magnitudes smaller (a few \AA{}$^2$)~\cite{Glotov01}.
Up to now, there is no definite explanation for these findings, albeit some possible phenomenological
reasons have been discussed~\cite{Jellinek99,Campbell00_coll,Campbell2003,Glotov01,Campbell2002}.
In addition, previous (at that time still approximate) microscopic Quantum Molecular Dynamics (QMD) calculations predicted the large fusion barriers~\cite{Knospe96}.
From these studies it is also well known that only very few mutual initial orientations of the colliding cages lead to fusion (without identifying them).
Anyway, fullerenes typically do {\bf not} fuse if they touch, even at high impact energies, and the question remains, why?
In this work, we provide a clear answer to this longstanding question.
\section{Microscopic results}
Motivated by our recent findings of the dominating role of the $A_g(1)$ breathing mode in $\text{C}_{60}${}-laser interaction~\cite{Laarmann07,Fischer2013}, we have reanalyzed fullerene-fullerene collisions with the help of the \emph{ab initio} nonadiabatic quantum molecular (NA-QMD) method~\cite{1Kunert03,Fischer2014_1,Fischer2014_2,Fischer2014_3}.
For systems as large as we are investigating here, NA-QMD is numerically more efficient than
its \emph{ab initio} QMD approximation~\cite{Fischer2014_1,Fischer2014_2,Fischer2014_3}.
In extension to previous studies~\cite{Jellinek99,Rohmund96,Knospe96,Rohmund96_jpb,Glotov01} we include a normal mode analysis~\cite{Zhang04} of the vibrational kinetic energy.
This method decomposes the total kinetic vibrational excitation energy $E_\text{vib}$ into the individual contributions of all 174 eigenmodes of C$_{60}$ as function of time $t$ according
to
\begin{equation*}
E_\text{vib}(t) = \sum_{i=1}^{60} \frac{m_c}{2} \dot{\vect{r}}_i^2 = \sum_{\nu=1}^{174}\frac{m_c}{2} \left( \sum_{i=1}^{60} \dot{\vect{r}}_i \vect{b}_{i\nu} \right )^2
\end{equation*}
with the atomic carbon mass $m_c$, the atomic velocities $\dot{\vect{r}}_i$ (in the molecular center of mass system without rotational components) and the normal mode eigenvectors $\vect{b}_{i\nu}$.
In fig.~\ref{fig:vib_analysis}, such an analysis, is shown for two central $\text{C}_{60}${}+$\text{C}_{60}${} collisions with the same impact energy of $E_\text{c.m.} = 104$~eV but for different initial orientations
of the clusters, leading in one case to scattering (fig.~\ref{fig:vib_analysis}(a)) and, in the other, to fusion (fig.~\ref{fig:vib_analysis}(b)).
In both cases, the extraordinary dominance of the $H_g(1)$ mode is obvious.
\begin{figure}
\includegraphics[width=0.49\columnwidth]{fig1a.eps}
\includegraphics[width=0.49\columnwidth]{fig1b.eps}
\caption{(color online) Normal mode analysis of central $\text{C}_{60}${}+$\text{C}_{60}${} collisions at an impact energy of $E_\text{c.m.}=104$~eV.
Shown are the vibrational kinetic energies $E_\text{vib}$ of all 174 internal normal modes as function of time $t$ for (a) typical scattering and (b) typical fusion events.
Distances of closest approach $R^\text{ret}$ are indicated by vertical, dashed lines.
The dominating role of the $H_g(1)$ mode (black thick lines) is apparent.
\label{fig:vib_analysis}}
\end{figure}
During approach, this mode is very strongly excited to a much higher degree than any other vibrational mode.
Its excitation energy of several eV is huge as compared to a single quantum of this mode of $\hbar \omega_\text{mode} \sim 34$~meV~\cite{Dong93} and, thus, its amplitude
is extremely large, as compared to a typical displacement of the elementary excitation (``giant'' $H_g(1)$ mode).
Consequently
at the distance of closest approach $R^\text{ret}$, a highly deformed \emph{oblate-oblate} configuration of the double cluster system is formed (see above illustrations in fig.~\ref{fig:vib_analysis}).
This clearly distinguishable state accommodates practically the whole impact energy into deformation (potential) energy, which is quantitatively shown in fig.~\ref{fig:qmd}(a), also for other impact energies.
Up to this stage of the collision, there is no appreciable difference to the other collision systems (nuclei, droplets), where at this ``critical distance'' the system loses immediately its memory and
the energy is dissipated into internal degrees of freedom (DOF) leading to a hot compound.
The fundamental difference to the other systems consists in the specific properties of the oblate-prolate mode in fullerenes and its special role in collisions.
First of all, among all vibrational modes, the $H_g(1)$ mode in $\text{C}_{60}${} has the largest oscillation period of $T = 122$~fs~\cite{Dong93}.
This is comparable with a typical collision time and therefore, once excited, this mode will \emph{not} lose immediately its memory, as clearly seen in figs.~\ref{fig:vib_analysis}(a) and~\ref{fig:vib_analysis}(b).
Second, the oblate-prolate mode is the only eigenmode which can couple directly to the relative motion via its elongated \emph{prolate-prolate} configuration, provided it survives the dissipative coupling to all the other internal modes.
This is exactly what happens in $\text{C}_{60}${}+$\text{C}_{60}${} collisions and results in the majority of cases in scattering, like a \emph{fission} process with the \emph{prolate-prolate} configuration at the scission point~\cite{Hasse1988} (note the pronounced excitation of the mode in fig.~\ref{fig:vib_analysis}(a) at $t \sim 165$~fs).
Only strong coupling to the internal DOF can prevent this mechanism, allowing for fusion (note the strongly damped prolate oscillation of the mode in fig.~\ref{fig:vib_analysis}(b) at $t \sim 165$~fs).
Thus, the competitive coupling of the oblate-prolate mode to the relative motion and to all the other (bath-like) vibrational DOF determines the reaction channel !
\begin{figure}
\begin{minipage}{\columnwidth}
\includegraphics[width=0.49\columnwidth]{fig2a.eps}
\includegraphics[width=0.49\columnwidth]{fig2b.eps}
\end{minipage}
\caption{(color online) Typical NA-QMD results for central $\text{C}_{60}$+$\text{C}_{60}$~collisions at impact energies
of $E_\text{c.m.} = 40$~eV (red), $60$~eV (blue), and $104$~eV (black), indicated as dotted
horizontal lines, and obtained by ensemble averaging over 20 different initial orientations.
(a): The potential (deformation) energy of the relative motion $U(R)$ in the entrance channel as function of the distance $R$.
The locations and lengths of the vertical arrows denote the distances of closest approach $R^\text{ret}$ and the total deformation energy $E^\text{ret} \equiv U(R^\text{ret})$ stored at $R^\text{ret}$, respectively.
The dashed line is the harmonic fit $U(R)= \frac{k}{2} \bigl( R-R^{\scriptscriptstyle 0} \bigr)^{\scriptscriptstyle 2}$ with $k = 0.12$~a.u. and $R^0= 18.5$~a.u.
(b): The kinetic energy of the relative motion $E_{kin}$ as function of time $t$ for scattering events.
The mean values of the final kinetic energies $\Delta E \approx 17$~eV and of the corresponding
energies dissipated into internal DOF $E_d(\infty)=E_\text{c.m.}-\Delta E$ are indicated by double arrows.
\label{fig:qmd}}
\end{figure}
The strength of the bath coupling is solely determined by the amount of energy stored in the mode at $R^\text{ret}$.
This coupling dominates, if the energy exceeds a certain limit which generally can happen only at appropriate large impact energies.
This explains (at least preliminarily and qualitatively) the high fusion barriers.
At a fixed impact energy just above the barrier $E_\text{c.m.} \gtrsim V_B$ (as in fig.~\ref{fig:vib_analysis}), only very rare and specific initial orientations of the clusters can lead to high $H_g(1)$ excitation energies , namely, those with the principal axes of the $H_g(1)$ mode aligned to the collision axis, ensuring maximal energy transfer, which is the case in the example shown in fig.~\ref{fig:vib_analysis}(b)
(note that $E_\text{vib}$ of the mode during approach in the case of fusion (fig.~\ref{fig:vib_analysis}(b)) is twice as large as compared to scattering (fig.~\ref{fig:vib_analysis}(a))).
This, finally, explains the low fusion cross sections and completes the present, new picture of the fusion mechanism.
It modifies also the hitherto existing interpretation of scattering, as a ``bouncing off'' mechanism~\cite{Jellinek99,Campbell00_coll,Strout93,Knospe96,Glotov01}, like in collisions between two soccer balls.
Instead, ``fission'' via the prolate-prolate configuration strongly suggests, that the final kinetic energy of the fragments is largely independent on impact energy.
This is nicely confirmed in the calculations and shown in fig.~\ref{fig:qmd}(b).
The collision scenario presented in figs.~\ref{fig:vib_analysis} and~\ref{fig:qmd} is characteristic for fullerene-fullerene reactions and qualitatively observed in our NA-QMD calculations also for the other combinations and finite, small impact parameters.
Despite its microscopic complexity, the mechanism is nevertheless simple and can be understood and described by ordinary \emph{macroscopic} concepts, as will be shown in the following.
\begin{figure}
\begin{minipage}{0.9\columnwidth}
\begin{minipage}{\columnwidth}
\includegraphics[width=\columnwidth]{fig3.eps}
\end{minipage}
\end{minipage}
\caption{(color online) Phenomenological collision model with two DOF, the distance between the centers $R(t)$ and the diameter $D(t)$ of the fullerenes:
Two collinear colliding springs (with spring constants $k_\text{mode}$ and masses $m$) interact during approach via a third, massless spring (spring constant~$k$ and initial length~$R_c^0$) located at the collision center. \label{fig:springs}}
\end{figure}
\section{Collision model}
The basic idea is to reduce drastically the {360-dimensional} scattering problem to a one-dimensional
one with only two, but relevant collective DOF, treated explicitly in a time-dependent fashion: the distance between the centers $R(t)$ (relative motion) and the diameters of the fullerenes $D(t)$ (aligned along their principal axes of the oblate-prolate mode).
Both are coupled via the contact distance $R_c = R - D$~(see fig.~\ref{fig:springs}).
The coupling to the other internal DOF will be treated implicitly in the exit channel only.
The macroscopic model is designed as follows:
(i) In the entrance channel, the system consists of two collinear colliding springs with initial lengths $D(t=0)=D^0$, spring constants $k_\text{mode}$ and masses $m = \frac{M}{2}$ at the
ends (describing the fullerenes with mass $M$ and their $H_g(1)$ modes).
Tightly located in between there is a third, massless spring with initial length $R_c^0$ and constant $k$, describing the repulsive potential $U$ during approach (remember fig.~\ref{fig:qmd}(a) and see fig.~\ref{fig:springs} and fig.~\ref{fig:model}(a)).
(ii) In the exit channel, the massless repulsive spring is replaced by a ``dissipative'' potential $U_d$,
which describes the coupling to all other internal vibrational DOF, and hence, controls the reaction channel (see fig.~\ref{fig:model}(b)).
\begin{figure}
\begin{minipage}{\columnwidth}
\begin{minipage}{\columnwidth}
\includegraphics[width=\columnwidth]{fig4.eps}
\end{minipage}
\end{minipage}
\caption{(color online) Schematic plot of the potential energy in the entrance channel $U(R_c)$ (left) and, for the case of scattering, in the exit channel $U_d(R_c)$ (right) as function of the contact distance $R_c$.
Dotted blue lines with directional arrows indicate a typical trajectory (idealized as a straight line in the exit channel) with impact energy $E_{c.m.}$ at the initial distance $R_c^0$ and potential energy $E^\text{ret}$ at the distance of closest approach $R_c^\text{ret}$.
The asymptotic value of the potential energy in the exit channel $U_d(R_c \rightarrow \infty) \equiv E_d(\infty)$ is also indicated.
\label{fig:model}}
\end{figure}
The coupled Newton equations in the entrance channel read
\begin{equation} \label{equ:newton}
\begin{split}
\mu \ddot R &= \left . -\frac{d U}{d R_c}\right |_{R_c=R-D}\\
\mu_\text{mode} \ddot D &= -\frac{d V_\text{mode}}{d D} + \frac{1}{2}\ \left . \frac{d U}{d R_c}\right |_{R_c=R-D}
\end{split}
\end{equation}
with the reduced masses $\mu =\frac{M}{2}$ and the mass constant $\mu_\text{mode}=\frac{M}{4}$.
The harmonic potentials are given by
$U(R_c) = \frac{k}{2} (R_c-R_{c}^0)^2$ with $k=0.12$~a.u. (see fig.~\ref{fig:qmd}(a))
and $V_\text{mode}(D) = \frac{k_\text{mode}}{2} (D-D^0)^2$ with the spring constant $ k_\text{mode} = \mu_\text{mode}\ \omega_\text{mode}^2=0.51$~a.u., obtained from the experimental frequency of the $H_g(1)$ mode in $\text{C}_{60}${} ($\omega_\text{mode}=273\ \text{cm}^{-1}$~\cite{Dong93}).
The equations of motion (EOM)~(\ref{equ:newton}) can be solved analytically (see appendix).
They describe the collision up to the distance of closest approach,
i.e., the classical returning point $R_c^\text{ret}$.
At this point the potential $U(R_c)$ for $R_c > R_c^\text{ret}$ is replaced by the ``dissipative'' one $U_d(R_c)$ which, in dependence on $E_\text{c.m.}$, controls the outcome.
It is therefore repulsive (leading in any case to scattering) or attractive
(leading usually to fusion, but not necessarily always).
Thus, it has the general form
\begin{align}
U_d(R_c) &= \left( E^\text{ret} - E_d(\infty) \right) f(R_c-R_c^\text{ret}) + E_d(\infty) \label{equ:pot_diss}
\end{align}
where the form factor $f$ must fulfill the conditions, $f(0)=1$ (ensuring the continuity of the potential at $R_c^\text{ret}$) and $f(\infty)=0$ (making sure that the maximal amount of dissipated energy cannot exceed $E_d(\infty)$, in the case of scattering).
In fact, the concrete radial dependence of the potential~(\ref{equ:pot_diss}) is not relevant, and thus, we choose a simple exponential form of $f(x)=\exp{\bigl(-\frac{x}{\Delta}\bigr)}$
with $\Delta =\bigl | \frac{R_c^\text{ret}-R_{c}^0}{2}\ (1-\frac{E_d(\infty)}{E^\text{ret}}) \bigr|$, which guarantees also continuity of the force at $R_c^\text{ret}$ in the case of scattering.
With this, all model parameters ($k$, $k_\text{mode}$, $\Delta E$) are fixed and the EOM~(\ref{equ:newton}) can be easily solved numerically.
The results are shown in fig.~\ref{fig:compare_springs_exp} and compared with NA-QMD calculations (for movies see \emph{www.dymol.org}).
The model reproduces nearly precisely the microscopic calculations for both scattering ($E_\text{c.m.} = 40, 60$~eV) and fusion ($E_\text{c.m.} = 150$~eV).
The ongoing oscillations of some quantities in the exit channel are the natural consequence of the absence of a damping mechanism for the $H_g(1)$ mode in the spring model.
The most impressive result, however, concerns the fusion barrier predicted by the model of $V_B = 85$~eV,
which is in excellent agreement with former (extremely expensive) QMD calculations~\cite{Knospe96} of $V_B = 80$~eV.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{fig5.eps}
\caption{Comparison between NA-QMD (dashed lines) and model calculations (solid lines)
for central $\text{C}_{60}$+$\text{C}_{60}$~collisions at three impact energies $E_\text{c.m.}$:
distance between the centers $R$ in a.u. (top),
kinetic energy of the relative motion $E_{kin}$ (middle),
and vibrational kinetic energy of the $H_g(1)$ mode $E_{vib}$ (bottom) in eV
as function of time $t$.
\label{fig:compare_springs_exp}}
\end{figure}
\section{Analytical solution for the fusion barrier}
The simplicity of the EOM~(\ref{equ:newton}) and the transparency of the ansatz~(\ref{equ:pot_diss})
encouraged us to further elaborate the model and to demonstrate its predictive power.
As mentioned already, the EOM~(\ref{equ:newton}) can be solved analytically, and, from this solution one can derive an approximate expression for the energy stored during approach $E^\text{ret}$ (see Fig~\ref{fig:model}(a)),
resulting in a linear dependence on the impact energy of $E^\text{ret} = \alpha( \kappa)\ E_\text{c.m.}$ with $\kappa$ the ratio of the spring constants
\begin{align}
\kappa &= \frac{k_\text{mode}}{k} = \frac{\frac{M}{4} \omega_\text{mode}^2}{k} \label{equ:kappa}
\end{align}
and the universal function $\alpha(\kappa)$ of
\begin{align}
\alpha(\kappa) &= \frac{1}{4 (\kappa^2+1)}\ \left[ \frac{\kappa-1-\sqrt{\kappa^2+1}}{\sqrt{\kappa+1+\sqrt{\kappa^2+1}}} \right. \notag
\\ & \times \left. \sin{(\frac{\pi}{2}\frac{\kappa+1+\sqrt{\kappa^2+1}}{\sqrt{2\kappa}})}
- \frac{\kappa-1+\sqrt{\kappa^2+1}}{\sqrt{\kappa+1-\sqrt{\kappa^2+1}}} \right]^2
\label{equ:alpha}
\end{align}
(see appendix).
The difference $(E^\text{ret} - E_d(\infty))$ determines the (positive or negative) slope of the potential $U_d$~(\ref{equ:pot_diss}) at the returning point $R_c^\text{ret}$.
Taking an idealized trajectory as shown in fig.~\ref{fig:model} (i.e., neglecting the acceleration of the mode near the barrier), the fusion condition simply reads $E^\text{ret} \stackrel{!}{=} E_d(\infty)$ at $E_\text{c.m.} = V_B$.
With $E_d(\infty)=E_\text{c.m.} - \Delta E$ (see fig.~\ref{fig:qmd}(b)),
the fusion barrier becomes
\begin{align}
V_B = \frac{\Delta E}{1-\alpha(\kappa)}.\label{equ:fusion_barrier}
\end{align}
For $\text{C}_{60}${}+$\text{C}_{60}${} collisions, this approximate expression gives $V_B= 82$~eV,
which is very close to the exact model value ($V_B= 85$~eV) obtained in the upper dynamical calculations.
To obtain a first insight about the qualitative trends for the other combinations
($\text{C}_{60}${}+$\text{C}_{70}${}, $\text{C}_{70}${}+$\text{C}_{70}${}), we ignore subtleties and use the same $k$ and $\Delta E$ values as obtained already for $\text{C}_{60}${}+$\text{C}_{60}${} by NA-QMD fine tuning.
The internal spring constant for $\text{C}_{70}${}, however, is carefully chosen and fixed again by experiment.
From the (partly) non-degenerate five $H_g(1)$ modes in $\text{C}_{70}${} ($E_2'$, $E_1''$, $A_1'$) an experimental
mean value of $\omega_\text{mode}(\text{C}_{70}) = 261\ \text{cm}^{-1}$ has been
reported~\cite{Wang95} giving the spring constant $k_\text{mode}(\text{C}_{70}) = 0.55$~a.u.
For the (slightly) asymmetric $\text{C}_{60}${}+$\text{C}_{70}${} collision a reasonable mean value of $k_\text{mode}(\text{C}_{60})$ and $k_\text{mode}(\text{C}_{70})$ of $k_\text{mode}(\text{C}_{60}/\text{C}_{70}) = 0.53$~a.u. is used.
With these parameters, the predicted fusion barriers from (\ref{equ:fusion_barrier}) are compared with high precision QMD values~\cite{Knospe96}, in table~\ref{tab:fusion_barriers} (first two rows).
The analytical model reproduces the right trend and delivers absolute values within 10\% accuracy.
Obviously, the $H_g(1)$ frequencies $\omega_\text{mode}$ and fullerene masses $M$, eq.~(\ref{equ:kappa}) determine the fusion barriers in fullerene-fullerene collisions.
This is strongly supported by comparing the predictions with experimental data.
In this case, the finite temperature ($T \approx 2000$~K) of the colliding fullerenes has to be taken into account~\cite{Rohmund96_jpb}.
This has been done in the former QMD calculations~\cite{Knospe96} and led to the (well known) perfect agreement with the experimental data (cf. third and fifth rows in table~\ref{tab:fusion_barriers}).
In the present model, a finite temperature can be naturally taken into account by reducing the mode frequencies $\omega_\text{mode}$.
Using an arbitrary common scaling factor for $k_\text{mode}$ of $0.85$, the predicted fusion barriers by eq.~(\ref{equ:fusion_barrier}) are in beautiful agreement with the experiment (fourth and fifth row in table~\ref{tab:fusion_barriers}).
\begin{table}[t]
\caption{Fusion barriers $V_B$ in~eV for various fullerene-fullerene combinations as predicted by our former QMD calculations~\cite{Knospe96} (first and third rows) and the present analytical model (second and fourth rows).
The experimental values~\cite{Rohmund96_jpb} are presented in the last row.}
\begin{tabular}{c|cc|cc|c}
& \multicolumn{2}{|c|}{$T=0$~K} & \multicolumn{3}{|c}{$T=2000$~K} \\
& QMD & model & QMD & model & exp. \\
\hline
$\text{C}_{60}$+$\text{C}_{60}$ & 80 & 82 & 60 & 65 & 60$\pm$1 \\
$\text{C}_{60}$+$\text{C}_{70}$ & 94 & 87 & 70 & 69 & 70$\pm$6.5 \\
$\text{C}_{70}$+$\text{C}_{70}$ & 104 & 93 & 75 & 73 & 76$\pm$4
\end{tabular}
\label{tab:fusion_barriers}
\end{table}
\section{Conclusion}
To summarize, \emph{ab initio} NA-QMD studies have finally cleared up the fusion mechanism in fullerene-fullerene collisions.
The excitation of a ``giant'' $H_g(1)$ mode explains both large fusion barriers and small fusion cross sections.
This \emph{microscopic} picture is non-ambiguously confirmed by a \emph{macroscopic} spring model which depicts clearly the physics, reproduces the NA-QMD results and the experimental fusion barriers \emph{quantitatively}.
We note finally, that a ``giant'' vibrational excitation of the $A_g(1)$ breathing mode in $\text{C}_{60}${} has been found recently in a time-resolved laser experiment~\cite{Laarmann07}.
The general investigation of large amplitude motion in fullerenes, including laser-induced \emph{fission}~\cite{Fischer2013}, could become an interesting new field of research.
\section{Appendix: Analytical solution in the entrance channel}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}
Inserting the harmonic potentials $U(R_c)$ and $V(D)$, the EOM (\ref{equ:newton}) can be written as
\begin{equation}\label{equ:eom}
\begin{split}
\mu \ddot R &= \left . -k\ (R_c-R_c^{0})\right |_{R_c=R-D} \\
\mu_\text{mode} \ddot D &= -k_\text{mode}\ (D-D^0) + \frac{k}{2}\ \left . \left( R_c-R_c^{0} \right )\right |_{R_c=R-D}
\end{split}
\end{equation}
which can be solved by making an exponential ansatz $e^{\imath \Omega t}$.
Doing so, the fundamental eigenfrequencies $\Omega_{1/2}$ of the system read
\begin{align}
\Omega_{1/2} &= \omega\ \underbrace{\sqrt{\kappa + 1 \pm \sqrt{\kappa^2+1}}}_{f_{1/2}(\kappa)} \label{equ:freq1}
\end{align}
with the frequency $\omega = \sqrt{\frac{k}{\mu}}$ and the force constant ratio $\kappa$ as defined in eq.~(\ref{equ:kappa}).\\
With the initial conditions
\begin{align*}
R(0) &\equiv R^{0} = D^{0}+R_{c}^0,& \dot R(0) &= v_\text{c.m.},& \\
D(0) &= D^{0},& \dot D(0) &= 0&
\end{align*}
the solution of (\ref{equ:eom}) is given by
\begin{align}\label{equ:analyt_solution1}
R(t) &= \frac{v_\text{c.m.}}{\omega} \sum_{i=1,2} a_i \sin{(\Omega_i t)} + R^{0},\\
D(t) &= \frac{v_\text{c.m.}}{\omega} \sum_{i=1,2} b_i \sin{(\Omega_i t)} + D^{0}\label{equ:analyt_solution2}
\end{align}
with the amplitudes $a_{1/2} = \frac{1}{2\ f_{1/2}(\kappa)}\ (1 \mp \frac{\kappa}{\sqrt{\smash[b]{\kappa^2+1}}})$
and $b_{1/2} = \mp \frac{1}{2\ f_{1/2}(\kappa)}\ \frac{1}{\sqrt{\smash[b]{\kappa^2+1}}}$.
With the analytical solution (\ref{equ:analyt_solution1}), (\ref{equ:analyt_solution2})
the potential energy at the returning point $E^\text{ret}$ can be calculated.
$E^\text{ret}$ is given by
\begin{align}
E^\text{ret} &\equiv U(R_c^\text{ret}) = \frac{k}{2} \left(R_c^\text{ret}-R_{c}^0\right)^2\label{equ:e_ret}
\end{align}
with $R_c^\text{ret} \equiv R_c(t^\text{ret})=R(t^\text{ret})-D(t^\text{ret})$.\\
The returning time $t^\text{ret}$ defined by~$\dot R(t^\text{ret})=0$ is approximated by $t^\text{ret} \approx \frac{\pi}{2 \Omega_2}$.
The approximation is justified since the second term of the sum in eq.~(\ref{equ:analyt_solution1}) dominates for the parameter range of $\kappa$ used here ($\frac{a_1}{\smash[t]{a_2}} \ll 1$).\\
Inserting the analytical solution (\ref{equ:analyt_solution1}), (\ref{equ:analyt_solution2}) in the definition (\ref{equ:e_ret}) and using the relation (\ref{equ:freq1})
we find
\begin{align*}
E^\text{ret} &= \alpha(\kappa)\ E_\text{c.m.}
\end{align*}
with the impact energy $E_\text{c.m.}=\frac{\mu}{2}\ v_\text{c.m.}^2$ and the coefficient $\alpha(\kappa)$
as defined before in eq.~(\ref{equ:alpha}).
\acknowledgments
We thank
Sebastian Schmidt (ETH Zurich) for many useful
comments and critical reading of the manuscript.
We gratefully acknowledge the allocation of computer resources from ZIH and MPI-PKS, Dresden and appreciate
the support of the DFG through Einzelverfahren.
|
\section{Introduction}\label{sec:intro}
Transient fluctuation theorems are exact relations restricting the statistics of work performed by externally controlled classical forces. While the considered system initially must be in a thermal equilibrium state, the subsequent forcing may drive it to out of equilibrium to states that cannot be described in terms of linear response theory. Exact relations of such form were pioneered by Bochkov and Kuzovlev \cite{Bochkov77} and the ones pertinent to our discussion bear the names of Jarzynski \cite{Jarzynski97} and Crooks \cite{Crooks99}.
The Crooks relation \cite{Crooks99},
\begin{equation}
p_{\Lambda}(w) = e^{-\beta(\Delta F - w)} p_{\bar{\Lambda}}(-w) \label{eq:crooks}
\end{equation}
relates the probability density function (pdf), $p_{\Lambda}(w)$, of work performed on a system with Hamiltonian $H[\lambda(t)]$ by the action of a generalized force $\lambda(t)$ that varies according to the protocol $\Lambda = \{ \lambda(t)|0\leq t \leq \tau \}$ to the work pdf of another process governed by the time-reversed protocol $\bar{\Lambda} = \{\lambda(\tau-t)|0 \leq t \leq \tau\}$\footnote{For the sake of simplicity we restrict ourselves to forces $\lambda(t)$ which transform evenly under time-reversal.}. Both processes start in a canonical equilibrium state at the same inverse temperature $\beta$ described by the density matrices $\rho_t = Z^{-1}(t) e^{-\beta H[\lambda(t)]}$ with $t=0$ and $\tau$ for the forward and the backward process, respectively. Here $Z(t) = \mbox{Tr}\, e^{-\beta H[\lambda(t)]}$, for $t=0$ and $\tau$ yields the partition functions which determine the free energy difference $\Delta F = -\beta^{-1} \ln (Z(\tau)/Z(0))$.
A straightforward integration of (\ref{eq:crooks}) with $e^{-\beta w}$ brought to its left hand side
leads to the Jarzynski equality \cite{Jarzynski97},
\begin{eqnarray}
\langle e^{-\beta w} \rangle = e^{-\beta \Delta F} \label{eq:jarzynski}.
\end{eqnarray}
Fluctuation theorems have been shown to be valid in a variety of situations ranging from open classical systems \cite{Jarzynski04}, closed \cite{Kurchan00,Tasaki00,Talkner07} and open quantum systems \cite{Crooks08,Talkner09,Campisi09}. For recent reviews of this topic \cite{Jarzynskirev11,Seifert12} (focus on classical systems) and \cite{Esposito09,Campisirev11,Campisireverr11,Hanggi15} (focus on quantum systems) can be consulted.
These theorems have also been verified experimentally for a variety of classical systems \cite{Hummer01,Liphardt01,Trepagnier04,Ciliberto05} and have also been used to determine free energy changes in molecular systems by measuring work statistics in controlled non-equilibrium processes \cite{Hummer01,Liphardt01}. Further details on this topic are available in the reviews \cite{Bustamante05,Ritort08}.
The situation regarding experimental verification of fluctuation theorems for quantum systems is still in its nascent stages. One central issue in the quantum context is that work is not an observable \cite{Talkner07}; it is process dependent and therefore the potential values that it may take in a single experiment cannot be represented as the eigenvalues of a hermitian operator acting on the Hilbert space of the considered system. This issue underpins the most common approach to determine work statistics for a quantum system, the so-called two-energy-measurement approach (TEMA) \cite{Talkner07,Campisirev11,Campisireverr11}: here the work supplied by a time dependent force $\lambda(t)$ during the time interval $0\leq t \leq \tau$ is defined as the difference between the system's energy at the final and initial times. The TEMA can be divided into the following steps: (\rmnum{1}) the system is prepared in a thermal equilibrium state at inverse temperature $\beta$ at $t=0$ with the initial Hamiltonian $H[\lambda(0)]$; (\rmnum{2}) a projective measurement of energy is performed yielding one of the eigenvalues $e_n(0)$ of $H[\lambda(0)]$ as possible outcome; due to the measurement the state of the system is projected to the associated energy eigenstate $\vert n; 0 \rangle$; (\rmnum{3}) subsequently it undergoes a unitary time evolution during $0 \leq t \leq \tau$ governed by the Hamiltonian $H[\lambda(t)]$ which changes in time according to the prescribed protocol $\Lambda$; (\rmnum{4}) at $t = \tau$ a second projective measurement of energy is performed; its outcome is an eigenvalue $e_m(\tau)$ of the final Hamiltonian $H[\lambda(\tau)]$. The work done in such a process is given by $w_e = e_m(\tau) - e_n(0)$. In general, projective measurements turn out to be difficult to perform in an experiment. Nonetheless proposals for implementing such a scheme using trapped ions was suggested in \cite{Huber08} and its experimental realisation was recently reported in \cite{Kim14}. In an indirect verification of the fluctuation theorems detailed in \cite{Dorner13,Mazzola13,Campisi13}, the experimental difficulties of projective measurements were circumvented by encoding the characteristic function of work in the reduced state of an ancillary system that interacts with the system of interest with a strength determined by the force protocol. An experiment, using NMR spectroscopy, of such a proposal was reported in \cite{Batalhao13}. A second way to address the difficulty of performing projective measurements is to ask if the replacement of the same by non-projective generalized measurements, which may be easier to implement in an experiment, preserves the fluctuation theorems \cite{Venkatesh14,Watanabe14} (see also \cite{Roncaglia14, deChiara14} for a recent proposal to measure work as the outcome of a single generalised measurement). With this line of enquiry we found in an earlier publication \cite{Venkatesh14} a no-go theorem restricting projective measurements of energy as the only ones within a large class of generalized measurements that satisfy both Crooks and Jarzynski equalities for arbitrary protocols. Although, for some specific forms of the generalized measurements, modified fluctuation theorems may be derived \cite{Watanabe14}.
In this work, we focus on yet another way to address the issue of measurement of work in quantum systems. Conventionally, in classical systems work can be determined in an incremental way by integrating up the supplied power which can be inferred from the instantaneous state of the system \cite{Bustamante05}. A direct extension of such a method to quantum systems is difficult since the system will have to be continuously monitored \cite{Misra77,Itano90}. It is known that a continuous monitoring of a quantum system can cause the freezing of coherent dynamics, a phenomenon known as the quantum Zeno effect \cite{Misra77,Itano90}. The central aim of this work is to explore in detail how and why the extension of power measurements to estimate the work done on quantum systems fails to provide
work statistics that satisfies the transient fluctuation theorems of Crooks \cite{Crooks99} and Jarzynski \cite{Jarzynski97}.
The paper is organized as follows. In section \ref{sec:powmeas} we define the quantum mechanical variable corresponding to the instantaneous power supplied to a system that experiences a force via a coupling to a generalized coordinate. We introduce an estimate of work computed by incremental projective measurements of the power in conjunction with the protocol, and explain how such an estimate differs from the usual TEMA based work both in magnitude and range of the possible values. In the limit of frequent measurements, the system's unitary evolution is frozen due to the quantum Zeno effect (we refer to such a regime of the dynamics as the ``Zeno limit" henceforth). We analyze the statistics of the power-based work in this limit and point out how the commutator of the power operator with the total Hamiltonian is crucial in deciding the merits of the work estimate. We also derive an inequality for the dissipated work estimated by the power measurement in the Zeno limit. Further, the results so far obtained are illustrated by the example of a two-level system undergoing an avoided crossing. In section \ref{sec:jointstat}, we allow for energy measurements at the beginning and the end of the force protocol additionally to the coordinate measurements necessary for the power measurements and compare the distributions of TEMA, $w_e$, and power-based work estimates $w_p$. In section \ref{sec:fintime} we relax the assumption of projective instantaneous measurements of the power operator and adopt a weak, continuous measurement of power to define the work. The treatment there is carried out within the framework of Stochastic Master Equations (SME) \cite{Jacobs06,WM}. Work statistics obtained from numerical solution of SMEs for the two-level system example are also studied. In section \ref{sec:conclusion} we conclude the paper. Appendices \ref{app:genpower} and \ref{app:smederiv} provide additional details omitted from the main text of the paper.
\section{Power measurement and the quantum Zeno effect} \label{sec:powmeas}
In classical mechanics, the energy-work relation can be invoked to define the work supplied to a thermally closed system\footnote{We denote a system as thermally closed if it does not exchange energy and/or particles with its environment. In order to be able to perform work on it,
it must though be possible to externally change some of the system parameters.} as the increase in its energy. First, let us consider a classical system whose energy is determined by the Hamiltonian $H^{\mathrm{cl}}[\mathbf{z},\lambda(t)]$ where $\mathbf{z}$ is a point in phase-space and $\lambda(t)$ is an external parameter which is varied in time leading to a change in the system's energy. Hence the work done by changing $\lambda(t)$ in the interval $0\leq t \leq \tau$ according to some specific protocol is given by \cite{Jarzynski97}
\begin{eqnarray}
w^{\mathrm{cl}} = H^{\mathrm{cl}} \sqlr{\mathbf{Z(\tau,\mathbf{z})},\lambda(\tau)} - H^{\mathrm{cl}}\sqlr{\mathbf{z},\lambda(0)} \label{eq:ediffwclass},
\end{eqnarray}
where $\mathbf{Z}(t,\mathbf{z})$ is the solution of Hamilton's equation of motion
\begin{eqnarray*}
\frac{d}{dt} \mathbf{Z} = \flrlr{H^{\mathrm{cl}}\sqlr{\mathbf{Z},\lambda(t)},\mathbf{Z}}
\end{eqnarray*}
at time $t$ for a trajectory with the initial condition $\mathbf{Z}(0,\mathbf{z}) = \mathbf{z}$. (\ref{eq:ediffwclass}) can be rewritten as an integral of the total time-derivative of the Hamiltonian which agrees with the partial time-derivative. This gives an equivalent expression for work,
\begin{eqnarray}
w^{\mathrm{cl}} = \int_0^{\tau} dt \frac{\partial H^{\mathrm{cl}}\sqlr{\mathbf{Z}(t,\mathbf{z}),\lambda(t)}}{\partial t} \label{eq:powintwclass},
\end{eqnarray}
where the integrand is the instantaneous power $L^{\mathrm{cl}}$ supplied to the system at time $t$, i.e.
\begin{eqnarray}
L^{\mathrm{cl}}(t,\mathbf{z}) &= \frac{\partial H\sqlr{\mathbf{Z}(t,\mathbf{z}),\lambda(t)}}{\partial t}\\
&=\frac{\partial H\sqlr{\mathbf{Z}(t,\mathbf{z}),\lambda(t)}}{\partial \lambda} \frac{d \lambda(t)}{d t}
\label{eq:classpow}.
\end{eqnarray}
The power-based work expression (\ref{eq:powintwclass}) is more convenient to determine work in experiments \cite{Hummer01}. Since the initial condition $\mathbf{z}$ is only specified in terms of a probability distribution (for, e.g., a canonical distribution if the system is initially in thermal equilibrium), work becomes a random quantity.
In quantum mechanics, both expressions (\ref{eq:ediffwclass}) and (\ref{eq:powintwclass}) can in principle be extended to operational definitions of work which, however, turn out to be no longer equivalent to each other. Here, we note that there have been earlier attempts \cite{Chernyak04,Engel07,Liu12,Wang13,Solinas13,Suomela14} to define work in quantum systems using expressions analogous to (\ref{eq:powintwclass}). In \cite{Chernyak04} this is done in the context of a driven harmonic oscillator interrupted by a small number of measurements of the coordinate and in \cite{Wang13} the same system is examined under the continuous quantum histories framework for the power operator. In the context of superconducting Cooper-pair box systems some lower order moments of a power-based work were considered in \cite{Solinas13} for a closed system and extended to include open systems in \cite{Suomela14}. In some of these attempts \cite{Engel07,Solinas13}, power-based work has been treated as a standard quantum mechanical observable (with a corresponding operator) in contrast to the view point in \cite{Talkner07}, which we adopt. One central aim of this work is to show that even a more careful implementation of the power-based work for quantum systems taking into account the process dependence will generally lead to qualitatively different statistics from the TEMA definition of work and generally fail to satisfy the fluctuation relations of Crooks and Jarzynski.
To that end, we consider the simplest possible situation where a single scalar parameter $\lambda(t)$ acts as a force on a system via a generalized coordinate $X$. In the absence of this force the system is described by the Hamiltonian $H_0$. In the quantum scenario, the generalised coordinate $X$ corresponds to a self-adjoint operator $\hat{X}$ acting on the system's Hilbert space. The total Hamiltonian is then given by:
\begin{eqnarray}
H[\lambda(t)] = H_0 + \lambda(t) \hat{X}. \label{eq:totham}
\end{eqnarray}
In analogy to the classical form of power (\ref{eq:classpow}) we define the power operator $L(t)$ as
\begin{eqnarray}
L(t) \equiv \frac{\partial H^{H}[\lambda(t)]}{\partial t} = \dot{\lambda}(t) \hat{X}(t) \label{eq:powerop},
\end{eqnarray}
where the superscript $H$ denotes the Heisenberg picture and the dot a time derivative. Since the generalised coordinate does not explicitly depend on time, we simply indicate the Heisenberg picture by the presence of the time-argument omitting the superscript $H$. In order to determine the power, one needs to perform a measurement of the generalized coordinate $\hat{X}$. The possible measurement outcomes, in a projective measurement \cite{vonneumann96}, are determined by the eigenvalues $x_{\alpha}$ of the generalized coordinate $\hat{X} = \sum_{\alpha} x_{\alpha} \Pi_{\alpha}^X$. For the sake of simplicity we assume that $\hat{X}$ has a non-degenerate discrete spectrum. Hence, the eigenprojection operators $\Pi_{\alpha}^X = \ket{\varphi_{\alpha}} \bra{\varphi_\alpha}$ are determined by the eigenfunction $\ket{\varphi_{\alpha}}$ of $\hat{X}$. In order to precisely capture the work as the integral of the power one should, in principle, continuously measure the generalized coordinate. This though will inevitably freeze the dynamics of the system in an eigenstate of $\hat{X}$. Provided that the force protocol is sufficiently slow compared to the unitary dynamics of the system, one can try to avoid the full halt of the dynamics by performing only a finite number of power measurements and approximate the integral by a discrete sum:
\begin{eqnarray}
w^{(N)}_p = \displaystyle \sum_{i=1}^N \dot{\lambda}(t_i) x_{\alpha_i} h, \label{eq:workpowbN}
\end{eqnarray}
where we assumed that $N$ measurements take place at regularly spaced times $t_i$ with $t_{i+1}-t_i = t_1 = h$ in the interval $\parlr{t_0 \equiv 0}\leq t \leq \parlr{t_{N+1} \equiv \tau}$. The first basic difference between the power-based and TEMA work estimates emerges from the varying allowed values of work that the two approaches produce. On the one hand, the set of possible TEMA work values is given by
\begin{eqnarray}
\mathcal{W}^{e} = \flrlr{w = e_m(\tau)-e_n(0)|m\in I(\tau),n\in I(0)}, \label{eq:workTMA}
\end{eqnarray}
where $I(t)$ is the set of indices labelling the spectrum of $H(t)=\sum_{n \in I(t)} e_n(t) \Pi_n(t)$. Here $\Pi_n(t)$ are operators projecting on the eigenstates corresponding to $e_n(t)$. Comparing with all possible power-based work values having the form given by (\ref{eq:workpowbN}), it is apparent that both the number of possible values and magnitudes of work are different in the two approaches. Indeed for a system with finite $D$-dimensional Hilbert space, the maximum number of possible work values from TEMA is $D^2$, whereas in the power measurement case with $N$ measurements it is given by $D^{N}$. Whereas the allowed TEMA work values are functions of all parameters entering the full Hamiltonian (\ref{eq:totham}), the power-based work values depend solely on the time-derivatives of the force at the measurement times. Only the probabilities which characterize the occurrence of these work values may depend on the other parameters of the system.
Still one could hope that the two estimates of work might be similar in a statistical sense. For this purpose we now compute the distribution for the work estimate in (\ref{eq:workpowbN}) to compare it with the distribution of the TEMA work values in terms of their cumulative probabilities. Beginning at $t=0$ with a canonical density matrix $\rho_0 = Z^{-1}(0) e^{-\beta H[\lambda(0)]}$ we obtain for the joint probability $\mathscr{P}_{\Lambda}(\mathbf{x})$ of finding the eigenvalues $\mathbf{x} = \parlr{x_{\alpha_1}, x_{\alpha_2}, \cdots, x_{\alpha_N}}$ in the $N$ measurements at times $t_1,t_2, \cdots, t_N$:
\begin{eqnarray}
\mathscr{P}_{\Lambda}(\mathbf{x}) = \mbox{Tr}\, V_{\Lambda}\parlr{\mathbf{x}} \rho_0 V_{\Lambda}^{\dagger}\parlr{\mathbf{x}} \label{eq:jointprobproj},
\end{eqnarray}
where
\begin{eqnarray}
V_{\Lambda}\parlr{\mathbf{x}} = \Pi_{\alpha_{N}}^X U_N \Pi_{\alpha_{N-1}}^X \cdots U_2 \Pi_{\alpha_{1}}^X U_1 \label{eq:jointprobform}.
\end{eqnarray}
Here $U_k = U_{t_k,t_{k-1}}(\Lambda)$ denotes the time-evolution operator for the system from time $t_{k-1}$ to $t_k$ under the influence of the protocol $\Lambda$. It is the solution to the Schr\"{o}dinger equation
\begin{eqnarray}
i\hbar \frac{\partial U_{t,s}}{\partial t} = H[\lambda(t)] U_{t,s}(\Lambda),
\label{U}
\end{eqnarray}
obeying the initial condition
\begin{eqnarray*}
U_{s,s}(\Lambda) = \mathbb{1}.
\end{eqnarray*}
As evident from (\ref{eq:jointprobform}), we have taken the measurements of the generalized coordinate to be projective (following von Neumann \cite{vonneumann96}). We will relax this assumption in Appendix \ref{app:genpower} and consider generalised measurements of the coordinate. Expressing the projection operators in terms of the eigenfunctions of $\hat{X}$, the joint probability becomes:
\begin{eqnarray}
\mathscr{P}_{\Lambda}(\mathbf{x}) = \displaystyle \prod_{k=1}^{N-1} \left | \bra{\varphi_{\alpha_{k+1}}} U_{k+1} \ket{\varphi_{\alpha_{k}}}\right |^2 \bra{\varphi_{\alpha_1}} U_1 \rho_0 U^{\dagger}_1 \ket{\varphi_{\alpha_1}} \label{eq:jointprobeig}.
\end{eqnarray}
The pdf $p_{\Lambda p}^{(N)} (w)$ of finding the value $w$ for the power-based work estimate $w_p^{(N)}$ is then given by:
\begin{eqnarray}
p_{\Lambda p}^{(N)} (w) =\displaystyle \sum_{\flrlr{{\alpha_i}}} \delta \parlr{w-\sum_{i=1}^{N} \dot{\lambda}(t_i)x_{\alpha_i}h} \mathscr{P}_{\Lambda}(\mathbf{x}). \label{eq:projfinmeaspowpdf}
\end{eqnarray}
For a large number of measurements, the Zeno effect \cite{Misra77,Itano90} causes a freezing of the system in the state corresponding to the outcome of the first measured eigenvalue of $\hat{X}$. This is a consequence of the fact that the transition probabilities $\left | \bra{\varphi_{\alpha_{k+1}}} U_{k+1} \ket{\varphi_{\alpha_{k}}}\right |^2$ between different eigenstates of $\hat{X}$ vanish as $h^2$. Hence for large values of $N$ (which correspond to small values of $h = \tau/(N+1)$), the joint probability in (\ref{eq:jointprobeig}) becomes
\begin{eqnarray}
\mathscr{P}_{\Lambda}(\mathbf{x}) = \displaystyle \prod_{k=1}^{N-1} \delta_{x_{\alpha_k},x_{\alpha_{k+1}}} \bra{\varphi_{\alpha_1}} \rho_0 \ket{\varphi_{\alpha_1}} + O(h).
\end{eqnarray}
Putting this asymptotic result into the expression in (\ref{eq:projfinmeaspowpdf}) we obtain
\begin{eqnarray}
p_{\Lambda p}^{(\infty)}(w) = \displaystyle \sum_{\alpha} \delta \parlr{w - \sqlr{\lambda(\tau)-\lambda(0)}x_{\alpha}}\bra{\varphi_{\alpha}}\rho_0 \ket{\varphi_{\alpha}}, \label{eq:projinfmeaspowpdf}
\end{eqnarray}
for the work pdf neglecting corrections of the order $h$. Note that in the above $h \rightarrow 0$ limit, the sum in the delta function specifying the work in (\ref{eq:projfinmeaspowpdf}) can be replaced by an integral yielding $\lim_{N \to \infty}\sum_{i=1}^{N} \dot{\lambda}(t_i)x_{\alpha_1}h = \sqlr{\lambda(\tau)-\lambda(0)}x_{\alpha_1}$.
Having obtained the pdf for the power-based estimate of work, we can immediately check if the Jarzynski equality is satisfied. To that end we have for the average of the exponentiated work the following expression:
\begin{eqnarray}
\langle e^{-\beta w} \rangle_p = \int dw\, p_{\Lambda p}^{\infty}(w) e^{-\beta w} = \frac{\mbox{Tr}\, e^{-\beta \sqlr{\lambda(\tau) -\lambda(0)} \hat{X}} e^{-\beta\sqlr{H_0 + \lambda(0) \hat{X}}}}{Z(0)} \label{eq:expbetawpowinf}.
\end{eqnarray}
For the Jarzynski equality to hold, the numerator of the right-hand side must coincide with the partition function $\mbox{Tr}\, e^{-\beta H[\lambda(\tau)]}$. However, this is only the case if the unperturbed Hamiltonian $H_0$ commutes with the generalized coordinate $\hat{X}$, i.e. if $[ H_0, \hat{X} ] = 0$, which is an atypical situation. Because in the commuting case, the forcing does not lead to transitions between different eigenstates of $H[\lambda(t)]$, the set of allowed TEMA work values is given by
\begin{equation}
\mathcal{W}^e = \{ (\lambda(\tau) - \lambda(0)) x_\alpha|\alpha \in I \}
\label{Wec}
\end{equation}
and hence coincides with the set of allowed power-based work values. This follows from the eigenvalues of $H\sqlr{\lambda(t)}$ taking the form $e_{\alpha}(t) = e_{\alpha} + \lambda(t) x_{\alpha}$, $\alpha \in I$ where
$I$ is the time-independent set labelling the eigenvalues $e_{\alpha}$ and $x_{\alpha}$ of $H_0$ and $\hat{X}$, respectively.
In general, when $H_0$ and $\hat{X}$ do not commute, the power-based work estimate does not conform with the Jarzynski equality\footnote{In \cite{Wang13} a similar conclusion was obtained using the continuous quantum histories framework for the model system of a center-shifted harmonic oscillator.}. In this case one finds with the Golden-Thompson inequality \cite{Golden65} $\mbox{Tr}\, e^{A}e^B \geq \mbox{Tr}\, e^{A+B}$ (with $A$ and $B$ being hermitian) in (\ref{eq:expbetawpowinf}) yielding
\begin{eqnarray*}
\langle e^{-\beta w} \rangle_p \geq \frac{\mbox{Tr}\, e^{-\beta \sqlr{\lambda(\tau) -\lambda(0)} \hat{X}-\beta\sqlr{H_0 + \lambda(0) \hat{X}}}}{Z(0)} = e^{-\beta \Delta F}.
\end{eqnarray*}
Thus the power-based work estimate satisfies the inequality
\begin{eqnarray}
\langle e^{-\beta \parlr{w-\Delta F}} \rangle_p \geq 1.
\end{eqnarray}
Hence, an estimate of the free energy based on the exponentiated power-based work average underestimates the true value
\begin{eqnarray}
\Delta F_p \equiv -\beta^{-1} \ln \langle e^{-\beta w} \rangle_p \leq \Delta F.
\end{eqnarray}
A more detailed comparison can be made on the basis of the power-based work pdf (\ref{eq:projinfmeaspowpdf}) and the corresponding TEMA work pdf $p_{\Lambda e}(w)$ given by
\cite{Campisirev11,Campisireverr11}:
\begin{eqnarray}
p_{\Lambda e}(w) = \sum_{m,n} \delta \parlr{w-e_m(\tau) + e_n(0)} p_{\Lambda}(m,n),
\label{pe}
\end{eqnarray}
where the joint probability $p_{\Lambda}(m,n)$ to find the eigenstates $\ket{n;0}$ and $\ket{m;\tau}$ in projective energy measurements at the beginning and the end of the force protocol, respectively, reads
\begin{eqnarray}
p_{\Lambda}(m,n) = \mbox{Tr}\, \Pi_m (\tau) U_{\tau,0}(\Lambda) \Pi_n(0) \rho_0 \Pi_n(0)U_{\tau,0}^{\dagger}(\Lambda).
\end{eqnarray}
In the above equation $\Pi_j(t)$ denotes the eigenprojection operators corresponding to the eigenvalue $e_j(t)$ of the Hamiltonian $H\sqlr{\lambda(t)}$ and $U_{\tau,0}(\Lambda)$ is the unitary time evolution operator from the beginning to the end of the force protocol defined in (\ref{U}).
We further elucidate the differences of the work pdfs, (\ref{eq:projfinmeaspowpdf}) and (\ref{pe}), which both are rather involved, by means of a simple example.
\subsection{Landau-Zener}
Next we want to illustrate the differences between the power-based estimates and the TEMA work values
and also the approach to the Zeno limit for a driven two-level system undergoing an avoided level-crossing.
The Hamiltonian for a two-level system driven through an avoided crossing, known as the Landau-Zener(-St\"{u}ckelberg-Majorana) model \cite{LZSM}, is given by
\begin{eqnarray}
H_{LZ}(t) = \Delta \sigma_x + \frac{v t}{2}\sigma_z \label{eq:LZmodel},
\end{eqnarray}
where $\sigma_j$ are the Pauli matrices. In this case the power operator is given by $L = v \sigma_z/2$. Thus the power measurement involves projective measurements in the $\sigma_z$ basis. Digressing from our previous convention, we follow the usual custom and consider a symmetric protocol about $t=0$ that runs between $-\tau/2 \leq t \leq \tau/2$. The Hamiltonian (\ref{eq:LZmodel}) is readily diagonalized yielding for the eigenvalues $e_j(t) = (-1)^j \sqrt{(vt/2)^2 + \Delta^2}$ with $j=1,2$. Hence, the possible TEMA work values are given by:
\begin{eqnarray}
\mathcal{W}^{e} = \flrlr{-E_{\mathrm{max}},0,E_{\mathrm{max}}}, \label{eq:weLZ}
\end{eqnarray}
with $E_{\mathrm{max}} = 2 \sqrt{ (v\tau/4)^2 + \Delta^2 }$, whereas the set of possible work values based on $N$ power measurements becomes
\begin{eqnarray}
\mathcal{W}^p = \flrlr{\frac{v \tau}{2(N+1)} g,\,\,\,\, g =-N,-N+2,\cdots , N}, \label{eq:wpLZ}
\end{eqnarray}
clearly showing a fundamental difference between the two approaches. We also note here that for this example the range of TEMA work bounds the possible values in the power measurement estimate. The range of the latter increases with the number $N$ reaching $\pm v \tau/2$ for $N\to \infty$. The magnitude of the maximum work value possible from the TEMA, $E_{\mathrm{max}}$, is always larger than $v \tau/2$ but in the limit of large enough $\tau$ such that $v \tau/4 \gg \Delta$, it approaches the latter.
According to (\ref{eq:projinfmeaspowpdf}), the work pdf for the LZ problem estimated by power measurements in the Zeno limit becomes
\begin{eqnarray}
p_{\Lambda p}^{(\infty)}(w) = \bra{z_+} \rho_0 \ket{z_+} \delta(w-v\tau/2) + \bra{z_-} \rho_0 \ket{z_-} \delta(w+v\tau/2) \label{eq:LZinfpowpdf},
\end{eqnarray}
where $\ket{z_{\pm}}$ denote the eigenstates of $\sigma_z$ and $\rho_0 = e^{-\beta H_{LZ}(-\tau/2)}/Z(-\tau/2)$ is the initial density matrix. On the other hand, the work pdf from the TEMA is given by \cite{Pekola13}
\begin{eqnarray}
p_{\Lambda e}(w) = p_e P_{LZ}\delta(w+E_{\mathrm{max}}) + (1-P_{LZ})\delta(w) + p_g P_{LZ} \delta(w-E_{\mathrm{max}}),
\end{eqnarray}
with $p_g = \parlr{1+e^{-\beta E_{\mathrm{max}}}}^{-1}$, $p_e = 1-p_g$ and $P_{LZ}= e^{-2 \pi \Delta^2/(\hbar v)}$ \footnote{Strictly speaking this expression is valid only for $\tau \rightarrow \infty$ but provides a very good approximation for large finite $\tau$ with $v \tau/4 \gg \Delta $.} denotes the LZ tunneling probability \cite{LZSM} from the ground state at $t=-\tau/2$ to the excited state at $t=\tau/2$.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{par1betahighzenoappv3.pdf}
\caption{Cumulative probability of work estimated from the sum of $N$ power measurements for the LZ problem with $v = 5 \Delta^2/\hbar$ and $\tau = 20 \hbar/\Delta$ ($N = 10$ - red, $N=10^2$ - blue, $N=10^3$ - green, and $N=10^4$ - black with ascending line thickness). The initial temperature is small $\beta E_{\mathrm{max}} = 10$. For large $N$, due to the Zeno effect, a distinct peak appears at the maximum value of $w$ (see text for discussion). For comparison the pdf computed from TEMA is also shown (cyan dashed line). The inset shows the Kolmogorov-Smirnov distance between the TEMA and power-based work estimate as a function of $N$.}
\label{fig:par1betahighzenoapp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{par1betalowzenoappv3.pdf}
\caption{Cumulative probability of work estimated from the sum of $N$ power measurements for the LZ problem for the same system sweep parameters (and same legends) as in figure \ref{fig:par1betahighzenoapp}. The initial temperature is large $\beta E_{\mathrm{max}} = 10^{-1}$. For large $N$, due to the Zeno effect, two distinct peaks appear at the maximum and minimum value of $w$ (see text for discussion). For comparison the cumulative probability computed from TEMA is also shown (cyan dashed line). The inset shows the Kolmogorov-Smirnov distance between the TEMA and power-based work estimate as a function of $N$. }
\label{fig:par1betalowzenoapp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{par1betahighTEMAwpower.pdf}
\caption{Cumulative probability of work from TEMA during a protocol for the LZ problem that is interrupted by $N$ measurements of $\sigma_z$. The system sweep parameters and legends are as in figure \ref{fig:par1betahighzenoapp}, in particular, the color code indicates the same number of power measurements $N$. The initial temperature is small $\beta E_{\mathrm{max}} = 10$. For comparison the cumulative probability computed from $N = 10^4$ measurements of power is also shown (cyan dashed line). The inset shows the Kolmogorov-Smirnov distance between the TEMA and power-based work estimate as a function of $N$.}
\label{fig:par1betahighTEMApower}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{par1betalowTEMAwpower.pdf}
\caption{Cumulative probability of work from TEMA during a LZ sweep that is interrupted with power measurements. System parameters (and legend) are as in figure \ref{fig:par1betahighTEMApower}. The initial temperature is large $\beta E_{\mathrm{max}} = 0.1$. }
\label{fig:par1betalowTEMApower}
\end{figure}
Figures \ref{fig:par1betahighzenoapp} and \ref{fig:par1betalowzenoapp} display the cumulative probabilities of work computed for different numbers of power measurements $N$ and for two different temperatures.
The cumulative probability $Q(w) = \int_{-\infty}^w dw' p(w)$ quantifies the probability to find a work value $w' \leq w$ based on the pdf $p(w)$. As a quantitative measure between two work distributions we use the Kolmogorov-Smirnov (KS) distance $D_{KS}$, corresponding to the maximum absolute difference between the respective cumulative probabilities, i.e. $D_{KS}= \max_w |Q_1(w) -Q_2(w)|$. For the parameter values chosen (see figure caption) the initial eigenstates of the LZ system have significant overlaps with the $\sigma_z$ eigenstates, i.e., $\ket{e_{\mp}(-\tau/2)}_{LZ} \approx \ket{z_{\pm}}$ (the minus sign on l.h.s stands for the ground state). As a result, in the low temperature example in figure \ref{fig:par1betahighzenoapp} one can see that as $N$ is increased, in agreement with (\ref{eq:LZinfpowpdf}), the largest jump of the cumulative probability occurs near $w \approx v \tau/2$.
Only a small jump at $w \approx -v\tau/2$ is visible due to the low initial occupation of the excited state. For the high temperature example in figure \ref{fig:par1betalowzenoapp}, two jumps of comparable height appear in the large $N$ limit as both eigenstates are occupied in the initial distribution. The cumulative probability from the TEMA is also plotted for reference and it is apparently quite different from the pdf for the power-based work estimate. The largest Kolmogorov-Smirnov distance between the TEMA and the power-based work distributions results in the Zeno limit.
In earlier work \cite{Campisi11}, it has been shown that though fluctuation theorems for TEMA work are robust to measurements during the protocol, the work statistics itself can be strongly modified. In this light it is interesting to compare the pdf for the power-based work estimate $p_{\Lambda p}^{(N)}(w)$ with the TEMA work pdf, $p_{\Lambda e}^{(N)}$, in the presence of ($N$) measurements of power during the protocol. Note that while computing $p_{\Lambda e}^{(N)}$, we sum over all possible results of the intermediate power measurements. In figures \ref{fig:par1betahighTEMApower} (low initial temperature) and \ref{fig:par1betalowTEMApower} (high initial temperature), the cumulative probability for the TEMA work pdf with varying number $N$ intermediate power measurements are displayed for the LZ problem. For comparison the cumulative probability for the power-based work estimate with $N=10^4$ measurements is also plotted. In a qualitative way, the two distributions approach each other in the limit of $N \rightarrow \infty$, but the KS distance (shown in the insets) increases with $N$. This apparent contradiction can be resolved by comparing the $N \rightarrow \infty$ limit of the TEMA work pdf with intermediate power measurements for the LZ problem given by
\begin{eqnarray}
p_{\Lambda e}^{\infty}(w) &=& \frac{1}{2}\sqlr{1+\left(\frac{v \tau}{2E_{\mathrm{max}}}\right)^2}\sqlr{p_g \delta(w-E_{\mathrm{max}}) + p_e \delta(w+E_{\mathrm{max}})} \label{eq:LZintmeaszenopdf}\\
&+& \frac{1}{2}\sqlr{1-\left(\frac{v \tau}{2E_{\mathrm{max}}}\right)^2} \delta(w) \nonumber
\end{eqnarray}
with the equivalent expression for the power-based estimate (\ref{eq:LZinfpowpdf}). In the low temperature case (hence $p_g \approx 1$) depicted in figure \ref{fig:par1betahighTEMApower}, the KS distance is maximised at large $N$, because the locations at which the cumulative probabilities perform the largest jumps differ. For the power-based estimate, from (\ref{eq:LZinfpowpdf}), the jump occurs at $w = v \tau/2$ whereas for the TEMA based work the jump occurs at $w = E_{\mathrm{max}} = 2\sqrt{(v\tau/4)^2+\Delta^2}$. Hence even for $N \rightarrow \infty$, only in the limit that $\Delta/(v \tau) \rightarrow 0$, the power-based work and the one from TEMA with intermediate measurements agree. Note that in the limit $\Delta/(v \tau) \rightarrow 0$, the power operator always commutes with the hamiltonian.
\section{Joint statistics of work from TEMA and power measurements} \label{sec:jointstat}
To further elucidate the differences between the TEMA and power-based work estimates, we consider a modified set-up that allows the simultaneous study of both approaches. In order to combine these two approaches, we consider a thought experiment where in addition to $N$ power-measurements of the type described in the previous section, also energy measurements at the beginning and the end of the force protocol are performed according to the TEMA scheme. The outcome of energies $e_n(0)$ and $e_m(\tau)$ at the beginning and the end of the force protocol and of a sequence $\mathbf{x} = (x_{\alpha_1}, x_{\alpha_2}, \cdots, x_{\alpha_N})$ of eigenvalues of the generalized coordinate $\hat{X}$ at the equally spaced times of measurement $t_1,t_2,\cdots, t_N$ occurs with the joint probability $\mathscr{P}_\Lambda(m,\mathbf{x},n)$ given by
\begin{eqnarray}
\mathscr{P}_\Lambda(m,\mathbf{x},n) &= \mbox{Tr}\, \sqlr{\Pi_m(\tau)V_{\Lambda}(\mathbf{x})\Pi_n(0)\rho_0 \Pi_n(0)V_{\Lambda}^{\dagger}(\mathbf{x})} \nonumber\\
&=\abs{\bra{m;\tau}U_{N+1}\ket{\varphi_{\alpha_N}}}^2\abs{\bra{\varphi_{\alpha_1}}U_{1}\ket{n;0}}^2 \nonumber\\
&\quad \times \prod_{k=1}^{N-1} \left | \bra{\varphi_{\alpha_{k+1}}} U_{k+1} \ket{\varphi_{\alpha_{k}}}\right |^2\bra{n;0}\rho_0\ket{n;0}
\label{eq:jointprobindx},
\end{eqnarray}
where, as before, $\Pi_n(t)$ denotes the projector on the eigenstate $\ket{n;t}$\footnote{We do not allow for degeneracy of the energy eigenvalues for the sake of notational simplicity.} of $H[\lambda(t)]$ belonging to the eigenenergy $e_n(t)$, and $V_\Lambda$ is defined in (\ref{eq:jointprobproj}).
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{corrfacjp_fin.pdf}
\caption{Correction factor $\langle e^{-\beta(w_e-w_p)}\rangle^{N}_{\bar{\Lambda}}$ in the integral fluctuation theorem (\ref{eq:jarwpow}) as a function of the number $N$ of power measurements for the LZ problem. The red dotted curve is for the parameters $v = 5 \Delta^2/\hbar$ and the black dashed line is for a faster sweep rate of $v = 40 \Delta^2/\hbar$ with $\tau = 20 \Delta/\hbar$. In both cases the initial temperature is low and satisfies $\beta E_{\mathrm{max}} = 10$. The solid red (thick) and black (thin) lines represent the correction factor computed in the $N\rightarrow \infty$ limit for the slow and fast sweep respectively.}
\label{fig:corrfacjp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{corwewpfnv.pdf}
\caption{Correlation between the TEMA work $w_e$ and power-based estimate $w_p$ for the LZ system is displayed as a function of the sweep velocity $v\tau/\Delta$. The sweep time is fixed at $\tau=20\hbar/\Delta$. The initial temperature, $\beta E_{\mathrm{max}}(v\tau/\Delta)=10$(low temperature, red solid) and $\beta E_{\mathrm{max}}(v\tau/\Delta)=0.1$ (high temperature, black dashed line), is scaled to ensure the same population distribution at different $v$.}
\label{fig:corwewpfnv}
\end{figure}
This joint probability and the according probability for the time-reversed process, $\mathscr{P}_{\bar\Lambda}(m,\bar{\mathbf{x}},n)$, satisfy a detailed balance-like relation of the form
\begin{equation}
\mathscr{P}_\Lambda(m,\mathbf{x},n)= e^{-\beta(\Delta F +e_n(0) -e_m(\tau))}\mathscr{P}_{\bar\Lambda}(n,\bar{\mathbf{x}},m),
\label{dbr}
\end{equation}
where $\bar{x} = \parlr{x_N,x_{N-1}, \cdots, x_1}$. It holds under two conditions. First, the Hamiltonian $H_0$ and the generalized coordinate $\hat{X}$ must be time reversal invariant meaning that both operators commute with the anti-unitary time reversal operator $\Theta$, $[H_0,\Theta]=[\hat{X},\Theta]=0$.\footnote{This condition can be modified for Hamiltonians depending on fields changing their parity under time reversal. For details see e.g. \cite{TMYH,AG}} This property allows one to express the time evolution running in the backward time direction in terms of a propagator proceeding along the forward arrow of time \cite{Campisireverr11,AG} in the following way:
\begin{eqnarray}
U^\dagger_{t,s}(\Lambda) = \Theta^\dagger U_{\tau-s,\tau-t}(\bar{\Lambda}) \Theta,
\label{tr}
\end{eqnarray}
The second condition requires that the forward and the backward processes start from canonical equilibrium states $e^{-\beta H[\lambda(0)]}/Z(0)$ and $e^{-\beta H[\lambda(\tau)]}/Z(\tau)$, respectively. For a detailed derivation we refer to \cite{Watanabe14a}. In terms of the conditional probability
\begin{equation}
\mathscr{P}_\Lambda(m,\mathbf{x}|n) = \mathscr{P}_\Lambda(m,\mathbf{x},n) e^{\beta(e_n(0) -F(0))}
\label{pfcmxn}
\end{equation}
with the free energy $F(0)$ of the initial state and the analogous conditional probability for the backward process
\begin{equation}
\mathscr{P}_{\bar{\Lambda}}(n,\bar{\mathbf{x}}|m) = \mathscr{P}_{\bar{\Lambda}}(n,\bar{\mathbf{x}},m) e^{\beta(e_m(\tau) -F(\tau))},
\label{pbcmxn}
\end{equation}
one obtains the microcanonical, detailed balance-like relation\footnote{If the energy eigenvalues $e_n(t)$ are degenerate the respective multiplicities $d_n(t)$ have to be taken into account as $ \mathscr{P}_\Lambda(m,\mathbf{x}|n) d_n(0)= \mathscr{P}_{\bar{\Lambda}}(n,\bar{\mathbf{x}}|m) d_m(\tau)$, \cite{TMYH}.}
\begin{equation}
\mathscr{P}_\Lambda(m,\mathbf{x}|n) = \mathscr{P}_{\bar{\Lambda}}(n,\bar{\mathbf{x}}|m)
\label{cdbr}
\end{equation}
generalizing the detailed balance-like relation that holds for the forward and backward conditional probabilities of energy in the absence of intermediate measurements \cite{TMYH}.
The joint pdf $\mathcal{P}^{N}_\Lambda(w_e,w_p)$ for the two work estimates immediately follows as:
\begin{eqnarray}
\mathcal{P}_{\Lambda}^{N}(w_e,w_p) &= \sum_{n,m,\flrlr{{\alpha_i}}} \delta \parlr{w_e-\parlr{e_m(\tau)-e_n(0)}} \delta \parlr{w_p-\sum_{i=1}^N \dot{\lambda}(t_i)x_{\alpha_i}h} \label{eq:jointprobwewp} \nonumber\\
& \quad \times \mathscr{P}_\Lambda(m,\mathbf{x},n).
\end{eqnarray}
With the according expression for the time-reversed process,
\begin{eqnarray}
\mathcal{P}^N_{\bar{\Lambda}}(w_e,w_p) &= \sum_{n,m,\flrlr{{\alpha_i}}} \delta \parlr{w_e+\parlr{e_m(\tau)-e_n(0)}} \delta \parlr{w_p+\sum_{i=1}^N \dot{\lambda}(t_i)x_{\alpha_i}h} \nonumber\\
& \quad \times \mathscr{P}_{\bar{\Lambda}}(n,\bar{\mathbf{x}},m),
\label{eq:jointprobwewptimerev}
\end{eqnarray}
in combination with the detailed balance-like relation (\ref{dbr}) we find a Crooks-type fluctuation theorem for the joint distribution of TEMA and power-based work:
\begin{eqnarray}
\mathcal{P}^N_{\bar{\Lambda}}(-w_e,-w_p) = e^{-\beta \parlr{w_e - \Delta F}}\mathcal{P}^N_{\Lambda}(w_e,w_p) \label{eq:crooksjointprob}.
\end{eqnarray}
In full agreement with the earlier observation that any number of intermediate projective measurements leaves the Crooks relation, and, consequently, the Jarzynski equality, unchanged while modifying the work statistics \cite{Watanabe14a,Campisi11}. One finds for the marginal distribution $ \mathcal{P}^N_{\Lambda,e}(w_e) = \int dw_p \mathcal{P}^N_{\Lambda}(w_e,w_p)$ for the TEMA work
\begin{equation}
\mathcal{P}^N_{\Lambda,e}(w_e) = e^{-\beta(\Delta F -w_e)} \mathcal{P}^N_{\bar{\Lambda},e}(-w_e).
\label{CN}
\end{equation}
i.e.\ the Crooks relation being fulfilled. Note that the presence of the unnoticed power measurements do have their impact rendering $p_{\Lambda e}(w)$ as defined in (\ref{pe}) different from $\mathcal{P}^N_{\Lambda,e}(w_e)$. Yet both work distributions satisfy the Crooks relation in agreement with \cite{CTH10,Campisi11,Watanabe14a}.
For the marginal power-based work pdf $ \mathcal{P}^N_{\Lambda,p}(w_p) =\int dw_e \mathcal{P}^N_{\Lambda}(w_e,w_p)$ only modified fluctuation theorems can be obtained. Integrating (\ref{eq:crooksjointprob}) over all values of $w_e$ one obtains
\begin{equation}
e^{-\beta(\Delta F - w_p)} \mathcal{P}^N_{\bar{\Lambda},p}(-w_p) = \mathcal{P}^N_{\Lambda,p}(w_p) \langle e^{-\beta (w_e - w_p)}|w_p \rangle^N_\Lambda ,
\label{mCp}
\end{equation}
where $\langle \cdot |w_p \rangle^N_\Lambda = \int dw_e \cdot \mathcal{P}^N_{\Lambda}(w_e,w_p)/ \mathcal{P}^N_{\Lambda,p}(w_p)$ denotes a conditional average. This conditional average of the exponentiated work difference determines the modification relative to the strict Crooks relation. In general it depends on the force protocol.
Changing $\Lambda$ into $\bar{\Lambda}$ and performing an integration over all power-based work values, one obtains
an integral fluctuation theorem in the form of a modified Jarzynski equality reading
\begin{eqnarray}
\langle e^{-\beta w_p} \rangle^N_{\Lambda} = e^{-\beta \Delta F} \langle e^{-\beta\parlr{w_e-w_p}}\rangle^N_{\bar{\Lambda}}, \label{eq:jarwpow}
\end{eqnarray}
where $\langle \,\,\cdot \,\, \rangle^N_{\Lambda} = \int dw_e dw_p \cdot \mathcal{P}^N_{\Lambda}(w_e,w_p)$ denotes the average over the joint pdf with $N$ power measurements. The correction factor is now determined by the full average of the exponentiated difference between TEMA and power-based work. In general, it is also a protocol dependent factor. In the limit of a large number of power measurements, we can use the approach in section \ref{sec:powmeas} and in a straightforward manner show
\begin{equation}
\langle e^{-\beta\parlr{w_e-w_p}}\rangle^{\infty}_{\bar{\Lambda}} = \frac{\mbox{Tr}\, e^{-\beta \sqlr{\lambda(\tau)-\lambda(0)}\hat{X}}e^{-\beta H\sqlr{\lambda(0)}}}{Z(\tau)} \label{eq:jointprobcorrfacinf}.
\end{equation}
Hence as we remarked in (\ref{Wec}) if $[H_0,\hat{X}] = 0$ is satisfied, the two estimates of work $w_e$ and $w_p$ coincide in the limit of a large number of power measurements (Zeno limit) and the correction factor is unity. Figure \ref{fig:corrfacjp} depicts this correction factor for the LZ problem for two speeds of sweeping (see figure captions for details). The correction factor behaves non-monotonically with $N$ and can be greater or less than $1$ for small $N$. For large $N$ it tends to the correct limiting value as shown in the figure. Moreover for the diabatic sweep $v = 40 \Delta/\hbar$, the correction factor is in general significantly smaller than the moderate sweep rate case of $v = 5 \Delta/\hbar$. This can be anticipated since in the diabatic sweep case with large $v$, except for a small interval around $t=0$ the total LZ Hamiltonian (\ref{eq:LZmodel}) approximately commutes with the power operator as the diagonal terms $ \pm v t/2$ dominate the off-diagonal coupling $\Delta$.
So far in this section we elucidated some detailed and integral fluctuation relations satisfied by joint TEMA and power-based work estimates and their respective marginals and also pointed to formal differences between these two estimates. One quantitative measure of the relation between the two estimates of work is the mutual information $\Delta I_{\Lambda}(w_e:w_p) \equiv \ln \frac{\mathcal{P}_{\Lambda}^N(w_e,w_p)}{\mathcal{P}_{\Lambda,e}^N(w_e)\mathcal{P}_{\Lambda,p}^N(w_p)}$. While this measure locally quantifies the interdependence of $w_e$ and $w_p$, we consider, the correlation function, as a global measure. It is defined as:
\begin{eqnarray}
\mathrm{Cor}(w_e,w_p) = \frac{\sigma_{w_e,w_p}}{\sigma_{w_e}\sigma_{w_p}} \label{eq:corrdefn},
\end{eqnarray}
where the covariance and standard deviation of the two variable are given by
\begin{eqnarray*}
\sigma_{w_e,w_p} &=& \langle w_e w_p \rangle^N_{\Lambda} - \langle w_e\rangle_{\Lambda}\langle w_p\rangle^N_{\Lambda},\\
\sigma_{w_{j}}^2 &=& \langle w_{j}^2\rangle^N_{\Lambda} - \parlr{\langle w_j\rangle^{N}_{\Lambda}}^2 \,\, ;j=p,e.
\end{eqnarray*}
Here, the averages are performed over the joint pdf given by (\ref{eq:jointprobwewp}). The formal expressions for the correlation in general are not very transparent and instead it is more illuminating to consider a specific example. We choose the LZ problem introduced in section \ref{sec:powmeas} and fix the number of power measurements at $N=100$. With a fixed time of sweep at $\tau= 20\hbar/\Delta$, we ask how the correlation between the two estimates of work varies as the velocity of sweep $v$ is changed. In figure \ref{fig:corwewpfnv}, we present our results for two thermal initial states. For each sweep rate the temperature is chosen such that $\beta E_{\mathrm{max}}(v,\Delta) = \{10,0.1\}$ corresponding to fixed populations of the ground state independent of the sweep rate. In figure \ref{fig:corwewpfnv} we can clearly see that as $v \tau/\Delta$ is increased, the two estimates become more and more correlated. This can again be understood as the effect of making the power term $\frac{vt}{2} \sigma_z$ much larger in magnitude compared to the time-independent part $\Delta \sigma_x$. This effectively renders the commutator between $[H_0,\hat{X}] \sim 0$ for most of the interval and hence the power-based work estimate agrees well with the TEMA based one. Secondly we also see that at the smaller value of $\beta E_{\mathrm{max}}$ corresponding to a larger temperature, since the system is more ``classical", the correlation between the two estimates is better.
In the next section we consider weak continuous measurements of power to estimate the work as opposed to projective measurements considered thus far in the paper.
\section{Weak continuous measurement of power} \label{sec:fintime}
In the previous sections we modelled the monitoring of the supplied power by means of projective measurements of the generalised coordinate $\hat{X}$. As we have shown, in the limit of large number of such measurements, the system dynamics is frozen in the basis of the generalized coordinate and the unitary dynamics generated by the driving plays no role. The work statistics in this limit also differs from the ones determined by the TEMA. In an attempt to mitigate this situation we consider a {\it weak} continuous measurement of the generalised coordinate and determine power from such a measurement. One might hope that in this case although the estimate of work will be affected by errors inherent in a weak measurement process, the measurement backaction will not be so overwhelming as to render the unitary dynamics moot. To this end in what follows we will use the theory of continuous quantum measurement developed in \cite{Contmeaspaps} for our specific situation of determining work statistics from power measurements.
Before considering a particular scenario of continuous coordinate measurements in more detail, we shortly discuss the measurement of the relevant coordinate by means of Gaussian Kraus operators \cite{Sudarshankraus} as a particular example of a weak instantaneous measurement. More general Kraus operators generating a positive operator valued measure (POVM) of the coordinate are discussed in the appendix \ref{app:genpower}.
\subsection{Instantaneous Gaussian coordinate measurements}
We now assume that the measurement of the generalized coordinate $\hat{X}$ is performed with measurement operators consisting of weighted sums of eigen-projection operators rather than of a single one. Choosing Gaussian weights we have
\begin{eqnarray}
M_{\alpha} &=& \frac{1}{\parlr{2 \pi \sigma^2}^{1/4}}\exp \sqlr{-\frac{(\alpha - \hat{X})^2}{4\sigma^2}} \label{eq:Gaussianmeas}\\
&=& \frac{1}{\parlr{2 \pi \sigma^2}^{1/4}}\sum_{n}\exp \sqlr{-\frac{\parlr{\alpha - x_n}^2}{4\sigma^2}} \Pi_{n}^{X},\nonumber
\end{eqnarray}
where $\alpha$ denotes the pointer state indicating the measured value of the coordinate,(Note that this choice is slightly different from the one adopted in Appendix \ref{app:genpower}, where we assume that the set of pointer states consists only of the eigenvalues of the generalized coordinate in contrast to the continuous range of $\alpha$ values in (\ref{eq:Gaussianmeas})) and $\sigma^2$ the variance of the error distribution of measured coordinate values. The work that can be estimated from $N$ generalised measurements of $\hat{X}$ equally spaced in time, is given by $w_p = \sum_{n} h \dot{\lambda}\parlr{t_n} \alpha_n$ and its pdf takes the form (see appendix \ref{app:genpower} for details)
\begin{eqnarray}
p_{\Lambda p}^{(N)} (w)=\displaystyle \int {\prod_{k=1}^N d \alpha_k} \,\, \delta \parlr{w-\sum_{k} h \dot{\lambda}\parlr{t_k} \alpha_k } \mbox{Tr}\, \mathcal{M}^\dagger \mathcal{M} \rho_0 \label{eq:gaussmeaspdf},
\end{eqnarray}
where $\mathcal{M} = M_{\alpha_{N}}(t_N) M_{\alpha_{N-1}}(t_{N-1}) \cdots M_{\alpha_{1}}(t_1)$. In the above, we have used the Heisenberg picture representation of the Gaussian measurement operators introduced in the appendix (\ref{eq:jointprobformgenheisform}). A natural way to define continuous measurements in this framework is to consider weak measurements characterized by a variance that increases proportionally to the inverse of the time $h$ between two measurements, i.e. as $\sigma^2 = 1/(8\kappa h)$ where $\kappa>0$ quantifies the measurement strength. Hence, with an increasing number of measurements less and less information is gained from a single measurement. When we take such a scaling for the variance, the measurement operator (\ref{eq:Gaussianmeas}) can be expanded as
\begin{eqnarray}
M_{\alpha} \propto \exp \sqlr{-2\kappa h (\alpha-\hat{X})^2} = I -2\kappa h (\alpha-\hat{X})^2 + O(h^2) \label{eq:gaussmeasopexpand}.
\end{eqnarray}
In appendix \ref{app:genpower} we take the generalised measurement operators to be independent of the time step. Since the unitary operators can always be written as $U_k = I + O(h)$, to lowest order in $h$ we obtain an expression for the pdf (\ref{eq:genmeaspower}) that depends only on the measurement operators and the initial state of the system, provided that the time-independent part of the Hamiltonian, $H_0$, is symmetric in the $\hat{X}$ basis. In contrast, for the weak measurement case (\ref{eq:gaussmeasopexpand}) implies that both the measurement and unitary operators enter (\ref{eq:gaussmeaspdf}) at the same order in $h$ and the system's state is unchanged to the lowest order. In this situation to the best of our attempts, we are not able to determine a simple expression for the pdf (\ref{eq:gaussmeaspdf}) in the $h \rightarrow 0$ limit. In the next subsection we follow another path to tackle this problem by focusing on the differential equation that describes a system that is under continuous measurement and is simultaneously driven by a change in the Hamiltonian. For the sake of completeness we note that although we cannot find a simple closed form expression for the pdf (\ref{eq:gaussmeaspdf}) in general, for the trivial case of the system Hamiltonian commuting with the measured coordinate $[H_0,\hat{X}] = 0$, we can easily compute the pdf as
\begin{eqnarray}
p_{\Lambda p}^{(\infty)} (w) &=& \frac{4 \kappa}{\pi \int_0^{\tau} dt \dot{\lambda}(t)^2} \nonumber\\
& \mbox{Tr}\, &\sqlr{\exp \flrlr{-\frac{4 \kappa}{\int_0^{\tau} dt \dot{\lambda}(t)^2} \parlr{w - \sqlr{\lambda(\tau) - \lambda(0)} \hat{X}}^2} \rho_0} \label{eq:gaussmeasinf}.
\end{eqnarray}
The only difference to the power-based work pdf for projective measurements, (\ref{eq:projinfmeaspowpdf}) is the replacement of the delta-function by a Gaussian weight under the trace with a variance depending on the force protocol. The average of the exponentiated work yields the expression
\begin{equation}
\langle e^{-\beta w_p} \rangle = e^{\frac{\beta^2}{16 \kappa } \int_0^\tau dt \dot{\lambda}^2(t)} \mbox{Tr}\, e^{-\beta\sqlr{\lambda(\tau)-\lambda(0)}\hat{X}} \rho_0 = e^{\frac{\beta^2}{16 \kappa } \int_0^\tau dt \dot{\lambda}^2(t)} e^{-\beta \Delta F},
\label{ebw}
\end{equation}
where the second equality on the the right-hand side follows from the commutation of $H_0$ and $\hat{X}$. Thus we find that even for the trivial situation with $[H_0,\hat{X}] = 0$ work defined by the integral of weak continuous measurement of power does not satisfy the Jarzynski relation.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{powpapLZmidparsmkap.pdf}
\caption{Cumulative probability estimated from continuous measurement of power in the LZ problem with $v = 5 \Delta^2/\hbar$, $\tau = 20 \hbar/\Delta$ (blue solid line) and measurement strength $\kappa = 0.001 \Delta/\hbar$. The initial temperature is small $\beta E_{\mathrm{max}} = 10$. The results were obtained from 10000-trajectory simulations of the SME (\ref{eq:contsme}). The shaded region represents the error in the computed pdf represented by the solid line (see text). The dashed blue line represents the pdf computed by TEMA.}
\label{fig:paradbbetahighkapsmLZ}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{powpapLZveradblgkap.pdf}
\caption{Cumulative probability of work estimated from continuous measurement of power in the LZ problem with a large measurement strength $\kappa = 4 \Delta/\hbar$. The system parameters are chosen as $v = 0.01 \Delta^2/\hbar$, $\tau = 20 \hbar/\Delta$ to correspond to an adiabatic LZ sweep. The initial temperature is large $\beta E_{\mathrm{max}} = 0.1$ for the red curve and smaller $\beta E_{\mathrm{max}} = 10$ for the blue curve.}
\label{fig:paradbkapzeno}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{powpapLZdiabkapzeno.pdf}
\caption{Cumulative probability of work estimated from continuous measurement of power in the LZ problem with a large measurement strength $\kappa = 4 \Delta/\hbar$. The system parameters are chosen as $v = 40 \Delta^2/\hbar$, $\tau = 20 \hbar/\Delta$ to correspond to a diabatic LZ sweep. The initial temperature is large $\beta E_{\mathrm{max}} = 0.1$ for the red curve and smaller $\beta E_{\mathrm{max}} = 10$ for the blue curve.}
\label{fig:pardiabkapzeno}
\end{figure}
\subsection{Continuous measurements}
In a final attempt we shall relax the assumption inherent in both the projective and the generalized measurement approach that a measurement consists in an instantaneous event that interrupts the unitary dynamics of the system. For that purpose we adopt a model proposed in \cite{Jacobs06}. The time evolution of the density matrix caused by the unitary dynamics in the presence of continually performed measurements of an observable $\hat{X}$ is described by the following non-linear stochastic master equation (SME)
\begin{eqnarray}
\dot{\rho}(t) &=& -\frac{i}{\hbar} \sqlr{H [\lambda(t)],\rho(t)} - \kappa [\hat{X},[\hat{X},\rho(t)]] \label{eq:contsme}\\
&+&\sqrt{2\kappa}\parlr{\hat{X} \rho(t) + \rho(t) \hat{X} - 2 \langle \hat{X}(t) \rangle \rho(t)} \xi(t) \nonumber,
\end{eqnarray}
where $\xi(t)$ denotes Gaussian white noise of unit intensity, i.e. $\langle \xi(t) \xi(s) \rangle = \delta(t-s)$. Apart from the first, Hamiltonian term on the right-hand side, the second and third terms are supposed to describe the impact of continuous measurements of the observable on the system's dynamics. The second term models the average influence of the measurements, and the third, nonlinear and random term accounts for the influence of an individual run of measurements. The meanfield-like non-linearity $2 \langle \hat{X} \rangle \rho$ guaranties the conservation of the normalization of the density matrix in the presence of the measurement-induced fluctuations. The average $\langle \hat{X} \rangle$ is performed with respect to the fluctuating density matrix $\rho$ and therefore it is itself a random quantity. In appendix \ref{app:smederiv}, for the sake of pedagogy, we reproduce a derivation of (\ref{eq:contsme}) following \cite{Jacobs06}. There we utilise Gaussian measurement operators with variances that are scaled inversely with the time step like earlier in this section.
According to the theory of continuous measurements \cite{Jacobs06}, the result $\alpha(t)$ of the measurement of $\hat{X}$ at the instant of time $t$ is given by
\begin{eqnarray}
\alpha(t) = \langle \hat{X} (t) \rangle + \frac{1}{\sqrt{8\kappa}} \xi(t) \label{eq:discoutcome},
\end{eqnarray}
and consequently, the work based on the measurement of the power becomes
\begin{eqnarray}
w_p^{\mathrm{c}} &= \int_0^{\tau} dt \dot{\lambda}(t) \alpha(t) \nonumber \\
&= \int_0^{\tau} \dot{\lambda}(t) \parlr{\langle \hat{X} \rangle(t) + \frac{1}{\sqrt{8\kappa}}\xi(t)} dt. \label{eq:workcontpow}
\end{eqnarray}
The SME is understood in the It\={o} sense. Therefore, the third term on the right-hand side of (\ref{eq:contsme}) disappears upon an average over realizations of the Gaussian white noise yielding for the averaged density matrix $\bar{\rho}$ the linear master equation
\begin{eqnarray}
\frac{d \bar{\rho}}{dt} &=& -\frac{i}{\hbar} \sqlr{H[\lambda(t)],\bar{\rho}} - \kappa [\hat{X},[\hat{X},\bar{\rho} \,]] \label{eq:contsmeavg},
\end{eqnarray}
It describes the state of the system in a non-selective measurement of the coordinate $\hat{X}$.
Before discussing the non-linear SME by means of numerical simulations of an example we shortly mention a linear SME that is obtained from (\ref{eq:contsme}) by disregarding the mean-field-type contribution proportional to the average $\langle \hat{X} \rangle$ \cite{LSE,LSEjacobs}. As a consequence, the normalization of the density matrix is no longer conserved. Even this linear equation is rather complicated for our scenario due to the explicit time dependence of the Hamiltonian. In general we find that it can be solved analytically only if $H_0$ and $\hat{X}$ commute. The resulting work pdf coincides as expected with that of the continuous action of weak instantaneous Gaussian measurements given by (\ref{eq:gaussmeasinf}).
As an example, we generated $10000$ realizations of the density matrix by solving the SME in (\ref{eq:contsme}), displayed in figures \ref{fig:paradbbetahighkapsmLZ}, \ref{fig:paradbkapzeno} and \ref{fig:pardiabkapzeno} for the LZ model (\ref{eq:LZmodel}) in the presence of continuous measurements of $\sigma_z$. For this purpose we used an implicit stochastic Runge-Kutta scheme of order $3/2$ \cite{KloedenPlaten,Steckbook} and checked for numerical convergence of single trajectories using the method of consistent Brownian paths \cite{Steckbook}. We also checked that, for the chosen time step, the density matrix remains normalised to $1$ to high accuracy at all times. We obtained the work pdf by computing a total of 10000 trajectories. For the LZ problem the work follows from (\ref{eq:workcontpow}) for each trajectory as
\begin{eqnarray}
w_p^{\mathrm{c}} = \frac{v}{2} \sqlr{\int_{0}^{\tau} dt \langle \sigma_z \rangle(t) + \frac{1}{\sqrt{8\kappa}} \int_0^{\tau} dt \,\xi(t)}. \label{eq:workdefncontLZ}
\end{eqnarray}
Evidently, the work values in a continuous measurement of $\sigma_z$ span the entire real line due to the additive contribution of the integral of Gaussian white noise. In order to estimate the work pdf from a finite number of simulations, we introduced a binning of the size $\Delta w_p = v h$ with the same time-step $h$ as used in the discretisation of the SME leading to the same set of possible work values as in (\ref{eq:wpLZ}) for $N = \tau/h-1$ measurements. Based on the independence of different trajectory simulations we could estimate the inherent statistical error of the work histograms as the sample variance of 100 blocks of 100 trajectories each.
When the measurement strength $\kappa$ is small such that $\kappa \tau \ll 1$, the work estimate (\ref{eq:workdefncontLZ}) is dominated by the noise term and the resulting pdf is approximately Gaussian distributed with zero mean. This feature is independent of the other system parameters such as $v$ and $\Delta$. In figure \ref{fig:paradbbetahighkapsmLZ} we plot the cumulative probability for the power-based work computed from solving the SME for the same LZ parameters as in section \ref{sec:powmeas}.
When $\kappa \tau \gtrsim 1$, in the work expression the contribution from $\langle \sigma_z \rangle(t)$ dominates over the noise. Thus the behaviour of the cumulative probability in figures \ref{fig:paradbkapzeno} and \ref{fig:pardiabkapzeno} can be well understood from the behaviour of $\langle \sigma_z \rangle(t)$. For instance, from (\ref{eq:workdefncontLZ}) this immediately explains why the range of allowed work values is comparable to the interval $[-v\tau/2,v\tau/2]$. Also note that for our choice of parameters in figures \ref{fig:paradbkapzeno} and \ref{fig:pardiabkapzeno}, we are also in the strong measurement regime of $\kappa > \Delta/\hbar$ for the LZ problem where the coherent dynamics rate $\Delta$ is trumped by the measurement backaction. It is known \cite{Gagen93,Haikka14,Novelli15} that for the LZ system under continuous measurement, the behaviour of single trajectory solutions of (\ref{eq:contsme}) goes from near unitary at very small measurement strength to the so-called ``random-telegraph'' dynamics at the strong measurement $\kappa>\Delta/\hbar$ regime (In \cite{Gagen93,Novelli15} the dynamics in the strong measurement regime is referred to as the quantum Zeno effect. In our work we reserve the latter term for the complete freezing of unitary dynamics achieved by repeated \textit{projective} measurements discussed in section \ref{sec:powmeas}). The random-telegraph behaviour is characterized by the population difference of diabatic states $\langle \sigma_z \rangle$ remaining localised either at $\pm 1$ and undergoing rapid transitions between the two values at different times during the evolution (see discussion in Sec. \Rmnum{3} B of \cite{Gagen93}). In the adiabatic sweep case depicted in figure \ref{fig:paradbkapzeno}, since $v \tau \ll \Delta$ the initial state is off-diagonal (with almost equal distribution amongst the $\pm 1$ eigenstates) in the diabatic i.e. $\sigma_z$ basis and throughout the dynamics the system Hamiltonian has a significant off-diagonal component in the diabatic basis i.e. $[H[\lambda(t)],\sigma_z] \neq 0$. Since there is equal contribution from both $\langle \sigma_z \rangle = \pm 1$ to (\ref{eq:workdefncontLZ}), the work pdf is centered about $w = 0$ in figure \ref{fig:paradbkapzeno}. Nonetheless the power-based cumulative probability does not capture the sharp jump at $w=0$ of the TEMA estimate. In the diabatic sweep in figure \ref{fig:pardiabkapzeno}, the system Hamiltonian and the initial state are almost diagonal in the diabatic basis giving cumulative probabilities that have sharp jumps at $w = v \tau/2$ (low initial temperature) or $w = \pm v \tau/2$ (high temperatures). Also in figure \ref{fig:pardiabkapzeno} the jumps in the power-based and TEMA estimates occur at approximately the same values of $w$. We note that this behaviour owes its explanation, as noted in section \ref{sec:powmeas}, to the fact that for our choice $v \tau/4 \gg \Delta$ the magnitude of maximum work possible in TEMA $E_{\mathrm{max}} \approx v \tau/2$ almost coincides with the location of the jumps in the power measurement cumulative probability.
In summary we find that the work pdf estimated even by weak continuous measurement of the power does not reproduce the behaviour of the TEMA estimate in general. For very weak measurement strengths $\kappa \tau \ll 1$, the estimate is noise dominated and has Gaussian behaviour irrespective of $v$ and $\Delta$. When $\kappa \tau \gtrsim 1$ (strong measurement regime) the pdf depends on the sweep rate $v$ (for a fixed value of $\tau$) which sets the extent to which the total Hamiltonian is off-diagonal in the measurement basis.
\section{Conclusion} \label{sec:conclusion}
This work serves as a detailed illustration of the difficulties involved in defining work in quantum systems in a manner analogous to classical systems. We considered the statistical properties of work, defined as the integral of supplied power, for a quantum system. We showed that even a careful definition of work in terms of repeated measurements of the system's instantaneous power leads to a statistics quite different from the usual two energy measurement approach (TEMA) for defining work. In the power-based approach in general even the Jarzynski equality or the Crooks relation do not hold. In the limit of a large number of projective measurements of the instantaneous power we found that the quantum Zeno effect leads to a freezing of the system's dynamics in the power operator's basis. The statistics of power-based work in this limit is very different from TEMA except for the trivial case when the power operator commutes with the system Hamiltonian at all times. Furthermore we carried out a detailed comparison of the power-based work and the TEMA work estimate by considering the joint pdf of both types of work estimates in a setting that combines both approaches. We obtained a Crooks type fluctuation theorem for the joint probability distribution and modified integral fluctuation theorems for the marginal power-based work distribution. We also studied the correlation between the two estimates of work using the joint probability for the LZ problem. Finally, relaxing the condition of projective, instantaneous measurements, weak continuous measurement of power was discussed within the formalism of stochastic master equation (SME). Using the Landau-Zener problem as an example, we determined the pdf of power-based work numerically and analysed its properties in the limiting cases of weak and strong measurement strengths. We found that since the power operator in general does not commute with the total Hamiltonian of the system, also within the framework of continuous measurements the power-based estimate is not able to reproduce the qualitative features of the TEMA work estimate.
Finally we would like to note that the treatment of continuous measurements in terms of an SME also leads to a Lindblad-type master equation, see (\ref{eq:contsmeavg}), similarly as obtained in various recent publications \cite{Esposito,Horowitz,Horowitz13,Hekking13} which are concerned with the definition of work and heat and the quest for fluctuation theorems in open quantum systems. There, the presence of terms in the equation of motion of the reduced density matrix describing energy-non-conserving effects is caused by the interaction with the environment. In contrast, in the present case it is solely the result of a weak but continuously acting measurement that leads to similar formal structures.
\ack
We thank Sebastian Deffner for useful discussions and the anonymous referee for useful comments. We acknowledge support by the Max Planck Society, the Korea Ministry of Education, Science and Technology (MEST), Gyeongsangbuk-Do, Pohang City, as well as by the Basic Science Research Program through the National Research Foundation of Korea funded by the MEST (No. 2012R1A1A2008028). P.T. thanks the Foundation for Polish Science (FPS) for providing him with an Alexander von Humboldt Polish Honorary Research Fellowship. G.W. also acknowledges support under Project Code (IBS-R024-D1).
|
\section{Introduction}
\
Let $S=\CC[x_1,\ldots,x_n]$ be the \emph{ring of polynomials} in $n$ variables and $\Omega=\bigoplus_{r\geq 0} \Omega^r$ the \emph{algebra of differential forms} of $S$ over $\CC$, where $\Omega^r$ denotes the module of $r$-differential forms. Since $\Omega^r$ has a natural structure of graded $S$-module, we will decompose it as $\Omega^r=\bigoplus_{\substack{b\geq 0}}\Omega^r(b)$, where we assign degree $+1$ to each $dx_i$. We stress the fact that this (second) grading is the one given by the \emph{Lie derivative} with respect to the \emph{radial vector field} $R=\sum_{i=1}^n x_i\frac{\partial}{\partial x_i}$, \emph{i.e.}, for $\tau\in\Omega^r(b)$ we have
\[
\LL_R(\tau) = b\ \tau
\]
as we will recall in eq. \ref{lie2}. In general, for a graded module, we will note in parenthesis the homogeneous component of the given degree; in the case of $\Omega$ it will always be the one given by the Lie derivative with respect to the radial vector field $R$.
\
For fixed $q,a$, suppose we are given a differential operator of order one $D:\Omega\rightarrow{}\Omega$, see Definition \ref{def-op-dif}, such that
\[
D(\Omega^r(b))\subset\Omega^{r+q}(b+a)
\]
for every $r,b\geq 0$.
\
Regarding $D$ as a morphism of $\CC$-vector spaces would probably lead to finding kernel and image of infinite dimension; on the other side, trying to compute these dimensions on each homogeneous component $\Omega^r(b)$ would force to know $D$ in complete detail. Instead, a module structure on $\Omega$ attached to $D$ would lead to more precise and computable information; such as a finite set of generators or a Hilbert polynomial associated to its kernel, or image, exposing discrete invariants.
\
Both sheaves of principal parts, see \cite{egaivIV}, and complexes of differential operators, see \cite{herrera-lieberman}, solve problems of linearization for $D$. In these works the solution is universal (in a categorical sense) and, as part of the construction, the operator $D$ changes as well as the domain. This fact makes the relation between $D$ and its linearized version difficult to track down.
\
Our approach to the problem of linearization is focused on the multiplicative action of $S$ over $\Omega$, leaving the operator $D$ invariant as well as the $\CC$-vector space structure on $\Omega$.
\
As a first example, let us consider the usual \emph{exterior differential} $d:\Omega\rightarrow{}\Omega$ and denote by $i_R$ the \emph{contraction} with the radial vector field $R$. As we will show later in Section \ref{ext-diff}, for $f\in S(c)$ and $\tau\in\Omega^r(b)$, where the grading is taken with the Lie derivative with respect to the radial vector field R, in this work we introduce the following action
\begin{equation}\label{ej1}
f\dt \tau := \frac{b}{b+c}\ \left(\ f\tau\ +\ \frac{1}{b}\ df\wedge i_R\tau \ \right)
\end{equation}
which makes the exterior differential $d:\Omega\rightarrow{}\Omega$ a morphism of $S$-modules (when $b,c=0$ we adopt the usual multiplication). Even if we take a different approach, this can be readily verified by direct computations.
\
In Section \ref{definition} we first give a formal definition of the deformations of the exterior algebra $\Omega$ over $S$ that we propose, Definition \ref{def-lin}, and prove a general condition condition for these modules to be finitely generated.
\
In Section \ref{theorem} we state our main result, Theorem \ref{teo1}, which is a classification of a distinguished class of differential operators that allow a linearized structure as in Definition \ref{def-lin}.
This classification exploits the decomposition of a differential operator in terms of a linear map plus a derivation which, in turn, can be decomposed as the Lie derivative plus a contraction with respect to vector valued differential forms. We give the details and definitions of this decomposition in Section \ref{definitions}
\
Finally, in Section \ref{applications} we present two examples of differential operators that allow the linearized structures that we defined.
\
\subsection{Geometric Motivation}
Even if our approach to the subject is purely algebraic, there is a geometric nature in our work that we would like to state here.
A codimension one foliation in projective space $\PP^{n-1}$ of degree $a-2$, $Fol^1(\PP^{n-1},$ $a-2)$, is given by a differential 1-form $\omega\in\Omega^1(a)$ such that it descends to projective space, \emph{i.e.}, $i_R\omega=0$, and such that verifies the Frobenious integrability condition $\omega\wedge d\omega = 0$. As it is shown in \cite{fji}, the Zariski tangent space to the space of such foliations can be parameterized by
\[
T_\omega Fol^1(\PP^{n-1},a-2) = \{ \eta\in\Omega^1(a)\ :\ i_R\eta=0 \text{ and } \omega\wedge d\eta+ d\omega\wedge \eta = 0\}.
\]
Using the second equation, C. Camacho and A. Lins-Neto, in \cite{caln}, define the following notion of regularity of an integrable, homogeneous, differential 1-form and prove an associated stability result.
By looking at $\omega$ as a homogeneous affine form in $\CC^{n}$, $\omega$ is said to be regular if for every $a<e$ the graded complex of homogeneous elements
\begin{equation}\label{reg-intro}
\xymatrix@R=0pt{
T(a-e) \ar[r]^-{} & \Omega^1(a) \ar[r]^-{} & \Omega^3(a+e)\\
X \ar@{|->}[r] & \LL_X(\omega) & &\\
& \save[] *{\eta\ }="s" \restore & \save[] +<40pt,0pt> *{\ \omega\vartriangle\eta:=\omega\wedge d\eta + d\omega\wedge \eta}="t" \restore & \ar@{|->}"s";"t"
}
\end{equation}
has trivial homology in degree 1, where we denote as $T:=(\Omega^1)^*$ to the module of \emph{vector fields} and assign degree $-1$ to each $\frac{\partial}{\partial x_i} := dx_i^*$.
\
Studying the function $a\longmapsto \varphi_{\omega}(a):=dim_\CC\left(Ker(\omega\vartriangle - )(a)\right)$, for every $a\in\NN$, it was clear the necessity of a better understanding of the differential operator $\omega\vartriangle - $ what led us to the present work (among other things, as we will briefly mention in \ref{reg-complex}).
\subsection*{Acknowledgements}
I would like to thank Matias del Hoyo, Alicia Dickenstein and Federico Quallbrunn for their suggestions and meaningful discussions.
\
\section{Basic Definitions and General Setting}\label{definitions}
\
Along this section we give the definitions of our main objects of study, which are derivations, differential operators and vector valued differential forms on $\Omega$. With these objects we state a decomposition theorem for a differential operator on $\Omega$, see Theorem \ref{decomp}. Such decomposition theorem, as far as we know, is missing in the literature.
\
Let us first recall some basic definitions, see \emph{e.g.} \cite{michor}:
\
A \emph{linear map of degree $q$} in $\Omega$ is a linear map $\varphi:\Omega\rightarrow{} \Omega$ such that $\varphi(\Omega^r)\subset \Omega^{r+q}$ for every $r\geq 0$.
\
A \emph{derivation of degree $q$} in $\Omega$ is a linear map $D_0:\Omega\rightarrow{}\Omega$ of degree $q$ such that
\[
D_0(\mu\wedge\tau) = D_0(\mu)\wedge\tau + (-1)^{rq}\ \mu\wedge D_0(\tau)
\]
for $\mu\in\Omega^r$ and $\tau\in\Omega^s$. If $D\left(\Omega^r(b)\right)\subset \Omega^{q+r}(b+a)$ we will say that $D$ has \emph{bidegree $(q,a)$} and denote by $\der^q(\Omega)$ and $\der^{(q,a)}(\Omega)$ the spaces of derivations of degree $q$ and bidegree $(q,a)$ respectively. In case that $D_0$ is $S$-linear we will say that $D_0$ is a \emph{linear derivation}.
\
We recall from the introduction that we will write as $T:=(\Omega^1)^*$ to the module of vector fields and assign degree $-1$ to each $\frac{\partial}{\partial x_i} := dx_i^*$.
In $\Omega$ we have the exterior differential $d$ which is a derivation of degree 1 and, for $X\in T$, the contraction $i_X$ which is a linear derivation of degree $-1$.
\
Let $\varphi,\psi:\Omega\rightarrow{} \Omega$ be two linear maps of degree $q$ and $r$ respectively. We define the \emph{(graded) Lie bracket} $[\varphi,\psi]$ as
\[
[\varphi,\psi]:=\varphi\circ\psi - (-1)^{qr}\psi\circ\varphi.
\]
\
Following Cartan's formula, see \emph{e.g.} \cite{warner}, the \emph{Lie derivative} with respect to $X$, $\LL_X$, can be computed as
\begin{equation*}
\LL_X = [i_X,d] = i_Xd + di_X
\end{equation*}
which is a derivation of degree $0$.
Taking $\tau\in\Omega^r(b)$, the Lie derivative with respect to the radial vector field $R$, verifies the well known formula, see \emph{e.g.} \cite{jou},
\begin{equation}\label{lie2}
\LL_R(\tau) = i_Rd\tau + di_R\tau = b \ \tau
\end{equation}
which decomposes in a unique way a differential form in a radial plus and exact form.
\
A linear derivation $D_0$ of degree $q$ is uniquely determined by its restriction to 1-forms $D_0:\Omega^1\rightarrow{} \Omega^{q+1}$, which can be viewed as an element $L\in\Omega^{q+1}\otimes T$.
\
For $q\geq -1$, we will say that $L\in\Omega^{q+1}\otimes T$ is a \emph{vector valued differential form} and note $\Omega^{q+1}_{T}$ to the space of such elements. For $\tau\in\Omega^r$, we define the contraction $i_L\tau\in\Omega^{q+r}$ by the formula
\begin{align}\label{contraction}
i_L\tau &(X_1,\ldots,X_{q+r}) := \nonumber \\
&= \frac{1}{(q+1)!(r-1)!}\sum\limits_{\sigma\in S_{q+r}} sign(\sigma) \ \tau\left(L\left(X_{\sigma(1)},\ldots,X_{\sigma(q+1)}\right),X_{\sigma(q+2)},\ldots,X_{\sigma(q+r)}\right)
\end{align}
for $X_1,\ldots,X_{q+r}\in T$ and $S_{q+r}$ the permutation group of $q+r$ elements.
\
We recall from \cite{michor} the following two propositions that give a classification of derivations:
\begin{prop}\label{prop}If $L\in\Omega^{q+1}_{T}$, then $i_L\in\der^q (\Omega)$ and any linear derivation is of this form.
\end{prop}
\begin{proof}
We just notice that we are using the the identification of a differential $r$-form as an alternating map $\left(T\right)^{\otimes r}\rightarrow{} S$, see \emph{e.g.} \cite{warner}. For the rest, we follow \cite[Chapter IV, 16.2, p.~192]{michor}.
\end{proof}
\
Using eq. 3 and Proposition \ref{prop} we can define the Lie derivative for a vector valued differential form, $K\in\Omega^{q}_{T}$, as
\[
\LL_K:=[i_K,d]\in\der^q(\Omega).
\]
\
\begin{prop}\label{decomp1} If $D_0\in \der^q(\Omega)$, then there exists unique $K\in\Omega^{q}_{T}$ and $L\in\Omega^{q+1}_{T}$ such that
\begin{equation*}\label{der}
D_0 = \LL_K+i_L.
\end{equation*}
\end{prop}
\begin{proof}
See \cite[Chapter IV, 16.3, p.~193]{michor}.
\end{proof}
\
For $\tau\in\Omega$ we will denote by $\lambda_{\tau}$ the endomorphism of \emph{left multiplication} by $\tau $ in $\Omega$. It is immediate to see that a linear map $\varphi:\Omega\rightarrow{} \Omega$ of degree $q$ is $\Omega$-linear, \emph{i.e.} $\varphi(\mu) = \varphi(1)\wedge\mu$, if and only if
\[
[\varphi, \lambda_\tau] = 0
\]
for every $\tau\in \Omega$. We then define:
\
A \emph{differential operator of order 1 and degree $q$}, or simply a \emph{differential operator of degree $q$}, in $\Omega$ is a linear map $D:\Omega\rightarrow{}\Omega$ of degree $q$ such that
\begin{equation}
\label{def-op-dif}
[\ [D, \lambda_\mu]\ ,\lambda_\tau] = 0
\end{equation}
for all $\mu,\tau\in \Omega$. If $D\left(\Omega^r(b)\right)\subset\Omega^{q+r}(b+a)$ we will say that $D$ has bidegree $(q,a)$ and denote by $\dif^{\,q}(\Omega)$ and $\dif^{(q,a)}(\Omega)$ the spaces of differential operators of degree $q$ and bidegree $(q,a)$ respectively.
\
Even if the following proposition is very well known, we add a proof just to show that the sign rule arising from the skew commutativity of $\Omega$ does not make any conflicts.
\
\begin{prop}\label{decomp2} Let $D\in\dif^q(\Omega)$. Then $D$ can be decomposed as
\[
D = \left(D - \lambda_{D(1)}\right) + \lambda_{D(1)}
\]
where $D - \lambda_{D(1)}\in\der^q(\Omega)$ and $\lambda_{D(1)}$ is a linear map.
\end{prop}
\begin{proof}
Evaluating at 1 the formula $[\ [D, \lambda_\mu]\ ,\lambda_\tau] = 0 $ we get
\begin{align*}
D(\mu\wedge\tau) + D(1)\wedge\mu\wedge\tau = D(\mu)\wedge\tau + (-1)^{qr} \mu\wedge D(\tau).
\end{align*}
And by subtracting $-2\ D(1)\wedge\mu\wedge\tau$ in both sides we see that
\begin{align*}
D(\mu\wedge\tau) - D(1)\wedge\mu\wedge\tau = \big(D(\mu)-D(1)\wedge\mu\big)\wedge\tau + (-1)^{qr} \mu\wedge\big( D(\tau)-D(1)\wedge\tau\big)
\end{align*}
showing that $D-\lambda_{D(1)}\in\der^q(\Omega)$.
\end{proof}
\
As a direct corollary of Proposition \ref{decomp1} and Proposition \ref{decomp2}, we have the following decomposition for a differential operator:
\begin{teo}\label{decomp} Let $D\in\dif^q(\Omega)$. Then $D$ can be written as
\[
D = \LL_K + i_L + \lambda_{\mu}
\]
for unique $K\in\Omega^{q}_{T}$, $L\in\Omega^{q+1}_{T}$ and $\mu\in\Omega^q$.
\end{teo}
\
\section{Deformations of the Exterior Algebra}\label{definition}
\
In Definition \ref{def-lin} we give formal definition of the deformations of $\Omega$ induced by a differential operator $D\in\dif^{(q,a)}(\Omega)$ and then, in Proposition \ref{accion-gen}, we give conditions to these modules to be finitely generated.
As eq. \ref{ej1} shows, these deformations have some denominators in the formula, that can be zero in low degrees. Because of that, we first need a technical definition that will allow us to avoid this situation.
\
Two graded $S$-modules $M$ and $N$ are said to be \emph{stably isomorphic} if there exists an $n_0\in\NN$ such that $M(k)\simeq N(k)$ for every $k\geq n_0$.
\
Let $\widetilde{\Omega}$ be an algebra of differential forms stably isomorphic to $\Omega$. Without loss of generality we can assume that
\[
\widetilde{\Omega} = \bigoplus_{\substack{r\geq 0\\b\geq n_r}} \Omega^r(b)
\]
for some $n_r\in\NN$.
\
\begin{deff}\label{def-lin} Let $D\in\dif^{(q,a)}(\widetilde{\Omega})$. For $f\in S(c)$ and $\tau\in\Omega^r(b)$ we define the following action of $S$ in $\Omega^r(b)$
\[
f\gt \tau\ =\ \alpha \ f\tau\ +\ \beta\ df\wedge i_R\tau
\]
where $\alpha = \alpha(r,b,c)$ and $\beta = \beta(r,b,c)$ verify the conditions
\begin{enumerate}
\item[a)] $\alpha(-,-,0)= 1$
\item[b)] $D(f\gt\tau) \ = \ f\ \gt D(\tau)$
\item[c)] $(gf)\gt \tau = g\gt(f\gt \tau)$.
\end{enumerate}
We will note $\Omega_D$ to $\widetilde{\Omega}$ under this action from $S$.
\end{deff}
\
In case such $\alpha$ and $\beta$ exists, $\Omega_D$ gets a structure of a graded $S$-module extending the usual multiplication from $\CC$ and
\[
\xymatrix{\Omega_D \ar[r]^-{D} & \Omega_D}
\]
is $S$-linear.
\
It is clear that the modules $\Omega^r_D$ are stably free of rank $\binom{n}{r}$. Next we give a condition for $\Omega_D$ to be finitely generated.
\
\begin{prop}\label{accion-gen} Let $\Omega^r_D = \bigoplus\limits_{\substack{b\geq n_r}} \Omega^r_D(b)$ with $n_r>n$ and let $\alpha(r,b,c)$ and $\beta(r,b,c)$ as in Definition \ref{def-lin}. If for every $b\geq n_r$ we have
\[
\alpha(r,b,1)^2-\beta(r,b,1)^2 \neq 0
\]
then $\Omega^r_D$ is finitely generated by elements of degree $n_r$.
\end{prop}
\begin{proof}Let us fix ($r$ and) $b$ and denote $\alpha = \alpha(r,b,1)$ and $\beta = \beta(r,b,1)$.
Consider the multi-index $\gamma = (\gamma_1,\ldots,\gamma_n)\in\NN_0$ such that $\sum\limits_{i=1}^n \gamma_i = g+1=b+1-r$ and let $I=\{i_1,\ldots,i_r\}\subset \{1,\ldots,n \}$.
\
Take $b\geq n_r$. Every element of $\Omega^r_D(b+1)$ it is a sum of elements of the form
\[
x^\gamma dx_I
\]
where we write $x^\gamma=\prod\limits_{i=1}^{n} x_i^{\gamma_i}$ and $dx_I = dx_{i_1}\wedge \ldots \wedge dx_{i_r}$.
\
Let $k\in\{1,\ldots,n\}$ such that $\gamma_k\neq 0$. Then we have the formula
\begin{align}\label{ec1}
x_k \gt \left(x^{\gamma-e_k} dx_I\right) &= \alpha\ x^\gamma dx_I + \beta\ \sum_{j=1}^r\ x^{\gamma-e_k+e_{i_j}}\ dx_k\wedge i_{\frac{\partial}{\partial x_{i_j}}}dx_I
\end{align}
where we note $e_i$ to the $i$-th canonical vector.
\
If $k\in I$, then eq. \ref{ec1} equals to $(\alpha + \beta)\ x^\gamma\ dx_I$. Since $\alpha^2-\beta^2\neq0$ we have
\begin{equation}\label{gen1}
\frac{1}{a+b}x_k \gt \left(x^{\gamma-e_k} dx_I\right) = x^\gamma\ dx_I
\end{equation}
when $k\in I$ and $\gamma_k\neq 0 $.
\
On the other side, assume that for all $k$ such that $\gamma_k\neq 0$ we have $k\notin I$. By hypothesis we know $b\geq n_r>n$. Replacing $b>n$ in the equation $b=g+r$ we have $g>n-r$. Then, necessarily there exists $k$ such that $\gamma_k\geq 2$.
\
Let $\ell\in\{1,\ldots,r\}$ such that $\ell\in I$ and $k$ such that $\gamma_k\geq 2$. We have
\begin{align}\label{ec2}
x_\ell &\gt \left(x^{\gamma-e_k}dx_k\wedge i_{\frac{\partial}{\partial x_{\ell}}}dx_I\right) = \nonumber\\
&= \alpha\ x^{\gamma-e_k+e_\ell}\ dx_k\wedge i_{\frac{\partial}{\partial x_{\ell}}}dx_I + \beta\sum\limits_{t\in I\cup \{k\}} x^{\gamma-e_k+e_{t}}\ dx_\ell\wedge i_{\frac{\partial }{\partial x_{t}}}\left( dx_k\wedge i_{\frac{\partial}{\partial x_{\ell}}}dx_I\right) = \nonumber\\
&= \alpha\ x^{\gamma-e_k+e_\ell}\ dx_k\wedge i_{\frac{\partial}{\partial x_{\ell}}}dx_I + \beta\ x^\gamma\ dx_I - \beta\sum\limits_{j=1}^r x^{\gamma-e_k+e_{i_j}}\ dx_k\wedge i_{\frac{\partial}{\partial x_{i_j}}}dx_I.
\end{align}
Operating with eq. \ref{ec1} and eq. \ref{ec2} we get
\begin{align}\label{ec3}
\alpha\bigg[x_k \gt \left(x^{\gamma-e_k}\ dx_I\right)\bigg] -\beta \bigg[ x_\ell \gt \left(x^{\gamma-e_k}\ dx_k\wedge i_{\frac{\partial}{\partial x_{\ell}}}dx_I\right)\bigg] = \nonumber\\
= (\alpha^2-\beta^2)\ x^\gamma \ dx_I + (\alpha + \beta)\beta\ \sum\limits_{j=1}^rx^{\gamma-e_k+e_{i_j}}\ dx_k\wedge i_{\frac{\partial}{\partial x_{i_j}}}dx_I.
\end{align}
Let us call $J=(I\backslash\{\ell\})\cup\{k\}$. In each of the terms of the right of eq. \ref{ec3} we have $k\in J$ and $\gamma_k\geq 2$. Then $(\gamma-e_k+e_{i_j})_k\neq 0 $ and we can clear all the right terms of eq. \ref{ec3} following eq. \ref{gen1}, obtaining $(\alpha^2-\beta^2) x^\gamma dx_I$.
\
This way, every element is generated by elements of the previous degree starting from $b+1\geq n_r$ and the result follows.
\end{proof}
\begin{rem} We would like to point out that we do not know what happens with the converse statement, \emph{i.e.}, in the case where $\alpha(r,b,1)^2-\beta(r,b,1)^2=0$.
\end{rem}
\begin{rem} It is worth mentioning that $\Omega^r_D\neq \left(\Omega^1_D\right)^{\wedge r}$. For this, it is enough to compute the products
\[
(z\gt dx)\wedge dy \qquad dx\wedge (z\gt dy)
\]
and see that they are different.
\end{rem}
\
\section{Classification of Differential Operators}\label{theorem}
\
In Theorem \ref{teo1} we classify which differential operators allow a linearization as the one in Definition \ref{def-lin} for a distinguished class of differential operators.
\
There is a distinguished vector valued differential form $Id\in\Omega^1_{T}$, which is the one arising from the identity map on $\Omega^1\rightarrow{} \Omega^1$; then, $Id $ takes the form $Id = \sum_{i=1}^n dx_i\otimes\frac{\partial}{\partial x_i}$. Following eq. \ref{contraction}, we can apply $i_{Id}$ to an $r$-product of 1-differential forms $\tau = \tau_1\wedge\ldots\wedge\tau_r$ and get the formula
\[
i_{Id}(\tau) = r\ \tau.
\]
Then, it is immediate that the usual exterior differential can be computed as
\[
\LL_{Id}(\tau) = d\tau.
\]
\
The space of derivations arising from the module structure of $\Omega$ on the space generated by $(\LL_{Id},i_{Id})$ it is given by $(\LL_{\omega_1\wedge Id},i_{\omega_2\wedge Id})_{\{\omega_1,\omega_2\in\Omega\}}$; this can be easily seen by the following equalities, see \cite[Chapter IV, 16.7, Theorem, p.~194]{michor},
\[
\LL_{\omega\wedge Id} = \omega\wedge\LL_{Id} + (-1)^q d\omega\wedge i_{ Id} \qquad \text{and}\qquad i_{\omega\wedge Id} = \omega\wedge i_{Id}\ ,
\]
where $\omega\in\Omega^r(q)$. This allow us to define:
\begin{deff}We define the space of differential operators of bidegree $(q,a)$ associated to $Id\in\Omega^1_{T}$ as
\[
\dif^{(q,a)}_{Id}(\Omega) = \left\{ \widetilde{\omega}_1\wedge\LL_{Id} + \widetilde{\omega}_2\wedge i_{Id} + \lambda_{\widetilde{\mu}}:\text{for some }\widetilde{\omega}_1\in\Omega^{q-1}(a), \widetilde{\omega}_2,\widetilde{\mu}\in\Omega^q(a)\right\}
\]
\end{deff}
\
\begin{deff}For $q\geq 1$, we define the set $\lin^{(q,a)}(\Omega)$ of \emph{linearizable differential operators} of bidegree $(q,a)$
\begin{align*}\label{lin}
\lin^{(q,a)}(\Omega) = \bigg\{&\omega_1\wedge\LL_{ Id} + \left(\frac{1}{q}d\omega_1+\omega_2\right)\wedge i_{ Id} + (t\ \lambda_{d\omega_1}+\lambda_\mu): \text{ for some } \\
& \omega_1\in\Omega^{q-1}(a)\text{ and } \omega_2,\mu\in\Omega^{q}(a) \text{ such that } \\
&\hspace{4cm} i_R\omega_1=i_R\omega_2 = i_R\mu=0\text{ and }t\in\CC\bigg\}.
\end{align*}
\end{deff}
\
\begin{rem} Notice that $\lin^{(q,a)}\subsetneq\dif^{(q,a)}$. This can be seen by using the decomposition of eq. \ref{lie2} together with the conditions $i_R\omega_1=i_R\omega_2 = i_R\mu=0$, that fixes the exact form of the forms $\widetilde{\omega}_1$, $\widetilde{\omega}_2$ and $\widetilde{\mu}$ to $0$, $\frac{1}{q}d\omega_1$, and $td\omega_1$, respectively.
\end{rem}
\
\begin{rem} One can turn $\lin^{(q,a)}(\Omega)$ in a $\CC$-vector space in the following way: for $D$ and $D'$ in $\lin^{(q,a)}(\Omega)$ defined as
\begin{align*}
D &= \omega_1\wedge\LL_{ Id} + \left(\frac{1}{q}d\omega_1+\omega_2\right)\wedge i_{ Id} + (t\ \lambda_{d\omega_1}+\lambda_\mu)\\
D' &= \omega_1'\wedge\LL_{ Id} + \left(\frac{1}{q}d\omega_1'+\omega_2'\right)\wedge i_{ Id} + (t'\ \lambda_{d\omega_1'}+\lambda_{\mu'})
\end{align*}
we define the addition as
\[\begin{aligned}
&D+D' := (\omega_1+\omega_1')\wedge\LL_{ Id} + \left(\frac{1}{q}\ d\big(\omega_1+\omega_1'\big)+\big(\omega_2+\omega_2'\big)\right)\wedge i_{ Id} + \\
&\hspace{6cm} + \bigg[(t+t')\ \lambda_{d(\omega_1+\omega_1')}+ \lambda_{(\mu + \mu')}\bigg].
\end{aligned}
\]
Also, there is no ambiguity in the way these differential forms are written, since $\frac{1}{q}d\omega_1+\omega_2$ and $t\ d\omega_1+\mu$ are the addition of a radial plus an exact term, for which eq. \ref{lie2} assures uniqueness of writing.
\end{rem}
\
With the following theorem we classify which differential operators arising from $ Id$ can be linearized.
\begin{teo}\label{teo1} Let $D\in\dif_{Id}^{(q,a)}(\widetilde{\Omega})$. Then there exists $\alpha$ and $\beta$ that verify the conditions of Definition \ref{def-lin} making $D$ an $S$-linear operator if and only if $D\in\lin^{(q,a)}(\widetilde{\Omega})$, for $q\geq 1$. If $D\in\lin^{(q,a)}(\widetilde{\Omega})$ is given by
\[
D(\tau) = \omega_{1}\wedge \LL_{Id} + \left(\frac{1}{q}\ d\omega_{1} + \omega_{2}\right)\wedge i_{Id} + (t \ \lambda_{d\omega_{1}} +\lambda_\mu)
\]
with $\omega_1\neq 0$, then $\alpha$ and $\beta$ can be chosen to be
\[
\alpha(r,b,c) := \frac{b-a\left(\frac{r}{q}+(-1)^{q}\,t\right)}{b+c-a\left(\frac{r}{q}+(-1)^{q}\,t\right)} \qquad \beta(r,b,c) := \frac{\alpha(r,b,c)}{b-a\left(\frac{r}{q}+(-1)^{q}\,t\right)}
\]
When $\omega_1=0$, the usual multiplication law can be used.
\end{teo}
\begin{proof} Take $D\in\dif_{Id}^{(q,a)}(\widetilde{\Omega})$. Then $D$ can be written as
\[
D = \omega_1\wedge \LL_{Id}+ \omega_2\wedge i_L + \lambda_\mu
\]
for some $\omega_1\in\Omega^{q-1}(a)$ and $\omega_2,\mu\in\Omega^q(a)$.
It will be convenient to decompose $\omega_1,\omega_2$ and $\mu$ in the following way
\begin{align*}
\omega_1 &= \omega_{1r} + \omega_{1d} \qquad \qquad \omega_2 = \omega_{2r} + \omega_{2d} + t_1 \ d\omega_{1r}\qquad \qquad \mu = \mu_{r} + \mu_{d} + t_2\ d\omega_{1r}
\end{align*}
where the subindex $r$ and $d$ denote radial and exact terms, and $\omega_{2d}$ and $\mu_{d}$ are linearly independent to $d\omega_{1r}$. For $\tau\in\Omega^r(b)$, we then have
\begin{align}
D(\tau) &= (\omega_{1r}+\omega_{1d})\wedge d\tau + \Big[ (rt_1+t_2) \ d\omega_{1r} + (r \omega_{2r}+\mu_{r}) + (r \omega_{2d}+\mu_{d})\Big]\wedge\tau \label{D-expr1}\\
D(\tau) &= (\omega_{1r}+\omega_{1d})\wedge d\tau + \Big[ (rt_1+t_2) \ d\omega_{1r} + \nu_r + \nu_d\Big]\wedge\tau \label{D-expr2}
\end{align}
where we are writing $\nu_r = r \omega_{2r}+\mu_{r}$ and $\nu_d = r \omega_{2d}+\mu_{d}$.
\
We now want to see under what conditions we have the equalities
\begin{enumerate}
\item[b)] $f\cdot D(\tau) = D(f\cdot\tau)$
\item[c)] $g\cdot(f\cdot \tau) = (gf)\cdot\tau$
\end{enumerate}
for the action defined in Definition \ref{def-lin}: $f\cdot \tau = \alpha\ f\ \tau + \beta\ df\wedge i_R\tau$.
\
From $b)$ we get
\begin{align}\label{equ1}
&f\cdot D(\tau) = f\cdot\bigg\{(\omega_{1r}+\omega_{1d})\wedge d\tau + \Big[(rt_1+t_2)\ d\omega_{1r} + \nu_r + \nu_d \Big]\wedge\tau\bigg\} = \nonumber\\
&= \alpha(r+q,b+a,c)\ f\ \bigg\{(\omega_{1r}+\omega_{1d})\wedge d\tau + \Big[(rt_1+t_2)\ d\omega_{1r} + \nu_r + \nu_d \Big]\wedge\tau\bigg\} + \nonumber\\
&\hspace{.2cm}+ \beta(r+q,b+a,c)\ df\wedge i_R\Big((\omega_{1r}+\omega_{1d})\wedge d\tau + \Big[(rt_1+t_2)\ d\omega_{1r} + \nu_r + \nu_d \Big]\wedge\tau\Big) = \nonumber\\
&= \alpha(r+q,b+a,c)\ f\ \bigg\{(\omega_{1r}+\omega_{1d})\wedge d\tau + \Big[(rt_1+t_2)\ d\omega_{1r} + \nu_r + \nu_d \Big]\wedge\tau\bigg\} + \nonumber\\
&\hspace{.2cm}+\beta(r+q,b+a,c)df\wedge\bigg\{ i_R\omega_{1d}\wedge d\tau + (-1)^{q-1}(\omega_{1r}+\omega_{1d})\wedge i_Rd\tau + \nonumber\\
&\hspace{.2cm} + \Big[a(rt_1+t_2)\ \omega_{1r}+i_R\nu_d\Big] \wedge \tau + (-1)^{q}\Big[(rt_1+t_2)\ d\omega_{1r} + \nu_r + \nu_d \Big]\wedge i_R\tau \bigg\}
\end{align}
and, writing $\alpha=\alpha(r,b,c)$,
\begin{align}\label{equ2}
&D(f\cdot\tau) = D\Big(\alpha\ f\ \tau + \beta\ df\wedge i_R\tau\Big) = \nonumber\\
&= \alpha \ (\omega_{1r}+\omega_{1d})\wedge df\wedge \tau + \alpha f \ (\omega_{1r}+\omega_{1d})\wedge d\tau - \beta \ (\omega_{1r}+\omega_{1d})\wedge df\wedge di_R\tau + \nonumber\\
&\hspace{1cm}+ \Big[(rt_1+t_2)\ d\omega_{1r} + \nu_r + \nu_d \Big]\wedge \Big(\alpha f\ \tau + \beta\ df\wedge i_R\tau\Big).
\end{align}
\
Taking the coefficients from every different term in eqs. \ref{equ1} and \ref{equ2}, we get the system
\[
\left\{
\begin{aligned}
&{\rm I})\ f\ \omega_{1r}\wedge d\tau: & \alpha(r+q,b+a,c)\ = \alpha(r,b,c)\ \\
&{\rm II})\ f\ \omega_{1d}\wedge d\tau: & \alpha(r+q,b+a,c)\ = \alpha(r,b,c)\ \\
&{\rm III})\ f\ d\omega_{1r}\wedge\tau: & \alpha(r+q,b+a,c)(rt_1+t_2)\ = \ \alpha(r,b,c)(rt_1+t_2)\\
&{\rm IV})\ f\ \nu_r\wedge\tau: & \alpha(r+q,b+a,c)\ = \alpha(r,b,c)\ \\
&{\rm V})\ f\ \nu_d\wedge\tau: & \alpha(r+q,b+a,c)\ = \alpha(r,b,c)\ \\
&{\rm VI})\ df\wedge i_R\omega_{1d}\wedge d\tau: & \beta(r+q,b+a,c)\ = 0\ \\
&{\rm VII})\ df\wedge\omega_{1r}\wedge i_Rd\tau: & (-1)^{q-1}\beta(r+q,b+a,c)\ = (-1)^{q-1}\beta(r,b,c)\ \\
&{\rm VIII})\ df\wedge \omega_{1d}\wedge i_Rd\tau: & (-1)^{q-1}\beta(r+q,b+a,c)\ = (-1)^{q-1}\beta(r,b,c)\ \\
&{\rm IX})\ df\wedge\omega_{1r}\wedge\tau: & \beta(r+q,b+a,c)a(rt_1+t_2)\ =\hspace{1cm}\\
& & = (-1)^{q-1}(\alpha(r,b,c) -b\beta(r,b,c))\ \\
&{\rm X})\ df\wedge i_R\nu_d\wedge\tau: & \beta(r+q,b+a,c)\ = 0\ \\
&{\rm XI})\ df\wedge d\omega_{1r}\wedge i_R\tau: & \beta(r+q,b+a,c)(-1)^{q}(rt_1+t_2) = \hspace{1cm}\\
& & = \beta(r,b,c)(-1)^{q}(rt_1+t_2)\\
&{\rm XII})\ df\wedge \nu_r\wedge i_R\tau: & (-1)^{q}\beta(r+q,b+a,c) \ =\ (-1)^{q}\beta(r,b,c)\\
&{\rm XIII})\ df\wedge\nu_d\wedge i_R\tau: & (-1)^{q}\beta(r+q,b+a,c) \ =\ (-1)^{q}\beta(r,b,c)
\end{aligned}
\right.
\]
which can be reduced to
\[
\left\{
\begin{aligned}
&{\rm XI})\ df\wedge d\omega_{1r}\wedge i_R\tau: & (-1)^{q-1}\beta(r+q,b+a,c)\ = (-1)^{q-1}\beta(r,b,c)\ \\
&{\rm IV})\ f\ \nu_r\wedge\tau: & \alpha(r+q,b+a,c)\ = \alpha(r,b,c)\ \\
&{\rm IX})\ df\wedge\omega_{1r}\wedge\tau: & \beta(r+q,b+a,c)a(rt_1+t_2)\ = \hspace{1cm}\\
& & = (-1)^{q-1}(\alpha(r,b,c) -b\beta(r,b,c))\ \\
&{\rm X})\ df\wedge i_R\nu_d\wedge\tau: & \beta(r+q,b+a,c)\ = 0
\end{aligned}
\right.
\]
\
It is clear the sufficiency of these equalities to accomplish $b)$. For the necessity we proceed as follows:
\begin{itemize}
\item if $i_R\tau = 0$, then the terms involving eqs. ${\rm I}),\ldots,{\rm X})$ must coincide, leaving the terms of eqs. ${\rm XI}),{\rm XII})$ and ${\rm XIII})$ apart which can be considered in another system. Since these equations are linearly independent by hypothesis, we get that eq. ${\rm XI})$ must be satisfied.
\item if $\tau=df\wedge\rho_d$, for some exact differential form $\rho_d$, then only eqs. ${\rm III}),{\rm IV})$ and ${\rm V})$ survive. Also they are linearly independent by hypothesis, then eq. ${\rm IV})$ must be also satisfied.
\item by the previous arguments, we can clear off eqs. ${\rm I})$ to ${\rm V})$, ${\rm VII}),{\rm VIII})$ and ${\rm XI})$,${\rm XII})$,\linebreak${\rm XIII})$. Taking now $\tau=\tau_d$ we get only eqs. ${\rm IX})$ and ${\rm X})$ which are also linearly independent.
\end{itemize}
\
Using eq. ${\rm IX})$ we can clear $\beta$ as
\[
\beta(r,b,c) = \frac{\alpha(r,b,c)}{b-(-1)^qa(rt_1+t_2)}.
\]
From this equality and eqs. ${\rm I})$ and ${\rm VII})$ we get the formula
\[
\frac{\alpha(r,b,c)}{b-(-1)^qa(rt_1+t_2)} = \frac{\alpha(r,b,c)}{(b+a)-(-1)^qa\big((r+q)t_1+t_2)\big)}
\]
from where we are able to clear $t_1$ as $\frac{(-1)^q}{q}$ and obtain the system
\begin{equation}\label{sis-1}
\left\{
\begin{aligned}
& \alpha(r,b,c)\ = \alpha(r+q,b+a,c)\ \\
& \beta(r,b,c) \ = \ \frac{\alpha(r,b,c)}{ b-a\left(\frac{r}{q}+(-1)^{q}\,t_2\right)}\\
\end{aligned}
\right.
\end{equation}
From eqs. ${\rm VI})$ and ${\rm X})$ we see that $\omega_{1d}=\nu_d=0$, since $\alpha$ and $\beta$ must be non trivial. Then recalling the expressions of eqs. \ref{D-expr1} and \ref{D-expr2}, the differential operator $D$ must be of the form
\begin{align*}
D(\tau) &= \omega_{1r}\wedge d\tau + \Big[ \left(\frac{r}{q}+t_2\right) \ d\omega_{1r} + \nu_r \Big]\wedge\tau\\
D(\tau) &= \omega_{1r}\wedge d\tau + r\left(\frac{1}{q}\ d\omega_{1r} + \omega_{2r}\right)\wedge\tau + (t_2 \ d\omega_{1r} +\mu_{r})\wedge\tau.
\end{align*}
This way, we have that
\[
D = \omega_{1r}\wedge \LL_{Id} + \left(\frac{1}{q}\ d\omega_{1r} + \omega_{2r}\right)\wedge i_{Id}+ (t_2 \ \lambda_{d\omega_{1r}} +\lambda_{\mu_{r}})
\]
showing that $D\in\lin^{(q,a)}(\widetilde{\Omega})$.
\
From $c)$ we have
\begin{align}\label{equ3}
g\cdot&(f\cdot\tau) = g\cdot\big(\alpha(r,b,c)\ f\ \tau + \beta(r,b,c)\ df\wedge i_R\tau\big) = \nonumber\\
&=\alpha(r,b+c,e)g\ \Big(\alpha(r,b,c)\ f\ \tau + \beta(r,b,c)\ df\wedge i_R\tau\Big) + \nonumber\\
&\hspace{1cm}+ \beta(r,b+c,e)\ dg\wedge i_R\Big(\alpha(r,b,c)\ f\ \tau + \beta(r,b,c)\ df\wedge i_R\tau\Big) = \nonumber\\
&= \alpha(r,b+c,e)\alpha(r,b,c)\ gf\ \tau + \alpha(r,b+c,e)\beta(r,b,c)\ g\ df\wedge i_R\tau \ + \nonumber\\
&\hspace{0cm}+ \beta(r,b+c,e)\alpha(r,b,c)\ f \ dg\wedge i_R\tau + \beta(r,b+c,e)\beta(r,b,c)c\ f\ dg\wedge i_R\tau
\end{align}
and
\begin{align}\label{equ4}
(gf)&\cdot\tau = \alpha(r,b,c+e)\ gf\ \tau + \beta(r,b,c+e)\ d(gf)\wedge i_R\tau = \nonumber\\
&= \alpha(r,b,c+e)\ gf\ \tau + \beta(r,b,c+e)g\ df\wedge i_R\tau + \beta(r,b,c+e)f\ dg\wedge i_R\tau.
\end{align}
\
Again, joining eqs. \ref{equ3} and \ref{equ4} as before we get the following system of equations
\begin{equation}\label{sis-2}
\left\{
\begin{aligned}
&{\rm I})\ gf\ \tau: &\alpha(r,b+c,e)\alpha(r,b,c)\ = \ \alpha(r,b,c+e)\\
&{\rm II})\ g\ df\wedge i_R\tau: &\alpha(r,b+c,e)\beta(r,b,c)\ = \ \beta(r,b,c+e)\\
&{\rm III})\ f\ dg\wedge i_R\tau: &\Big(\beta(r,b+c,e)\alpha(r,b,c)+\beta(r,b+c,e)\beta(r,b,c)c\Big)\ = \\
& & =\ \beta(r,b,c+e)
\end{aligned}
\right.
\end{equation}
which implies condition $c)$.
\
For the necessity we can assume $i_R\tau=0$ from where we get eq. ${\rm I})$. Removing that equation from the system, we can choose $f$ and $g$ linearly independent and we are done.
\
Putting together eqs. \ref{sis-1} and \ref{sis-2} we get the conditions
\[
\mathcal{S}:\ \left\{
\begin{aligned}
&\alpha(r,b,c)\ = \ \alpha(r+2,b+a,c) \ \\
& \alpha(r,b,c)\ = \ \frac{\alpha(r,b,c+e)}{\alpha(r,b+c,e)} \\
&\beta(r,b,c) \ = \ \frac{\alpha(r,b,c)}{ b-a\left(\frac{r}{q}+(-1)^{q}\,t_2\right)}\\
\end{aligned}
\right.
\]
\
The first two equations suggest a linear relation between the first two coordinates, for which the denominator of the third equation propose a formula for that. The second equation suggest a multiplicative relation between the last two coordinates.
\
As stated in the theorem, a formula that satisfies system $\mathcal{S}$ can be given by
\[
\alpha(r,b,c) := \frac{b-a\left(\frac{r}{q}+(-1)^{q}\,t_2\right)}{b+c-a\left(\frac{r}{q}+(-1)^{q}\,t_2\right)} \qquad
\beta(r,b,c) := \frac{\alpha(r,b,c)}{b-a\left(\frac{r}{q}+(-1)^{q}\,t_2\right)}
\]
which clearly verifies condition $a)\alpha(-,-,0)=1$ of Definition \ref{def-lin}.
\
For the case $\omega_1=0$, we can choose $\alpha=1$ and $\beta=0$ which reduces to the usual multiplication.
\end{proof}
\begin{rem} A more general formula to the one given in the previous theorem can be given by, for an appropriate $t$,
\[
\alpha(r,b,c) := \frac{F\Big(L\big(b\big)-L\left(a\left(\frac{r}{q}+(-1)^{q}\,t\right)\right)}{F\Big(L\big(b+c\big)-L\left(a\left(\frac{r}{q}+(-1)^{q}\,t\right)\right)} \quad
\beta(r,b,c) := \frac{\alpha(r,b,c)}{ b-a\left(\frac{r}{q}+(-1)^{q}\,t\right)}
\]
where $F$ is any function and $L$ is a linear function.
\end{rem}
\
\section{Applications}\label{applications}
\
Along this section we show two different applications of the formula for deformations that we gave in Definition \ref{def-lin}.
The first one shows how to linearize the usual exterior differential as we mentioned in eq. \ref{ej1}. The second one is related to the regularity complex of Camacho and Lins-Neto given in the introduction, see eq. \ref{reg-intro}.
\
\subsection{Exterior differential}\label{ext-diff}
\
\begin{deff}\label{def-d} For $f\in S(c)$ and $\tau\in\Omega^r(b)$, we define
\begin{equation*}
f\dt \tau := \frac{b}{b+c}\ \left(\ f\tau\ +\ \frac{1}{b}\ df\wedge i_R\tau \ \right)
\end{equation*}
when $b$ or $c$ is not null and define $f\cdot_d\tau = f\tau$ when $b,c=0$. We denote by $\Omega^r_d=\Omega^r$ and $\Omega_{d}=\bigoplus_{r\geq 0} \Omega^r_d$ to the $\CC$-vector spaces with this action from $S$.
\end{deff}
\
We then have:
\begin{prop}\label{acc-3} The exterior differential
\[
\xymatrix{\Omega_d \ar[r]^d & \Omega_d}
\]
is a morphism of $S$-modules and the $S$-modules $\Omega^r_d$ are finitely generated.
\end{prop}
\begin{proof}
It is clear that $d \in\lin^{(1,0)}(\Omega)$. Then, following Theorem \ref{teo1}, we get the formula proposed in Definition \ref{def-d}.
\
To see that $\Omega_d$ is finitely generated we have that $\alpha^2(r,b,1)-\beta^2(r,b,1)=0$ if and only if
\[
\left(\frac{b-1}{b}\right)^2 - \left(\frac{b-1}{b^2}\right)^2 = \frac{(b-1)^2(b^2-1)}{b^4} = 0.
\]
\[
\left(\frac{b}{b+1}\right)^2 - \left(\frac{b}{(b+1)^2}\right)^2 = \frac{b^2[(b+1)^2-1]}{(b+1)^4} = 0.
\]
Then, the conditions of Proposition \ref{accion-gen} are verified for $b\geq 1$ and the result follows.
\end{proof}
\
\subsection{Regularity complex}\label{reg-complex}
\
In \cite{moli} we introduce a long complex $C^\bullet(\omega)$ of differential operators from the short complex of Camacho and Lins-Neto, see \cite[Section III.I, p.~17]{caln} or eq. \ref{reg-intro}. In the same work we give $\CC$-linear isomorphisms of $C^\bullet(\omega)$ to an $S$-linear complex and use it to obtain geometric information of the singular locus of the foliation defined by $\omega$ as well as its relation with first order unfoldings of $\omega$. Here we expose a different approach to linearize the complex $C^\bullet(\omega)$.
\
Let us recall from \cite[Section 6.1, p.~19]{moli} the following definition:
\
Let $\omega\in\Omega^1(a)$ such that $i_R\omega = 0$ and $\omega\wedge d\omega = 0$. We define the differential operator $\omega\vartriangle \ \in\dif^{(2,a)}(\Omega)$ as
\[
\xymatrix@R=0pt@C=50pt{ \Omega^r \ar[r]^{\omega\vartriangle} & \Omega^{r+2} \\
\save[] *{\tau\ }="s" \restore & \save[] +<50pt,0pt> *{\ \omega\vartriangle\tau:=\omega\wedge d\tau + \kappa(r)\ d\omega\wedge \tau}="t" \restore & \ar@{|->}"s";"t"
}
\]
where $\kappa(r) := \frac{r+1}{2}$.
\
Using \cite[Proposition 6.1.2, p.~19]{moli} we know that $\omega\vartriangle$ defines a differential of a complex of $\CC$-vector spaces which allow us to define the graded complex $C^\bullet(\omega)$ as
\[
\xymatrix{
C^\bullet(\omega):& T\ar[r]^-{\omega\vartriangle} & \Omega^1 \ar[r]^-{\omega\vartriangle} & \Omega^3 \ar[r]^-{\omega\vartriangle} & \ldots & &
}
\]
where the 0-th differential is defined as $\omega\vartriangle X:=\LL_X(\omega)= i_Xd\omega + di_X\omega$.
\
\begin{deff}\label{def-t} For $f\in S(c)$ and $\tau\in\Omega^r(b)$ we define the action
\begin{equation*}
f\vt \tau := \frac{1}{b+c-\kappa(r)a}\ \Big[\ \big(b-\ \kappa(r)a\big) f\tau\ +\ df\wedge i_R\tau\ \Big]
\end{equation*}
and denote by $\Omega^r_{\omega\vartriangle}= \bigoplus_{b> \kappa(r)a }\Omega^r(b)$ and $\Omega_{\omega\vartriangle}=\bigoplus_{r\geq 0}\Omega^r_{\omega\vartriangle}$ to the $\mathbb{C}$-vector spaces with this action from $S$.
For $f\in S(c)$ and $X\in T(b)$ we also define
\[
f\vt X = \frac{b}{b+c} \ fX
\]
and denote by $T_{\omega\vartriangle}=\bigoplus\limits_{b\geq 1}T(b)$ to the $\mathbb{C}$-vector space with this action from $S$.
\end{deff}
\
We then have:
\begin{prop} The complex
\[
\xymatrix@R=0pt@C=35pt{
C^\bullet_{\omega\vartriangle}(\omega):& T_{\omega\vartriangle} \ar[r]^-{\omega\vartriangle} & \Omega^1_{\omega\vartriangle} \ar[r]^-{\omega\vartriangle} & \Omega^3_{\omega\vartriangle} \ar[r]^-{\omega\vartriangle} & \ldots\\
}
\]
is a complex of $S$-modules and the $S$-modules $\Omega^r_{\omega\vartriangle}$ and $T_{\omega\vartriangle}$ are finitely generated.
\end{prop}
\begin{proof}
Writing $\omega\vartriangle\ $as
\[
\omega\vartriangle\ = \omega\wedge\LL_{Id} + \frac{1}{2}d\omega\wedge i_{Id} + \frac{1}{2}\lambda_{d\omega}
\]
it is clear that $\omega\vartriangle\ \in\lin^{(2,a)}(\Omega)$. Then, following Theorem \ref{teo1}, we get the formula proposed in Definition \ref{def-t}.
\
It is also clear that $T_{\omega\vartriangle}$ is finitely generated. For $\Omega^r_{\omega\vartriangle}$ we have that $\alpha^2(r,b,1)-\beta^2(r,b,1)=0$ if and only if
\begin{align*}
\left(\frac{b-\kappa(r)a}{b+1-\kappa(r)a}\right)^2-\left(\frac{1}{b+1-\kappa(r)a}\right)^2 = \frac{\big(b-\kappa(r)a\big)^2-1}{(b+1-\kappa(r)a)^2} = 0.
\end{align*}
Then, the conditions of Proposition \ref{accion-gen} are verified for $b>1+\kappa(r)a$ and the result follows.
\end{proof}
|
\section{Introduction}
\label{sec:intro}
Let $L$ be a link in $S^3$. The unlinking number of $L$ is the minimum number
of crossing changes to convert a diagram of $L$ to a diagram of the unlink,
where the minimum is taken over all diagrams. This is the obvious
generalisation to links of the unknotting number, and one should expect that
many of the same methods should apply to compute it for examples.
There have been several successful applications of Donaldson's diagonalisation theorem and
Heegaard Floer homology to the calculation of unknotting numbers, for example \cite{cl,osu1,unknotting,Gu1,altu1}. It is an interesting problem,
which we begin to address here, to generalise these techniques to the case of links.
The
systematic study of unlinking number for links with more than one component
seems to have been initiated by Kohn in \cite{kohn}, in which he computed
unlinking numbers for all but 5 prime, nonsplit, 2-component links which admit
diagrams with 9 crossings or less. In this paper we determine the unlinking
number for these remaining examples and provide a complete table of unlinking
numbers for prime nonsplit links with crossing number at most 9.
The main result of this paper is a generalisation of a theorem of Cochran and
Lickorish \cite{cl}, and of a refinement due to the second author and Strle
\cite{imm}, to the case of links with more than one component. We choose an
orientation on a given link and consider the sum $\sigma+\eta$ of the classical
link signature and nullity. This sum is equal to $k-1$ for the $k$-component
unlink; it increases by 2 or stays constant when a positive crossing is
changed, and decreases by 0 or 2 when a negative crossing is changed. Thus if
$\sigma+\eta$ is less than $k-1$ for a given orientation on a $k$-component
link $L$, then any set of crossing changes converting $L$ to the unlink must
include changing at least $(-\sigma-\eta+k-1)/2$ positive crossings. We show
that if the number of positive crossings changed does not exceed this minimum
then the double branched cover $Y$ of $L$ bounds a smooth definite 4-manifold
$W$ which constrains the linking form of $Y$, and which leads to
obstructions coming from Donaldson's diagonalisation theorem and Heegaard Floer
theory. Note that the trace of a homotopy associated to a sequence of crossing
changes from $L$ to the the unlink gives rise to an immersion of $k$ disks in
the 4-ball bounded by $L$, with one double point for each crossing change.
\begin{maintheorem}
\label{thm:mainthm} Let $L$ be an oriented $k$-component link in $S^3$ with
signature $\sigma$ and nullity $\eta$, and suppose that $L$ is the oriented
boundary of a properly normally immersed union of $k$ disks in the 4-ball
with $p$ positive double points and $n$ negative double points. Then
\begin{equation}
\label{eq:pineq}
p\ge\frac{-\sigma-\eta+k-1}2.
\end{equation}
If equality holds in \eqref{eq:pineq}
then the double cover $Y$ of $S^3$ branched along $L$ bounds a smooth negative-definite
4-manifold $W$ with
$$b_2(W)=2n-\sigma=2(p+n)+\eta-k+1,$$
for which the restriction map
$$H^1(W;{\mathbb Q})\to H^1(Y;{\mathbb Q})$$
is an isomorphism.
Moreover $H_2(W;{\mathbb Z})$ contains $p+n$ pairwise disjoint spherical classes of self-intersection $-2$, whose images in $H_2(W;{\mathbb Z})/{\it Tors}$ span a primitive sublattice.
\end{maintheorem}
We denote by $c^*(L)$ the \emph{4-ball crossing number} of a link $L$ in $S^3$. This is the minimal number of double points in a properly normally immersed union of disks in $D^4$ bounded by the link. The conditions imposed by
Theorem \ref{thm:mainthm} on the linking form of $Y$ are sufficient to determine the unlinking number and 4-ball crossing number of four of Kohn's remaining examples. Recall that a 2-component link in $S^3$ has two Murasugi signatures corresponding to two choices of \emph{quasi-orientation} (orientation up to overall reversal).
\begin{mainexample}
\label{eg:kohn}
Let $L$ be a 2-component link in $S^3$ with $c^*(L)<3$. Suppose that the double branched cover $Y$ has finite cyclic homology group.
Then one of the statements below holds:
\begin{enumerate}
\item $\det(L)=2t^2$ with $t\in{\mathbb Z}$, or
\item $\det (L)$ is divisible by $4$ and at least one of the signatures of $L$ is in $\{-1,1\}$, or
\item $\det (L)$ is divisible by $16$.
\end{enumerate}
In particular, the links
$$9^2_3=L9a30,\, 9^2_{15}=L9a15,\, 9^2_{27}=L9a17,\, 9^2_{31}=L9a2$$
have unlinking number and 4-ball crossing number $3$.
\end{mainexample}
We refer in Example \ref{eg:kohn} to links using both Rolfsen and Thistlethwaite names.
A proof that the unlinking number of $L9a30$ is 3 was given by Kanenobu and Kanenobu in \cite{KK}.
The following corollary is obtained by combining Theorem \ref{thm:mainthm} with Donaldson's diagonalisation theorem \cite{don} and a construction of Gordon and Litherland \cite{gl}.
\begin{maincorollary}
\label{cor:alt}
Let $L$ be an oriented nonsplit alternating link with $k$ components and signature $\sigma$, and suppose that $L$ bounds a properly normally immersed union of $k$ disks in the 4-ball with $p=\frac{-\sigma+k-1}2$ positive double points and $n$ negative double points.
Let $m$ be the rank of the positive-definite Goeritz lattice $\Lambda$ associated to an alternating diagram of $L$. Then $\Lambda$ admits an embedding in the standard integer lattice $Z={\mathbb Z}^{m+2n-\sigma}={\mathbb Z}^{m+2(p+n)-k+1}$, with $p+n$ pairwise orthogonal vectors of square $2$ which span a primitive sublattice in the orthogonal complement of $\Lambda$ in $Z$.
\end{maincorollary}
The final unknown example from Kohn's table, as well as the links in Example \ref{eg:kohn}, may be settled using Corollary \ref{cor:alt}. An alternative method involves the use of Heegaard Floer correction terms of the double branched cover. We illustrate the use of both methods in Section \ref{sec:eg}.
\begin{mainexample}
\label{eg:L9a10}
The link $9^2_{36}=L9a10$ has unlinking number and 4-ball crossing number $3$.
\end{mainexample}
Combining Examples \ref{eg:kohn} and \ref{eg:L9a10} with previous results of Kauffman-Taylor \cite{kt}, Kohn \cite{kohn}, Kawauchi \cite{kaw14} and Borodzik-Friedl-Powell \cite{bfp} leads to the complete Table \ref{table:u} of unlinking numbers of nonsplit prime links with crossing number at most 9. We also show that for all links in Table \ref{table:u}, the unlinking number is equal to the 4-ball crossing number.
\vskip2mm \noindent{\bf Acknowledgements.}
We are grateful to Maciej Borodzik, Stefan Friedl and Mark Powell for bringing the unknown values in Kohn's table to our attention in their paper \cite{bfp}.
The first author thanks the University of Glasgow for its hospitality.
The second author acknowledges the influence of his earlier joint work with Sa\v{s}o\ Strle, especially \cite{imm}.
We thank Mark Powell for helpful comments on an earlier draft, and the anonymous referee for helpful suggestions.
\section{Proofs of Theorem \ref{thm:mainthm} and Corollary \ref{cor:alt}}
For $L$ a link in $S^3$ we denote the two-fold
branched cover of $L$ by $\TwoF{L}$. Let $F$ be a connected
Seifert surface for $L$.
Following Trotter and Murasugi \cite{trotter,murasugi},
the \emph{signature} $\sigma(L)$ of the link $L$ is the signature of the symmetrised Seifert pairing
\begin{align*}
H_1(F) \times H_1(F) &\rightarrow {\mathbb Z}\\
a, b &\mapsto \lk(a_+,b) + \lk(a, b_+),
\end{align*}
where $\lk(a,b)$ denotes the linking number of $a$ and $b$ in $S^3$ and
$b_+$ denotes the cycle obtained by pushing $b$ in the positive normal direction.
The \emph{nullity} $\eta(L)$ of the link $L$ is the dimension of the nullspace of the symmetric form above. This is also equal to the first Betti number of $\dbc{L}$ \cite{kt}. We note that if $-L$ denotes the mirror of $L$ then $\sigma(-L)=-\sigma(L)$ and $\eta(-L)=\eta(L)$.
The following lemma is based on \cite[Proposition 2.1]{cl} and \cite[Theorem 5.1]{st}.
\begin{lemma}
Let $L$ be a link in $S^3$.
\label{lem:crossingchange}
Suppose that $L$ and $L'$ are oriented links in $S^3$ such that $L'$ is obtained from $L$ by changing a single negative crossing. Then
$$\sigma(L')\pm\eta(L')\in\{\sigma(L)\pm\eta(L),\sigma(L)\pm\eta(L)-2\},$$
where the choice of $\pm$ is consistent in all three instances above.
Similarly if $L'$ is obtained from $L$ by changing a single positive crossing, then
$$\sigma(L')\pm\eta(L')\in\{\sigma(L)\pm\eta(L),\sigma(L)\pm\eta(L)+2\}.$$
In either case,
$$|\eta(L')-\eta(L)|\le1.$$
\end{lemma}
\begin{proof}
We take a diagram for $L$ containing a negative crossing, such that changing
this crossing yields $L'$. The Seifert algorithm, followed by tubing together components if necessary, gives a connected surface $F$; we may
choose a basis for $H_1(F)$ represented by loops, exactly one of which passes
through the distinguished crossing. We see that the matrix of the symmetrised
Seifert pairing in this basis is the same for $L'$ as for $L$ save for one
diagonal entry which is reduced by 2. This will leave the signature and
nullity of the pairing on a codimension one sublattice unchanged. The last
eigenvalue may also be unchanged in sign, or may change from positive to
negative or zero, or from zero to negative. The first assertion of the lemma
follows; the second follows on applying the first to $-L$ and $-L'$, and the
third follows easily by restricting consideration to the nullity.
\end{proof}
A surface with boundary is normally properly immersed in $D^4$ if its boundary (respectively, interior) is contained in the boundary (resp., interior) of the 4-ball and its only singularities are normal double points, locally modelled on two transversely intersecting planes in ${\mathbb R}^4$. Double points may be given a sign by taking an arbitrary orientation on the surface and comparing an oriented basis for the tangent plane to one sheet of the surface at the singularity followed by an oriented basis on the other sheet, to the ambient orientation.
\begin{lemma}
\label{lem:ineq}
Let $L$ be an oriented $k$-component link in $S^3$ with signature $\sigma$ and nullity $\eta$, and suppose that $L$ bounds a properly normally immersed union of
$k$ disks with $p$ positive double points and $n$ negative double points.
Then
$$n\ge\frac{\sigma-\eta+k-1}2 \quad\mbox{ and }\quad p\ge\frac{-\sigma-\eta+k-1}2.$$
\end{lemma}
\begin{proof}
From \cite[Proposition 2.1]{imm} the immersed disks bounded by $L$ may be
isotoped so they are given by a concordance, followed by the trace of a regular
homotopy corresponding to changing $p$ positive and $n$ negative crossings,
followed by a nullconcordance. Since signature and nullity are
concordance invariants \cite{kt}, the inequalities now follow from the previous lemma
and the fact that the $k$-component unlink has signature zero and nullity
$k-1$.
\end{proof}
The following proposition will be used to prove Theorem \ref{thm:mainthm}, and
may be used to give another proof of Lemma \ref{lem:ineq}.
\begin{proposition}
\label{prop:cl}
Let $L$ be a $k$-component oriented link in $S^3 = \partial D^4$. Suppose $L$
bounds normally properly immersed disjoint oriented disks in
$D^4$ with $p$ positive and $n$ negative self-intersections. Then
there exists an oriented smooth $4$-manifold $W$
with boundary $\partial W = \TwoF{L}$ such that
\begin{enumerate}
\item the second Betti number of $W$ satisfies $b_2(W) \leq 2(p+n) + 1-k + \eta(L)$;
\item the manifold $W$ has signature $\sigma(W) = -2n + \sigma(L)$;
\item the inclusion $\partial W \subset W$ induces
an injection $H^1(W; {\mathbb Q}) \hookrightarrow H^1(\partial W; {\mathbb Q})$, with cokernel dimension equal to the nullity of the intersection form on $H_2(W)$;
\item there exist $p+n$ classes in $H_2(W;{\mathbb Z})$ represented by pairwise disjoint spheres of self-intersection $-2$. The images of these in $H_2(W;{\mathbb Z})/{\it Tors}$ span a primitive sublattice; in other words the quotient of $H_2(W;{\mathbb Z})/{\it Tors}$ by this sublattice is torsion-free.
\end{enumerate}
\end{proposition}
The proof is an adaptation to the case of links of arguments given in \cite{cl,imm}.
\begin{proof}
By forming connected sums with $\overline{\mathbb{CP}^2}$ and resolving the singularities as
in \cite[Lemma 3.4]{cl}, we obtain an embedded surface $\Delta$ in
\begin{align*}
C := D^4 \# \left( \underset{p+n}{\#} \overline{\mathbb{CP}^2} \right)
\end{align*}
whose boundary is the link $L$ in $\partial C$ and each of whose components
is a disk.
The Euler characteristic of $C$ is $\chi(C) = 1 + p +n$.
The first homology of the complement of $\Delta$ in $C$ with $\zz/2\zz$ coefficients has a basis consisting of the meridians of
the component disks of $\Delta$. Thus we may take the double cover $W=\dbc{C,\Delta}$ of $C$
branched along $\Delta$.
By definition we have $\partial W = \TwoF{L}$.
\begin{claim}
The Euler characteristic of $W$ is $\chi(W) = 2(p+n) -k +2$.
\end{claim}
This can be seen as follows: we have the equality
\begin{align*}
\chi(C) = \chi(C \setminus \Delta) + \chi(\Delta) =\chi(C \setminus \Delta) + k.
\end{align*}
Therefore the double cover of $C \setminus \Delta$ has
Euler characteristic $2(\chi(C)-k)$. The manifold $W$ is obtained
by gluing $\Delta$ back in again. We obtain
\begin{align*}
\chi(W) = 2( \chi(C) - k) + k = 2(p + n) -k +2.
\end{align*}
\begin{claim}
The groups $H^1(W,\partial W;\mathbb{Q})\cong H_3(W;\mathbb{Q})$ are trivial.
\end{claim}
We show that $H_3(W; \zz/2\zz)$ vanishes. By Poincar\'{e} duality, $H^1(C, \partial C; \zz/2\zz) = 0$.
Then note that for a set $F \subset C$ such that
$\partial C \cup F$ is connected, we get
$$
H^0(C;\zz/2\zz)\overset{\cong}{\rightarrow}H^0(\partial C \cup F; \zz/2\zz)\rightarrow H^1(C, \partial C \cup F; \zz/2\zz)\rightarrow H^1(C; \zz/2\zz)=0.
$$
This implies that $H^1(C, \partial C \cup \Delta; \zz/2\zz) = 0$.four-
From \cite[Theorem 1]{Lee95} and
dualising, we obtain the following exact sequence:
$$
0=H^1(C, \partial C; \zz/2\zz)\rightarrow H^1(W, \partial W; \zz/2\zz)\rightarrow H^1(C, \partial C \cup \Delta; \zz/2\zz) = 0.
$$
Therefore $H^1(W, \partial W; \zz/2\zz) = 0$ holds.
By Poincar\'{e} duality we obtain directly that $H_3(W; \zz/2\zz) = 0$.
\begin{claim}
The inclusion $\partial W \subset W$ induces an injection $H^1(W;{\mathbb Q}) \hookrightarrow H^1(\partial W; {\mathbb Q})$, whose cokernel is isomorphic to the nullspace of the intersection form on $H_2(W;{\mathbb Q})$.
In particular, the inequality $\eta(L)=b_1(\partial W) \geq b_1(W)$ holds.
\end{claim}
The long exact sequence of the pair $(W, \partial W)$ with ${\mathbb Q}$ coefficients gives
\begin{align*}
0\rightarrow H^1(W) \rightarrow H^1(\partial W) \rightarrow H^2(W,\partial W) \rightarrow H^2(W),
\end{align*}
with the last map given in appropriate bases by the matrix of the intersection
pairing. The claim follows immediately.
Combining these claims, we obtain an upper bound for the second Betti number
$b_2(W)$. The Euler characteristic of $W$ is $\chi(W) = 1 - b_1(W) + b_2(W)$.
This gives
\begin{align*}
b_2(W) &= b_1(W) - 1 + 2(p + n) -k +2 \\
&= 2(p+n) + 1 - k + b_1(W) \\
&\leq 2(p+n) +1 -k + \eta(L).
\end{align*}
We proceed to calculate the signature $\sigma(W)$.
This is done exactly as in \cite{cl}. For the reader's convenience we
repeat the argument.
\begin{claim}
The signature of $W$ is $\sigma(W) = -2n+\sigma(L)$.
\end{claim}
We pick a connected Seifert surface $F_{-L}$ of the link $-L$ with interior pushed into $-D^4$.
Glue to obtain a closed $4$-manifold $(\widetilde{C}, F) := (C, \Delta) \cup_{(S^3, L)} (-D^4, F_{-L})$.
We consider the double cover $\widetilde{W}=\TwoF{\widetilde{C},F}$ of $\widetilde{C}$ branched along $F$.
This manifold has a decomposition $\widetilde{W} = W \cup_{\TwoF{L}} X_L$, where $X_L$ is the double
cover $\TwoF{F_{-L}}$ of $-D^4$ branched along $F_{-L}$. By \cite{viro,kt} the signature of $X_L$ is $-\sigma(L)$, and by Novikov additivity
\cite{index}, we have
\begin{align*}
\sigma(\widetilde{W}) = \sigma(W) + \sigma(X_L).
\end{align*}
Furthermore the $G$-signature theorem \cite{index,Gsig} allows us to express $\sigma(\TwoF{\widetilde{C},F})$ as
\begin{align*}
\sigma(\widetilde{W}) = 2 \sigma(\widetilde{C}) - \frac{1}{2}\left ( [F] \cdot [F]\right )
\end{align*}
The self-intersection of $F$ is $[F]\cdot [F] = -4p$; this is because each intersection point in the original immersed disks is replaced by two fibres of the disk bundle over $S^2$ which is the punctured $\overline{\mathbb{CP}^2}$ used to resolve the singularity, and these two disks have the same orientation in the case of a positive double point and opposite in the case of a negative double point. Thus $[F]$ has twice the generator of the second homology of this punctured $\overline{\mathbb{CP}^2}$ as a summand in the case of a positive double point, and zero times this generator otherwise.
We obtain the equations
\begin{align*}
\sigma(W) - \sigma(L)
& = 2 (\sigma(\widetilde{C}) +p) = 2(\sigma(C) +p) = -2n\\
\Rightarrow \sigma(W) &= -2n + \sigma(L).
\end{align*}
The existence of $p+n$ pairwise disjoint spherical classes $x_1,\dots,x_{p+n}$ of square $-2$
follows as in \cite[Theorem 2.2]{imm} from the fact that the double cover of a
disk bundle with Euler number $-1$ branched along two fibres is a disk bundle of
Euler number $-2$. It remains to see that their images in $H_2(W;{\mathbb Z})/{\it Tors}$ span a primitive sublattice.
Let $\alpha \in H_2(W;{\mathbb Z})$ and
$$m\alpha=\sum b_i x_{i}+\beta,$$
where $\beta$ is a torsion class and $m$ a nonzero integer. Denote
the induced map in homology of the double branched cover by
$\pi \colon H_2(W;{\mathbb Z}) \rightarrow H_2(C;{\mathbb Z})$. The classes
$\pi(x_i)$ are pairwise orthogonal of square $-1$. Thus we obtain
\begin{align*}
\pi(m\alpha) \cdot \pi(x_i) = -b_i
\end{align*}
and therefore $m$ divides each $b_i$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:mainthm}]
Suppose as in the theorem that $L$ is an oriented $k$-component link in $S^3$
with signature $\sigma$ and nullity $\eta$, and that $L$ bounds a properly
normally immersed union of $k$ oriented disks in the 4-ball with $p$
positive double points and $n$ negative double points.
By Proposition \ref{prop:cl}
there exists a smooth 4-manifold $W$ bounded by $Y=\dbc{L}$ with
$$2(p+n)+1-k+\eta\ge b_2(W)\ge-\sigma(W)=2n-\sigma.$$
This inequality in fact recovers
\eqref{eq:pineq}, and equality in \eqref{eq:pineq} thus proves that $W$ is
negative-definite. Proposition \ref{prop:cl} also tells us that the inclusion of $Y$
into $W$ induces an isomorphism $H^1(W; {\mathbb Q}) \cong H^1(Y; {\mathbb Q})$, and of the
existence of $p+n$ disjoint spheres of self-intersection $-2$ whose images in $H_2(W;{\mathbb Z})/{\it Tors}$ span
a primitive sublattice.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:alt}]
Recall that a link diagram $D$ determines two Goeritz matrices $G$ and $G'$
(see \cite{gl} for details and Section \ref{sec:eg} for an example). It is shown
in \cite{gl} that these are the intersection forms of the double covers of
$D^4$ branched along the pushed-in black surface coming from either of the two
chessboard colourings of the regions of $S^2\setminus D$; denote these
simply-connected 4-manifolds by $X$ and $X'$. The union $X\cup
-X'=\dbc{S^4,F}$ of these Gordon-Litherland manifolds is equal to the double
cover of $S^4=D^4\cup -D^4$ branched along the surface $F$ obtained by pushing
one of these surfaces into each hemisphere of the 4-sphere. A local argument
at the crossings shows that if $D$ is alternating then $\pm\dbc{S^4,F}$ is a
connected sum of complex projective planes, with one summand for each crossing (the union of
black and white subsurfaces in the neighbourhood of a crossing gives a
punctured $\mathbb{RP}^2$, whose double branched cover is a punctured $\pm\mathbb{CP}^2$; all
signs are the same if the diagram is alternating).
Thus for an alternating diagram one of the Goeritz matrices is
positive-definite and the other is negative-definite. This recovers the
well-known fact that the nullity of a nonsplit alternating link is zero, since
the Goeritz matrices present $H_1(\dbc{L};{\mathbb Z})$.
Now suppose $L$ satisfies the hypotheses of the Corollary and that $X$ is the
positive-definite Gordon-Litherland manifold associated to an alternating
diagram of $L$, with intersection lattice $\Lambda=(H_2(X;{\mathbb Z}),G)$ of rank
$m$. By Theorem \ref{thm:mainthm}, $\dbc{L}$ bounds a
negative-definite manifold $W$ with $b_2(W)=2n-\sigma$,
and with a primitive sublattice of $H_2(W;{\mathbb Z})/{\it Tors}$ spanned by $p+n$
pairwise orthogonal classes of square $-2$. Gluing $X$ to $-W$ by
a diffeomorphism of their boundaries gives a smooth closed positive-definite
manifold $\widetilde{X}$.
By Donaldson's diagonalisation theorem \cite{don} we know that
$(H_2(\widetilde{X};{\mathbb Z})/{\it Tors},Q_{\widetilde{X}})$ is isometric
to the standard integer lattice $Z={\mathbb Z}^{m+2n-\sigma}$.
Now the Mayer-Vietoris exact sequence
shows that we have an inclusion
$$\Lambda\oplus (H_2(-W;{\mathbb Z})/{\it Tors},Q_{-W})\subset Z,$$
as required.
\end{proof}
We also note the following generalisation of \cite[Corollary 3.21, second inequality]{kt} and
\cite[Corollary 3]{kohn}. The lower bound on the number of double points also
follows easily from Lemma \ref{lem:ineq}. The case of links with nullity zero
is given in \cite[Corollary 4.3]{kaw14}, and indeed the case of nonzero nullity may also be derived from results in \cite{kaw14}.
\begin{lemma}
\label{lem:det}
Let $L$ be a $k$-component link $L$ in $S^3$ with nullity $\eta$. Then
$$c^*(L)\ge k-1-\eta.$$
If $c^*(L)\le k-1$, then $\det L=2^{k-1}c^2$ for some $c\in{\mathbb Z}$.
\end{lemma}
\begin{proof}
From Proposition \ref{prop:cl} there exists a smooth 4-manifold $W$ bounded by $Y=\Sigma_2(L)$ with
$$2c^*(L)-(k-1-\eta)\ge b_2(W)\ge c^*(L).$$
This yields the lower bound on $c^*(L)$. If the nullity $\eta$ is nonzero then the determinant is zero.
If $c^*(L)=k-1$ and $\eta=0$, we find $b_2(W)=c^*(L)$ and, again from Proposition \ref{prop:cl}, $H_2(W;{\mathbb Z})$ has a basis of pairwise orthogonal classes with self-intersection $-2$. It follows that the absolute value of the intersection form on $H_2(W;{\mathbb Z})$ is $2^{k-1}$. From the long exact sequence in cohomology of the pair $(W,Y)$ we see that
$$\det L=|H^2(Y;{\mathbb Z})|=2^{k-1}c^2,$$
where $c$ is the order of the image of the torsion subgroup of $H^2(W;{\mathbb Z})$ in $H^2(Y;{\mathbb Z})$.
\end{proof}
It is also useful to generalise another basic observation of Kohn \cite[inequality (1)]{kohn}. This may be used recursively.
\begin{lemma}
\label{lem:sublinks}
Let $L=A\cup B$ be an oriented link in $S^3$ which is a union of sublinks $A$ and $B$. Then
$$u(L)\ge u(A)+u(B)+\left|\lk(A,B)\right|,$$
where $\lk(A,B)$ denotes the linking number between $A$ and $B$, and
$$c^*(L)\ge c^*(A)+c^*(B)+\left|\lk(A,B)\right|.$$
\end{lemma}
\begin{proof}
The first inequality follows since crossing changes involving only a sublink $A$ do not change the link type of $B$ or the linking number between $A$ and $B$, and crossing changes involving both $A$ and $B$ change $\lk(A,B)$ by one and do not change the link type of $A$ or $B$. The second inequality also uses \cite[Proposition 2.1]{imm} which tells us that immersed disks in $D^4$ bounded by $L$ are isotopic to a concordance followed by crossing changes followed by a nullconcordance, and the fact that linking number is a link concordance invariant.
\end{proof}
\section{Linking forms and Heegaard Floer correction terms}
\label{sec:rhod}
In this section we recall some obstructions to a 3-manifold $Y$ bounding a 4-manifold with a given intersection form, with specific attention to the case that $Y$ is the double branched cover of a link in $S^3$. For more details on this material see, e.g., \cite{mac}.
If $M$ is an orientable manifold of dimension at most 4, then there is a free and transitive action of $H^2(M;{\mathbb Z})$ on the set $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(M)$ of \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi structures on $M$.
For the rest of this section we let $Y$ be a rational homology 3-sphere, i.e., a closed oriented 3-manifold with $b_1(Y)=0$. There is then a natural inclusion of the set of spin structures on $Y$ into $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)$, and the spin structures are the fixed points of the involution given by conjugation of \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi structures.
The $\rho$-invariant of $Y$ (see for example \cite{AbsGrHF}) is the function
$$\rho:\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)\to {\mathbb Q}/2{\mathbb Z}$$
defined as follows. For each $\mathfrak t\in\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)$ we choose a \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi 4-manifold $(X,\mathfrak s)$ with a fixed diffeomorphism $\partial X\cong Y$ taking
$\mathfrak s|_{\partial X}$ to $\mathfrak t$, and define
\begin{equation}
\label{eq:rhodef}
\rho(Y,\mathfrak t)=\frac{c_1(\mathfrak s)^2-\sigma(X)}{4}\pmod2.
\end{equation}
This is well-defined due to Novikov additivity of the signature and the fact that for any closed \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi 4-manifold $(X,\mathfrak s)$ we have $c_1(\mathfrak s)^2\equiv\sigma(X)\pmod8$. Also note that this changes sign under orientation-reversal, i.e., $$\rho(-Y,\mathfrak t)=-\rho(Y,\mathfrak t).$$
Given an $n\times n$ matrix $Q$ with integer entries, let
$$\Char(Q)=\{\xi\in{\mathbb Z}^n\,:\,\xi_i\equiv Q_{ii}\pmod2,\,i=1,\dots n\}.$$
Then there is a free and transitive action of ${\mathbb Z}^n$ on $\Char(Q)$ given by $(x,\xi)\mapsto\xi+2x$ and correspondingly a free and transitive action of the group ${\mathbb Z}^n/Q{\mathbb Z}^n$ on $\Char(Q)/2Q{\mathbb Z}^n$.
The following proposition is proved in \cite{mac}, and follows from the fact that Chern classes of \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi structures on a 4-manifold $X$ are characteristic covectors for the intersection form of $X$, together with careful consideration of the long exact sequence in cohomology of the pair $(X,\partial X)$. It is also observed in \cite{mac} that the $\rho$-invariant is a quadratic enhancement of the linking form, and that the following obstruction to bounding $Q$ is equivalent to that coming from the linking form of $Y$.
\begin{proposition}
\label{prop:rho}
Suppose a rational homology 3-sphere $Y$ is the boundary of a smooth 4-manifold
$W$ with intersection form represented by an $n\times n$ matrix $Q$, with
$\left|\det Q \right|=\delta$.
Then the order of $H=H^2(Y;{\mathbb Z})$ is $\delta t^2$, and there exists a subgroup $T<H$ of order $t$ and an affine monomorphism
$$\phi:\Char(Q)/2Q{\mathbb Z}^n\to\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/T$$
which is equivariant with respect to the involutions coming from multiplication by $-1$ on $\Char(Q)$ and conjugation of \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi structures, such that
\begin{align}
\dfrac{\xi^TQ^{-1}\xi-\sigma(Q)}{4}&\equiv \rho(Y,\mathfrak t)\pmod2\label{eq:rhocong}
\end{align}
for all $\xi\in\Char(Q)$ and $\mathfrak t\in\phi([\xi])$.
\end{proposition}
In \cite{AbsGrHF}, Ozsv\'{a}th\ and Szab\'{o}\ associate a rational number $d(Y,\mathfrak t)$, called a \emph{correction term}, to each \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi structure $\mathfrak t$ on $Y$. They showed that the reduction of $d(Y,\mathfrak t)$ modulo 2 is $\rho(Y,\mathfrak t)$, and also that the correction terms give additional constraints on negative-definite \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi manifolds bounded by $(Y,\mathfrak t)$. These constraints give rise to the following refinement of Proposition \ref{prop:rho}, which is proved in \cite[Theorem 2.2]{qhs}.
\begin{proposition}
\label{prop:d}
Suppose a rational homology 3-sphere $Y$ is the boundary of a smooth 4-manifold $W$ with negative-definite intersection form represented by an $n\times n$ matrix $Q$, with $\left|\det Q\right|=\delta$.
Then the order of $H=H^2(Y;{\mathbb Z})$ is $\delta t^2$, and there exists a subgroup $T<H$ of order $t$, and a ${\mathbb Z}/2{\mathbb Z}$-equivariant affine monomorphism
$$\phi:\Char(Q)/2Q{\mathbb Z}^n\to\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/T$$
satisfying
\begin{align}
\dfrac{\xi^TQ^{-1}\xi+n}{4}&\le d(Y,\mathfrak t)\label{eq:dineq}\\
\mbox{and }\quad
\dfrac{\xi^TQ^{-1}\xi+n}{4}&\equiv d(Y,\mathfrak t)\pmod2\label{eq:dcong}
\end{align}
for all $\xi\in\Char(Q)$ and $\mathfrak t\in\phi([\xi])$.
\end{proposition}
We note that in Propositions \ref{prop:rho} and \ref{prop:d}, the subgroup $T$ is the image in $H^2(Y;{\mathbb Z})$ of the torsion subgroup of $H^2(X;{\mathbb Z})$ under the map induced by restriction to the boundary.
It is often helpful to consider the fixed points of the involutions on $\Char(Q)/2Q{\mathbb Z}^n$ and $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/T$. In particular each coset of $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/T$ which contains a spin structure is such a fixed point, and from \cite[Lemma 2.4]{qhs} these are the only fixed cosets unless there exists an element $\alpha$ in $H^2(Y;{\mathbb Z})/ T$ of order 4 with $2\alpha\in T$. In particular if the order of $T$ is odd then each fixed coset in $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/T$ contains a spin structure. If $Y=\Sigma_2(L)$ is the double cover of a link in $S^3$ then there is a bijection due to Turaev \cite{turaev} between quasi-orientations on $L$ (orientations up to overall reversal) and spin structures on $Y$. The following lemma is taken from \cite{lconc} and \cite{qa}.
\begin{lemma}
\label{lem:dbc}
Let $\mathfrak t$ be the spin structure on $Y=\Sigma_2(L)$ associated to an orientation on a link $L$ in $S^3$, and let $\sigma(L)$ be the signature of $L$ with the same orientation.
Then
$$\rho(Y,\mathfrak t)\equiv-\sigma(L)/4\pmod2.$$
If $L$ is nonsplit and alternating, or indeed quasi-alternating, then
$$d(Y,\mathfrak t)=-\sigma(L)/4.$$
\end{lemma}
\begin{proof}
Let $X$ be the double cover of $D^4$ branched along a pushed-in Seifert surface $F$ for the oriented link $L$. By \cite[Proposition 3.3]{lconc} there is a unique spin structure $\mathfrak s$ on $X$, and the restriction of $\mathfrak s$ to $Y$ is $\mathfrak t$. By \cite[Lemma 1.1]{kt} or \cite{viro} the signature of $X$ is equal to the signature of the oriented link. The formula for the $\rho$-invariant then follows from \eqref{eq:rhodef} since the first Chern class of the spin structure $\mathfrak s$ is zero.
The equality for quasi-alternating links is proved in \cite{qa}.
\end{proof}
\section{Examples}
\label{sec:eg}
In this section we will complete Kohn's table of unlinking numbers for two-component links with crossing number 9 or less by showing that each of
$$9^2_3=L9a30,\, 9^2_{15}=L9a15,\, 9^2_{27}=L9a17,\, 9^2_{31}=L9a2,\,9^2_{36}=L9a10$$
has unlinking number 3. In fact we will show that none of them bound a pair of normally properly immersed disks in the 4-ball with two double points. Each of these may be unlinked by changing 3 crossings in their minimal diagrams, as may be readily seen from Figures \ref{fig:alllinks} and \ref{fig:L9a10}.
\subsection{The links $L9a2$, $L9a15$, $L9a17$, $L9a30$}
The following lemma implies the statement in Example \ref{eg:kohn}.
\begin{figure}[htbp]
\begin{center}
\labellist
\small\hair 3pt
\pinlabel $L9a2$ at 210 620
\pinlabel $L9a15$ at 700 620
\pinlabel $L9a17$ at 210 40
\pinlabel $L9a30$ at 700 40
\endlabellist
\includegraphics[width=0.70\textwidth]{L9a_2_15_17_30}
\caption {The links $L9a2$, $L9a15$, $L9a17$, and $L9a30$.}
\label{fig:alllinks}
\end{center}
\end{figure}
\begin{lemma}
\label{lem:kohn}
For any oriented link in $S^3$, half the absolute value of the signature is a lower bound for the 4-ball crossing number.
Let $L$ be a 2-component link in $S^3$. Suppose that $H=H^2(\Sigma_2(L);{\mathbb Z})$ is finite cyclic. If $c^*(L)<3$, then at least one of the following holds:
\begin{enumerate}[(i)]
\item $\det(L)=2t^2$, or
\item $\det(L)$ is divisible by 4, and at least one of the signatures of $L$ has absolute value 1, or
\item $\det(L)$ is divisible by 16.
\end{enumerate}
\end{lemma}
\begin{proof}
The well-known fact that the 4-ball crossing number is bounded below by half of the absolute value of
the signature follows as in the proof of Lemma \ref{lem:ineq}, since signature
is a concordance invariant and changes by at most 2 when a crossing changes,
and since the signature of the unlink is zero.
Now suppose $L$ is a 2-component link and $H=H^2(\Sigma_2(L);{\mathbb Z})$ is finite
cyclic. Since $H$ is finite, the nullity of $L$ is zero. It follows from
Lemma \ref{lem:crossingchange} that for any orientation of $L$, the signature
$\sigma$ is odd. Thus if $c^*(L)<3$ then the signatures of $L$, for either
quasi-orientation, are elements of $S=\{-3,-1,1,3\}$.
By assumption, the link $L$ bounds a normally immersed pair of disks in $D^4$
with $p$ positive and $n$ negative double points, with $p+n=2$ (if $c^*(L)<2$
we introduce extra double points).
We claim that after possibly reflecting and for some choice of orientation, we have equality in
\eqref{eq:pineq}. This is immediate if $L$ has signature $-3$, and if $L$ has
signature $3$ then we can reflect. If $L$ has signature 1, then $n \geq 1$ by Lemma \ref{lem:ineq}.
Thus either $p=0$ and we have equality in \eqref{eq:pineq} for $L$ or $p=1$ and
we have equality in \eqref{eq:pineq} for $-L$.
Assume then that $L$ is oriented with equality in \eqref{eq:pineq}. By Theorem \ref{thm:mainthm}, $Y=\Sigma_2(L)$ bounds a 4-manifold $W$ whose
intersection form, in some basis, is given by
$$Q=\begin{pmatrix}-a & b & c\\ b &
-2 & 0\\ c & 0 & -2 \end{pmatrix}.$$
Since this is negative-definite and
presents a finite cyclic group, we may assume after change of basis that either
$a=b=1$ and $c=0$, or $b=c=1$ and $a>1$. In the first case $Q$ has determinant $-2$. From the long exact sequence of the pair $(W,Y)$
we conclude that $H$ has order $2t^2$ for some integer $t$ (for more details see for example \cite[Lemma 2.1]{qhs}).
Now suppose that the intersection matrix of $W$ is represented by
$$Q=Q_a=\begin{pmatrix}-a & 1 & 1\\ 1 & -2 & 0\\ 1 & 0 & -2
\end{pmatrix},$$
for some $a>0$. This presents ${\mathbb Z}/(4a-4){\mathbb Z}$. It follows that $H$ has order $(4a-4)t^2$ for some integer $t$; in particular $\det(L)$ is divisible by 4. If $t$ is even then $\det(L)$ is divisible by 16. Suppose instead that $t$ is odd. It follows from \cite[Lemma 2.4]{qhs} that both cosets of $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/T$ which are fixed under the
involution coming from conjugation in fact contain a spin structure. Thus we
can combine Proposition \ref{prop:rho} and Lemma \ref{lem:dbc} to compute the
signatures of $L$ using $Q$. Fixed points of the involution on
$\Char(Q)/2Q({\mathbb Z}^n)$ are given by characteristic covectors $\xi$ with
$$\xi=-\xi+2Qx,$$ or in other words $\xi=Qx$ for some $x\in{\mathbb Z}^n$. There are
two such cosets for $Q_a$: if $a$ is odd they are represented by $(1,-2,0)$ and $(1,0,-2)$, both with square $\xi^TQ_a^{-1}\xi=x^TQ_ax=-2$, and if $a$ is even they are represented by $(0,0,0)$ with square 0 and $(2,-2,-2)$ with square $-4$. Thus by Proposition \ref{prop:rho} if $a$ is odd then the spin structures on $Y$ both have $\rho$-invariant $(-2+3)/4=1/4$ and if $a$ is even their $\rho$-invariants are $3/4$ and $-1/4$. In either case we find that at least one of the signatures of $L$ has absolute value 1.
\end{proof}
The double branched cover of the link $L9a30$ has cyclic second cohomology group of order 30. The links $L9a2$, $L9a15$, and $L9a17$ all have signature $\sigma=-3$ for both choices of orientation, possibly after reflecting, and $H^2(\Sigma_2(L);{\mathbb Z})\cong{\mathbb Z}/40{\mathbb Z}$. Applying Lemma \ref{lem:kohn} we conclude that each has $u(L)=c^*(L)=3$.
\subsection{The link $L9a10$ using Donaldson's diagonalisation theorem}
We take $L$ to be the link $L9a10$, oriented and reflected so that the signature is $1$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=\textwidth]{L9a10both.eps}
\caption {The link $L9a10$, with orientation and chessboard colouring. The link on the left has signature $-1$; its mirror on the right has signature $1$.}
\label{fig:L9a10}
\end{center}
\end{figure}
First consider the link $-L$ as in the first diagram of Figure \ref{fig:L9a10}, which has signature $-1$. Using the chessboard colouring shown in the figure and following \cite{gl}, the Goeritz matrix is
$$\left(
\begin{matrix} 5&-1&-1\\-1&4&-2\\-1&-2&4 \end{matrix}
\right).$$
Applying \eqref{eq:pineq} to $-L$ leaves two
possibilities: a normally immersed pair of disks in $D^4$ bounded by $-L$ may
have one double point of each sign, or two positive double points. There are 4 embeddings of the Goeritz lattice $\Lambda$ into the standard integer lattice ${\mathbb Z}^{6}$,
up to automorphisms of $\Lambda$ and ${\mathbb Z}^6$. These can be found by hand; they can also be found using the
\texttt{OrthogonalEmbedings} command in GAP \cite{gap} which implements an algorithm due to Plesken \cite{plesken}. The complete list is as follows:
$$\begin{aligned}
&\{2e_1+e_2,-e_2+e_3+e_4+e_5,-e_2-e_3-e_4-e_5\},\\
&\{2e_1+e_2,-e_1+e_2+e_3+e_4,-e_2-e_3+e_5+e_6\},\\
&\{e_1+e_2+e_3+e_4+e_5,e_1-e_2-e_3+e_6,-e_1+e_2-e_3-e_6\},\\
&\{e_1+e_2+e_3+e_4+e_5,e_1-e_2-e_3+e_6,-e_1+e_4-e_5-e_6\}.\\
\end{aligned}$$
None of these has a pair of orthogonal square two vectors in the orthogonal complement. Corollary \ref{cor:alt} then implies we do not have $p=n=1$. Thus if $-L$ bounds a pair of immersed disks in the 4-ball with two double points then they must both be positive.
We now consider $L$, as in the second diagram of Figure \ref{fig:L9a10}, with signature $1$. Reflecting, we see from the preceding paragraph that if $L$ bounds a pair of immersed disks in the 4-ball with two double points then both are negative, so that we again have equality in \eqref{eq:pineq}. The Goeritz matrix is
$$\left(
\begin{matrix} 2&-1&0&0&-1&0\\-1&2&-1&0&0&0\\0&-1&3&-1&0&0\\
0&0&-1&3&-1&-1\\-1&0&0&-1&3&0\\0&0&0&-1&0&2 \end{matrix}
\right)$$
and the complete list of embeddings in ${\mathbb Z}^{9}$ is
$$\begin{aligned}
&\{e_1+e_2,-e_2+e_3,-e_3+e_4+e_5,-e_5+e_6+e_7,-e_1-e_4+e_5,-e_7+e_8\},\\
&\{e_1+e_2,-e_2+e_3,-e_3+e_4+e_5,-e_5+e_6+e_7,-e_1-e_7+e_8,-e_4+e_5\},\\
&\{e_1+e_2,-e_2+e_3,-e_3+e_4+e_5,-e_5+e_6+e_7,-e_1-e_7+e_8,-e_6+e_9\},\\
&\{e_1+e_2,-e_2+e_3,-e_3+e_4+e_5,-e_5+e_6+e_7,-e_1-e_7+e_8,-e_7-e_8\},\\
&\{e_1+e_2,-e_2+e_3,-e_1+e_2+e_4,-e_4+e_5+e_6,-e_2-e_3+e_4,-e_5+e_7\}.\\
\end{aligned}$$
Each of the first four do not admit a pair of orthogonal square two vectors in the orthogonal complement. The last embedding is immediately seen to have image in ${\mathbb Z}^7$, with orthogonal complement in ${\mathbb Z}^9$ isomorphic to
$$\langle3\rangle\oplus{\mathbb Z}^2,$$
which does contain a pair of orthogonal square two vectors. However these span a finite-index sublattice of ${\mathbb Z}^2$ which is not primitive. We conclude from Corollary \ref{cor:alt} that $L$ does not bound a pair of immersed disks in $D^4$ with two double points, and that $u(L)=3$.
\subsection{The link $L9a10$ using Heegaard Floer correction terms}
The determinant of $L=L9a10$ is 48, and the double branched cover has cyclic homology group. As in the argument above using Donaldson's theorem, we apply Theorem \ref{thm:mainthm} to both $-L$, which has signature $-1$ for either quasi-orientation, and to $L$ which has both signatures equal to $1$.
We begin by arguing as in the proof of Lemma \ref{lem:kohn}. We need to obstruct both $Y=\Sigma(L)$ and $-Y$ from bounding a 4-manifold $W$ with intersection form represented by the matrix
$$Q_a=\begin{pmatrix}-a & 1 & 1\\ 1 & -2 & 0\\ 1 & 0 & -2
\end{pmatrix}.$$
This has determinant $4a-4$. If it is bounded by $\pm Y$ then $4a-4$ divides $48$ with quotient a square from which it follows that $a$ is either 4 or 13. The $\rho$-invariants of the spin structures of $\pm Y$ are both $\mp1/4$ by Lemma \ref{lem:dbc}. This obstructs $-Y$ from bounding $Q_4$ and it obstructs $Y$ from bounding $Q_{13}$, as in the proof of Lemma \ref{lem:kohn}: if a rational homology sphere $M$ with cyclic homology group bounds $Q_{4}$ then at least one of the spin structures of $M$ has $\rho$-invariant in $\{-1/4,3/4\}$, and if $M$ bounds $Q_{13}$ then at least one of its spin structures has $\rho$-invariant $1/4$.
We complete the proof by using $\rho$-invariants to show that $-Y$ cannot bound $Q_{13}$, and by using $d$-invariants to show that $Y$ cannot bound $Q_4$. To do this we need to compute the $d$-invariants of $Y$, which also give the $\rho$-invariants by reducing modulo 2.
The entire set of correction terms of the double branched cover of a nonsplit alternating link can be obtained, using a computer, from the Goeritz matrix of the link as in \cite{osu1}. We briefly recall how to do this.
Given a negative-definite symmetric integer matrix $G=(g_{ij})$, we partition the set of characteristic covectors $\xi=(\xi_1,\dots,\xi_n)$ with $g_{ii}\le \xi_i<-g_{ii}$ into cosets of $\Char(G)/2G{\mathbb Z}^n$, and use the Smith normal form of $G$ to record the affine group structure of $\Char(G)/2G{\mathbb Z}^n$. We also maximise the quantity
$(\xi^TQ^{-1}\xi+n)/4$ on each of these finite sets of coset representatives; this gives a function
$$m_G:\Char(G)/2G{\mathbb Z}^n\to{\mathbb Q}.$$
Proposition \ref{prop:d} can then be restated with the left hand side of inequality \eqref{eq:dineq} replaced by $m_G([\xi])$. By \cite[Proposition 3.2]{osu1}, if $G$ is the negative-definite Goeritz matrix of a nonsplit alternating diagram of $L$, then in fact $m_G$ computes the correction terms of $\Sigma_2(L)$, or in other words, there is an \emph{isomorphism} $\phi$ as in Proposition \ref{prop:d} with equality in \eqref{eq:dineq}.
In our case we
find the correction terms of $Y$ are, in cyclic order starting at a spin structure,
$$\left\{
\arraycolsep=1.4pt\def2.2{2.2}
\begin{array}{rrrrrrrrrrrr}
-\dfrac14,&\dfrac{17}{48},&\dfrac{1}{6},&-\dfrac{13}{16},&-\dfrac{7}{12},&-\dfrac{55}{48},&-\dfrac{1}{2},&-\dfrac{31}{48},&\dfrac{5}{12},&\dfrac{11}{16},&\dfrac{1}{6},&-\dfrac{55}{48},\\
-\dfrac{5}{4},&-\dfrac{7}{48},&\dfrac{1}{6},&-\dfrac{5}{16},
&\dfrac{5}{12},&\dfrac{17}{48},&-\dfrac{1}{2},&-\dfrac{7}{48},&-\dfrac{7}{12},&\dfrac{3}{16},&\dfrac{1}{6},&-\dfrac{31}{48},\\
-\dfrac{1}{4},&-\dfrac{31}{48},&\dfrac{1}{6},&\dfrac{3}{16},&-\dfrac{7}{12},&-\dfrac{7}{48},&-\dfrac{1}{2},&\dfrac{17}{48},&\dfrac{5}{12},&-\dfrac{5}{16},
&\dfrac{1}{6},&-\dfrac{7}{48},\\
-\dfrac{5}{4},&-\dfrac{55}{48},&\dfrac{1}{6},&\dfrac{11}{16},&\dfrac{5}{12},&-\dfrac{31}{48},&-\dfrac{1}{2},&-\dfrac{55}{48},&-\dfrac{7}{12},&-\dfrac{13}{16},&\dfrac{1}{6},&\dfrac{17}{48}
\end{array}
\right\}.$$
If we compute the full $\rho$-invariant of $Q=Q_{13}$, for example by computing $m_Q$ as above (though in fact to compute the $\rho$-invariant one just needs an arbitrary representative from each coset of $\Char(Q)/2Q{\mathbb Z}^n$), we find that for $\xi=(3,2,0)$ we have $m_Q([\xi])=-1/12$. Since the reduction modulo 2 of the $d$-invariants gives the $\rho$-invariants of $Y$, we see that $1/12$ is not the $\rho$-invariant of any \ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi structure on $Y$. By Proposition \ref{prop:rho} we conclude that $-Y$ does not bound $Q_{13}$.
Now consider the form $Q=Q_4$. Computing as above we find the values of $m_Q$ are, in cyclic order starting at a fixed point of the involution on $\Char(Q)/2Q{\mathbb Z}^3$,
$$\left\{
\arraycolsep=1.4pt\def2.2{2.2}
\begin{array}{rrrrrrrrrrrr}
&-\dfrac{1}{4},&\dfrac{1}{6},&-\dfrac{7}{12},&-\dfrac{1}{2},&\dfrac{5}{12},&\dfrac{1}{6},\dfrac34,&\dfrac{1}{6},&\dfrac{5}{12},&-\dfrac{1}{2},&-\dfrac{7}{12},&\dfrac{1}{6}
\end{array}
\right\}.$$
There is a unique affine monomorphism $\phi$ from $\Char(Q)/2Q{\mathbb Z}^3$ to $\ifmmode{{\rm Spin}^c}\else{${\rm spin}^c$\ }\fi(Y)/({\mathbb Z}/2{\mathbb Z})$ satisfying the conditions of Proposition \ref{prop:rho}. Labelling the second list above from $0$ to $11$ and the list of $d$-invariants from $0$ to $47$, the only possibility is
$$\phi:i\mapsto \{2i,2i+24\pmod{48}\}.$$
This however (just!) fails to satisfy \eqref{eq:dineq} since $3/4>-5/4$.
We conclude that in fact $Y$ does not bound $Q_4$, and that $u(L)=c^*(L)=3$.
\section{Unlinking numbers of prime links with 9 crossings or less}
Table \ref{table:u} contains the unlinking numbers and 4-ball crossing numbers of all prime nonsplit links with crossing number 9 or less and with more than one component. Unknotting numbers and 4-ball crossing numbers of prime knots with 9 or fewer crossings are available from \cite{knotinfo}.
With the exception of the links in Examples \ref{eg:kohn} and \ref{eg:L9a10}, the unlinking numbers were known previously due to work of Kauffman-Taylor, Kohn, Kawauchi, and Borodzik-Friedl-Powell \cite{kt, kohn, kaw14,bfp}, and in many cases these were also known to be equal to the 4-ball crossing number.
For all links in the table, the unlinking number is realised by changing a subset of the crossings in the minimal diagram given in \cite{knotatlas}. That the minimal number of crossings in each case is equal to the 4-ball crossing number follows from applying one or more of the results in this paper.
The required lower bound is provided by Lemma \ref{lem:det} for the links $L5a1$, $L6a4$, $L7a1$, $L7a3$, $L7a4$, $L7a6$, $L8a1$, $L8a8$, $L8a9$, $L8a16$, and $L9aN$ for
$$N\in \{1,3,4,8,9,18,20,21,22,25,26,27,35,38,40,42\}.$$
The bound comes from Lemma \ref{lem:ineq} for $L9a14$, $L9a29$ and $L9a36$, and from Lemma \ref{lem:kohn} for $L9a2$, $L9a15$, $L9a17$, and $L9a30$. The bound for $L9a10$ is obtained in Section \ref{sec:eg}. For all remaining links in the table the required bound may be obtained using Lemma \ref{lem:sublinks}.
The fact that the invariants $c^*(L)$ and $u(L)$ coincide in this table is due to the small crossing number restriction. There are nontrivial nonsplit prime slice links with
10-crossing diagrams (for example $L10n32$ and $L10n36$ in \cite{linkinfo,knotatlas}) for which these invariants will clearly differ.
\begin{table}[ht]
\begin{center}
\small
\begin{tabular}[t]{lc}
\toprule
Link $L$ & $u(L)$\\
\midrule
$L2a1$ &1\\
$L4a1$ &2\\
$L5a1$ &1\\
$L6a1$ &2\\
$L6a2$ &3\\
$L6a3$ &3\\
$L6a4$ &2\\
$L6a5$ &3\\
$L6n1$ &3\\
$L7a1$ &2\\
$L7a2$ &3\\
$L7a3$ &2\\
$L7a4$ &2\\
$L7a5$ &1\\
$L7a6$ &2\\
$L7a7$ &3\\
$L7n1$ &3\\
$L7n2$ &1\\
$L8a1$ &2\\
$L8a2$ &1\\
$L8a3$ &3\\
$L8a4$ &1\\
$L8a5$ &3\\
$L8a6$ &2\\
$L8a7$ &3\\
$L8a8$ &2\\
$L8a9$ &2\\
$L8a10$ &3\\
$L8a11$ &3\\
$L8a12$ &4\\
$L8a13$ &4\\
$L8a14$ &4\\
$L8a15$ &3\\
\bottomrule
\end{tabular}%
\qquad
\begin{tabular}[t]{lcc}
\toprule
Link $L$ & $u(L)$\\
\midrule
$L8a16$ &3\\
$L8a17$ &4\\
$L8a18$ &4\\
$L8a19$ &2\\
$L8a20$ &4\\
$L8a21$ &4\\
$L8n1$ &3\\
$L8n2$ &1\\
$L8n3$ &4\\
$L8n4$ &4\\
$L8n5$ &2\\
$L8n6$ &4\\
$L8n7$ &4\\
$L8n8$ &4\\
$L9a1$ &2\\
$L9a2$ &3\\
$L9a3$ &2\\
$L9a4$ &2\\
$L9a5$ &3\\
$L9a6$ &4\\
$L9a7$ &3\\
$L9a8$ &2\\
$L9a9$ &2\\
$L9a10$ &3\\
$L9a11$ &3\\
$L9a12$ &4\\
$L9a13$ &3\\
$L9a14$ &3\\
$L9a15$ &3\\
$L9a16$ &3\\
$L9a17$ &3\\
$L9a18$ &2\\
$L9a19$ &2\\
\bottomrule
\end{tabular}%
\qquad
\begin{tabular}[t]{lcc}
\toprule
Link $L$ & $u(L)$\\
\midrule
$L9a20$ &2\\
$L9a21$ &1\\
$L9a22$ &2\\
$L9a23$ &4\\
$L9a24$ &2\\
$L9a25$ &2\\
$L9a26$ &2\\
$L9a27$ &1\\
$L9a28$ &4\\
$L9a29$ &3\\
$L9a30$ &3\\
$L9a31$ &2\\
$L9a32$ &4\\
$L9a33$ &3\\
$L9a34$ &2\\
$L9a35$ &2\\
$L9a36$ &3\\
$L9a37$ &2\\
$L9a38$ &1\\
$L9a39$ &2\\
$L9a40$ &2\\
$L9a41$ &2\\
$L9a42$ &2\\
$L9a43$ &4\\
$L9a44$ &4\\
$L9a45$ &3\\
$L9a46$ &2\\
$L9a47$ &3\\
$L9a48$ &4\\
$L9a49$ &4\\
$L9a50$ &3\\
$L9a51$ &4\\
$L9a52$ &3\\
\bottomrule
\end{tabular}%
\qquad
\begin{tabular}[t]{lcc}
\toprule
Link $L$ & $u(L)$\\
\midrule
$L9a53$ &2\\
$L9a54$ &3\\
$L9a55$ &4\\
$L9n1$ &3\\
$L9n2$ &2\\
$L9n3$ &1\\
$L9n4$ &4\\
$L9n5$ &2\\
$L9n6$ &2\\
$L9n7$ &3\\
$L9n8$ &2\\
$L9n9$ &3\\
$L9n10$ &2\\
$L9n11$ &2\\
$L9n12$ &4\\
$L9n13$ &1\\
$L9n14$ &2\\
$L9n15$ &4\\
$L9n16$ &4\\
$L9n17$ &2\\
$L9n18$ &4\\
$L9n19$ &4\\
$L9n20$ &4\\
$L9n21$ &4\\
$L9n22$ &4\\
$L9n23$ &3\\
$L9n24$ &3\\
$L9n25$ &2\\
$L9n26$ &3\\
$L9n27$ &1\\
$L9n28$ &3\\%linking numbers & Whitehead sublink
&\\
&\\
\bottomrule
\end{tabular}
\end{center}
\medskip
\caption{Unlinking numbers and 4-ball crossing numbers of prime nonsplit links with 9 or fewer crossings.}
\label{table:u}
\end{table}
\clearpage
\bibliographystyle{amsplain}
|
\section{The interplay between neutron particle physics and cosmology}
There is a profound connection between cosmology -- the science of the infinitely large --
and particle physics -- the science of the infinitely small.
The primordial creation of the light nuclei is an example:
a detailed knowledge of nuclear interactions is needed to predict the outcome of the Big Bang nucleosynthesis.
Conversely, cosmological observations have a great impact on our understanding of the laws of Nature at the most fundamental level.
The existence of Dark Matter, Dark Energy and the matter-antimatter asymmetry provide three glimpses of what is hiding beyond the Standard Model of particle physics.
It is likely that future progress in particle physics will proceed from the solution of one of these cosmological puzzles, or vice versa.
The deep connection between particle physics and cosmology is particularly manifest in neutron particle physics,
when we use low energy neutrons to investigate the fundamental interactions and symmetries.
The work presented in the next chapters builds on this tradition.
We will contemplate the possibility to probe models of Dark Energy with neutron experiments.
Before developing on this original idea, we review in the present chapter the traditional interplay between neutron particle physics and cosmology.
A comprehensive treatment of this topic can be found in the excellent review of \textcite{Dubbers2011}.
See also \textcite{Dubbers2014} for a shorter and more recent overview.
There are two fundamental properties of the neutron connected to the physics of the early Universe.
First, the value of the {\bf neutron lifetime} is a basic input for the theory of Big Bang nucleosynthesis.
Second, the search for the {\bf neutron electric dipole moment} is key to elucidate the origin of the baryon asymmetry of the Universe.
We will explain these two subjects, after a brief review of neutron particle physics and cosmology.
\subsection{Free up neutrons}
Neutrons make up about half the mass of ordinary objects, like bicycles, cheese or elephants.
Even an empty bottle, that is, full of air, contains more than $10^{20}$ neutrons per cm$^3$.
These neutrons are bound into nuclei since billions of years.
On the contrary free neutrons do not exist naturally because they decay with a radioactive period of ten minutes.
It takes a significant effort to extract neutrons from nuclei; it is much more difficult than extracting electrons from atoms.
Indeed the binding energy of neutrons in nuclei is of the order of 1~MeV whereas the chemical binding energy is only about a few electron-Volts.
Thus, no chemical reaction would set neutrons free: a nuclear reaction is needed.
\para{Neutron sources. }
About twenty large nuclear installations dedicated to neutron production are presently being operated over the world.
These installations fall into two categories: fission nuclear reactors and spallation sources.
The reactor of the Institut Laue Langevin (ILL) in Grenoble belongs to the first category.
Neutrons are produced by fission in a compact cylindrical core (containing 9 kg of weapons grade uranium, 93 \% $^{235}$U) and moderated in heavy water.
Each fission ejects on average 2.5 neutrons with an energy of 2~MeV.
The reactor design optimizes the thermal neutron flux, reaching $10^{15}$~cm$^{-2}$~s$^{-1}$ with a reactor thermal power of 58~MW.
In a typical nuclear reactor used to produce electricity with a thermal power of 3000~MW, the neutron flux is about $10^{14}$~cm$^{-2}$~s$^{-1}$.
Among the 246 operating research reactors listed by the International Atomic Energy Agency,
46 of them are high flux reactors (above $10^{14}$ neutrons cm$^{-2}$~s$^{-1}$),
11 of them feature a cold neutron source and operate instruments available to external users (see Table \ref{listReactors}).
\begin{table}
\center
\caption{
High flux reactors producing beams of cold neutrons.
The number of instruments available for outside users is indicated.
}
\scriptsize
\begin{tabular}{lllll}
\hline
Reactor & City & Started & Th. Power & Instruments \\
\hline
ILL & Grenoble (France) & 1971 & 58 MW & 45 \\
HFIR & Oak Ridge (USA) & 1965 & 85 MW & 12 \\
CARR & Bejing (China) & 2010 & 60 MW & 12 \\
FRM II & Munich (Germany) & 2004 & 20 MW & 19 \\
HANARO & Daejon (Korea) & 1995 & 30 MW & 8 \\
WWR-M & Gatchina (Russia) & 1959 & 18 MW & 19 \\
NIST & Gaithersburg (USA) & 1967 & 20 MW & 28 \\
JRR-3M & Tokai (Japan) & 1990 & 20 MW & 26 \\
BRR & Budapest (Hungary) & 1959 & 10 MW & 11 \\
OPAL & Sydney (Australia) & 2006 & 20 MW & 7 \\
BER II & Berlin (Germany) & 1973 & 10 MW & 22
\end{tabular}
\label{listReactors}
\end{table}
The ultracold neutron facility at the Paul Scherrer Institute (PSI) in Switzerland uses a neutron source based on the spallation process.
A proton beam, with an energy of 600~MeV and an intensity of 2.5~mA is shot at a lead target.
When a proton collides with a heavy nucleus (lead in this case), it ejects about 20 neutrons with an energy of about 20~MeV along with other nuclear fragments.
The neutrons are then moderated in a heavy water tank surrounding the spallation target.
Spallation sources are now competing with reactors as neutron factories, Table \ref{listSpallation} lists the spallation sources
equipped with neutron instruments available to external users.
\begin{table}
\center
\caption{
Neutron spallation sources serving instruments available to users.
}
\scriptsize
\begin{tabular}{llll}
\hline
Spall. Source & City & Beam Power & Instruments \\
\hline
SINQ @PSI & Villigen (Switzerland) & 1.5 MW & 20 \\
SNS @ORNL & Oak Ridge (USA) & 1.4 MW & 25 \\
JSNS @KEK & Tsukuba (Japan) & 0.3 MW & 17 \\
ISIS @RAL & Oxford (UK) & 0.2 MW & 36 \\
Lujan @LANSCE & Los Alamos (USA) & 0.1 MW & 13
\end{tabular}
\label{listSpallation}
\end{table}
What is the purpose of investing so much efforts and money for building sources of free neutrons?
What is special about free neutrons that cannot be done with the neutrons inside nuclei?
Beyond doubt the main desirable feature of the neutron is its electrical neutrality
\footnote{The particle data group \cite{PDG} evaluation of the neutron charge is $(-2 \pm 8) \times 10^{-22}$ electron charges, in agreement with electrical neutrality.}.
Even atoms, which have no net electric charge, interact mainly with electromagnetic interactions due to their relatively large polarisability.
Atomic processes are essentially unaffected by nuclear or gravitational interactions and all phenomena we experience in everyday life are ultimately due to basic electromagnetic interactions (except for the rather trivial effect of weight).
Using neutrons one can probe matter using non electromagnetic interactions.
Studying the diffraction of neutrons at a sample material,
solid-state physicists are able to obtain information about the structure of the sample
which is complementary to the information they get using electromagnetic probes such as X-ray diffraction.
Particle physicists also find neutron's sensitivity to non-electromagnetic forces quite useful to do fundamental physics.
Studying free neutron beta decay we measure the fundamental parameters of the weak interaction.
Measuring neutron bounces we explore gravity with quantum objects.
The search for the neutron electric dipole moment constitutes a crucial test of the time reversal invariance in elementary processes.
One can also use neutrons to search for new, yet undiscovered interactions.
All these experiments would be impossible with charged particles for which the Coulomb force overwhelms any other force by many orders of magnitude.
We shall come back to these subjects in the rest of this manuscript.
At the ILL, 38 instruments are dedicated to the study of material structure, condensed matter and magnetism and 7 instruments are dedicated to quantum, nuclear and particle physics.
For a broad exploration of the physics of slow neutrons, see \textcite{Byrne1994}.
\para{Wave particle duality. }
Experiments very often benefit from using slow neutrons.
Those neutrons result from the thermalisation of the primary fast neutrons in a thermal or cold moderator.
In a moderator, a neutron will dissipate its energy by collision until it reaches the kinetic energy ($kT = 25$~meV at room temperature) of the molecules.
The De-Broglie wavelength $\lambda$ of a neutron with kinetic energy $E$ is
\begin{equation}
\lambda = \frac{2 \pi \hbar}{\sqrt{2 m E}} = 0.18 \ {\rm nm} \times \sqrt{\frac{25 \ {\rm meV}}{E}},
\end{equation}
where $m = 939.6$~MeV/c${^2}$ is the neutron mass.
It is remarkable that the wavelength of thermal neutrons corresponds to the typical distance between atoms in solid matter.
Thus, whereas fast neutrons behave essentially as particles, cold neutrons could behave like waves.
E.~Fermi was the first to realize that slow neutrons could undergo optical phenomena such as reflection and refraction at surfaces.
It can be shown (see for example \textcite{Golub1991}) that matter acts as a uniform medium with a potential energy for slow neutrons called the \emph{Fermi potential}.
The Fermi potential is given by the following expression:
\begin{equation}
V_F = 4 \pi \ \frac{\hbar^2}{2 m} \ b \, n,
\label{FermiPot}
\end{equation}
where $b$ is the bound coherent neutron scattering length of the nuclei constituting the material and $n$ is its number density.
If the material is heterogeneous one must sum the potentials of all nuclear species composing the material.
For most materials the Fermi potential is positive, i.e. repulsive
\footnote{
This is surprising because the strong nuclear interaction responsible for the Fermi potential is attractive: it holds neutrons inside nuclei.
For an explanation of this apparent paradox, see for example \textcite{Pignol2009}.};
it is of the order of $10^{-7}$~eV, much smaller than the kinetic energy of thermal neutrons.
For example, natural nickel is a material with a relatively high Fermi potential which amounts to $250$~neV.
Neutrons approaching a surface at grazing incidence could be reflected by the Fermi potential of the material.
Total reflection occurs when the kinetic energy associated with the velocity normal to the surface is smaller than the potential barrier \eqref{FermiPot}.
In practice this phenomenon is at play in neutron guides.
It is possible to transport neutrons from the core of the reactor where they are produced to an experimental hall situated at a distance of up to 100~m,
using evacuated rectangular tubes, with a cross section of typically 100~cm$^2$, made up of plates coated with nickel or with a multilayer of nickel and titanium.
Thermal neutrons with an energy of $E = 25$~meV are reflected off a nickel surface if the angle of incidence $\theta$
satisfies the condition $\sin^2 \theta < V_F / E$, that is, $\theta < 0.2 \deg$.
For colder neutrons, the efficiency of the guiding is better and a beam of cold neutrons transported in a guide can have a larger angular divergency.
\para{Ultracold neutrons. }
\textcite{Zeldovich1959} realized that neutrons with total kinetic energy lower than $250$~neV
should be reflected at any angle of incidence and therefore could be stored in material bottles.
These storable neutrons were called \emph{ultracold neutrons (UCNs)} by the pioneer workers in the field.
It is not easy to get ultracold neutrons, because their proportion in the energy spectrum of thermal neutrons
(when neutrons are thermalised in a moderator at room temperature) is only $10^{-11}$.
The first storage of ultracold neutrons was reported by \textcite{Groshev1971}.
The group of Russian physicists stored on average 2 neutrons at a time in a copper cylindrical bottle (diameter 14~cm and length 174~cm).
The capacity of the bottle to store neutrons was quantified by a storage time of about 30~s (compare to the neutron beta decay lifetime of 880~s).
The kinetic energy of ultracold neutrons corresponds to a temperature as low as a few mK.
However, it is important to understand that ultracold neutrons can be stored in bottles at room temperature.
The neutrons do not thermalise with the walls of the bottle, like photons of visible light do not thermalise when reflecting off a mirror.
In fact the specular (i.e. mirror) reflection of ultracold neutrons and visible photons share many similarities, because the wavelength of UCNs is of the order of 100~nm,
very close to the wavelength of visible light.
In both cases the wavelength is much larger than the lattice spacing of atoms in the matter of the walls.
It means that the particle interacts with a large number of atoms in the wall, it is almost blind to the thermal motion of individual atoms.
Using appropriate wall materials, UCNs can undergo as many as $10^4$ specular reflections before being inelastically scattered or captured by a single nucleus.
The ability to store neutrons for a long period was immediately recognized as a great opportunity for fundamental physics, for two main reasons.
First, ultracold neutrons provide a direct method to measure the neutron lifetime by simply storing neutrons for a certain time and counting the survivors.
Second, longer observation times enhance the sensitivity of detection schemes like the Rabi or Ramsey resonance techniques, following the uncertainty relation $\delta E \times T \geq \hbar$,
where $\delta E$ is the precision of the measured energy and $T$ is the observation time.
A 1~m long apparatus observes a neutron spin in a cold beam for typically 10~ms, whereas in a storage experiment observation times longer than 100~s could be achieved.
The measurement of the neutron electric dipole moment (nEDM) could thus benefit from an improvement of the sensitivity by four orders of magnitude.
Since 1970 the techniques to produce and handle ultracold neutrons were greatly improved to pursue these two goals:
a precise measurement of the neutron lifetime and the search for the nEDM.
We will come back to these two topics after taking a cosmological detour.
\subsection{The standard picture of the early Universe}
The modern picture of the Big Bang was initiated by G. Lema\^itre, supported by the measurement of the expansion of the Universe by \textcite{Hubble1929}.
At that time, the modern theory of gravity, Einstein's general relativity, was just invented.
Within this framework it was possible to consider the Universe as a physical system interacting with its content.
Assuming the cosmological principle - the Universe is homogeneous and isotropic on large scales -
the cosmic evolution of a spatially flat Universe is described by the Robertson-Walker metric
\begin{equation}
\label{RobertsonWalker}
ds^2 = dt^2 - a(t) \left( dx^2 + dy^2 + dz^2 \right)
\end{equation}
where $a(t)$ is the dimensionless scale factor defined in such a way that at the present time $t_0$ the scale factor is unity $a(t_0) = 1$.
In our expanding Universe, the Hubble parameter $H(t) = \dot{a}(t)/a(t)$ is positive.
In the cosmic past, all distances were smaller by the factor $a(t)$.
For example, light emitted at time $t$ with a wavelength $\lambda$ is observed now with a larger wavelength $\lambda_0 = \lambda / a(t) = (z+1) \lambda$.
The redshift $z = \frac{1}{a(t)} -1$, being actually observable, is often preferred to $t$ as a variable to indicate the ticking of the cosmic clock.
Assuming that the Universe is filled with an homogeneous fluid with an energy density $\rho$ and a pressure $p$,
Einstein's equations of general relativity (with a vanishing cosmological constant) reduce to the Friedmann equations
\begin{eqnarray}
& & H^2 = \frac{8 \pi G}{3} \rho \label{F1} \\
& & \frac{d \rho}{dt} + 3 H (\rho + p) = 0.
\label{F2}
\end{eqnarray}
In the early Universe, the expansion was dominated by radiation, i.e. photons and relativistic particles.
In this case the pressure is related to the energy density as $p = \rho/3$.
Then from the Friedmann equations we deduce $\rho \propto a^{-4}(t)$, $a(t) \propto \sqrt{t}$ and $H(t) = \frac{1}{2 t}$.
Going backwards in time (higher $z$) the energy density of the radiation $\rho \propto (z+1)^4$ was higher.
Assuming that the radiation was in thermal equilibrium, one can associate a temperature $T$ to the radiation according to
\begin{equation}
\rho = \frac{\pi^2}{30} N(T) T^4,
\end{equation}
where $N(T)$ counts the effective number of relativistic degrees of freedom ($+1$ per boson polarization state, $+7/8$ per fermion state).
Combining with the Friedmann equations we can relate the age of the Universe $t$ to the temperature $T$
\begin{equation}
\label{Tvs_t}
\ t = \sqrt{\frac{45}{16 \pi^3 G N(T)}} \frac{1}{T^2} = \frac{2.4 \ {\rm s}}{\sqrt{N(T)}} \ \frac{1 \ {\rm MeV}^2}{T^2}.
\end{equation}
The Universe was hotter in the past.
For example, at the beginning of the Big Bang nucleosynthesis the temperature was $T \approx 1$~MeV,
the effective numbers of degrees of freedom was $N(T) = 29/4$ (only photons and three families of neutrinos),
thus the age of the Universe was $t \approx 0.5$~s.
One can also relate the photon temperature to the redshift,
\begin{equation}
\label{Tvs_z}
T = (z+1) T_0 \left( \frac{N(T_0)}{N(T)} \right)^{1/4},
\end{equation}
where $T_0 = 2.7255(6)$~K is the present temperature of the Cosmic Microwave Background (CMB).
Note that Eq. \eqref{Tvs_t} is valid only in the radiation dominated epoch whereas Eq. \eqref{Tvs_z} holds for the whole thermal history of the Universe.
\begin{table*}
\center
\caption{
The history of the Universe: standard scenario.
Values in bold are taken from \textcite{PDG}.
}
\begin{tabular}{lllll}
\hline
& Transition & Temperature & Time & Redshift $z$ \\
\hline
\multirow{8}{*}{\includegraphics[width=0.07\linewidth]{thermometer}}
& Inflation ends (reheating) & $\sim 10^{15}$~GeV & $\sim 10^{-37}$~s & $\sim 10^{28}$ \\
& Electroweak phase transition & $170$~GeV & $10^{-11}$~s & $10^{15}$ \\
& QCD phase transition & $170$~MeV & $30 \, \mu$s & $10^{12}$ \\
& Big Bang nucleosynthesis & $80$~keV & $150$~s & $3 \times 10^8$ \\
& Matter radiation equality & $1$~eV & $30000$~Yr & ${\bf 3360(70)}$ \\
& CMB decoupling (recombination) & $0.3$~eV & ${\bf 372000}$~Yr & ${\bf 1090.2(7)}$ \\
& First stars (reionization) & $30$~K & ${\bf 460}$~MYr & ${\bf 11}$ \\
& Today & ${\bf 2.7255(6)}$~K & ${\bf 13.81(5)}$~GYr & $0$
\end{tabular}
\label{historyUniverse}
\end{table*}
Very different energy scales of the primordial plasma were successively at play in the thermal history of the Universe.
This leads to a fascinating and fruitful connection between cosmology and particle physics that relate phenomena at the highest energies with phenomena at the earliest epoch.
Before exploring more specifically the relevance of low energy neutron physics, which is a subfield of particle physics, we shall shortly review the present standard scenario of the hot Big Bang (see Table \ref{historyUniverse}).
For a more complete review on the status of the standard cosmological model, see \textcite{Bartelmann2010}.
It is believed that the Universe started with a phase of accelerated expansion driven by a scalar field called the inflaton \cite{Guth1981}.
What happened before is even more speculative because it is related with physics at the Planck scale and involves the great problem of unifying gravity with quantum mechanics.
The inflaton field is supposed to be coupled to Standard Model particles and the energy density of the scalar field would be converted into particles by the end of the inflation.
This \emph{reheating} process would start the thermal history of the Universe with an initial temperature of presumably about $10^{15}$~GeV.
Then the Universe cooled down while expanding and a succession of phase transitions occurred.
At temperatures corresponding to the electroweak scale ($T \approx 200$~GeV),
the Higgs boson field expectation value condensed from the symmetric phase $\langle \phi \rangle = 0$ down to the non-symmetric phase $\langle \phi \rangle = 246$~GeV.
We will come back to the electroweak phase transition in section \ref{EDM}.
Next, at a temperature below $T = 100$~MeV, quarks and gluons condensed into nucleons during the QCD phase transition.
The Universe was then populated by photons, neutrinos, relativistic electrons and positrons, and traces of protons and neutrons.
The nucleons are leftovers from a tiny asymmetry between quarks and antiquarks, the so-called baryon asymmetry of the Universe (BAU) which was generated earlier.
Further condensation of nucleons into nuclei became possible at the temperature of $T \approx 80$~keV during the Big Bang nucleosynthesis (BBN).
We will explain the connection between the BBN and the measurement of the neutron lifetime in the next section.
The radiation dominated period of the Universe ended at a temperature of $T = 1$~eV,
when the energy density of dark matter became larger than that of radiation.
Shortly after, at $T = 0.3$~eV, electrons and protons combined into hydrogen atoms and the Universe became transparent to photons.
When we observe the CMB today we see in fact photons emitted during the recombination at the redshift of $z\approx 1000$.
In the matter dominated Universe, the primordial density fluctuations grew until star formation began at the very recent redshift of $z=11$.
In the late luminous Universe the expansion is again accelerating, perhaps driven by a new substance called Dark Energy.
We will develop in the next chapters some possible ways to address the Dark Energy problem with neutron experiments.
\subsection{Big Bang nucleosynthesis and the neutron lifetime}
\para{The Big Bang nucleosynthesis. }
\textcite{Gamow1948} formulated the hypothesis that nuclear fusion in the early Universe is responsible for the formation of nuclei.
He argued that the primordial nucleosynthesis started at a temperature of $T_{\rm BBN} = 100$~keV, when the radiative dissociation of deuterium stopped.
The age of the Universe was then a few minutes, which he considered to be the timescale associated with nucleosynthesis.
Conveniently, this is also the order of magnitude of the neutron lifetime, which was poorly known in 1948.
Then, he required the mean time for the reaction $p + n \rightarrow d + \gamma$ to be about 2 minutes.
Indeed, if the mean time was much longer then no complex nuclei would have been formed during the BBN.
On the other hand if the mean time was much shorter then there would be no deuterium left (subsequent fusion reactions happen to be faster than deuterium formation).
Quantitatively this educated guess, sometimes referred to as the Gamow criterion, can be expressed as
\begin{equation}
n_b \ \sigma v \times 2 \ {\rm min } \approx 1,
\end{equation}
where $\sigma \approx 10^{-29}$~cm$^2$ is the fusion cross section, $v \approx 5 \times 10^8$~cm/s is the thermal velocity of the protons and neutrons at the temperature $T_{\rm BBN}$ and $n_b$ is the number density of protons and neutrons.
With the above criterion Gamow got $n_b \approx 10^{18}$~cm$^{-3}$ corresponding to a mass density of $\rho_b(t_{\rm BBN}) \approx 10^{-6}$~g/cm$^3$.
Following this line of argument one can predict the present temperature of the CMB, $T_0 = T_{\rm BBN} \sqrt[3]{\rho_b(t_0)/\rho_b(t_{\rm BBN})}$,
by using the scaling laws $T(t) \propto a(t)^{-1}$ and $\rho_b(t) \propto a(t)^{-3}$.
Using the estimate $\rho_b(t_0) \approx 10^{-30}$~g/cm$^3$ obtained from the Hubble rate,
\textcite{Alpher1949} predicted $T_0 \approx 5$~K.
The Big Bang model was unambiguously confirmed in 1965 when Penzias and Wilson discovered the CMB.
Building on the seminal work of Gamow, a more sophisticated theory of the BBN was worked out.
A key quantity is the ratio of baryon to photon number densities $\eta = n_b/n_\gamma$.
For a given value of $\eta$ the theory can predict the relative abundance of the four light elements produced during the BBN, namely D, $^3$He, $^4$He and $^7$Li.
From the observed relative abundances of each of these elements one can extract $\eta$.
The most sensitive probe of $\eta$ comes from the measurement of the deuterium to proton ratio in extragalactic clouds, D/H$ = (2.53 \pm 0.04) \times 10^{-5}$,
from which the following result is obtained \cite{Cooke2014}
\begin{equation}
\label{etaBBN}
\eta_{\rm BBN} = (6.0 \pm 0.1) \times 10^{-10}.
\end{equation}
The study of the temperature anisotropies in the CMB provides
an independent determination of the baryon density of the Universe at the recombination.
The Planck satellite \cite{Planck2014} has produced the result
\begin{equation}
\label{etaCMB}
\eta_{\rm CMB} = (6.05 \pm 0.07) \times 10^{-10},
\end{equation}
in excellent agreement with the BBN value.
In order to compute the abundances from the BBN theory, the neutron lifetime $\tau_n$ needs to be known with sufficient accuracy,
because the number of neutrons available for fusion at $T_{\rm BBN}$ depends on $\tau_{\rm n}$.
According to state-of-the-art calculations by \textcite{Coc2014}, the sensitivity of the deuterium fraction to the values of the input parameters is
\begin{equation}
\frac{\Delta ({\rm D}/{\rm H})}{{\rm D}/{\rm H}} = -1.6 \, \frac{\Delta \eta}{\eta} + 0.4 \, \frac{\Delta \tau_{\rm n}}{\tau_{\rm n}}.
\end{equation}
The parametric uncertainty due to the error on the neutron lifetime becomes negligible compared to the observational error if $\Delta \tau_{\rm n} \ll 30$~s.
We will explain in the next paragraph how the needed sub-percent accuracy has been achieved in the 1990's.
\para{The measurement of the neutron lifetime. }
The neutron decays due to the weak interaction into a proton, an electron and an antineutrino with a period of about ten minutes.
It is the simplest case of nuclear beta decay.
Given its importance in cosmology and particle physics, the neutron lifetime has been measured by more than 20 experiments.
There are two distinct experimental approaches to measure the neutron lifetime $\tau_{\rm n}$.
\begin{enumerate}
\item {\bf The beam method}.
A detector records the decay products in a well-defined part of a neutron beam.
A neutron beam is indeed radioactive due to beta decay and the rate of appearance in the beam of the decay products is
\begin{equation}
- \frac{d N}{dt} = \frac{N}{\tau_{\rm n}}.
\end{equation}
This method requires (i) a determination of the number of neutrons $N$ in the beam,
that is, an absolute determination of the neutron flux and (ii) a detector for decay products with a well calibrated registration efficiency to measure $dN/dt$.
\item {\bf The bottle method}.
A bottle is filled with UCNs. After a certain waiting time $t$, the trapped neutrons are emptied into a detector to count the number of remaining neutrons $N(t)$.
One repeats this operation for various times $t$ and the storage time $\tau_{\rm stor}$ is extracted from a fit of the storage curve
\begin{equation}
N(t) = N(0) \, e^{-t/\tau_{\rm stor}}.
\end{equation}
The efficiency of the neutron counter does not need to be known accurately, but the losses due to absorption or inelastic scattering during collisions with the walls must be carefully controlled in order to determine $\tau_{\rm n}$ from the storage time $\tau_{\rm stor}$.
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=0.97\linewidth]{neutronLifetime}
\caption{
Results of the latest neutron lifetime measurements with the beam method (\textcolor{blue}{$\CIRCLE$}), material bottle method (\textcolor{red}{$\blacksquare$}) and magnetic bottle method (\textcolor{magenta}{$\diamondsuit$}).
Dashed lines correspond to measurements which have been withdrawn or corrected later on.
}
\label{neutronLifetime}
\end{figure}
The current situation of the measurements is summarized in Fig. \ref{neutronLifetime}.
References to the publications of each experiment can be found in the recent review by \textcite{Wietfeldt2011},
except for the recent reevaluation of the latest beam result by \textcite{Yue2013}.
Concerning the beam method, one can choose to measure either the electron or the proton activity of the beam (it would be silly to try to measure the neutrino activity).
Although early experiments (not represented in Fig. \ref{neutronLifetime}) detected the electrons coming out of the beam,
modern measurements count the protons.
Combining the two most recent measurements we obtain
\begin{equation}
\tau_{\rm n}^{\rm beam} = ( 888.0 \pm 2.1 ) \, {\rm s}.
\label{lifetimeBeam}
\end{equation}
Let us now come to the storage method where one measures the storage time $\tau_{\rm stor}$ of UCNs in a material bottle,
which is a combination of the beta decay lifetime and the lifetime due to losses at wall collisions : $1/\tau_{\rm stor} = 1/\tau_{\rm n} + 1/\tau_{\rm wall}$.
Most of the time neutrons are specularly reflected at wall collisions, but there is a small probability of capture or up-scattering.
The loss probability can be of the order of $\mu \approx 10^{-5}$, using walls coated with hydrogen-free oil such as fomblin for example.
In a typical material bottle with mean distance $\lambda \approx 30$~cm, UCNs with velocity $v \approx 3$~m/s will collide with the walls at a frequency $f = v/\lambda \approx 10$~Hz.
Then we expect $\tau_{\rm wall} = 1/f \mu \approx 10^4$~s, which corresponds to a 10~\% correction of the beta decay lifetime.
The usual strategy to control this effect is to measure the storage time for various geometries of the bottle, with different volume to surface ratio,
and extrapolate to the ideal case of vanishing wall collision frequency.
In some experiments the bottle is cooled down to reduce the neutron losses due to up-scattering.
The five UCN measurements performed at the ILL produced the combined value
\begin{equation}
\tau_{\rm n}^{\rm bottle} = ( 879.6 \pm 0.8 ) \, {\rm s}.
\label{lifetimeBottle}
\end{equation}
There is a $3.8 \sigma$ discrepancy between the beam result \eqref{lifetimeBeam} and the bottle result \eqref{lifetimeBottle} that needs to be resolved.
To alleviate the issue of the losses at wall collisions which are not fully understood yet, new projects aim at confining UCNs in a magnetic bottle.
The neutron is a spin 1/2 particle with a magnetic moment $\mu_n = - 60$~neV/T.
Thus a magnetic ``wall'' of the order of 1~T acts as a repulsive potential of height 60~neV for low field seekers neutrons
(i.e. neutrons with spin parallel to the magnetic field).
The measurement of the neutron lifetime is still an active field of research.
Several teams in Europe and in the US are currently attempting to obtain a reliable measurement with an accuracy of 0.1~s.
\subsection{The matter antimatter asymmetry and the neutron electric dipole moment}
\label{EDM}
Apparently our Universe is made up of matter, not antimatter.
Not a single complex antinucleus like antihelium has been detected in cosmic rays.
No excess of gamma radiation resulting from the annihilation of antimatter with matter is observed.
Most cosmologists believe that the matter dominance over antimatter extends to at least the whole visible Universe.
In fact, the imbalance is tiny.
When the baryons were in equilibrium with the rest of the plasma in the early Universe just before the QCD phase transition,
for every billion baryon-antibaryon pair there was one spare baryon.
It should be noted that we quantify the matter-antimatter asymmetry by the baryon asymmetry.
For sure there are also more electrons than antielectrons in the Universe today.
However, an excess of antineutrinos over neutrinos in the cosmic neutrino background could perhaps compensate for the electron-positron asymmetry.
Therefore we do not know for certain that an excess of matter over antimatter exists for leptons.
As we have seen, the asymmetry parameter $\eta = n_b/n_\gamma$ deduced from considerations about the primordial nucleosynthesis \eqref{etaBBN}
agrees with the one deduced from the microwave background \eqref{etaCMB}.
Given that the two methods rely on two completely separated epochs in the Universe, the agreement is remarkable.
\para{The Sakharov conditions to generate the baryon asymmetry. }
We have solid evidence that the baryon asymmetry exists since before the Big Bang nucleosynthesis.
Is it merely an initial condition of the Big Bang tuned to allow intelligent life to emerge?
In the context of inflation this idea is no longer plausible, because inflation would have tremendously diluted the initial baryon density.
Well before inflation was invented, \textcite{Sakharov1967} proposed that the asymmetry could have been generated dynamically from an initially symmetric state.
He outlined three conditions that are necessary for this to be possible:
\begin{enumerate}
\item The baryon number should not be conserved.
\item The Universe should at some time depart from thermal equilibrium.
\item The discrete symmetries C (charge conjugation) and CP (charge conjugation combined with parity transformation) should be violated.
\end{enumerate}
See \textcite{Bernreuther2002} for a pedagogical introduction to baryogenesis, the hypothetical process in the early Universe that meets all these conditions to generate the baryon asymmetry.
\para{Electroweak baryogenesis. }
The second Sakharov condition suggests that the baryon asymmetry was generated during a phase transition in the early Universe before the primordial nucleosynthesis.
Electroweak baryogenesis, a mechanism that biases the baryon number during the electroweak phase transition,
is one of the most attractive among the many proposed realizations of baryogenesis.
In fact, all three Sakharov conditions are qualitatively fulfilled by the standard electroweak theory although it fails quantitatively to predict the correct baryon asymmetry.
Nevertheless, electroweak baryogenesis is still a viable scenario in extensions of the Standard Model (SM) of particle physics.
New physics is required at or close to the electroweak scale, this makes the scenario testable - and falsifiable - by current or planned experiments.
Let us take a closer look at the Sakharov conditions in the context of the electroweak phase transition.
Surprisingly, the SM does accommodate baryon number violation,
induced by non-perturbative effects associated with the nontrivial structure of the SU(2) gauge fields vacuum.
We do not observe baryon number violation in the laboratory because the process requires a quantum tunneling through a large energy gap with an extremely small probability.
However, \textcite{Kuzmin1985} discovered that baryon number violation processes called \emph{sphalerons} were frequent in the early Universe when the temperature was high enough to overcome the energy barrier.
In fact, the first Sakharov condition does not demand physics beyond the Standard Model.
\begin{figure}
\centering
\includegraphics[width=0.97\linewidth]{EWPT_SM}
\includegraphics[width=0.97\linewidth]{EWPT_BSM}
\caption{
Effective Higgs potential for different temperatures at the electroweak phase transition.
Top: calculation within the SM, a smooth cross-over occurs at $T = 170$~GeV.
Bottom: calculation within the $\phi^6$ extension, a first order phase transition occurs at $T = 120$~GeV.
}
\label{EWPT}
\end{figure}
Next, the second Sakharov condition deserves a discussion about the electroweak phase transition, when the Higgs field acquired a non-zero expectation value.
When coupled to the thermal bath at temperature $T$, the effective potential of the SM Higgs field $\phi$ can be written
as a sum of the bare potential and a thermal correction:
\begin{equation}
\label{HiggsPotential}
\mathcal{V}_{\rm eff}(\phi, T) = - \frac{\mu^2}{2} \phi^2 + \frac{\lambda}{4} \phi^4 + \mathcal{V}_{\rm th}(\phi, T).
\end{equation}
At zero temperature the thermal correction to the potential is absent ($\mathcal{V}_{\rm th} = 0$) and the potential \eqref{HiggsPotential} has the famous Mexican hat shape
with a minimum at $\phi = \sqrt{\mu^2/\lambda} = 246$~GeV.
At high temperature the thermal correction $\mathcal{V}_{\rm th} \neq 0$ modifies the potential in such a way that the minimum of the potential is $\phi = 0$.
When the temperature decreased, the field $\phi$ condensed from the symmetric state $\phi=0$ to the state $\phi = 246$~GeV, breaking spontaneously the electroweak symmetry.
Successful baryogenesis requires that the phase transition be first order: the field must change discontinuously from $\phi = 0$ to $\phi >0$.
Figure \ref{EWPT} (Top) shows the modification of the potential close to the phase transition at $T = 170$~GeV.
The formulas for the effective potential are taken from the thesis of \textcite{Fromme2006},
where I specified the value of the Higgs mass to $m_h = 126$~GeV conforming with the recent discovery at the Large Hadron Collider.
We see that the minimum of the potential smoothly moves away from zero, there is no sufficient departure from thermal equilibrium and no baryogenesis in the Standard Model.
Now, a non-standard Higgs potential could lead to a first order electroweak phase transition.
As an example, we could add to the potential \eqref{HiggsPotential} a non-renormalisable operator of the form $\frac{1}{8 \Lambda} \phi^6$.
We plot in Fig. \ref{EWPT} the modification of the effective potential in the case $\Lambda = 700$~GeV, using again the formulas in \textcite{Fromme2006}.
In this case the minimum of the potential suddenly changes from $\phi = 0$ to $\phi = 140$~GeV at the critical temperature of $T_{\rm cr} = 120$~GeV, this is a first order phase transition.
If such a transition did take place in the early Universe, bubbles of true vacuum started to nucleate and expand in a sea of false vacuum.
In such a boiling environment, the Universe could have been driven sufficiently out of equilibrium for baryogenesis to be possible.
A first order electroweak phase transition can be realized in more sophisticated extensions of the scalar sector of the SM,
such as the two-Higgs-doublet model or supersymmetric extensions.
Since these models modify the physics at the electroweak scale, they are testable at particle colliders. .
Last, electroweak baryogenesis requires CP violation at the electroweak scale.
It turns out that the CP violation contained in the SM, from phase $\delta$ of the quark mixing matrix, is not strong enough to generate the baryon asymmetry.
Thus CP-violating new physics is required to satisfy the third Sakharov condition as well.
Such new physics is best probed by low energy precision experiments searching for electric dipole moments (EDM) of particles.
In summary, the failure of the SM to allow for the electroweak baryogenesis is a hint toward the presence of new physics lying just above the electroweak scale that may be discovered both by collider experiments and EDM searches.
\begin{figure}
\centering
\includegraphics[width=0.77\linewidth]{EDM_loop}
\caption{
One loop diagram contributing to the quark EDM.
The quark EDMs then transfer almost directly to the neutron EDM.
}
\label{EDM_loop}
\end{figure}
\para{CP violation and neutron EDM. }
The existence of a nonzero EDM for a spin 1/2 particle such as the neutron would imply the violation of the CP symmetry.
In the Standard Model, the induced neutron EDM expected from the $\delta$ phase is tiny, $d_n \approx 10^{-32} \ e$~cm.
This value is to be compared to the current best limit obtained at the ILL by \textcite{Baker2006}:
\begin{equation}
|d_n| < 3 \times 10^{-26} \ e \, {\rm cm \ (90 \, \% \ C.L.)}.
\end{equation}
Therefore, improvements of the neutron EDM measurement are motivated by the potential discovery of a new source of CP violation beyond the SM.
As for any low energy observable, the new physics at a high energy scale manifests itself via quantum loops involving virtual particles.
Figure \ref{EDM_loop} shows a Feynman diagram contributing to a quark EDM via the CP-violating vertices of heavy scalars and fermions with masses $M$.
Generically the neutron EDM induced by such a loop amounts to \cite{Pospelov2005}
\begin{equation}
d_n \approx \left( \frac{1 \, {\rm TeV}}{M} \right)^2 \times \sin \phi \times 10^{-25} \ e \, {\rm cm}
\end{equation}
where $M$ is the mass of the particles running in the loop.
In this case the heavy particles couple strongly with the quark
(such as the SUSY coupling between quark, squark and gluinos) with a CP-odd vertex multiplied by $\sin \phi$.
The CP-odd part usually originates from the imaginary part of some parameter in the Lagrangian and $\phi$ would then correspond to the CP-violating phase of that specific parameter.
Thus, considering natural CP violation ($\sin \phi \approx 1$) in the new heavy sector, the neutron EDM is sensitive to new physics at the multi-TeV scale.
Searches for permanent electric dipole moments of other particles (protons, deuterons, muons, atoms, molecules) are complementary probes of CP violation above the electroweak scale.
Many experimental efforts are under way, they have been compiled recently by \textcite{Kirch2013}.
Below we concentrate on the search for the neutron EDM.
\para{Measuring the neutron EDM. }
The neutron EDM measurement is based on the analysis of the Larmor precession frequency of neutrons, stored in a volume permeated with electric and magnetic static fields either parallel or antiparallel.
For such configurations, the precession frequency $f_n$ reads
\begin{equation}
h f_n = \mu_n B \pm d_n E
\end{equation}
where $\mu_n$ and $d_n$ are the magnetic and electric dipole moments of the neutron and $h$ is Planck's constant.
The frequency difference of these two configurations gives directly access to the neutron EDM: $d_n = h \Delta f_n / 4E$.
To measure the precession frequency, we use Ramsey's method of separated oscillatory fields which provides a statistical precision per cycle of
\begin{equation}
\sigma(d_n) = \frac{\hbar}{2 \alpha E T \sqrt{N}}
\end{equation}
where $T$ is the precession time, $\alpha$ is the visibility (related to the polarization of the neutrons) and $N$ the total number of detected neutrons.
Using stored ultracold neutrons, the precession time can be set to $T = 200$~s, five orders of magnitude
longer as compared to the first experiment performed by \textcite{Smith1957} using a neutron beam!
The main experimental challenge in current experiments consists in achieving a magnetic field homogeneity at the
level of $10^{-4}$ over a volume of typically 20 $\ell$
while maintaining a temporal stability of about $10^{-7}$ over 100 s.
These requirements are necessary to control the subtle systematic effects (see for example \textcite{Pignol2012}).
Atomic magnetometry and magnetic shielding techniques are therefore at the core of such a measurement.
\begin{figure}
\centering
\includegraphics[width=0.97\linewidth]{OILL}
\caption{
The nEDM apparatus installed at the PSI.
The precession chamber sits in a large cylindrical vacuum chamber (with front end open on the picture), surrounded by a four-layer mumetal magnetic shield.
The UCN guide injects the neutrons from the bottom.
High voltage is applied using the HV feed-trough connected to the top electrode.
}
\label{OILL}
\end{figure}
As of 2015 there are six projects worldwide competing for improving the measurement of the neutron EDM, all at different stages.
\begin{enumerate}
\item A PNPI group is currently operating a room-temperature double-chamber spectrometer at the ILL \cite{Serebrov2014}.
Later the apparatus will be moved back to PNPI where a new UCN source will be built.
\item Another PNPI group pursues a completely different experimental method.
It exploits inter-atomic electric fields in crystals that are $10^4$ times stronger than electric fields produced by charging electrodes as in all other experiments.
However the interaction time is much shorter because it is a neutron scattering technique.
The result obtained recently at the ILL \cite{Fedorov2010} is two orders of magnitude less precise than the conventional UCN method.
\item A German project at Munich \cite{Altarev2012} is waiting for the construction of a new UCN source at the reactor FRM-2.
The newly built magnetically shielded room performs very well \cite{Altarev2014}.
This large room is designed to host a double-chamber room-temperature experiment.
\item A Canada-Japan collaboration plans to build a superthermal UCN source and a room-temperature EDM apparatus at TRIUMF Vancouver,
based on the prototype operated at RNCP Osaka \cite{Masuda2012,Matsuta2013,Masuda2014}.
\item An ambitious project is pursued by the US community based on the original proposal of \textcite{Golub1994}.
The idea is to run the experiment in superfluid $^4$He, that serves both as a superthermal UCN source and a precession chamber.
Traces of spin-polarized $^3$He will be injected in the volume to both detect the neutrons and act as a comagnetometer \cite{Ye2009}.
The project is still in the R\&D phase and will start real data taking at SNS Oak Ridge in 2020 at best.
\item A European collaboration (Belgium, France, Germany, Poland, Switzerland and UK) is running an experiment at the PSI UCN source \cite{Baker2011}.
We are currently operating an upgraded version of the RAL-Sussex spectrometer (see picture Fig. \ref{OILL}) which has produced the lowest experimental limit for the neutron EDM
and that we moved from the ILL to the PSI in 2009.
In the same time we have started the design of n2EDM, a next generation room-temperature double-chamber apparatus in view of its delivery around 2018.
With the current apparatus we aim at slightly improving the present limit on the neutron EDM whereas with n2EDM we aim at a precision of better than $10^{-27} \, e$~cm.
In addition, we have produced several spin-off scientific results with the apparatus
(i) a search for neutron to mirror-neutron oscillations \cite{Ban2007,AltarevPRD2009}
(ii) a sensitive test of Lorentz invariance \cite{Altarev2009,Altarev2010,Altarev2011}
(iii) a measurement of the neutron magnetic moment with an uncertainty of 0.8 ppm \cite{Afach2014}
(iv) a search for Axionlike particles \cite{Afach2015}.
\end{enumerate}
Although the experiments are getting more and more difficult
(in the 1950s and 1960s two or three persons could run an experiment, nowadays author lists rarely count less than 20 people)
prospects are good to improve the accuracy on the neutron EDM by more than an order of magnitude in the next decade.
\bigskip \begin{center} \photon \end{center} \bigskip
In conclusion, experiments with neutrons address the cosmological question of the origin of the baryonic matter.
Today we understand how protons and neutrons combined in the early Universe to form nuclei.
Experiments measuring the neutron lifetime contributed a fare share of the successful prediction of the Big Bang nucleosynthesis.
However we still do not know how the neutrons and protons were created in the first place, because we do not understand the baryon asymmetry of the Universe.
Planned measurements of the neutron EDM will contribute to solve this question.
\clearpage
\section{Chameleon Dark Energy}
\label{Chameleon}
The 2011 Nobel Prize in physics was awarded to S. Pelmutter, B. Schmidt and A. Riess
for the discovery of the late acceleration of the expansion of the Universe.
Although the acceleration is now established as a fact, it remains a great puzzle.
We do not know the nature of the Dark Energy responsible for the acceleration.
The two other cosmological puzzles, the origin of the baryon asymmetry and the nature of the Dark Matter, are thought to be related to physics at high energy, beyond the electroweak scale.
On the contrary, the energy scale associated to the Dark Energy has the peculiar value of 2~meV.
It suggests that there is still new physics in the infrared to be discovered, perhaps with low energy precision experiments using neutrons?
\subsection{The accelerated expansion of the Universe}
\begin{figure}
\centering
\includegraphics[width=0.88\linewidth]{Hubble1929}
\includegraphics[width=0.99\linewidth]{Perlmutter1999}
\caption{
Top: diagram taken from \textcite{Hubble1929} representing the velocity (deduced from the redshift) of nearby galaxies as a function of distance.
Bottom: supernovae Hubble diagram taken from \textcite{Perlmutter1999} representing the magnitude (indicating the distance) as a function of the redshift.
Notice that the highest velocity of $1000$~km/s in the 1929 diagram corresponds to the redshift $z = 0.003$ which is smaller than the smallest redshift in the supernovae diagram.}
\label{Hubble}
\end{figure}
A linear relation between the distance and radial velocity among galaxies, $v = H_0 d$, was first established by \textcite{Hubble1929}.
The first Hubble diagram is shown in Fig. \ref{Hubble}.
Incidentally, due to an error in the distance evaluations, Hubble derived an expansion rate of $500$~km/s/Mpc which is much larger than the most recent determination by \textcite{Planck2014}:
\begin{equation}
H_0 = 67.3 \pm 1.2 \quad {\rm km/s/Mpc}.
\end{equation}
Hubble's linear law was discovered using observations of nearby galaxies.
At much larger distances, the relation is not expected to be linear anymore.
The expansion of the Universe was thought to slow down because of the attractive power of gravity.
The deceleration of the Universe was then searched for by drawing a Hubble diagram with more distant objects.
Explosions of type IA supernovae are
(i) very bright, they can be seen at cosmological distances
(ii) very reproducible, they all have almost the same intrinsic brightness.
A sufficient number of supernovae observations using the Hubble space telescope and ground based telescopes were reported by two teams \cite{Riess1998,Perlmutter1999}.
The supernovae Hubble diagram of one team is shown in Fig. \ref{Hubble}.
Surprisingly, it was found that the expansion of the Universe accelerates.
Let us analyse this incredible finding in the framework of a flat Universe
whose expansion is driven by an homogeneous substance with an energy density $\rho$ and a pressure $p$ (i.e. a perfect fluid).
We assume an \emph{equation of state} of the substance of the form $p = w \rho$.
From the Friedmann equations \eqref{F1},\eqref{F2} one derives the acceleration:
\begin{equation}
\frac{\ddot{a}}{a} = - \frac{4 \pi G}{3} \rho \, (1+3 w).
\end{equation}
A substance made up of non-relativistic matter ($w=0$) or radiation ($w=1/3$) would decelerate the expansion.
Instead, a positive acceleration requires $w < -1/3$.
The mysterious substance driving the expansion must have a negative pressure.
This substance is generically called Dark Energy.
\subsection{The $\Lambda$CDM concordance model}
In the standard model of cosmology, the acceleration of the expansion is attributed to a perfect fluid with $w = -1$.
According to the second Friedmann equation \eqref{F2},
the energy density of this fluid is constant.
It does not dilute with the expansion, it is a cosmological constant.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{AgeUniverse}
\caption{
The age of the Universe calculated in the flat $\Lambda$CDM model as a function of $\Omega_\Lambda$.
The vertical orange band represents the Dark Energy fraction reported by \textcite{Planck2014}.
\label{ageUniverse}
}
\end{figure}
In the late Universe, the expansion is driven by two distinct perfect fluids: the cosmological constant and some non-relativistic matter (radiation became unimportant a few million years after the Big Bang).
Hence the name ``$\Lambda$CDM'' for the standard model, which stands for a cosmology with a cosmological constant and cold dark matter (baryonic matter accounts for only a fraction of the non-relativistic matter).
We note $\rho_{M,0}$ and $\rho_{\Lambda,0}$ the present value of the energy density associated with the matter and the cosmological constant,
and $\rho_{c,0} = \frac{3 H_0^2}{8 \pi G}$ the critical density.
We use the standard notation $\Omega_M = \rho_{M,0} / \rho_{c,0}$ and $\Omega_\Lambda = \rho_{\Lambda,0}/\rho_{c,0}$,
which represent the fraction of the total energy density in the form of matter and cosmological constant.
For a flat Universe, Eq. \eqref{F1} considered at the present epoch implies $\Omega_M + \Omega_\Lambda = 1$.
From the second Friedmann equation \eqref{F2} the past values of the densities are
\begin{equation}
\rho_M(t) = \frac{\rho_{M,0}}{a(t)^3} \quad , \quad \rho_\Lambda(t) = \rho_{\Lambda,0}.
\end{equation}
Then, Eq. \eqref{F1} can be brought to the form
\begin{equation}
\left( \frac{\dot{a}}{a} \right)^2 = H_0^2 \left( \frac{\Omega_M}{a^3} +\Omega_\Lambda \right).
\end{equation}
This is the equation describing the expansion rate in the flat $\Lambda$CDM model.
Let us calculate the age of the Universe, by transforming the previous equation into:
\begin{equation}
1 = \frac{\dot{a} \, \sqrt{a}}{H_0 \sqrt{\Omega_M + a^3 \Omega_\Lambda}}.
\end{equation}
Next, we integrate the equation from $t=0$ to $t=t_0$ ($t_0$ is the age of the Universe).
Recall that in our conventions $a(t_0) = 1$.
We get
\begin{equation}
t_0 = \frac{1}{H_0} \int_0^1 \frac{\sqrt{a} \, da}{\sqrt{\Omega_M + a^3 \Omega_\Lambda}} = \frac{2}{3 H_0} \frac{ {\rm asinh} \sqrt{\Omega_\Lambda / \Omega_M }}{\sqrt{\Omega_\Lambda}}.
\end{equation}
The age of the Universe is plotted in Fig. \ref{ageUniverse} as a function of $\Omega_\Lambda$.
We can compare it to the age of the oldest stars, estimated to at least 11~Gyr \cite{Bartelmann2010}.
The Einstein-de Sitter model, i.e. the flat matter dominated Universe with $\Omega_M = 1$ and $\Omega_\Lambda = 0$,
is excluded because the Universe cannot be younger than the stars.
The long-standing disagreement between the Hubble expansion rate and the age of the oldest stars has finally been resolved in the $\Lambda$CDM model with the introduction of the cosmological constant.
In addition to the supernovae and the age of oldest stars, the acceleration of the Universe is now supported by several other independent cosmological observations including baryon acoustic oscillations, weak gravitational lensing and counts of galaxy clusters.
These observational probes of cosmic acceleration were recently reviewed in \textcite{Weinberg2013} and future prospects are described.
Since the discovery of the accelerated expansion, the precision of the cosmological data increased spectacularly, in particular the measurements of the CMB.
The $\Lambda$CDM model, which emerged at the end of the last century as a concordance cosmology, is still able to describe all observations.
$\Lambda$CDM fits indicate that the energy budget of the Universe today is:
\begin{equation}
\Omega_b = 5 \, \% \quad \Omega_{\rm DM} = 27 \, \% \quad \Omega_\Lambda = 68 \, \%,
\end{equation}
where the non-relativistic matter has a baryonic and non-baryonic component ($\Omega_M = \Omega_{\rm DM} + \Omega_b$).
These three components driving the expansion of the Universe today are associated with the three major puzzles in cosmology.
It is often said that the baryonic content of the Universe is the only part which is well understood.
In fact, the very presence of baryons results from the initial matter-antimatter asymmetry of the Universe, it cannot be explained by the standard theory.
Next, the microscopic description of the non-baryonic Dark Matter is lacking.
It could be made up of massive weakly interacting particles a.k.a. WIMPS.
It could also well be something completely different like a Bose-Einstein condensate of very light scalar particles like Axions.
Last, the microscopic nature of the Dark Energy is completely unknown.
Contrary to Dark Matter, no well motivated extensions of the Standard Model of particle physics provide a natural candidate for the Dark Energy.
\subsection{The cosmological constant problem}
In natural units ($\hbar = c = 1$), the energy density associated with the Dark Energy is
\begin{equation}
\label{LAMBDA}
\rho_\Lambda = \Omega_\Lambda \, \rho_{c,0} = (2.2 \, {\rm meV})^4.
\end{equation}
The fact that the energy scale associated with the Dark Energy is so curious constitutes the cosmological constant problem.
In the $\Lambda$CDM model, the Dark Energy is simply the energy density of the vacuum, or equivalently a cosmological constant.
In principle the vacuum energy is expected to receive contributions from several sources.
At the classical level, any scalar field $\phi$ contributes to the vacuum energy by the amount $\mathcal{V}(\phi_0)$ where $\mathcal{V}(\phi)$ is the energy density function of the field and $\phi_0$ is the vacuum value of the field.
One could argue that the zero energy is arbitrary, we can choose to define it as the state of lowest energy for all fields in the Universe.
All local phenomena do not depend on this choice of the zero energy, there is no reason that it should have such a dramatic gravitational effect on cosmological scales.
However, in the case of phase transitions, the effect of zero energy cannot be ignored by simply declaring that the zero energy cannot gravitate.
During the electroweak symmetry breaking mentioned in section \ref{EDM}, the energy density associated with the Higgs field changes, by an amount as large as $(100 \, {\rm GeV})^4$.
The vacuum energy was much larger before the electroweak phase transition.
This difference in vacuum energy is definitely a physical meaningful quantity, according to straightforward application of the general relativity theory in a purely classical context.
If differences in vacuum energy certainly gravitate, what determines the actual value of the vacuum energy?
It could perfectly correspond to an intrinsic property of the Universe that should be treated as another free parameter.
Why then is it adjusted so that the vacuum energy after the electroweak phase transition is almost zero, but not quite?
It seems that Nature decided to ``lift'' the Mexican hat potential of the Higgs field in a very precise, fine-tuned amount.
In addition, the vacuum energy receives contributions from quantum fluctuations induced by virtual particles.
Quantum field theories predict that every bosonic degree of freedom of frequency $\omega$ has a zero-point energy of $\hbar \omega /2$.
Since there are in principle an infinite number of degrees of freedom in the quantum fields,
the total vacuum energy is technically infinite,
unless the contribution from the fermionic and bosonic degrees of freedom compensate each other.
If physics can be described by an effective local quantum field theory up to the Planck scale,
\begin{equation}
m_{\rm Pl} = \sqrt{\frac{\hbar c}{8 \pi G}}= 2.4 \times 10^{18} \, {\rm GeV},
\end{equation}
dimensional analysis indicate that the vacuum energy induced by quantum fluctuations should be of the order of $m_{\rm Pl}^4 \approx 10^{120} \rho_\Lambda$.
This is definitely the worse theoretical prediction ever.
Supersymmetric compensation between the fermionic and bosonic quantum fluctuations above the SUSY breaking scale
could reduce the disagreement between the expected and observed vacuum energy from $10^{120}$ down to $10^{58}$, still enormous.
In summary, the vacuum energy can be separated as a sum of two terms: $\rho_\Lambda = \rho_{\Lambda, \rm class} + \rho_{\Lambda, \rm quant}$.
The term induced by quantum fluctuations $\rho_{\Lambda, \rm quant}$ cannot be calculated in a self consistent way in quantum field theory but it is expected to be huge from dimensional analysis.
The classical term $\rho_{\Lambda, \rm class}$ is the energy density of the scalar fields in the ground state, it is a free parameter of the theory.
At the end, the vacuum energy can perfectly be considered as a free parameter in the present incomplete theories of particle physics and gravity.
However, there is a naturalness problem very similar to the hierarchy problem of particle physics
in the sense that ``bare'' value has to be extraordinarily fine-tuned to almost compensate for the huge quantum contribution.
Well before the discovery of the accelerating Universe, \textcite{Weinberg1987} discussed an anthropic explanation to the cosmological constant problem.
He gave an upper bound on the vacuum energy density:
\begin{equation}
\label{Weinberg}
\rho_\Lambda \lesssim (10 \, {\rm meV})^4.
\end{equation}
The argument goes as follows.
If the vacuum energy is too large, the accelerated phase of the expansion starts too early.
Then the inhomogeneities of the mass density do not have time to grow sufficiently before getting stretched away by the accelerated expansion.
Only if the Weinberg bound \eqref{Weinberg} is satisfied can the perturbations grow to form stars and galaxies.
Otherwise the Universe would consist in an homogeneous fluid in accelerated expansion, quite unfriendly for life.
Imagining that many universes exist, with a different value for the cosmological constant in each,
it is not surprising that we happen to live in a seemingly very special universe where life is possible at all.
Indeed the observed vacuum energy density \eqref{LAMBDA} is not particularly fine-tuned with respect
to the range of possible values in the anthropic sense given by \eqref{Weinberg}.
This anthropic explanation is further supported by the landscape of string theory:
there is a huge number of possible vacua of the string theory associated with the many many different ways to compactify the extra dimensions.
One might believe that all these versions of string theory ``exist'' somehow,
and intelligent life can emerge only in a tiny fraction of a vast number of possible versions.
The fine-tuning of the parameters could then be just an illusion.
In addition, the eternal inflation scenario supports the plausibility that many worlds can ``exist'' in a multiverse.
In these theories, the Universe is inflating due to the energy density of a scalar field -- the inflaton -- which is in a false vacuum.
The inflaton decays in the true vacuum, stopping the inflationary phase.
But this process happens only locally, forming a bubble of true vacuum surrounded by the rest of the Universe which is still inflating.
In some versions of the theory, the part of the Universe which is still inflating grows fast enough to accommodate for the continuous formation of bubbles
(the speed of the bubble walls could be slower than the exponential expansion of the false vacuum space).
In this case the inflationary phase of the Universe lasts forever in some regions of the Universe and causally disconnected bubbles are produced on and on.
Different bubbles having different cosmological constants is not strictly speaking a prediction of eternal inflation,
but if inflation is due to Planck-scale physics or stringy effects, who knows?
We might be living in one out of a vast number of bubbles, where the vacuum energy density is such that it does not forbid the appearance of intelligent life.
\subsection{The quintessence as Dark Energy}
The anthropic explanation of the cosmological constant problem is both elegant and fascinating.
However it rests on speculative ideas, to say the least.
This explanation is a source of great inspiration for imagination and philosophy,
unfortunately it gives poor insight for further observations and experiments.
It is thus necessary to explore other possibilities.
We can take the view that Dark Energy is a dynamical process revealing new physics at the energy scale $\rho_\Lambda^{1/4} \approx 1$~meV.
A review of the possible dynamics of Dark Energy and its observational consequences can be found in \textcite{Copeland2006}.
Although none of these models really solve the cosmological constant problem in a top-down way,
they are useful bottom-up approaches to guide future observations and experiments to refine our knowledge of Dark Energy.
The most popular routes to a dynamical extension of the simple $\Lambda$CDM parametrization are quintessence models.
Akin to inflationary models, the accelerated expansion is due to the energy density of a scalar field.
The scalar field responsible for the acceleration of the Universe during the inflation era is called the inflaton,
whereas the hypothetical scalar field making up the Dark Energy driving the late acceleration is called the quintessence.
There are even models explaining both the early inflation and the late acceleration of the Universe using a single scalar field \cite{Peebles1999}.
Let us review the theory of the gravitational dynamics of a scalar field.
The action for gravity and the quintessence field $\varphi$ with the so-called \emph{minimal coupling} is
\begin{equation}
S = \int d^4 x \sqrt{-g} \left( - \frac{m_{\rm Pl}^2}{2} \, \mathcal{R} + \mathcal{L}_\varphi \right),
\end{equation}
where $g = \det(g_{\mu \nu})$ is the determinant of the metric tensor, $\mathcal{R} = g^{\mu \nu} R_{\mu \nu}$ is the trace of the Ricci curvature tensor $R_{\mu \nu}$ and
\begin{equation}
\mathcal{L}_\varphi = \frac{1}{2} g^{\mu \nu} \partial_\mu \varphi \partial_\nu \varphi - \mathcal{V}(\varphi)
\end{equation}
is the Lagrangian of the scalar field with a generic potential $\mathcal{V}(\varphi)$.
Einstein's equations governing the response of the metric to the scalar field
are obtained from the variational principle $\delta S / \delta g^{\mu \nu} = 0$:
\begin{equation}
G_{\mu \nu} = 8 \pi G \, T_{\mu \nu}^\varphi,
\end{equation}
where $G_{\mu \nu} = R_{\mu \nu} -1/2 g_{\mu \nu} \mathcal{R}$ is Einstein's tensor and
\begin{eqnarray}
T_{\mu \nu}^\varphi & = & 2 \frac{\delta \mathcal{L}_\varphi}{\delta g^{\mu \nu}} - g_{\mu \nu} \mathcal{L}_\varphi \\
& = & \partial_\mu \varphi \partial_\nu \varphi - \frac{1}{2} g_{\mu \nu} g^{\alpha \beta} \partial_\alpha \varphi \partial_\beta \varphi + g_{\mu \nu} \mathcal{V}(\varphi)
\end{eqnarray}
is the stress-energy tensor of the scalar field.
In the homogeneous Universe, the scalar field is uniform in a Robertson-Walker metric \eqref{RobertsonWalker}.
The problem reduces to only two degrees of freedom $a(t)$ and $\varphi(t)$ and
tensors take a diagonal form:
\begin{eqnarray}
g_{\mu \nu} & = & {\rm diag} \left( 1, - a^2 \right) \\
g^{\mu \nu} & = & {\rm diag} \left( 1, - 1/a^2 \right) \\
G_{\mu \nu} & = & {\rm diag} \left( 3H^2, a^2 [-3H^2 - 2\dot{H} ] \right) \\
T_{\mu \nu}^\varphi & = & {\rm diag} \left( \frac{1}{2} \dot{\varphi}^2 + \mathcal{V}(\varphi) , a^2 [ \frac{1}{2} \dot{\varphi}^2 - \mathcal{V}(\varphi) ] \right)
\end{eqnarray}
The $00$ component of Einstein's equations reduces then to the Friedmann equation
\begin{equation}
\label{FriedmannPhi}
H^2 = \frac{8 \pi G}{3} \left( \frac{1}{2} \dot{\varphi}^2 + \mathcal{V}(\varphi) \right).
\end{equation}
In addition, the evolution of the scalar field is given by the variational principle $\delta S / \delta \varphi = 0$, which in the homogeneous Universe reduces to
\begin{equation}
\label{KleinGordonPhi}
\ddot{\varphi} + 3 H \dot{\varphi} + \mathcal{V}'(\varphi) = 0.
\end{equation}
The coupled dynamics of the expansion (i.e. the time evolution of $H = \dot{a}/a$ and $\varphi$)
is completely specified by the two equations \eqref{FriedmannPhi} and \eqref{KleinGordonPhi}.
It is useful to notice that these two equations are equivalent to the two Friedman equations \eqref{F1},\eqref{F2} with
$\rho = \dot{\varphi}^2/2 + \mathcal{V}(\varphi)$ the energy density of the quintessence and $p = \dot{\varphi}^2/2 - \mathcal{V}(\varphi)$ its pressure.
The equation of state parameter
\begin{equation}
w = \frac{p}{\rho} = \frac{\dot{\varphi}^2/2 - \mathcal{V}(\varphi)}{\dot{\varphi}^2/2 + \mathcal{V}(\varphi)}
\end{equation}
is a dynamical quantity that can evolve with time.
Different cosmologies can be built depending on the potential $\mathcal{V}(\varphi)$.
If $\varphi$ corresponds to a standard massive scalar field, such as the Axion, the potential $\mathcal{V}(\varphi) = \frac{1}{2} \mu^2 \varphi^2$ is parabolic.
For the harmonic potential there is equipartition between the kinetic energy $\dot{\varphi}^2/2$ and the potential energy $\mathcal{V}(\varphi)$ and therefore $w=0$.
In this case $\varphi$ is a candidate for pressurless cold Dark Matter.
It is also possible to build models which mimic a cosmological constant with $w \approx -1$, where the field does not develop significant kinetic energy.
Such a model has been first proposed by \textcite{Ratra1988}, with the potential given by
\begin{equation}
\label{RatraPeebles}
\mathcal{V}(\varphi) = M_\Lambda^4 \left( \frac{M_\Lambda}{\varphi} \right)^n
\end{equation}
where $M_\Lambda$ is a new mass scale and $n$ is a positive exponent called the Ratra-Peebles index.
As a generic feature of quintessence models, a departure from $w=-1$ is expected.
Therefore there is a chance that future supernovae surveys could detect a deviation from the $\Lambda$CDM model where $w=-1$.
Now, is there another way to reveal the presence of the quintessence field $\varphi$?
If the field were coupled to ordinary matter, then it would mediate a fifth force.
It is then very appealing to design laboratory experiments searching for this new force.
\subsection{The chameleon}
A quintessence field coupled with matter mediates a new force with a range extending up to cosmological scales.
At first it was generally assumed that the coupling must be very small, much smaller than the gravitational strength,
in order to satisfy the stringent tests of the equivalence principle, as well as the tests of general relativity in the Solar system.
However, \textcite{KhouryPRL2004,KhouryPRD2004} discovered that the coupled quintessence scalar theory
automatically features a very efficient mechanism to suppress the force called the chameleon mechanism.
Then, \textcite{Brax2004} explored the cosmological consequences of the chameleon theory to find that it is a nice viable Dark Energy candidate.
Let us explain the chameleon screening mechanism.
We start by considering the Lagrange density of the scalar field $\varphi$ coupled with a fermion $\psi$ (with a mass $m$) in a Minkowski spacetime:
\footnote{The term $\beta m/m_{\rm Pl} \varphi \ \bar{\psi} \psi$
originates from a \emph{conformal coupling} of the scalar field to the fermion.
The conformal coupling consists in replacing the metric $g_{\mu \nu}$ in the fermion Lagrange density by a function of the scalar field $A^2(\varphi) g_{\mu \nu}$ : $\mathcal{L}_m (A^2(\varphi)g_{\mu \nu}, \psi)$.
A popular choice of the conformal coupling function is $A(\varphi) = e^{\beta \varphi / m_{\rm Pl}}$.
Since the excursions of the field always satisfies $\beta \varphi / m_{\rm Pl} \ll 1$ in concrete cosmological or laboratory situations, we can approximate $A(\varphi) = 1 + \beta \varphi / m_{\rm Pl}$.
It finally produces the Yukawa coupling term $\beta m/m_{\rm Pl} \varphi \ \bar{\psi} \psi$.
}
\begin{equation}
\label{lagrangien}
\mathcal{L} = \frac{1}{2} \partial_\mu \varphi \ \partial^\mu \varphi - \mathcal{V}(\varphi) - \beta \frac{m}{m_{\rm Pl}} \ \varphi \ \bar{\psi} \psi.
\end{equation}
The dimensionless constant $\beta$ corresponds to the strength of the coupling relative to gravity.
From the Lagrange density we can deduce in principle how $\varphi$ affects a particle, and conversely, how the presence of particles generates a field.
\begin{enumerate}
\item A fermion evolving in a field $\varphi({\bf r})$ is affected by the potential
\begin{equation}
V({\bf r}) = \beta \frac{m}{m_{\rm Pl}} \ \varphi({\bf r}),
\end{equation}
corresponding to a force:
\begin{equation}
{\bf F} = - \beta \frac{m}{m_{\rm Pl}} \ \boldsymbol{\nabla} \varphi({\bf r}).
\end{equation}
\item The field $\varphi({\bf r})$ generated by a distribution of fermions is given by the Euler-Lagrange equation
\begin{equation}
\partial_\mu \left( \frac{\partial \mathcal{L}}{\partial(\partial_\mu \varphi)} \right) - \frac{\partial \mathcal{L}}{\partial \varphi} = 0.
\end{equation}
\end{enumerate}
More specifically, the Euler-Lagrange equation deduced from the Lagrange density \eqref{lagrangien} reads
\begin{equation}
\label{KG1}
\partial_\mu \partial^\mu \varphi + \mathcal{V}'(\varphi) + \beta \frac{ \rho }{m_{\rm Pl}} = 0,
\end{equation}
where $\rho$ is the mass density of the fermions (we have assumed the non-relativistic limit $m \bar{\psi} \psi = \rho$).
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{chameleonPotential}
\caption{
The black dashed line represents the runaway potential $\mathcal{V}(\varphi)$.
The solid lines represent the effective potential $\mathcal{V}_{\rm eff}(\varphi)$
where the environment density $\rho$ is low (blue) and where the density is high (red).
\label{chameleonPotential}
}
\end{figure}
It is then apparent that, within an environment of density $\rho$,
the dynamics of the chameleon is not governed by $\mathcal{V}(\varphi)$,
but rather by the effective potential
\begin{equation}
\mathcal{V}_{\rm eff} = \mathcal{V}(\varphi) + \beta \frac{\rho}{m_{\rm pl}} \ \varphi,
\end{equation}
in the sense that Eq. \eqref{KG1} can be written in the form $\partial_\mu \partial^\mu \varphi + \mathcal{V}_{\rm eff}'(\varphi) = 0$.
Let us recall that the equation governing a basic massive scalar field (with a potential $\frac{1}{2} \mu^2 \varphi^2$) is the Klein-Gordon equation
$\partial_\mu \partial^\mu \varphi + \mu^2 \varphi = 0$, where $\mu$ is the mass of the elementary excitations of the field.
In this well-known case, the field generated by a static point source (if the field is coupled to fermions, this corresponds to a source term in the right hand side of the Klein-Gordon equation)
takes the form of a Yukawa potential $\varphi(r) \simeq e^{-\mu r}/r$.
It means physically that the field mediates an interaction between fermions with a range $1/\mu$ (in S.I. units, $\lambda = \frac{\hbar c}{\mu c^2}$).
A very massive field will mediate a short-range force, whereas a light field can travel long distances.
We now come back to the more complicated situation of the coupled quintessence field.
We will discuss the consequences of the specific features of the effective potential depicted in Fig. \ref{chameleonPotential}.
Assuming that the density $\rho$ is uniform everywhere, then there is a static and uniform solution to Eq. \eqref{KG1},
given by the minimum of the effective potential $\mathcal{V}_{\rm eff}'(\varphi_{\rm min}) = 0$.
In the case of the Ratra-Peebles potential \eqref{RatraPeebles} the minimum of the effective potential reads
\begin{equation}
\label{phiMin}
\varphi_{\rm min} = M_\Lambda \left( \frac{n M_\Lambda^3 m_{\rm pl}}{\beta \rho} \right)^{\frac{1}{n+1}}.
\end{equation}
For small perturbations of the field around this uniform value
(induced by an extra mass on top of the environment density $\rho$ for example),
we can approximate the effective potential around the minimum by
\begin{equation}
\mathcal{V}_{\rm eff}(\varphi) \approx \mathcal{V}_{\rm eff}(\varphi_{\rm min}) + \frac{1}{2} (\varphi - \varphi_{\rm min})^2 \ \mathcal{V}_{\rm eff}''(\varphi_{\rm min}).
\end{equation}
We can attribute an effective mass $\mu$ for the perturbations of the field, with
\begin{equation}
\label{effectiveMass}
\mu^2 = \mathcal{V}_{\rm eff}''(\varphi_{\rm min}) = \beta \frac{\rho}{m_{\rm pl}} \frac{n+1}{\varphi_{\rm min}}.
\end{equation}
The effective mass is associated with the curvature of the effective potential at the minimum.
Clearly, what is happening here is that the mass of the field is density-dependent,
in such a way that the range of the force mediated by the field shrinks in a high density environment.
For this reason, the scalar field is called ``chameleon'':
it adapts its properties to the environment to evade observation.
Consider for example the Sun as a source of the chameleon field.
Since the density in the interior of the Sun is quite high, the chameleon is effectively massive in the Sun.
Therefore, a test particle situated outside the Sun is not affected by the inner part of the Sun because the field
created by the mass inside the Sun cannot propagate to the outside.
In fact, it is shown in \textcite{KhouryPRL2004,KhouryPRD2004} that when the coupling $\beta$ is large enough,
then only a thin shell at the surface of a large body such as the Sun or the Earth contributes to the force on a test particle outside the body.
Then, in a more elaborate analysis, \textcite{Mota2007} concluded that strongly coupled chameleons,
i.e. $\beta \gg 1$, are not ruled out by terrestrial and solar system tests of gravity.
\subsection{1D solutions of the chameleon equation}
In order to gain more insight about the properties of the chameleon field,
we shall now solve explicitly the chameleon equation in the case of the simple one-dimensional problem
illustrated on Fig. \ref{chameleon1D}.
We consider a plate of material (to be specific, a plate of aluminium with thickness $2 d = 1$~mm) surrounded by perfect vacuum.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{chameleon1D}
\caption{
Sketch of the one-dimensional problem.
\label{chameleon1D}
}
\end{figure}
We consider a static solution in one dimension, therefore
$\partial_\mu \partial^\mu \varphi = \frac{\partial^2 \varphi}{\partial t^2} - \Delta \varphi = - \frac{d^2 \varphi(x)}{dx^2}$.
In this case the chameleon equation \eqref{KG1} becomes
\begin{equation}
\frac{d^2 \varphi}{dx^2} = \mathcal{V}_{\rm eff}'(\varphi)
\end{equation}
In addition, the symmetry of the problem implies $\varphi(-x) = \varphi(x)$.
We will therefore solve for $x>0$ with the boundary condition $\varphi'_0 = 0$.
For simplicity we treat the case with $n=2$ but other Ratra-Peebles indices can be treated similarly.
\para{Solution in vacuum. }
In the vacuum ($X>0$), the chameleon equation reads
\begin{equation}
\frac{d^2 \varphi}{dX^2} = - 2 \ \frac{M_\Lambda^6}{\varphi^3}.
\end{equation}
we can find a family of solutions with different values $\varphi_S$ at the boundary:
\begin{equation}
\label{vacSol1}
\varphi(X) = \varphi_S \sqrt{ 1 + X/X_S }
\end{equation}
with $X_S$ satisfying
\begin{equation}
\label{vacSol2}
X_S = \frac{\sqrt{2}}{4} \frac{\varphi_S^2}{M_\Lambda^3}.
\end{equation}
Notice the remarkable expression of the derivative at the boundary
\begin{equation}
\varphi_S' = \sqrt{2} \ \frac{M_\Lambda^3}{\varphi_S}.
\end{equation}
Next, we deal with the inside of the plate ($0<x<d$).
No exact solution exists for the problem,
we need to consider two asymptotic regimes.
\para{Solution inside the plate: linear regime. }
We start by considering what we call the linear regime,
where $\varphi_0 \gg \varphi_{\rm min}$.
We will derive {\rm a posteriori} the corresponding condition for the parameter $\beta$.
We recall that $\varphi_{\rm min}$ is the value of the field
minimizing the effective potential inside the plate, it is given by Eq. \eqref{phiMin}.
In this regime, the effective potential is dominated by the coupling term and we can neglect the self-interaction term $\mathcal{V}(\varphi)$.
Therefore, the chameleon equation reduces to a Poisson equation :
\begin{equation}
\frac{d^2 \varphi}{dx^2} \approx \beta \frac{\rho}{m_{\rm pl}} = \frac{2 M_\Lambda^6}{\varphi_{\rm min}^3}.
\end{equation}
with the solution
\begin{equation}
\varphi(x) = \varphi_0 + \frac{M_\Lambda^6}{\varphi_{\rm min}^3} x^2
\end{equation}
We now need to connect this solution to the vacuum solution at the boundary by equalizing the value of the field and the derivative.
We get the solution
\begin{equation}
\frac{\varphi_0}{M_\Lambda} = \frac{1}{\sqrt{2} (M_\Lambda d)}\left( \frac{\varphi_{\rm min}}{M_\Lambda} \right)^3 - (M_\Lambda d)^2 \left( \frac{M_\Lambda}{\varphi_{\rm min}} \right)^3.
\end{equation}
Now we can work out {\it a posteriori} the domain of validity of the linear approximation.
We see from the solution that the condition $\varphi_0 \gg \varphi_{\rm min}$ is equivalent to the condition
$M_\Lambda d \ll (\varphi_{\rm min}/M_\Lambda)^2$, or, expressed in terms of the basic parameters:
\begin{equation}
\label{linearRegime}
{\rm linear \ regime: } \quad \quad \beta \ll \frac{m_{\rm pl} M_\Lambda^3}{\rho} \ \frac{1}{(M_\Lambda d)^{3/2}}.
\end{equation}
Let us evaluate numerically this condition in the case of our aluminium plate.
Equations are expressed in natural units, so we will express all masses in eV and all distances in eV$^{-1}$.
We will consider the energy scale of the chameleon potential to correspond to the Dark Energy scale:
\begin{equation}
M_\Lambda = \rho_\Lambda^{1/4} = 2.4 \times 10^{-3} \ {\rm eV}.
\end{equation}
The Planck mass amounts to $m_{\rm pl} = 2.4 \times 10^{27}$~eV.
The half-thickness of the plate is $d = 0.5 \ {\rm mm} = 2.5 \times 10^{3} \ {\rm eV}^{-1}$
and the density of aluminium is $\rho = 2.7 \ {\rm g/cm}^3 = 1.2 \times 10^{19} \ {\rm eV}^4$.
We find that the solution is described by the linear regime when
\begin{equation}
\beta \ll 0.2.
\end{equation}
The linear regime does not describe strongly coupled chameleons (i.e. $\beta \gg 1 $) in the presence of ordinary objects of normal size!
\para{Solution inside the plate: saturated regime. }
In the opposite regime, which we call the saturated regime,
the field inside the plate is very close to $\varphi_{\rm min}$.
We recall that $\varphi_{\rm min}$ would be the uniform value of the field
in the situation where the plate with density $\rho$ fills the entire space uniformly.
In the saturated regime we can write $\varphi(x) = \varphi_{\rm min} + \phi$,
with $\phi \ll \varphi_{\rm min}$.
In this case, the chameleon equation reduces to the Klein-Gordon equation
\begin{equation}
\frac{d^2 \phi}{dx^2} = \mu^2 \phi
\end{equation}
where the effective mass $\mu$ satisfies Eq. \eqref{effectiveMass}, that is:
\begin{equation}
\mu^2 = \beta \frac{\rho}{m_{\rm pl}} \frac{3}{\varphi_{\rm min}} = \frac{6 M_\Lambda^6}{\varphi_{\rm min}^4}.
\end{equation}
At this point it is useful to notice that the opposite of
the linear regime condition Eq. \eqref{linearRegime} can be written in the form:
\begin{equation}
\label{saturatedRegime}
{\rm saturated \ regime: } \quad \quad \mu d \gg 1.
\end{equation}
Physically, it means that in the saturated regime the range of the force mediated by the chameleon is much shorter than the thickness of the plate.
The solution of the Klein-Gordon equation is:
\begin{equation}
\phi(x) = \phi_0 \cosh (\mu x).
\end{equation}
The continuity of $\varphi$ and $\varphi'$ at the boundary provides a quadratic equation for $\phi_0$:
\begin{equation}
\phi_0 \mu \sinh(\mu d) \left( \varphi_{\rm min} + \phi_0 \cosh(\mu d) \right) = \sqrt{2} M_\Lambda^3,
\end{equation}
or equivalently, we get a quadratic equation for $\varphi_S = \varphi_{\rm min} + \phi_0 \cosh \mu d$:
\begin{equation}
(\varphi_S - \varphi_{\rm min} ) \varphi_S = \sqrt{2} \frac{M_\Lambda^3}{\mu} \ \frac{\cosh \mu d}{\sinh \mu d}.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{solution1D}
\caption{
Solution of the 1D chameleon equation with $n=2$.
The half-thickness of the aluminium plate is $d = 0.5$~mm.
The blue curve ($\beta = 0.05$) is calculated using the linear approximation.
The black ($\beta=1$) and red ($\beta=20$) curves are calculated using the saturated approximation.
\label{solution1D}
}
\end{figure}
\para{Interpretation of the solution. }
The solution of the chameleon equation on the right side of the plate
is plotted on Fig. \ref{solution1D}, for three different values of the coupling constant $\beta$.
The field inside the plate is ``attracted'' to small values.
Outside the plate, the field wants to grow to minimize the potential $\mathcal{V}(\varphi)$.
For strongly coupled chameleons, the field inside the plate takes a nearly uniform value $\varphi = \varphi_{\rm min}$.
When the field is saturated inside the bulk, the plate acts as a screen:
if a source mass moves on the left side of the plate, the field will be modified on the left side but it will remain saturated to $\varphi_{\rm min}$ inside the plate.
Therefore the solution on the right side of the plate will remain unaffected.
The plate shields the right side from what happens on the left side.
This screening mechanism makes the chameleon field difficult to detect by laboratory experiments.
Another important feature of the chameleon is the saturation of the field.
We see in Fig. \ref{solution1D} that the solution outside the plate becomes independent of $\beta$ for large values of $\beta$.
To better understand this, let us discuss the gradient of the field $\frac{d \varphi}{dx}$ at the immediate vicinity of the plate, say $X = 10 \ \mu$m.
The gradient of the field is analogous to an electric field (which is the gradient of the electric potential).
The force acting on a test mass is proportional to $\frac{d \varphi}{dx}$ in the same way as the force acting on a charged particle is proportional to the electric field.
There is no saturation effect for the electric field, in the sense that an electrically charged plate produces at the vicinity of the surface an electric field proportional to the charge density of the plate.
Fig. \ref{solution1Dgrad} shows the gradient of the chameleon field as a function of $\beta$.
In the linear regime, the gradient grows linearly with $\beta$ (or $\rho$) as for the electric field (hence the name of the linear regime).
On the contrary, in the saturated regime, the gradient saturates to a finite asymptotic value.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{solution1Dgrad}
\caption{
Gradient of the chameleon field at $X=10 \ \upmu {\rm m}$ from the surface of the aluminium plate, as a function of $\beta$.
The blue curve is calculated using the linear approximation, the red curve is calculated using the saturated approximation.
\label{solution1Dgrad}
}
\end{figure}
As a matter of fact, when searching for strongly coupled chameleons,
one should think of solid matter as a zero boundary condition for the field $\varphi$.
In this case the field in the vicinity of the surface of any plate has been derived by \textcite{Brax2011} for any Ratra-Peebles index $n$.
In S.I. units it reads:
\begin{equation}
\label{OnePlate}
\varphi(X) = M_\Lambda \left( \frac{2+n}{\sqrt{2}} \ \frac{M_\Lambda}{\hbar c} \ X \right)^{2/(2+n)}.
\end{equation}
In the particular case $n=2$ we recover Eq. \eqref{vacSol1} in the limit $\varphi_S = 0$.
This solution is independent of the coupling strength $\beta$, because of the saturation property we just discussed.
The problem of the ``capacitor'', i.e. two parallel plates separated by a thickness $2 R$ of vacuum,
has been considered by \textcite{Ivanov2013}.
They proposed the approximate analytical solution
\begin{equation}
\label{TwoPlates}
\varphi(x) = M_\Lambda \left( \frac{2+n}{2 \sqrt{2}} \ \frac{R M_\Lambda}{\hbar c} \ \left[ 1 - \frac{x^2}{R^2} \right] \right)^{2/(2+n)}
\end{equation}
which is exact for $n=2$.
In this equation $x=0$ refers to the middle position of the capacitor.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{solution1Dbubble}
\caption{
Solution of the chameleon equation for $n=1$ (black curves) and $n=2$ (red curves).
Solid lines correspond to the two-plates solution Eq. \eqref{TwoPlates} with $2 R = 1$~cm.
Dashed lines correspond to the one-plate solution Eq. \eqref{OnePlate}.
\label{solution1Dbubble}
}
\end{figure}
The one-plate and two-plate chameleon profiles are displayed in Fig. \ref{solution1Dbubble}.
What is the physical reality of those chameleon profiles?
Consider a test particle of mass $m$ in the vicinity of a plate, or inside a two-plate capacitor,
the chameleon profile generates a potential energy $\beta m / m_{\rm pl} \varphi(x)$.
We will discuss later suitable neutron experiments to probe these chameleon fields.
It is important to note that a macroscopic test mass will not be subjected to the same potential energy.
For strong couplings $\beta$, the core of the test mass will be shielded from the external chameleon field.
It is sometimes useful to think of experiments searching for a new force as made of two components: a source (sometimes called an attractor) and a probe (sometimes called a detector).
In the case of strongly coupled chameleons, a macroscopic source such as a plate produces a saturated chameleon field, independent of $\beta$.
Also, the response of a macroscopic probe will show some saturation effect,
whereas the response of a particle such as a neutron will be proportional to the coupling strength $\beta$.
\subsection{Searching for the chameleon in the lab}
\para{Probing new forces with neutrons. }
Since the existence of new interactions are generic predictions of theories beyond the standard model,
they are actively searched for in a great variety of experiments probing scales from subatomic to astronomical distances (see \textcite{Antoniadis2011} for a recent review).
It is however quite recently that slow neutrons were recognized as interesting probes of new forces.
In particular, neutron scattering experiments were found to be sensitive to new Yukawa forces of nanometric range \cite{Nesvizhevsky2008,Pokotilovski2006}.
At the micrometer scale, the measurement of quantum states of neutrons bouncing over a mirror provides some constraints,
however they are not as stringent as those derived from fifth-force searches using macroscopic bodies.
In addition, neutrons can be sensitive to spin-dependent forces in the sub-millimeter range (induced by new bosons of mass $\mu > 10^{-3} \ {\rm eV}$).
Limits on CP violating spin-dependent forces induced by the exchange of light Axionlike particles were first set by bouncing neutrons
\cite{Baessler2007,Baessler2009,Jenke2014}.
These limits were quickly superseded by experiments measuring the spin-precession of ultracold neutrons
\cite{Serebrov2010,Afach2015}.
Now the neutron limits compete with experiments using hyper-polarized $^3$He \cite{Petukhov2010}.
At shorter range, a neutron diffraction experiment provides an interesting constraint on Axionlike particles \cite{Voronin2009}.
There are also neutron beam experiments probing new spin-dependent interactions mediated by spin 1 bosons
\cite{Piegsa2012,Yan2013}.
\para{Neutrons probing strongly coupled chameleons. }
Due to the peculiarity of the force induced by the chameleon field,
the constraints on generic short range Yukawa force do not generally apply.
In this case, it is quite attractive to consider the neutron as a probe of strongly coupled chameleons
because of the absence of the saturation effect.
The case of the neutron bouncer is particularly interesting.
Ultracold neutrons can bounce above an horizontal plate because
they are specularly reflected by the Fermi potential when they fall on the plate (which we call a mirror).
It is possible to prepare neutrons almost at rest (concerning the vertical motion), so that the bouncing height is only a fraction of a millimeter.
In this case the energy of the vertical motion is quantized,
as demonstrated experimentally for the first time at the Institut Laue Langevin by \textcite{Nesvizhevsky2002}.
We \cite{Brax2011} argued that the chameleon field Eq. \eqref{OnePlate} would be detectable in this system and deduced the upper limit $\beta < 10^{11}$.
Next, \textcite{Jenke2011} realized the first spectroscopy of the quantum states by inducing resonant transitions.
The chameleon field would shift the frequency of the resonances and the upper limit $\beta < 6 \times 10^8$ was deduced \cite{Jenke2014}.
In part \ref{sectionGRANIT} we will dwell on this topic and give the prospects of the GRANIT experiment.
Besides quantum states of bouncing neutrons, there is a second route: neutron interferometry.
\textcite{Pokotilovski2013} proposed to build a Llyod's mirror interferometer for very cold neutrons to probe the chameleon.
Such a method could in principle be sensitive to chameleon couplings down to $\beta \sim 10^{7}$ but the cold neutron interferometer
technique has yet to be developed.
As an alternative, we \cite{BraxPignol2013} proposed to use triple Laue-case interferometers with slow
neutrons that have been operated routinely for decades in several neutron facilities.
A first experiment has been performed in summer 2013 at the Institut Laue Langevin \cite{Lemmel2015}.
We will describe it in part \ref{sectionInterferometry}.
The constraints in the chameleon parameter space (Ratra-Peebles index $n$ and matter coupling $\beta$) from these neutron experiments
are shown in Fig. \ref{exclusion} together with constraints from non-neutron experiments.
\para{Other laboratory searches for chameleons. }
Aside from neutron experiments, there are of course a few other means to probe the chameleon.
\begin{itemize}
\item Using torsion pendulum with exquisite sensitivity
(such as \textcite{Kapner2007}) one searches for an anomalous torque between a rotating attractor disk and a torsion pendulum separated by a fraction of a millimeter.
The detector and the attractor are both disks with a diameter of a few centimeters, they are then subjected to the chameleon screening mechanisms in the case $\beta \gg 1$.
\textcite{Upadhye2012} undertook the very complicated task of deriving the limits on the chameleon couplings from this experiment.
For $n=1$ the excluded range is $10^{-2} < \beta < 10$.
Couplings larger than 10 are allowed!
\item Measurements of the Casimir force could also be sensitive to chameleons \cite{Brax2007}.
Existing experiments are sensitive only in the region of the parameter space where $M_\Lambda > 0.1~$eV.
Future experiments could be sensitive to the interesting case $M_\Lambda = 2.4 \times 10^{-3}$~eV for a wide range of coupling $\beta$.
\item Atomic precision tests in the hydrogen atom can reveal the presence of a scalar field coupled to protons and electrons.
\textcite{BraxBurrage2011} derived the limit $\beta < 10^{14}$ from existing data.
\item An atom-interferometry experiment using ultracold cesium atoms \cite{Hamilton2015} has been reported (while this manuscript was nearly finished).
They obtained the limit $\beta < 2 \times 10^4$ for $n=1$.
\item In addition to the coupling to matter (the term $\beta m / m_{\rm Pl} \varphi \bar{\psi} \psi$), the chameleon field can be coupled to photons with the Lagrange density term
\begin{equation}
\frac{1}{4} \frac{\beta_\gamma}{m_{\rm Pl}} \ \varphi \ F_{\mu \nu} F^{\mu \nu}.
\end{equation}
This case leads to a unique experimental signature.
Due to the $\gamma \gamma \varphi$ vertex, chameleons can be produced by shining an intense laser in a region permeated by an intense magnetic field (take a spare dipole magnet from a giant particle accelerator).
In this context, a chameleon is an elementary quantum excitation of the field $\varphi$.
The production region is a vacuum chamber, the produced chameleon has a very low mass.
The walls of the vacuum chamber act as a repulsive potential for the chameleon, because the mass of the chameleon is large inside the walls.
Therefore the chameleons can be stored in the chamber (provided that the coupling to matter is strong enough, in practice it requires $\beta > 10^4$).
After switching off the laser, the trapped chameleons can be converted back to photons through the reverse of the process that formed them, one would then measure an \emph{afterglow}.
A dedicated experiment called CHASE (Chameleon Afterglow Search) was performed at Fermilab \textcite{Steffen2010}.
No afterglow was observed, the experiment sets the limit $\beta_\gamma < 10^{11}$.
\end{itemize}
\para{Modified gravity. }
We conclude this part by mentioning that the chameleon theory is one
amongst several known (theoretically) screening mechanisms (see \textcite{Khoury2013} for a pedagogical overview).
In the general framework of modified gravity, the Einstein-Hilbert action of general relativity is modified by
adding a scalar field $\varphi$.
There are many possible choices for the potential $\mathcal{V}(\varphi)$, for the coupling to the standard model $A(\varphi)$, for non-standard kinetic terms, etc.
To understand the different possibilities we follow \textcite{Brax2013} and write the Lagrange density linearised around a background configuration $\varphi_0$:
\begin{equation}
\mathcal{L} = \frac{Z(\varphi_0)}{2} \partial_\mu \varphi \partial^\mu \varphi - \frac{1}{2} \mu^2(\varphi_0) \varphi^2 + \frac{\beta(\varphi_0)}{m_{\rm Pl}} \varphi \rho.
\end{equation}
In general the background configuration $\varphi_0$ depends on the environment density $\rho$,
a property which could bring about screening.
We can see that there are three general classes of screening mechanisms.
\begin{itemize}
\item The mass $\mu(\varphi_0)$ becomes large in a dense environment. This is the chameleon mechanism.
\item $Z(\varphi_0)$ becomes large in a dense environment.
A special case of this is the \emph{Galileon} field.
\item The coupling $\beta(\varphi_0)$ becomes small in a dense environment.
The Damour-Polyakov mechanism belongs to this class.
The \emph{symmetron} is another example.
\end{itemize}
\bigskip \begin{center} \photon \end{center} \bigskip
The comprehensive study of these theories, their cosmological implications, and the design of laboratory experiments
is a vivid ongoing field of research (see Fig \ref{spiresCount}).
The hunt is on for the non-gravitational interaction of dark energy.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{spiresCount}
\caption{
Number of publications per year containing in the title the keywords
``dark energy'' (black), ``cosmological constant'' (green), ``quintessence'' (red), ``chameleon'' (blue), ``galileon'' (pink).
The counting was done using inspirehep.net.
\label{spiresCount}
}
\end{figure}
\clearpage
\section{Neutron interferometry constrains the chameleon}
\label{sectionInterferometry}
This part describes a search for the chameleon using neutron interferometry.
The experiment has been reported also in \textcite{Lemmel2015}.
\subsection{Neutron interferometry}
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{interferometerSketch}
\caption{
General principle of neutron interferometry.
\label{interferometerSketch}
}
\end{figure}
The first neutron interferometer, analogous to a Mach-Zehnder interferometer for light, has been realized by \textcite{Rauch1974}.
The principle of neutron interferometry is depicted in Fig. \ref{interferometerSketch}.
A monochromatic neutron beam with a wavelength of $\lambda = 0.27$~nm (or wavenumber of $k = 2 \pi / \lambda = 23$~nm$^{-1}$)
is split by Bragg diffraction into two coherent beams using a single-crystal silicon plate.
Then part of these two beams are recombined using two additional similar parallel plates.
The neutron detectors measure the flux resulting from the interference of neutrons going through path I and path II.
A phase flag (usually an aluminium plate with variable angle)
is introduced in path II to record the interferogram,
i.e. the neutron flux in detectors H and O as a function of the phase flag position:
\begin{equation}
F_O = A_O + B_O \cos (\xi_{\rm flag}) ; \ F_H = A_H - B_H \cos (\xi_{\rm flag}),
\end{equation}
where $\xi_{\rm flag}$ is a function of the phase flag position.
Then, a sample is introduced in path I and a new interferogram is recorded:
\begin{equation}
F_O = A_O + B_O \cos (\xi_{\rm flag} + \xi) ; \ F_H = A_H - B_H \cos (\xi_{\rm flag} + \xi).
\end{equation}
The phase $\xi$ due to the sample is expressed as an integral of the potential $V(x)$ along the neutron beam in the sample \cite{Rauch2000}:
\begin{equation}
\xi = - \frac{m}{k \hbar^2} \int V(x) dx.
\label{phase}
\end{equation}
Recording interferograms with and without sample one can extract the phase shift due to the sample.
In order to probe the chameleon,
we \cite{BraxPignol2013} proposed to use a rectangular vacuum cell as a sample.
When evacuated, a chameleon field $\varphi$ with a bubble-like profile will appear in the cell.
Since the neutron has an energy $V = \beta m/m_{\rm pl} \ \varphi$ in the chameleon field,
the phase associated with the chameleon bubble reads:
\begin{equation}
\xi = - \frac{m}{k \hbar^2} \beta \frac{m}{m_{\rm pl}} \int \varphi(x) dx .
\end{equation}
The cell needs to be evacuated in order to build up a chameleon field ``bubble'' in the cell,
because even air at atmospheric pressure is too dense.
Indeed, from the calculations presented in the previous part of this document,
we can estimate that the chameleon field is in the saturated regime in a few cm of air for strong coupling ($\beta \gg 1$).
The bubble solution \eqref{TwoPlates} cannot build up in air, instead, the solution would be nearly uniform in the cell.
Note that the neutron interferometers are generally operated in air.
The search for the chameleon field calls for a dedicated experiment with a vacuum cell.
Interestingly, it is possible in principle to switch off and on the chameleon bubble by adding or removing a small amount of gas in the cell.
We can profit from this effect in an experiment, because this is a rather unique signature which is easy to exploit in a practical setup.
If $\beta = 10^8$, a pressure of $10^{-2}$~mbar of helium (about a million times less dense than air!) is enough to suppress the bubble, as we will calculate later.
\subsection{The experiment}
A dedicated experiment to search for the chameleon field was performed at the S18 instrument (see \textcite{Kroupa2000} for a description) at ILL in summer 2013.
We used the largest single-crystal interferometer available (see picture Fig. \ref{interferometer})
with a beam separation of $50$~mm.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{interferometer}
\caption{
Picture of the large interferometer made out of a single-crystal silicon used in the S18 experiment.
\label{interferometer}
}
\end{figure}
A vacuum chamber specially designed for the chameleon search was built by the Atominstitut group in Vienna.
The device, depicted in Fig. \ref{sketchS18} consists of an aluminium vacuum chamber with inner dimensions $40 \times 40 \times 94 \ {\rm mm}^3$ and two air
chambers sitting alongside the vacuum chamber.
The whole box can be moved sidewards for swapping the vacuum cell between beam path I and beam path II and to probe different beam trajectories within the vacuum cell.
The vacuum chamber is connected to a gas system through a hole of diameter $5$~mm on the top of the cell.
The gas system comprises a pressure gauge, a motorized leak valve connected to a helium bottle and pumps.
A turbomolecular pump is running continuously while a controlled amount of helium is let in through the leak valve.
This way, the pressure of helium in the cell can be controlled in the range $10^{-4} \ {\rm mbar} < P < 10^{-2} \ {\rm mbar}$.
To get higher pressure the turbomolecular pump needs to be disconnected.
\begin{figure}
\centering
\includegraphics[trim=5cm 8cm 5cm 8cm,clip=true,width=0.93\linewidth]{sketchS18}
\caption{
Top view of the setup.
The chamber box (blue) can be moved by a remote controlled robot,
allowing the beams to pass at different positions labeled by 'a' to 'n'.
\label{sketchS18}
}
\end{figure}
During data-taking, we recorded many interferograms while varying the parameters of the cell (helium pressure and beam position).
It takes about half an hour to record one interferogram.
A major difficulty concerns the phase drifts due to environmental factors such as temperature gradients, air flow and vibrations.
The use of such a big single-crystal interferometer is very delicate since it is more sensitive to these nuisances compared to smaller interferometers.
To compensate for environmental phase drifts, interferograms were recorded in parallel.
It means that the phase flag was rotated to the first angular position
and neutrons were counted for a certain amount of time for each parameter setting.
Then the phase flag is rotated to the next position and neutrons are counted again for all parameter settings etc.
We recorded data with two different modes: the ``profile mode'' and the ``pressure mode''.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{interferogram}
\caption{
Interferograms recorded in position 'd'.
The plots show the intensity in the detector O as a function of the path length difference created by the rotating phase flag.
\label{interferogram}
}
\end{figure}
\para{Pressure mode. }
In the pressure mode, the beam II passes in the middle of the cell (position 'd').
We recorded four interferograms in parallel with four different helium pressures ($2.4 \times 10^{-4}, \ 7.1 \times 10^{-4}, \ 2.7 \times 10^{-3}, \ 1.1 \times 10^{-2}$~mbar).
A subset of these interferograms is shown in Fig. \ref{interferogram}.
\para{Profile mode. }
In the profile mode, a fixed value of the helium pressure is established in the cell.
Interferogram with different positions of the cell are recorded in parallel, to look for the bubble profile.
More specifically, we recorded interferograms at positions 'a', 'd' and 'g', that we will refer later as
\emph{center position} $(y=0,z = -4 \ {\rm mm})$ and \emph{side position} $(y = \pm 15 \ {\rm mm}, z = -4 \ {\rm mm})$.
This mode allows to probe a wider domain of pressure because the pressure needs not be changed rapidly.
However, there is a possible systematic effect in this mode associated with the variation of the wall thickness of the cell.
The data in the profile mode require a position dependent phase correction based on a precise mapping of the wall thickness.
\subsection{Expected signal}
To analyse the data, we need to calculate the field $\varphi$ inside the cell for different helium pressure.
The field satisfies the chameleon equation $\Delta \varphi = \mathcal{V}_{\rm eff}'(\varphi)$.
Since no exact analytical solution exists for this problem, a numerical scheme was employed.
To simplify the problem, we considered the chameleon equation in two dimensions $y,z$ transverse to the beam.
Indeed, the longitudinal size of the cell ($l = 94$~mm) is longer than the transverse size ($a = 40$~mm), we can assume that the field is approximately uniform in the $x$ direction.
We solve for the field in units of $M_\Lambda$, i.e. we introduce the dimensionless field $F(y,z) = \varphi(y,z) / M_\Lambda$.
The numerical method adopted to calculate the solution is described in Appendix \ref{A1}.
Figure \ref{bubble} shows selected results of such calculations.
\begin{figure}
\centering
\includegraphics[width=0.77\linewidth]{bubble_beta5e7_10m2mbar}
\includegraphics[width=0.77\linewidth]{bubble_beta5e7_10m3mbar}
\includegraphics[width=0.77\linewidth]{bubble_vac}
\caption{
Calculations of the chameleon bubble profile transverse to the neutron beam for different helium pressure in the cell.
We plot $F(y,z) = \varphi(y,z) / M_\Lambda$, for $n=1$ and $\beta = 5 \times 10^7$.
Top: pressure $=10^{-2}$~mbar, middle: pressure $=10^{-3}$~mbar, bottom: in vacuum.
\label{bubble}
}
\end{figure}
\para{Expected chameleon phase-shift. }
From the calculations of the transverse profile $F$, we can extract the expected phase signal
\begin{eqnarray}
\xi & = & - \frac{m}{k \hbar^2} \beta \frac{m}{m_{\rm pl}} \int M_\Lambda F \ dx \\
& = & - \frac{m}{k \hbar^2} \beta \frac{m}{m_{\rm pl}} M_\Lambda F \ l_{\rm eff}.
\end{eqnarray}
We have introduced the effective length $l_{\rm eff}$ to take into account the edge effects in the direction longitudinal to the beam.
To estimate the effective length of the bubble in the longitudinal direction, we performed a 3D lattice calculation of the bubble.
The 3D calculation is much more demanding in terms of CPU time as compared to the 2D calculation, in particular it does not permit to use a fine grid neither to repeat the calculation for multiple change of parameters.
Calculating with a grid size of $0.5$~mm the empty bubble for $n=1$ we found an effective length of $l_{\rm eff} = 84$~mm (compare to the physical length of $l = 94$~mm).
At the end, we found the following numerical relation between the expected phase shift (in degree) of the bubble $\xi_{\rm exp}$ and the 2D profile $F$:
\begin{equation}
\xi = - 4.7 \times 10^{-9} \ {\rm deg} \times \beta \times F.
\end{equation}
The 2D profile $F$ is a function of the transverse position $y,z$ of the beam with respect to the cell,
the helium pressure $p$ in the cell and the coupling $\beta$.
\para{Nuclear phase-shift. }
There is in principle an expected phase shift proportional to the pressure due to the Fermi potential of the helium.
Combining \eqref{FermiPot} and \eqref{phase} we find that this phase amounts to
\begin{equation}
\xi_{\rm nucl} = - \lambda \ l \, b \, n_{\rm He} = - 0.1 \ {\rm deg} \times \frac{P}{1 \ {\rm mbar}},
\end{equation}
where $\lambda = 0.27$~nm is the neutron wavelength, $l = 94$~mm is the longitudinal length of the chamber, $b = 3.26$~fm is the bound neutron scattering length of $^4$He and $n_{\rm He}$ is the number density of atoms.
For pressures below $10^{-2}$~mbar, as in the actual experiment, this phase shift is too small to be detected.
We will ignore this effect.
\subsection{The results}
The final results of the S18 experiment are shown in Fig. \ref{resultsS18}.
In the profile mode, there is no significant phase difference between the center and side position of the beam, even when the cell is evacuated to $10^{-4}$ mbar of helium.
The data does not show the bubble profile associated with a chameleon field.
In the pressure mode, where the beam always probes the center of the cell, there is no significant phase difference when the helium pressure changes.
The data is in agreement with the ``no chameleon'' hypothesis.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{profile_mode}
\includegraphics[width=0.93\linewidth]{pressure_mode}
\caption{
Results of the experiment at S18.
Top: phase difference between center and side position in the cell as a function of the helium pressure (profile mode).
Bottom: phase difference between low pressure and 0.011 mbar (pressure mode).
In both cases the orange line represents the theoretical expectation for $n=1$ and $\beta = 5 \times 10^7$.
\label{resultsS18}
}
\end{figure}
To derive a limit on the coupling $\beta$, we calculate the $\chi^2$:
\begin{equation}
\label{chi2}
\chi^2(\beta) = \sum_i \frac{(\xi(\beta)_i - \zeta_i)^2}{\sigma_i^2}
\end{equation}
where the sum goes over all 8 data points $\zeta_i \pm \sigma_i$ shown in Fig. \ref{resultsS18} thus combining the profile and pressure modes.
Figure \ref{chi2S18} shows $\chi^2(\beta)$ calculated for $n = 1, 2, 3, 4$.
The minimum of $\chi^2(\beta)$ is always obtained for $\beta = 0$.
To derive the exclusion limits (at a confidence level of $95$ \%) we use the standard criterion
\begin{equation}
\chi^2(\beta_{\rm lim}) = \chi^2_{\rm min} + 1.96^2,
\end{equation}
the limits are reported in Table \ref{limitsS18} and also in Fig. \ref{exclusion}.
For $n=1$, the limit is already a factor of 30 better than the current limit obtained with Gravity Resonance Spectroscopy \cite{Jenke2014}.
\bigskip \begin{center} \photon \end{center} \bigskip
A new experiment with the interferometer is scheduled for summer 2015.
By improving the contrast of the interferometer and increasing the statistics, we can gain at least a factor of 10 in sensitivity.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{chi2S18}
\caption{
Calculated $\chi^2(\beta)$ (according to Eq. \eqref{chi2}) for $n=1,2,3,4$.
The horizontal orange dashed line corresponds to the limit at a confidence level of 95 \%.
\label{chi2S18}
}
\end{figure}
\begin{table}
\center
\caption{
Limits on the chameleon coupling $\beta$ (at 95 \% C.~L.) derived from the S18 experiment at ILL.
}
\begin{tabular}{lll}
$n$ & ~ ~ & $\beta_{\rm lim}$ \\
\hline
1 & ~ ~ & $1.6 \times 10^7$ \\
2 & ~ ~ & $4.6 \times 10^7$ \\
3 & ~ ~ & $1.2 \times 10^8$ \\
4 & ~ ~ & $3.4 \times 10^8$ \\
\hline
\end{tabular}
\label{limitsS18}
\end{table}
\clearpage
\section{Bouncing neutrons with GRANIT}
\label{sectionGRANIT}
\subsection{Quantum states of bouncing neutrons}
If the kinetic energy of an incident neutron is less than the Fermi potential of a material wall (or floor, or ceiling)
then the neutron will be reflected off the surface for any angle of incidence.
This is the definition of an ultracold neutron.
For example, glass has a Fermi potential of 90 neV.
It reflects neutrons up to an energy of $E = 90$~neV, or a velocity of 4 m/s.
The reflection is almost perfectly specular, i.e. mirror-like.
The probability of non-specular (diffuse) reflection off a well-polished glass mirror can be less than $10^{-3}$ \cite{Nesvizhevsky2007}.
The Fermi potential of glass corresponds to a height in the gravity field of $h = 88$~cm through the relation $E = m g h$, where $m$ is the neutron mass and $g = 9.806$~m/s$^2$ is the acceleration in the lab.
It means that a neutron dropped from a height of less than a meter will bounce above a glass floor.
The bouncing period $T = \sqrt{8 h/g}$ is of the order of a second, like any object bouncing with a height of about a meter.
Now, if we consider a shorter bouncing height, or a higher bouncing frequency, the simple classical description of the bouncing ball might fail.
We will see that when the bouncing frequency enters the audio range (above about 100 Hz), corresponding to a bouncing height of about $10 \ \upmu {\rm m}$, a quantum description is needed.
\para{The quantum bouncer. }
The vertical motion of a neutron bouncing above a mirror nicely realizes the academic problem of a particle confined in a potential well.
The well consists in the gravitational potential $V(z) = m g z$ pulling the particle down and a perfect mirror $V(0) = + \infty$
\footnote{The mirror potential is the Fermi potential of glass ($\approx 100$~neV), it can be considered infinite because it is much larger than the energies of the quantum levels ($\approx 1$~peV).}
pushing the particle up.
According to the quantum description, the vertical motion has a discrete energy spectrum.
The bound states energies $E_k$ and the corresponding wavefunctions $\psi_k(z)$ satisfy the stationary Schr\"odinger equation
\begin{equation}
\label{schrodinger}
- \frac{\hbar^2}{2m} \frac{d^2}{dz^2} \psi_k + m g z \psi_k = E_k \psi_k, \quad \psi_k(0) = 0.
\end{equation}
The problem of finding the bound states energies and wavefunctions is solvable and we are going to solve it.
By dimensional analysis we define the characteristic height of the quantum bouncer:
\begin{equation}
\label{z0}
z_0 = \left( \frac{\hbar^2}{2 m^2 g} \right)^{1/3} = 5.87 \ \upmu {\rm m}.
\end{equation}
Defining the dimensionless height $Z = z/z_0$, the stationary Schr\"odinger equation takes the form of the \emph{Airy equation}
\begin{equation}
\frac{d^2 \psi_k}{dZ^2} + (\epsilon_k - Z) \psi_k = 0,
\end{equation}
where $\epsilon_k = E_k / mgz_0$.
The general solution of this second order linear differential equation is a linear combination of the two Airy functions Ai and Bi:
$\psi(Z) = C \ {\rm Ai}(Z-\epsilon) + D \ {\rm Bi}(Z - \epsilon)$.
The function Bi diverges at $Z = + \infty$ therefore the physically acceptable solutions are those with $D=0$.
Next, the condition $\psi_k(0) = 0$ at the boundary enforces the quantification of the energy levels ${\rm Ai}(-\epsilon_k) = 0$.
Therefore the energy levels of the stationary states are given by
\begin{equation}
E_k = mg z_0 \ \epsilon_k
\end{equation}
where
\begin{equation}
\epsilon_k = \{ 2.338, 4.088, 5.521, 6.787, \cdots \}
\end{equation}
is the sequence of the negative zeros of the Airy function.
Numerically the energy levels are in the range of $1 \ {\rm peV} = 10^{-12}$~eV ($E_1 = 1.41$~peV, $E_2 = 2.46$~peV $\cdots$),
five orders of magnitude below the kinetic energy of ultracold neutrons.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{wavefunctions}
\caption{
Representation of the probability density functions $|\psi_k(z)|^2, k = 1, 2, 3, 4$ for the first four energy states of the neutron bouncer.
The four functions are translated in the vertical axis according to the energy of the state.
The grey zone is classically forbidden.
\label{wavefunctions}
}
\end{figure}
To each quantum state $\ket{k}$ corresponds the wavefunction
\begin{equation}
\psi_k(z) = C_k \ {\rm Ai}( z/z_0 - \epsilon_k),
\end{equation}
where $C_k$ is a normalization coefficient.
The probability density functions $|\braket{z}{k}|^2 = |\psi_k(z)|^2$ associated with the wavefunctions are plotted in Fig. \ref{wavefunctions}.
As in any 1D potential well, the number of nodes (the places where the wavefunction vanishes) of the $k^{\rm th}$ quantum state is $k$.
Let us compare the neutron quantum bouncer to the hydrogen atom, a more familiar case of a bound system.
The typical energy of the neutron bouncer is 13 orders of magnitude smaller than 13.6~eV, the binding energy of the electron in the ground state.
The typical size of the wavefunction is 5 orders of magnitude larger than 0.05~nm, the size of the hydrogen atom.
These very unusual scales arise due to the extraordinary weakness of gravity as compared to the Coulomb force.
Another important difference concerns the stability of the excited levels.
An excited state of the hydrogen atom decays by spontaneous emission of a photon with a time scale of the order of 1~ns.
The neutron bouncer could also in principle decay by spontaneous emission of gravitons.
However, \textcite{Pignol2007} estimated the timescale of this effect to $10^{77}$~s, i.e. 60 orders of magnitude longer than the age of the Universe.
Again, this is due to the extreme weakness of gravity.
\para{Observing the quantum states. }
It is usually believed that the weakness of gravity is not the experimentalist's friend.
In the case of the neutron bouncer, however, it is a great asset.
It makes the wavefunctions very large, almost visible to the naked eye.
This property has been used in a series of experiments started in 1999 at the PF2 ultracold neutron source of the ILL reactor
\cite{Nesvizhevsky2002,Nesvizhevsky2003,Nesvizhevsky2005}.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{flowThrough1}
\includegraphics[width=0.93\linewidth]{DiscoveryILL}
\caption{
Top: scheme of the method used to discover the quantum states of the neutron bouncer.
Bottom: Figure taken from \textcite{Nesvizhevsky2002}, reporting the flux in the detector as a function of the height of the slit.
The solid curve is the classical fit to these data.
\label{DiscoveryILL}
}
\end{figure}
The setup which served to observe for the first time the quantum states of the neutron bouncer is sketched in Fig. \ref{DiscoveryILL}.
It is a flow-through experiment: ultracold neutrons with horizontal velocity are coming from the left, they bounce on top of a horizontal, well-polished glass mirror.
They are then counted in a gaseous detector filled with $^3$He.
A flat neutron absorber is installed on top of the horizontal mirror, parallel to it.
It forms a slit, with a height $h$ that can be precisely regulated with a micrometer accuracy.
The experiment consists in recording the neutron flux passing through as a function of the height of the slit.
The result obtained in this experiment is shown in Fig. \ref{DiscoveryILL}.
When the slit height is $h < 15 \ \upmu {\rm m}$, no neutron passes through the slit, contrary to the classical expectation.
In the quantum picture, the neutron wavefunction is a linear combination of the stationary quantum states.
A given state ``decays'' in the slit if its wavefunction overlaps significantly with the absorber.
When the absorber is lower than $15 \ \upmu {\rm m}$, every state, including the ground state $\ket{1}$, overlaps with the absorber and the slit becomes completely opaque to neutrons.
This is precisely what was observed.
A quantitative analysis of the transmission curve allows to extract the spatial parameters of the quantum states.
The difficulty of the analysis lies in the choice of a model for the absorber.
\textcite{Nesvizhevsky2005} parametrized the transmission curve by introducing the critical height of the first quantum state $h_1 = \epsilon_1 z_0$.
It corresponds to the turning point of a classical bouncer with vertical energy $E_1$.
They report $h_1^{\rm exp} = (12.2 \pm 1.8_{\rm syst} \pm 0.7_{\rm stat}) \upmu {\rm m}$, the systematic uncertainty takes into account the error in the slit size calibration and the finite accuracy of the absorber model.
In terms of the parameter $z_0$, the extracted value $z_0^{\rm exp} = (5.2 \pm 0.8) \upmu {\rm m}$ agrees with the theoretical value $z_0^{\rm th} = 5.87 \ \upmu {\rm m}$.
There is a more direct way to access the spatial structure of the quantum states, by taking a photograph.
Instead of just counting the neutrons at the exit of the mirror, a position sensitive detector with a micrometric resolution could be used.
As proposed by \textcite{Nesvizhevsky2000}, one should aim at observing the node of the second quantum state
(at a height of $10 \ \upmu {\rm m}$) which provides a clear signature, instead of looking at the ground state which does not display a specific pattern.
Selecting the first few levels could be done with a slit: a horizontal mirror and an absorber on top.
At the exit of the slit, we place a second horizontal mirror (no absorber on top).
The second mirror is placed $13.5 \ \upmu {\rm m}$ below the first mirror, in order to increase the vertical energy of the neutron bouncing on the second mirror.
This way, the ground state population will be suppressed and the population of the second quantum state enhanced.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{photo}
\caption{
Histogram of the neutron vertical position measured with the nuclear track position sensitive detector.
A dip corresponding to the first node of the excited wavefuctions is clearly visible.
The figure is taken from \textcite{Pignol2009}.
\label{photo}
}
\end{figure}
A first version of this experiment has been performed at the PF2 ultracold neutron source in 2005.
A plastic nuclear track detector coated with a thin $^{235}$U conversion layer was used.
Absorption of an ultracold neutron in the uranium induces a fission.
One of the two fission products is emitted in the plastic and the ionization produces a latent track.
After removing the conversion layer and etching the plastic with NaOH, the tracks become visible with an optical microscope.
The result of this measurement is shown in Fig. \ref{photo}.
A fit of the data allows to extract the populations of the quantum states, the spatial resolution of the detector, the background level, the height offset and the $z_0$ parameter \cite{Pignol2009}.
The result is $z_0 = (6.0 \pm 0.2) \ \upmu {\rm m}$, it is more accurate than the transmission-through-the-slit method.
This first measurement was not ideal because the populations of quantum states up to $\ket{10}$ was quite high.
Improvements are definitely possible, by refining the purity of the selection of the state $\ket{2}$.
We aim at taking better pictures with the GRANIT instrument.
\para{Gravity Resonance Spectroscopy. }
Imagine we prepare an initial quantum state $\ket{k}$.
It is possible to induce a transition $k \rightarrow l$ to the state $\ket{l}$
with a periodic excitation of frequency $f_{kl} = (E_k - E_l)/ 2 \pi \hbar$.
We define\footnote{Notice that a classical ball dropped from the height $z_0$ bounces with a frequency of $\pi \times f_0$. }
\begin{equation}
f_0 = \frac{m g z_0}{2 \pi \hbar} \approx 145 \ {\rm Hz},
\end{equation}
the transition frequencies write
\begin{equation}
f_{k l} = f_0 ( \epsilon_k - \epsilon_l ).
\end{equation}
For the first few quantum states, the transition frequencies lie in the audio-range,
for example we expect a resonance associated to the $1 \rightarrow 2$ transition at the frequency $f_{21} = 254$~Hz.
Conveniently, the audio frequency range is easily accessible by electrical as well as mechanical oscillators.
This allows direct experimental access to the energy spectrum of the neutron quantum bouncer, a technique called \emph{Gravity Resonance Spectroscopy} (GRS).
Let us recall the quantum mechanics of a Rabi resonance, at the basis of the GRS.
Assume we apply the external harmonic perturbation $V(t) = \hat{V} \cos(\omega t)$,
with the excitation frequency $\omega/2 \pi$ close to a transition frequency $f_{k l}$.
It means that the detuning $\delta \omega = \omega - 2 \pi f_{k l}$ should be small compared to the angular frequencies of the transitions in the spectrum.
Assume also that the bouncer is prepared initially in the sate $\ket{\psi(0)} = \ket{k}$.
In this case, the two level approximation holds and the system can be described by $\ket{\psi(t)} = a(t) \ket{k} + b(t) \ket{l}$.
Solving the time-dependent Schr\"odinger equation
\begin{equation}
i \hbar \frac{d}{dt} \ket{\psi(t)} = \left( \frac{\hat{p}^2}{2m} + mg \hat{z} \right) \ket{\psi(t)} + \hat{V} e^{i \omega t} \ket{\psi(t)}
\end{equation}
we deduce the quantum amplitudes $a(t)$ and $b(t)$.
Finally the transition probability $P_{k \rightarrow l}(t) = |b(t)|^2$ is given by the \emph{Rabi formula}:
\begin{equation}
P_{k \rightarrow l}(t) = \frac{\sin^2 (\sqrt{\delta \omega ^2 + \Omega^2} \ t /2)}{1 + \delta \omega ^2 / \Omega^2},
\end{equation}
where
\begin{equation}
\Omega = \frac{1}{\hbar} \bra{l} \hat{V} \ket{k}
\end{equation}
is the Rabi angular frequency that characterizes the strength of the excitation.
The duration $t$ available for the excitation is an important parameter.
A longer duration is desirable because it corresponds to a narrower resonance line.
In flow through mode, neutrons pass a transition region of typically 16~cm at a speed of 4~m/s, corresponding to a duration of
$t_0 = 40$~ms (these parameters are those of the system built for the first stage of GRANIT).
With the duration of the excitation fixed, one should set the excitation strength such that $\Omega t_0 = \pi$.
Therefore, the transition probability is maximum at resonance: $P_{k \rightarrow l}(t_0) = 1$ when $\delta \omega = 0$.
The measurement consists in recording the presence of the state $\ket{l}$ (or the disappearance of the state $\ket{k}$) for different values of the excitation frequency $\omega/2 \pi$.
It forms a resonance curve with a peak at $\omega/2 \pi = f_{k l}$.
The width (FWHM) of the peak is $\Delta f = \Omega / \pi = 1 / t_0$.
There are two practical possibilities to do the excitation (i) using a vibrating mirror (ii) using a magnetic force.
GRS with vibrating mirror has been realized at the PF2 UCN beamline by \textcite{Jenke2011,Jenke2014}.
The technique with a magnetic excitation is in preparation with the GRANIT instrument.
We will come back to this subject in detail later.
\para{Testing the equivalence principle. }
We learn at school that it is impossible to tell the mass of a body by observing how it falls down.
Newton's law of motion for a falling body is $m_i a = m_g g$,
where $a$ is the acceleration of the body,
$g$ is the gravity strength,
$m_i$ is the inertial mass, $m_g$ is the gravitational mass.
If we assume that both masses are equal then we find that the acceleration of all bodies is the same, equal to $g$.
The hypothesis ($m_i = m_g$) is called the {\bf weak equivalence principle},
the observable consequence is called {\bf the universality of free fall}.
Surprisingly, the weak equivalence principle does not imply the universality of free fall in the case of the quantum bouncer.
This is because the mass in the Schr\"odinger equation \eqref{schrodinger} does not simplify, contrary to Newton's law of motion.
By measuring the properties of the quantum bouncer we can tell the mass.
Indeed, from Eq. \eqref{z0} we derive $m = \hbar/\sqrt{2 g z_0^3}$.
Using the result of the photograph Fig. \ref{photo} we measure the neutron mass:
$m c^2 = 910 \pm 50 \ {\rm MeV}$.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{EPtest}
\caption{
Inertial mass versus gravitational mass of the neutron.
The diagonal line corresponds to the weak equivalence principle $m_i = m_g$.
The pink band represents the $1 \sigma$ region deduced from the analysis of the photograph of the quantum states.
The blue dashed line is an iso-$E_0$ line.
\label{EPtest}
}
\end{figure}
It is therefore of interest to investigate the question of the equivalence principle with the quantum bouncer in more details.
Distinguishing the inertial and gravitational masses, the Schr\"odinger equation becomes
\begin{equation}
\label{schrodinger2}
- \frac{\hbar^2}{2m_i} \frac{d^2}{dz^2} \psi_k + m_g g z \psi_k = E_k \psi_k, \quad \psi_k(0) = 0.
\end{equation}
In this framework, the size of the wavefunctions is governed by the parameter
\begin{equation}
z_0 = \left( \frac{\hbar^2}{2 m_i m_g g} \right)^{1/3}.
\end{equation}
It means that the measurements of the spatial properties of the quantum states are sensitive to the product $m_i m_g$.
Figure \ref{EPtest} interprets the result of the first photograph in the $m_i, m_g$ plane.
Next, the energies of the quantum levels are $E_k = E_0 \epsilon_k$ with
\begin{equation}
E_0 = m_g g z_0 = \left( \frac{m_g^2}{m_i} \frac{g^2 \hbar^2}{2} \right)^{1/3}.
\end{equation}
It is remarkable that the measurement of the energy levels is sensitive to another combination of the masses, namely $m_g^2/m_i$.
Then, not only can we tell the mass of the neutron by observing the quantum bouncer,
but we can measure \emph{separately} the inertial mass $m_i$ and gravitational mass $m_g$.
For this, we need to measure the spatial feature of the quantum states to extract $z_0$ on the one hand
and to perform the spectroscopy of the quantum states with GRS to extract $E_0$ on the other hand.
\subsection{Effects of the chameleon field}
Since the energy difference between the first quantum states is as low as about $1$~peV,
the neutron quantum bouncer is a sensitive system to search for hypothetical new forces acting on the neutron at the vicinity of the mirror.
Such a new force could be induced by the presence of the chameleon field.
Since this force is attractive, it would shrink the wavefunctions of the quantum states and enlarge the energy spectrum.
We have argued in \textcite{Brax2011} that extremely large coupling $\beta > 10^{11}$
would lead to new bound states at a distance of less than $2 \ \upmu {\rm m}$
which is ruled out by the first transmission-through-the-slit experiment.
The limit $\beta < 10^{11}$ is practically independent of the Ratra-Peebles index $n$.
Next, \textcite{Jenke2014} analyzed their experiment with the vibrating mirror to derive limits on the chameleon coupling.
They quote the limit $\beta > 6 \times 10^8$, also practically independent of $n$.
Let us now analyse quantitatively how the properties of the quantum bouncer are modified in the presence of the chameleon field.
As we have seen in part \ref{Chameleon}, if the chameleon is strongly coupled with matter ($\beta \gg 1$),
a scalar field $\varphi(z)$ will be generated above the mirror.
We have seen that the field $\varphi(z)$ is saturated: it is independent of $\beta$ as long as $\beta \gg 1$, it is also independent of the mass density of the mirror.
When no absorber is placed above the mirror the scalar field is given by Eq. \eqref{OnePlate}.
Then, the neutron feels the potential
\begin{equation}
V(z) = m g z + \beta \frac{m}{m_{\rm Pl}} \varphi(z).
\end{equation}
The chameleon term in the potential has the form
\begin{equation}
\delta V (z) = \beta \frac{m}{m_{\rm Pl}} \varphi(z) = \beta v_n ( z / \lambda )^{\alpha_n}
\label{deltaV}
\end{equation}
with
\begin{equation}
v_n = \frac{m}{m_{\rm Pl}} M_\Lambda \left( \frac{2+n}{\sqrt{2}} \right)^{\alpha_n} \approx
\left( \frac{2+n}{\sqrt{2}} \right)^{\alpha_n} \times 10^{-21} \ {\rm eV},
\end{equation}
where $\lambda = \hbar c / M_\Lambda = 82 \ \upmu {\rm m}$ and $\alpha_n = 2/(2+n)$.
The task is now to calculate the shift of the energy states
and the shrinking of the wavefunctions induced by the perturbation $\delta V$.
We skip the details of the calculations (see appendix \ref{A2}) and we discuss the results.
\para{Shrinking of the wavefunctions. }
At first order in perturbation theory, the correction of the wavefunctions is
\begin{equation}
\psi_k(z) = \psi_k^{(0)}(z) + \beta v_n (z_0/\lambda)^{\alpha_n} \sum_{l \neq k} \frac{ O_{kl}(\alpha_n) }{E_k - E_l} \psi_l^{(0)}(z).
\end{equation}
In Fig. \ref{wavefunctionsChameleon}, we plot the modification of the second state for the coupling $\beta = 10^9$
using the results in the appendix \ref{A2} (the sum goes up to $l=5$).
The effect of the chameleon is a shrinking of the wavefunction, as expected for an attractive force.
The node of the wavefunction $h^0_2$ (which we will try to measure in GRANIT) is displaced to lower $z$.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{wavefunctionsChameleon}
\caption{
Wavefunction of the second quantum state $\psi_2(z)$.
The black curve corresponds to the unperturbed case (no chameleon).
The red curve is the wavefunction in the presence of the chameleon with $n=2$ and $\beta = 10^9$, calculated at the first order of perturbation theory.
\label{wavefunctionsChameleon}
}
\end{figure}
More quantitatively, a numerical study in the case $n=2$ shows that the displacement of the node is a linear function of $\beta$ for $\beta < 10^9$:
\begin{equation}
h^0_2 = ( 10.3 - \beta \times 7.7 \times 10^{-10} ) \ \upmu {\rm m}
\end{equation}
A precision of $0.1 \ \upmu {\rm m}$, which is what we could reasonably get in forthcoming measurements with GRANIT,
will correspond to $\beta \approx 10^8$.
This method is less sensitive than the gravity resonance spectroscopy, as we will see.
\para{Shifts of the transition frequencies. }
At first order in perturbation theory, the correction of the energy levels is
\begin{equation}
\delta E_k = \beta v_n (z_0/\lambda)^{\alpha_n} O_{kk}(\alpha_n),
\end{equation}
and the correction of the transition frequencies is
\begin{eqnarray}
\nonumber
\delta f_{kl} & = & \frac{\delta E_k - \delta E_l}{2 \pi \hbar} \\
& = & \beta \frac{v_n}{2 \pi \hbar} (z_0/\lambda)^{\alpha_n} ( O_{kk}(\alpha_n) - O_{ll}(\alpha_n) ).
\end{eqnarray}
For example, consider the shift of the resonance line $2 \rightarrow 1$ (expected at $f_{21} = 254$~Hz) due to the chameleon with the Ratra-Peebles index $n=2$.
We find numerically
\begin{equation}
\delta f_{21} = \beta \times 3.7 \times 10^{-8} \ {\rm Hz}.
\end{equation}
As we will see later, we aim at a precision better than $1$~Hz for the measurement of the $2 \rightarrow 1$ transition frequency
with the flow-through resonance setup in GRANIT.
This measurement will therefore be sensitive to a chameleon coupling of $\beta \approx 2 \times 10^{7}$.
It is possible in principle to improve the precision on the transition frequency by increasing the interaction time $t$.
We recall that the width of the resonance is given by $\Delta f = 1/t$,
and the precision is approximately given by $\delta f = \Delta f / \sqrt{N}$, where $N$ is the number of counts in the resonance.
Instead of a flow-through setup, one needs to trap the neutrons in quantum states.
Ultimately, the precision is limited by the beta decay lifetime of the neutron $\tau_n \approx 880$~s.
Therefore the ultimate precision is $\delta f \approx 1/(\tau_n \sqrt{N}) \approx 10^{-5}$~Hz (taking $N = 10000$).
In conclusion, the measurement of the spectrum of the quantum states could potentially be sensitive to chameleon couplings down to $\beta = 10^2$ (a long way to go, though...).
\subsection{The GRANIT instrument}
GRANIT is an instrument dedicated to the study of the neutron quantum bouncer, now in the commissioning phase.
For a description of the instrument see \textcite{SchmidtWellenburg2009,Baessler2011a,Baessler2011b,Roulier2015,Roulier2015b}.
GRANIT is located at the level C of the ILL reactor in Grenoble.
As shown in Fig. \ref{GRANIT}, the instrument comprises a superthermal $^4$He source
(fed by the monochromatic cold neutron beam H172A),
producing ultracold neutrons to be used in the spectrometer located in a ISO 5 class clean room.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{GRANIT1}
\includegraphics[width=0.93\linewidth]{GRANIT2}
\caption{
Top: schematics of the GRANIT instrument at the ILL H172A beam line.
Bottom: picture of the spectrometer inside the clean room.
\label{GRANIT}
}
\end{figure}
No attempt to describe the instrument in technical details will be made here, except for a few selected items.
We will first describe the ultracold neutron source.
Next, the setup for the magnetic excitation of resonant transitions will be explained.
\para{The ultracold neutron source. }
Previous experiments observing the neutron quantum bouncer were all performed at the PF2 instrument,
the ``first generation'' ultracold neutron source situated at the level D of the ILL reactor.
``Next generation'' sources can potentially deliver many more ultracold neutrons,
and GRANIT operates now a source based on the SUN1 prototype described in \textcite{Zimmer2011,Piegsa2014}.
The concept of the source has been proposed by \textcite{Golub1977}:
a volume filled with superfluid $^4$He is irradiated with a beam of cold neutrons.
Neutrons can undergo inelastic scattering in the medium, losing energy while emitting a phonon.
Given the dispersion relation of the phonons in superfluid helium,
it turns out that a neutron having a wavelength of exactly $\lambda^* = 0.89$~nm
transfers its entire energy to the phonon and comes at rest in the superfluid.
Neutrons with a wavelength not too far from $\lambda^*$ come almost at rest.
The process produces ultracold neutrons, with a volumic production rate (number of UCN produced per unit volume and unit time) of
\begin{equation}
P = \left. \frac{d \phi}{d \lambda} \right|_{\lambda^*} \times (3.0 \pm 0.2) \times 10^{-9} \ {\rm nm} \ {\rm cm}^{-1},
\end{equation}
where $d \phi / d \lambda$ is the differential flux in the cold neutron beam.
This theoretical estimate is valid for neutrons created in the velocity range $0 < v < 5.6 \ {\rm m/s}$ \cite{Piegsa2014},
corresponding to the storage capability of a trap made with stainless steel walls
\footnote{
Our trap is made of BeO walls with a Fermi potential of $257$~neV,
but a small portion is made of stainless steel with a Fermi potential of $184$~neV only.
Since superfluid helium has a Fermi potential of 18~neV, the threshold energy for trapped neutrons is $E < 184 - 18$~neV,
corresponding to a critical velocity of 5.6~m/s.
}.
When the reactor power is $58$~MW, H172A delivers $3 \times 10^9$ neutrons per second with a wavelength of 0.89~nm
(the differential flux has a Gaussian shape with a RMS $\sigma = 0.021(1)$~nm),
in a square section of $7 \times 7 \ {\rm cm}^2$ \cite{Roulier2015b}.
In the continuity of the guide, the source itself consists in a tube with square section (inner dimensions $7\times7\times100 \ {\rm cm}^3$) with BeO walls filled with superfluid helium.
The differential flux inside this production volume is reduced by absorption of cold neutrons in windows ($\times 0.9$) and by the divergence of the beam ($\times 0.8$),
resulting in an effective flux of $d\phi/d\lambda |_{\rm eff} = 8 \times 10^8 \ {\rm cm}^{-2} {\rm s}^{-1} {\rm nm}^{-1}$.
The UCN production inside our source is then expected to occur at a rate of $P \approx 2.3$~UCN/cm$^3$/s.
In total the 4900~cm$^3$ should produce about $10,000$ ultracold neutrons per second.
Those ultracold neutrons produced in the superfluid helium are stored by the BeO walls of the production volume.
To extract the neutrons, the volume is connected to a chimney (a stainless steel tube) going up to a UCN valve.
When the valve is closed UCNs accumulate in the volume.
Above the valve there is a UCN guide going up, then a 90 degree bent, followed by an horizontal stainless steel guide connected to the spectrometer inside the clean room.
For efficient accumulation of UCNs inside the source, the storage time $\tau$ should be as long as possible
in order to maximize the UCN density $\rho = P \tau$ at saturation.
The loss rate $1/\tau$ inside the source has several contributions:
\begin{equation}
1/\tau = 1/\tau_n + 1/\tau_{\rm wall} + 1/\tau_{^3{\rm He}} + 1/\tau_{\rm up}
\end{equation}
where $\tau_n \approx 880$~s is the beta decay lifetime,
$\tau_{\rm wall}$ corresponds to the losses at wall collision,
$\tau_{^3{\rm He}}$ is due to the absorption by residual traces of $^3$He in the bath of normally pure $^4$He,
and $\tau_{\rm up}$ corresponds to upscattering by thermal excitations in superfluid helium.
Inconveniently, this last contribution forces us to operate the source at temperatures below $1$~K.
According to the theoretical estimation by \textcite{Golub1979}, the temperature dependent upscattering rate is
\begin{equation}
1/\tau_{\rm up} = A e^{-E^*/T} + B \ T^7 + C T^{3/2} e^{- \Delta / T}
\end{equation}
where the first term corresponds to one-phonon absorption with $A = 500 \ {\rm s}^{-1}$,
the second term corresponds to two-phonon process with $B = 0.008 \ {\rm s}^{-1}$,
the last term corresponds to roton-phonon process with $C = 18 \ {\rm s}^{-1}$.
Helium temperature $T$ is expressed in Kelvin,
$E^* = 12$~K is the phonon energy corresponding to $\lambda^*$ and $\Delta = 8.6$~K.
The time $\tau_{\rm up}$ is plotted as a function of the temperature in Fig. \ref{upscattering} (top).
We see the dramatic effect of the temperature in Fig. \ref{upscattering} (bottom):
we measured the flux of UCNs at the exit of the source as a function of the temperature.
The flux is significantly reduced when the temperature is above 1~K.
Even at low temperature we only count 400 UCNs per second.
Some losses are due to absorption of neutrons in the vacuum separation window (a $15 \ \upmu {\rm m}$ thick titanium foil situated in the horizontal guide)
and the entrance window of the detector (another $15 \ \upmu {\rm m}$ thick titanium foil).
Also, the storage time at low temperature is only about $30$~s, which is not very long compared to the emptying time of $13$~s \cite{Roulier2015}.
There are also certainly some more losses which are being investigated.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{upscattering}
\includegraphics[width=0.96\linewidth]{UCNvsT}
\caption{
Top: Theoretical estimation of the lifetime due to upscattering $\tau_{\rm up}$ in superfluid helium,
as a function of the temperature.
Bottom: Flux of UCN at the measured at the exit of the horizontal extraction guide (a tube with inner diameter 3~cm)
as a function of the temperature of the helium (with the UCN valve completely open).
\label{upscattering}
}
\end{figure}
Let us describe a little bit what's inside the spectrometer.
As shown in Fig. \ref{GRANIT3}, the horizontal guide is connected
to a cylindrical storage volume labeled ``copper antichamber''
(in the test mentioned before, a detector was placed instead of the antichamber).
The $30$~cm long cylinder has a diameter of $4$~cm, it is made of pure copper (Fermi potential $168$~neV).
A small aperture in the antichamber leads to a semi-diffusive slit made of a bottom reflective mirror and a top diffusing mirror.
Both mirrors are coated with diamond like carbon.
The idea is that neutrons with too large vertical velocity in the slit are diffused back in the antichamber.
On the contrary, neutrons with small vertical velocity, those bouncing on the bottom mirror without touching the roof, are transmitted.
In a first test of this system conducted in December 2014, the opening of the semi-diffusive slit was set to $127 \ \upmu {\rm m}$,
preparing the first 20 quantum states.
A $^3$He counter with an aluminium entrance window was placed immediately at the exit of the slit.
Unfortunately due to a cryogenic failure it was only possible to get to $T = 1.3$~K for the temperature of the source.
In this configuration, we counted $0.13$ neutrons per second out of the semi-diffusive slit.
This is not yet as good as the experiments performed at PF2 (see Fig. \ref{DiscoveryILL}).
By fixing the cryogeny and using a detector with improved efficiency for soft UCNs (i.e. with a titanium entrance window for example) things will get better.
Progress is under way to deliver a decent flux of ultracold neutrons at the exit of the slit.
\begin{figure}
\centering
\includegraphics[width=0.99\linewidth]{GRANIT3}
\caption{
Scheme of the connection of the source to the spectrometer.
Not shown: container of the superfluid, thermal screens, vacuum chamber of the source,
lead biological shield, wall of the clean room, vacuum chamber of the spectrometer, support of the copper antichamber and semi-diffusive slit.
\label{GRANIT3}
}
\end{figure}
\para{The magnetic excitation. }
Now I describe how we plan to measure the resonant transitions using the magnetic excitation as proposed in
\textcite{Pignol2009,Kreuz2009,Pignol2014,Baessler2015}.
An oscillating magnetic field gradient will generate the vertical force on the neutron to excite the system.
Indeed, the motion of a neutron in a magnetic field ${\bf B}$ is governed by the potential $s \mu |{\bf B}|$
\footnote{This is valid when the variation of the magnetic field is slow compared to the Larmor frequency.
In this case, called \emph{the adiabatic regime}, the spin follows the direction of the magnetic field and the neutron trajectory and the spin dynamics are decoupled.},
where $\mu = 60$~neV/T is the magnetic moment and $s=1$ for ``spin up'' neutrons and $s=-1$ for ``spin down'' neutrons.
Classically, a vertical force is applied on the neutron by the field gradient $\partial_z |{\bf B}|$
(an homogeneous field does not exert any force, it acts only on the spin dynamics).
Let us now assume a magnetic excitation of the form $|{\bf B}| = \beta z \cos \omega t$,
the corresponding quantum mechanical potential reads
\begin{equation}
\hat{V}(t) = s \mu \beta \hat{z} \cos \omega t.
\end{equation}
This is what we need to excite resonant transitions between the quantum states of the neutron bouncer.
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{sketchMagneticTransition}
\caption{
Sketch of the flow through setup.
UCNs enter from the semi-diffusive slit on the left ; they fall down the step to depopulate the ground state (1),
16~cm long transition region (2),
9~cm long filter (3), and detector (4).
\label{sketchMagneticTransition}
}
\end{figure}
The sketch of the setup is shown in Fig. \ref{sketchMagneticTransition}.
The spectroscopy method is conceptually the same as the famous 1939 Rabi resonance experiment.
It is performed with four steps
(1) state preparation,
(2) resonant transition,
(3) state analysis, and
(4) detection of the transmitted flux.
\begin{enumerate}
\item UCNs are first prepared in an excited state by going down
a step (1) of height $15 \ \upmu {\rm m}$.
The population of the first state $\ket{1}$ is suppressed as compared to the populations of the excited states $\ket{2}, \ket{3}, \cdots$.
\item Next, transitions between quantum states are excited with a periodic magnetic field gradient.
The length of the transition region $L = 16$~cm corresponds to an average passage time $t_0 = 40$~ms
\footnote{We expect the $v_x$ velocity along the beam to be distributed with a mean value of 4~m/s and a standard deviation of 1.5~m/s.}.
Two different schemes could be implemented in principle: the AC excitation ad the DC excitation.
In the DC mode, the field gradient is static and spatially oscillating in the $x$ direction with a period of $d = 1$~cm.
In this case, only neutrons with specific horizontal velocities meet the resonance condition.
The deexcitation $2 \rightarrow 1$ is expected to occur at a frequency of $f_{21} = 254$~Hz corresponding to the resonant horizontal velocity of $v_{21} = d f_{21} = 2.54$~m/s (in the $3 \rightarrow 1$ case we expect $v_{31} = 4.62$~m/s).
In the AC mode, the field gradient is spatially uniform and oscillating in time.
One would then find the resonances by directly scanning the excitation frequency.
\item A second horizontal mirror above the main mirror serves as a state filter.
For a slit opening of about $25 \ \upmu {\rm m}$, only neutrons in the ground state are accepted and higher quantum states are absorbed.
\item Neutrons are finally detected at the exit of the filter.
In the AC mode one should see a resonance in the transmitted flux as a function of the excitation frequency.
In the DC mode, one needs to measure the horizontal velocity of transmitted neutrons.
This is achieved by measuring the height of the neutrons after a 30~cm long free fall.
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=0.93\linewidth]{wireArray}
\caption{
Picture of the wire array for the magnetic excitation.
It is made of four modules, each one holding 32 adjacent wires with a section of 1~mm$^2$ each and a length of 30~cm in the $y$ direction.
Adjacent wires are separated by a gap of 0.25~mm.
\label{wireArray}
}
\end{figure}
We need to apply an oscillating gradient with an amplitude of $\beta = 0.22$~T/m to induce the $2 \rightarrow 1$ transition and $\beta = 0.74$~T/m for $3 \rightarrow 1$ \cite{Pignol2014}.
This is done with an array of 128 copper wires with square section arranged as shown in Fig. \ref{sketchMagneticTransition}.
The array will be placed above the horizontal mirror in region (2) at a distance of 0.8~mm from the mirror.
The real system shown in Fig. \ref{wireArray}.
Electrical connectors are arranged so that the following 8-periodic pattern current could be applied:
$I_1, I_2, I_3, I_4, -I_1, -I_2, -I_3, -I_4, I_1, \dots$ and thus the magnetic field produced by the array will be 1~cm periodic in the $x$ direction.
It is possible to tune the currents in the 4 circuits $I_1, I_2, I_3, I_4$ to create a homogeneous gradient at the surface of the mirror,
as shown in Fig. \ref{gradient}.
The configuration of the gradient depends on the external magnetic field:
by applying a strong external field (of the order of 2~mT) on top of the field created by the wire array,
one generates an oscillating gradient in the $x$ direction.
The wire array is therefore a versatile device to generate the field gradient that can be used for the AC as well as for the DC excitation modes.
\begin{figure}
\centering
\includegraphics[angle=90,width=0.93\linewidth]{gradient}
\caption{
Calculation of the magnetic field gradient $\partial_z |{\bf B}|$ produced at the surface of the main mirror
with currents set to $I_1 = I_4 = 1.4$~A, $I_2 = I_3 = 3.5$~A and without any external field applied.
\label{gradient}
}
\end{figure}
We will apply currents up to 5~A in the circuits, corresponding to Joule heating of up to 25~W
(the resistance of each circuit is $0.3 \ \Omega$).
Since the wire array will be operated in vacuum, an active cooling system is necessary to dissipate the heat.
We built a heat exchanger with circulating cold gaseous nitrogen and tested it successfully in vacuum.
In the AC mode, we apply AC currents in the wires.
Note that the gradient $\beta$ is not a linear function of the currents.
In particular, the gradient always satisfies $\beta > 0$, because the field $|{\bf B}|$ is always weaker away from the wire.
As a result, the excitation frequency (i.e. the frequency of $\beta(t)$ seen by the neutrons)
is twice the driving frequency (i.e. the frequency of the AC current in the wires).
The complete system of excitation has been simulated in \textcite{Pignol2014} by solving the time-dependent Schr\"odinger equation
with realistic function $\beta(t)$ taking into account the non-linearities.
The expected resonance curve is shown in fig. \ref{resonance}.
We see a double resonance, one corresponding to spin up neutrons, the other to spin down.
This is a shift called \emph{Stern-Gerlach shift} \cite{Baessler2015} caused by the unavoidable constant (time-independent) component of $\beta(t)$.
Fortunately this shift cancels when averaging the two resonant frequencies.
With a statistics of 10,000 neutrons counted during a scan of the driving frequency between $50$~Hz and $200$~Hz,
it will be possible to extract the resonance frequency $f_{21}$ with a precision better than $1$~Hz.
\begin{figure}
\centering
\includegraphics[angle=90,width=0.93\linewidth]{resonance}
\caption{
Calculation of the expected resonance curve in the AC mode of excitation.
The probability of the $2 \rightarrow 1$ transition is plotted as a function of the driving frequency.
\label{resonance}
}
\end{figure}
\bigskip \begin{center} \photon \end{center} \bigskip
In the long term, increasing the precision of the transition frequencies can in principle be achieved by storing the neutrons in the quantum states
for long times compared with the flow-through arrangement.
One would use a trap consisting of an horizontal mirror and vertical walls, the excitation (vibration or magnetic) could be applied during a longer storage period.
The control of the environmental noises such as vibrations is an issue because it could provoke unwanted transition between quantum states,
but the lifetime of the quantum states in the trap could reach a few tens of seconds \cite{Pignol2009,Codau2012}.
The filling and emptying of such a trap with limited losses of neutrons remains an open problem.
Another interesting route for the future is the development of position sensitive detectors with micrometric resolution to ``see'' the quantum states.
We have mentioned an attempt to take pictures of the quantum states using a uranium-coated plastic nuclear track detector.
It achieves a good resolution of $2 \ \upmu {\rm m}$, but it lacks real-time readout.
More recently, pixel detectors based on CCD (charge coupling device) coated with $^6$Li or $^{10}$B have shown to offer both a good spatial resolution
and real-time readout capability \cite{Kawasaki2010}.
Pursuing in this way will certainly lead to clean measurements of the wavefunctions.
In conclusion, progress are foreseen both in the spectroscopy of the quantum bouncer
and in the measurement of the spatial features of the quantum states.
These will permit to probe new forces acting on the neutron and
to test the equivalence principle in a quantum context.
\newpage
\section{Conclusion}
It is fascinating that experiments at low energy with neutrons can address three big questions about the Universe:
\begin{enumerate}
\item Does the standard theory of primordial nucleosynthesis predict the correct amount of helium and deuterium?
\item Was the asymmetry between matter and antimatter generated during the electroweak phase transition?
\item Is Dark Energy a dynamical field interacting with matter?
\end{enumerate}
The measurement of the neutron lifetime helps to answer question 1.
Experiments will continue until the inconsistencies of the present data are resolved.
The answer to the second question is either yes or no.
By improving the measurement of the neutron electric dipole moment we could be able to reach a definitive answer in the next decade.
The third question is the main focus of this work.
We have looked in details at the strongly coupled Khoury-Weltman chameleon.
The present constraints are compiled in Fig. \ref{exclusion}.
We have identified two methods to probe the chameleon with neutrons.
The first is neutron interferometry. A pilot experiment performed in 2013 was described in this work.
The second is using the quantum states of bouncing neutrons.
Ultimately, it has the potential to explore a significant part of the parameter space.
Searching for Dark Energy in the lab is a new and rapidly developing field of research.
Less than ten years after the chameleon was proposed,
a major part of the parameter space is already excluded by experiments searching for deviations of the inverse square law of gravity at short distances,
neutron experiments, and very recently atom-interferometry.
Besides the chameleon, there are other theoretical ideas related to Dark Energy to be explored in the lab.
For sure there are other neutron experiments to be invented.
\begin{figure}
\centering
\includegraphics[width=1.\linewidth]{exclusion}
\caption{
Exclusion regions (95 \% C.L.) in the chameleon parameter space (Ratra-Peebles index $n$ and matter coupling $\beta$).
The blue zone is excluded by gravitational resonance spectroscopy of the quantum bouncer \cite{Jenke2014}.
The red dashed line corresponds to the sensitivity of the flow-though setup in GRANIT,
the red dotted line corresponds to the ultimate sensitivity.
The orange zone is excluded by the neutron-interferometry experiment \cite{Lemmel2015}.
The grey zone is excluded by the Eot-Wash experiment to search for a short range force \cite{Upadhye2012}.
The hatched region is excluded by the atom-interferometry experiment \cite{Hamilton2015}.
\label{exclusion}
}
\end{figure}
\clearpage
|
\section{Discussion}
\label{sec:discussion}
This paper presented a neural hypothesis for implementing translations of
temporal and 1-d spatial memory states so that future events can be quickly
anticipated without destroying the current state of memory. The hypothesis
assumes that time cells and place cells observed in the hippocampus represent
time or position as a result of a two-layer architecture that encodes and
inverts the Laplace transform of external input. It also assumes that
sequential translations to progressively more distant points in the future
occur within each cycle of theta oscillations. Neurophysiological
constraints were imposed as phenomenological rules rather than as emerging
from a detailed circuit model. Further, imposing scale-invariance and
coherence in translation across memory nodes resulted in Weber-Fechner
spacing for the representation of both the past (spacing of $s_n$ in the
memory nodes) and the future (the relationship between $\delta$ and
$\theta_o$). Apart from providing cognitive flexibility in accessing a
timeline of future predictions at any moment, the computational mechanism
described qualitative features of phase precession in the hippocampus and in
the ventral striatum. Additionally, we have also pointed out certain
distinctive features of the model that can be tested with existing technology.
\subsection{Computational Advantages}
The property of the $\T{}$ layer that different nodes represent the stimulus
values from various delays (past moments) is reminiscent of a shift register
(or delay-line or synfire chain). However, the two layer network encoding and
inverting the Laplace transform of stimulus has several significant
computational advantages over a shift register representation.
(i) In the current two-layer network, the spacing of $s$ values of the nodes
can be chosen freely. By choosing exponentially spaced $s$-values
(Weber-Fechner scaling) as in eq.~\ref{eq:thetas}, the $\T{}$ layer can
represent memory from exponentially long timescales compared to a shift
register with equal number of nodes, thus making it extremely
resource-conserving. Although information from longer timescales is more
coarse-grained, it turns out that this coarse-graining is optimal to
represent and predict long-range correlated signals \cite{ShanHowa13}.
(ii) The memory representation of this two layer network is naturally
scale-invariant (eq.~\ref{eq:old_big_T}). To construct a scale-invariant
representation from a shift register, the shift register would have to be
convolved with a scale-invariant coarse-graining function at each moment,
which would be computationally very expensive. Moreover, it turns out that any
network that can represent such scale-invariant memory can be identified with
linear combinations of multiple such two layer networks \cite{Shan15}.
(iii)
Because
translation can be trivially performed when we have access to the Laplace
domain, the two layer network enables translations by an amount
$\delta$ without sequentially visiting the intermediate states $<\delta$. This can be
done by directly changing the connection strengths locally
between the two layers as prescribed by diagonal $\mathbf{R}_{\delta}$ operator for any chosen
$\delta$.\footnote{In this paper we considered sequential translations of
various values of $\delta$, since the aim was to construct an
entire future timeline rather than to discontinuously jump to a
distant future state.} Consequently the physical time taken for the
translation can be decoupled from the magnitude of translation. One could
imagine a shift register performing a translation operation by an amount
$\delta$ either by shifting the values sequentially from one node to the next
for $\delta$ time steps or by establishing non-local connections between far
away nodes. The latter would make the computation very cumbersome because it
would require every node in the register to be connected to every other node
(since this should work for any $\delta$), which is in stark contrast with the
local connectivity required by our two layer network to perform any
translation.
Many previous neurobiological models of phase precession have been proposed
\cite{LismJens13,MehtEtal02,BurgEtal07,Hass12}, and many assume that
sequentially activated place cells firing within a theta cycle result from
direct connections between those cells \cite{JensLism96}, not unlike a synfire
chain. Although taking advantage of the Laplace domain in the two
layer network to perform translations is not the only possibility, it seems
to be computationally powerful compared to the obvious alternatives.
\subsection{Translations without theta oscillations}
Although this paper focused on sequential translation within a theta cycle,
translation may also be accomplished \emph{via} other neurophysiological
mechanisms. Sharp wave ripple (SRW) events last for about 100~ms and are
often accompanied by replay events--sequential firing of place cells
corresponding to locations different from the animal's current
location \cite{DaviEtal09,DragTone11,FostWils06,PfeiFost13,JadhEtal12}. Notably,
experimentalists have also observed preplay events during SWRs, sequential activation of
place cells that correspond to trajectories that have never been
previously traversed, as though the animal is planning a future path
\cite{DragTone11,OlafEtal15}.
Because untraversed trajectories could not have been used to learn and build
sequential associations between the place cells along the trajectory, the
preplay activity could potentially be a result of a translation operation on
the overall spatial memory representation.
Sometimes during navigation, a place cell corresponding to a distant goal
location gets activated \cite{PfeiFost13}, as though a finite distance
translation of the memory state has occurred. More interestingly, sometimes a
reverse-replay is observed in which place cells are activated in reverse order
spreading back from the present location \cite{FostWils06}. This is suggestive
of translation into the past (as if $\delta$ was negative), to implement a
memory search. In parallel, there is behavioral evidence from humans that
under some circumstances memory retrieval consists of a backward scan through
a temporal memory representation \cite{Hack80,Hock84,SingEtal15} (although
this is not neurally linked with SWRs). Mathematically, as long as the
appropriate connection strength changes prescribed by the $\mathbf{R}_{\delta}$ operator can
be specified, there is no reason translations with negative $\delta$ or
discontinuous shift in $\delta$ could not be accomplished in this framework.
Whether these computational mechanisms are reasonable in light of the
neurophysiology of sharp wave ripples is an open question.
\subsection{Multi-dimensional translation}
This paper focused on translations along one dimension. However it would be
useful to extend the formalism to multi-dimensional translations. When a rat
maneuvers through an open field rather than a linear track, phase precessing
2-d place cells are observed \cite{SkagEtal96}. Consider the case of an
animal approaching a junction along a maze where it has to either turn left or
right. Phase precessing cells in the hippocampus indeed predict the direction
the animal will choose in the future \cite{JohnRedi07}. In order to
generalize the formalism to 2-d translation, the nodes in the network model
must not be indexed only by $s$, which codes their distance from a landmark,
but also by the 2-d orientation along which distance is calculated. The
translation operation must then specify not just the distance, but also
the instantaneous direction as a function of the theta phase. Moreover,
if translations could be performed on multiple non-overlapping
trajectories simultaneously, multiple paths could be searched in
parallel, which would be very useful for efficient decision making.
\subsection{Neural representation of predictions}
The computational function of $\p{\delta}$ (eq.~\ref{eq:p}) is to represent an
ordered set of events predicted to occur in the future. Although we focused
on ventral striatum here because of the availability of phase precession data
from that structure, it is probable that many brain regions represent future
events as part of a circuit involving frontal cortex and basal ganglia, as
well as the hippocampus and striatum
\cite{SchuEtal97,FerbShap03,TanaEtal04,FeieEtal06,MainKepe09,TakaRoes11,YounShap11a}.
There is evidence that theta-like oscillations coordinates the activity in
many of these brain regions
\cite{JoneWils05,LansEtal09,vanWEtal10,FujiBuzs11}. For instance, 4~Hz
oscillations show phase coherence between the hippocampus, prefrontal cortex
and ventral tegmental area (VTA), a region that signals the presence of
unexpected rewards \cite{FujiBuzs11}. A great deal of experimental work has
focused on the brain's response to future rewards, and indeed the
phase-precessing cells in fig.~\ref{fig:MeerRedi} appear to be predicting the
location of the future reward. The model suggests that $\p{\delta}$ should
predict any future event, not just a reward. Indeed, neurons that appear to
code for predicted stimuli have been observed in the primate inferotemporal
cortex \cite{SakaMiya91} and prefrontal cortex \cite{RainEtal99}. Moreover,
theta phase coherence between prefrontal cortex and hippocampus are
essential for learning the temporal relationships between stimuli
\cite{BrinMill15}.
So, future predictions could be widely distributed throughout the brain.
\section{Mathematical model}
\label{sec:math}
In this section we start with a basic overview of the two layer memory model
and summarize the relevant details from previous work
\cite{ShanHowa12,ShanHowa13,HowaEtal14} to serve as a background. Following
that, we derive the equations that allow the memory nodes to be coherently
time-translated to various future moments in synchrony with the theta
oscillations. Finally we derive the predictions generated for various future
moments from the time-translated memory states.
\subsection{Theoretical background}
\label{ssec:mathback}
The memory model is implemented as a two-layer feedforward network
(fig.~\ref{fig:TILT}b) where the $\ti{}$ layer holds the Laplace transform of
the recent past and the $\T{}$ layer reconstructs a temporally fuzzy estimate
of past events \cite{ShanHowa12,ShanHowa13}. Let the stimulus at any time
$\tau$ be denoted as $\ensuremath{\mathbf{f}}{}(\tau)$. The nodes in the $\ti{}$ layer are leaky
integrators parametrized by their decay rate $s$, and are all independently
activated by the stimulus. The nodes are assumed to be arranged w.r.t.~their
$s$ values. The nodes in the $\T{}$ layer are in one to one correspondence
with the nodes in the $\ti{}$ layer and hence can also be parametrized by the
same $s$. The feedforward connections from the $\ti{}$ layer into the $\T{}$
layer are prescribed to satisfy certain mathematical properties which are
described below. The activity of the two layers is given by
\begin{eqnarray}
\frac{d}{d \tau} \ti{}(\tau,s) &=& -s \ti{}(\tau,s) + \ensuremath{\mathbf{f}}{}(\tau) \label{eq:teq} \\
\T{}(\tau,s) &=& [\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}} ] \, \ti{}(\tau,s) \label{eq:Teq}
\end{eqnarray}
By integrating eq.~\ref{eq:teq}, note that the $\ti{}$ layer encodes the
Laplace transform of the entire past of the stimulus function leading up to
the present. The $s$ values distributed over the $\ti{}$ layer represent
the (real) Laplace domain variable. The fixed connections between the $\ti{}$
layer and $\T{}$ layer denoted by the operator $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ (in eq.~\ref{eq:Teq}), is
constructed to reflect an approximation to inverse Laplace transform. In
effect, the Laplace transformed stimulus history which is distributed over
the $\ti{}$ layer nodes is inverted by $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ such that a fuzzy (or coarse
grained) estimate of the actual stimulus value from various
past moments is represented along the different $\T{}$ layer nodes.
More precisely, treating the $s$ values nodes as continuous, the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator can be succinctly expressed as
\begin{equation}
\T{}(\tau,s) = \frac{(-1)^k}{k!} s^{k+1} \ti{}^{(k)} (\tau,s) \equiv [\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}] \, \ti{}(\tau,s)
\label{eq:Lk}
\end{equation}
Here $\ti{}^{(k)}(\tau,s)$ corresponds to the $k$-th derivative of
$\ti{}(\tau,s)$ w.r.t.~$s$. It can be proven that $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator executes an
approximation to the inverse Laplace transformation and the approximation
grows more and more accurate for larger and larger values of $k$
\cite{Post30}. Further details of $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ depends on the $s$ values chosen for
the nodes \cite{ShanHowa13}, but these details are not relevant for this
paper as the $s$ values of neighboring nodes are assumed to be close enough
that the analytic expression for $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ given by eq.~\ref{eq:Lk} would be
accurate.
To emphasize the properties of this memory representation, consider the
stimulus $\ensuremath{\mathbf{f}}{}(\tau)$ to be a Dirac delta function at $\tau=0$. From
eq.~\ref{eq:teq}~and~\ref{eq:Lk}, the $\T{}$ layer activity following the
stimulus presentation ($\tau >0$) turns out to be
\begin{equation}
\T{} (\tau, s) = \frac{s}{k!} \left[s\tau \right]^k
e^{-\left[s\tau \right]}
\label{eq:old_big_T}
\end{equation}
Note that nodes with different $s$ values in the $\T{}$ layer peak in activity after different delays following the stimulus; hence the $\T{}$ layer nodes behave like {time cells}. In particular, a node with a given $s$ peaks in activity at a time $\tau=k/s$ following the stimulus. Moreover, viewing the activity of any node as a distribution around its appropriate peak-time ($k/s$), we see that the shape of this distribution is exactly the same for all nodes to the extent $\tau$ is rescaled to align the peaks of all the nodes. In other words, the activity of different nodes of the $\T{}$ layer represent a fuzzy estimate of the past information from different timescales and the fuzziness associated with them is directly proportional to the timescale they represent, while maintaining the exact same shape of fuzziness. For this reason, the $\T{}$ layer represents the past information in a {scale-invariant} fashion.
This two-layer memory architecture is also amenable to represent
one-dimensional spatial memory analogous to the representation of temporal
memory in the $\T{}$ layer \cite{HowaEtal14}. If the stimulus $\ensuremath{\mathbf{f}}{}$ is
interpreted as a landmark encountered at a particular location in a
one-dimensional spatial arena, then the $\ti{}$ layer nodes can be made to
represent the Laplace transform of the landmark treated as a spatial function
with respect to the current location. By modifying eq.~\ref{eq:teq} to
\begin{equation}
\frac{d}{d \tau} \ti{}(\tau,s) = v \left[ -s \ti{}(\tau,s) + \ensuremath{\mathbf{f}}{}(\tau) \right] ,
\label{eq:timespace}
\end{equation}
where $v$ is the velocity of motion, the temporal dependence of the $\ti{}$ layer activity can be converted to spatial dependence.\footnote{ Theoretically, the velocity here could be an animal's running velocity in the
lab maze or a mentally simulated human motion while playing video games.} By employing the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator on this modified $\ti{}$ layer activity (eq.~\ref{eq:timespace}), it is straightforward to construct a layer of nodes (analogous to $\T{}$) that exhibit peak activity at different distances from the landmark. Thus the two-layer memory architecture can be trivially extended to yield {place-cells} in one dimension.
In what follows, rather than referring to translation operations separately on spatial and temporal memory, we shall simply consider time-translations with an implicit understanding that all the results derived
can be trivially extended to 1-d spatial memory representations.
\subsection{Time-translating the Memory state}
\label{sec:newstuff}
The two-layer architecture naturally lends
itself for {time-translations} of the memory state in the $\T{}$ layer,
which we shall later exploit to construct a timeline of future predictions.
The basic idea is that if the current state of memory represented in the
$\T{}$ layer is used to anticipate the present (via some prediction
mechanism), then a time-translated state of $\T{}$ layer can be used to predict
events that will occur at a distant future via the same prediction
mechanism. Time-translation means to mimic the $\T{}$ layer
activity at a distant future based on its current state. Ideally translation
should be {non-destructive}, not overwriting the current activity in
the $\ti{}$ layer.
Let $\delta$ be the amount by which we intend to time-translate the state of
$\T{}$ layer. So, at any time $\tau$, the aim is to access
$\T{}(\tau+\delta,s)$ while still preserving the current $\ti{}$ layer
activity, $\ti{}(\tau,s)$. This is can be easily achieved because the $\ti{}$
layer represents the stimulus history in the Laplace domain. Noting that the
Laplace transform of a $\delta$-translated function is simply the product of
$e^{-s \delta}$ and the Laplace transform of the un-translated function, we see
that
\begin{eqnarray}
\ti{}(\tau+\delta,s) &=& e^{-s\delta} \ti{}(\tau,s) \label{eq:translate}
\end{eqnarray}
Now noting that $\T{}(\tau+\delta,s)$ can be obtained by employing the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$
operator on $\ti{}(\tau+\delta,s)$ analogous to eq.~\ref{eq:Lk}, we obtain the
$\delta$-translated $\T{}$ activity as
\begin{eqnarray}
\T{\delta}(\tau,s) \equiv \T{}(\tau+\delta,s) &=& [ \ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}] \, \ti{}(\tau+\delta ,s) \nonumber \\
&=& \left[\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}} \cdot \mathbf{R}_{\delta} \right] \, \ti{}(\tau,s)
\label{eq:Rd}
\end{eqnarray}
where $\mathbf{R}_{\delta}$ is just a diagonal operator whose rows and columns are indexed by
$s$ and the diagonal entries are $e^{-s \delta}$. The $\delta$-translated
activity of the $\T{}$ layer is now subscripted by $\delta$ as $\T{\delta}$ so
as to distinguish it from the un-translated $\T{}$ layer activity given by
eq.~\ref{eq:Lk} without a subscript. In this notation the un-translated state
$\T{}(\tau,s)$ from eq.~\ref{eq:Lk} can be expressed as $\T{0}(\tau,s)$. The
time-translated $\T{}$ activity can be obtained from the current $\ti{}$ layer
activity if the connection weights between the two layers given by $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ is
modulated by $\mathbf{R}_{\delta}$. This computational mechanism of time-translation can be implemented as a neural mechanism in the brain, by imposing certain phenomenological constraints and physical principles.
{\bf \emph{Observation 1:}}
Anatomically, along the dorsoventral axis of the hippocampus, the width of
place fields systematically increases from the dorsal end to the
ventral end \cite{JungEtal94,KjelEtal08}. Fig.~\ref{fig:theta} schematically
illustrates this observation by identifying the $s$-axis of the two-layer
memory architecture with the dorso-ventral axis of the hippocampus,
such that the scales represented by the nodes are monotonically arranged. Let
there be $N+1$ nodes with monotonically decreasing $s$ values given by $s_o$,
$s_1$, \ldots $s_N$.
\begin{figure}
\includegraphics[width=0.4\textwidth]{newfigs/TravelingWave.png}
\caption{Traveling theta wave along the $s$ axis.
The x-axis is real time. Each point along the dorsoventral axis corresponds
to a different value of $s_n$. The curvy blue lines show the theta
oscillation for several different values of $s$. Lines~1~and~2
connect the positions where the local phases $\theta_s$ are 0 and
$\pi$ respectively.
\label{fig:theta}}
\end{figure}
{\bf \emph{Observation 2:}}
The phase of the theta oscillations along the axis is non-uniform,
representing a traveling wave from the dorsal to ventral part of the hippocampus
with a net phase shift of $\pi$ \cite{LubeSiap09,PateEtal12}.
The oscillations in fig.~\ref{fig:theta} symbolize the local
field potentials at different locations of the $s$-axis. The local phase of
the oscillation at any position on the $s$-axis is denoted by $\theta_s$,
which ranges from $-\pi$ to $+\pi$ by convention. However, as a reference we
denote the phase at the top (dorsal) end as $\theta_o$ ranging from $0$ to $2
\pi$, with the understanding that the range $(\pi, 2 \pi)$ is mapped on to
$(-\pi, 0)$. The $x$-axis in fig.~\ref{fig:theta} is time within a theta
oscillation labeled by the phase $\theta_o$.
In this convention, the value of $\theta_s$ discontinuously jumps from $+\pi$
to $-\pi$ as we move from one cycle of oscillation to the next. In
fig.~\ref{fig:theta}, the diagonal (solid-red) line labeled `2' denotes all
the points where this discontinuous jump happens. The diagonal (dashed) line
labeled `1' denotes all the points where $\theta_s=0$. It is straightforward
to infer the relationship between the phase at any two values of $s$. Taking
the nodes to be uniformly spaced anatomically, the local phase $\theta_s$ of the $n$-th
node is related to $\theta_o$ (for $0<\theta_s<\pi$) by\footnote{ Since the $s$
values of the nodes are monotonically arranged, we can interchangeably use $s$
or $n$ as subscritpts to $\theta$. }
\begin{equation}
\theta_s/\pi= \theta_o/\pi - n/N .
\label{eq:thetasnophi}
\end{equation}
{\bf \emph{Observation 3:}} Synaptic weights in the hippocampus are modulated
periodically in synchrony with the phase of theta
oscillation \cite{WyblEtal00,SchaEtal08}. Based on this observation, we impose
the constraint that the connection strengths between the $\ti{}$ and $\T{}$
layers at a particular value of $s$ depend only on the local phase of the
theta oscillations. Thus the diagonal entries in the $\mathbf{R}_{\delta}$ operator should
only depend on $\theta_s$. We take these entries to be of the form $\exp{
(-\Phi_s (\theta_s)) } $, where $\Phi_s$ is any continuous function of
$\theta_s \in (-\pi, +\pi)$. Heuristically, at any moment
within a theta cycle, a $\T{}$ node with a given $s$ value will be
roughly translated by an amount $\delta = \Phi_s(\theta_s)/s$.
{\bf \emph{Principle 1:}} \emph{Preserve Scale-Invariance}
Scale-invariance is an extremely adaptive property for a memory to have; in
many cases biological memories seem to exhibit scale-invariance
\cite{BalsGall09}.
As the untranslated $\T{}$ layer activity already exhibits scale-invariance, we impose the constraint
that the time-translated states of $\T{}$ should also exhibit scale-invariance. This consideration requires the behavior of every node to follow the same pattern with respect to their local theta phase. This amounts
to choosing the functions $\Phi_s$ to be the same for all $s$, which we shall refer to as $\Phi$.
{\bf \emph{Principle 2:}} \emph{Coherence in translation}
Since the time-translated memory state is going to be used to make predictions
for various moments in the distant future, it would be preferable if all the nodes
are time-translated by the same amount at any moment within a theta
cycle. If not, different nodes would contribute to predictions for different
future moments leading to noise in the prediction. However, such a
requirement of global coherence cannot be imposed consistently along with the
principle~1 of preserving scale-invariance.\footnote{This is easily seen by
noting that each node will have a maximum translation inversely
proportional to its $s$-value to satisfy principle~1.} But in the light of
prior work \cite{HassEtal02,Hass12} which suggest that retrieval of
memory or prediction happens only in one half of the theta cycle,\footnote{
This hypothesis follows from the observation that while both synaptic
transmission and synaptic plasticity are modulated by theta phase,
they are out of phase with one another. That is, while certain synapses are
learning, they are not communicating information and vice versa. This
led to the hypothesis that the phases where plasticity is
optimal are specialized for encoding whereas the phases where transmission is
optimal are specialized for retrieval. }
we
impose the requirement of coherence only to those nodes that are all in the
positive half of the cycle at any moment. That is, $\delta= \Phi(\theta_s)/s$
is a constant along any vertical line in the region bounded between the
diagonal lines~1~and~2 shown in fig.~\ref{fig:theta}. Hence for all nodes
with $0<\theta_s< \pi$, we require
\begin{equation}
\Delta \left(\Phi\left(\theta_s\right) / s \right) = \Delta \left(\Phi\left(\theta_o - \pi n/N\right) / s_n \right) = 0.
\label{eq:deltathetas}
\end{equation}
For coherence as expressed in eq.~\ref{eq:deltathetas} to hold at all values
of $\theta_o$ between 0 and $2 \pi$, $\Phi(\theta_s)$ must be an exponential
function so that $\theta_o$ can be
functionally decoupled from $n$; consequently $s_n$ should also have an
exponential dependence on $n$. So the general solution to
eq.~\ref{eq:deltathetas} when $0<\theta_s<\pi$ can be written as
\begin{eqnarray}
\Phi(\theta_s)&=& \ensuremath{\Phi_o} \exp{[b \theta_s]} \label{eq:thetas} \\
s_n &=& s_o (1+c)^{-n} \label{eq:thetasn}
\end{eqnarray}
where $c$ is a positive number. In this paper, we shall take $c \ll 1$, so
that the analytic approximation for the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator given in terms of the
$k$-th derivative along the $s$ axis in eq.~\ref{eq:Lk} is valid.
Thus the requirement of coherence in time-translation implies that the $s$
values of the nodes---the timescales represented by the nodes---are spaced out
exponentially, which can be referred to as a Weber-Fechner scale, a commonly
used terminology in cognitive science. Remarkably, this result strongly
resonates with a requirement of the exact same scaling when the predictive
information contained in the memory system is maximized in response to
long-range correlated signals \cite{ShanHowa13}. This feature allows this
memory system to represent scale-invariantly coarse grained past information
from timescales exponentially related to the number of nodes.
The maximum value attained by the function $\Phi( \theta_s)$ is at $\theta_s=\pi$, and the maximum value is $\ensuremath{\Phi_{\textnormal{max}}} = \ensuremath{\Phi_o} \exp{[b \pi] }$, such that $\ensuremath{\Phi_{\textnormal{max}}}/\ensuremath{\Phi_o} = s_o/s_N$ and $b=(1/\pi) \log{(\ensuremath{\Phi_{\textnormal{max}}}/\ensuremath{\Phi_o})}$.
To ensure continuity around $\theta_s=0$, we take the eq.~\ref{eq:thetas} to
hold true even for $\theta_s \in (-\pi,0)$. However, since notationally
$\theta_s$ makes a jump from $+\pi$ to $-\pi$, $\Phi(\theta_s)$ would exhibit
a discontinuity at the diagonal line~2 in fig.~\ref{fig:theta} from $\ensuremath{\Phi_{\textnormal{max}}}$
(corresponding to $\theta_s=\pi $) to $\ensuremath{\Phi_{\textnormal{min}}}=\ensuremath{\Phi_o}^2/\ensuremath{\Phi_{\textnormal{max}}}$
(corresponding to $\theta_s=-\pi$).
Given these considerations, at any instant within a theta cycle, referenced by
the phase $\theta_o$, the amount $\delta$ by which the memory state is
time-translated can be derived from eq.~\ref{eq:thetasnophi} and
\ref{eq:thetas} as
\begin{equation}
\delta(\theta_o) = (\ensuremath{\Phi_o}/s_o) \exp{[b \theta_o]} .
\label{eq:deltatheta}
\end{equation}
Analogous to having the past represented on a Weber-Fechner scale, the
translation distance $\delta$ into the future also falls on a Weber-Fechner
scale as the theta phase is swept from 0 to $2 \pi$. In other words, the
amount of time spent within a theta cycle for larger translations is
exponentially smaller.
To emphasize the properties of the time-translated $\T{}$ state, consider the
stimulus to be a Dirac delta function at $\tau=0$. From eq.~\ref{eq:Rd}, we
can express the $\T{}$ layer activity analogous to eq.~\ref{eq:old_big_T}.
\begin{equation}
\T{\delta} (\tau, s) \simeq \frac{s}{k!} \left[s\tau+\Phi\left(\theta_s\right)\right]^k
e^{-\left[s\tau+ \Phi\left(\theta_s\right)\right]}
\label{eq:bigT}
\end{equation}
Notice that eqs.~\ref{eq:thetasnophi} and \ref{eq:deltatheta} specify a unique
relationship between $\delta$ and $\theta_s$ for any given $s$. The
r.h.s.~above is expressed in terms of $\theta_s$ rather than $\delta$ so as to shed
light on the phenomenon of {phase precession}.
Since $\T{\delta}(\tau,s)$ depends on both $\tau$ and $\theta_s$ only via the
sum $[ s\tau+ \Phi\left(\theta_s\right) ] $, a given node will show identical
activity for various combinations of $\tau$ and $\theta_s$.\footnote{While
representing timescales much larger than the period of a theta cycle, $\tau$
can essentially be treated as a constant within a single cycle. In other
words, $\theta_s$ and $\tau$ in eq.~\ref{eq:Rd} can be treated as
independent, although in reality the phases evolve in real time.}
For instance, a node would achieve its peak activity when $\tau$ is
significantly smaller than its timescale $(k/s)$ only when $\Phi(\theta_s)$ is
large---meaning $\theta_s \simeq +\pi$. And as $\tau$ increases towards the
timescale of the node, the peak activity gradually shifts to earlier phases
all the way to $\theta_s \simeq -\pi$. An important consequence of imposing
principle~1 is that the relationship between $\theta_s$ and $\tau$ on any
iso-activity contour is scale-invariant. That is, every node behaves similarly when
$\tau$ is rescaled by the timescale of the node. We shall
further pursue the analogy of this phenomenon of phase precession with
neurophysiological findings in the next section (fig.~\ref{fig:ph-pr}).
\subsection{Timeline of Future Prediction}
At any moment, $\T{\delta}$ (eq.~\ref{eq:bigT}) can be used to predict the
stimuli expected at a future moment. Consequently, as $\delta$
is swept through within a theta cycle, a timeline of future predictions can be
simulated in an orderly fashion, such that predictions for closer events occur
at earlier phases (smaller $\theta_o$) and predictions of distant events occur
at later phases. In order to predict from a time-translated state
$\T{\delta}$, we need a prediction mechanism.
For our purposes, we consider here a very simple form of learning and
prediction, \emph{Hebbian association}. In this view, an event is learned (or
an association formed in long term memory) by increasing the connection
strengths between the neurons representing the currently-experienced
stimulus and the neurons representing the recent past events ($\T{0}$).
Because the $\T{}$ layer activity contains temporal information about the preceding stimuli, simple associations between $\T{0}$ and the current stimulus
are sufficient to encode and express well-timed predictions \cite{ShanHowa12}.
In particular, the term Hebbian implies that the change in each connection
strength is proportional to the product of pre-synaptic activity---in this
case the activity of the corresponding node in the $\T{}$ layer---and
post-synaptic activity corresponding to the current stimulus.
Given that the associations are learned in this way, we define the prediction of a
particular stimulus to be the scalar product of its association strengths with
the current state of $\T{}$. In this way, the scalar product of association
strengths and a translated state $\T{\delta}$ can be understood as the
future prediction of that stimulus.
Consider the thought experiment where a
conditioned stimulus \textsc{cs} is consistently followed by another
stimulus, \textsc{a} or \textsc{b}, after a time $\tau_o$. Later when
\textsc{cs} is repeated (at a time $\tau=0$), the subsequent activity in the
$\T{}$ nodes can be used to generate predictions for the future occurrence of
\textsc{a} or \textsc{b}. The connections to the node corresponding to
\textsc{a} will be incremented by the state of $\T{0}$ when \textsc{a} is
presented; the connections to the node corresponding to
\textsc{b} will be incremented by the state of $\T{0}$ when \textsc{b} is
presented. In the context of Hebbian learning, the prediction for the
stimulus at a future time as a function of $\tau$ and $\tau_o$ is obtained as
the sum of $\T{\delta}$ activity of each node multiplied by the learned
association strength ($\T{0}$):
\begin{equation}
\p{\delta}(\tau,\tau_o) = \sum_{n=\ell}^{N} \T{\delta}\left(\tau,s_n\right)
\ \T{0}\left(\tau_o,s_n\right)/ s_{n}^{w}.
\label{eq:pthetatau}
\end{equation}
The factor $s_n^w$ (for any $w$) allows for differential association
strengths for the different $s$ nodes, while still preserving the scale
invariance property. Since $\delta$ and $\theta_o$ are monotonically related
(eq.~\ref{eq:deltatheta}), the prediction $\p{\delta}$ for various future moments
happens at various phases of a theta cycle.
Recall that all the nodes in the $\T{}$ layer are coherently time-translated
only in the positive half of the theta cycle. Hence for computing future
predictions based on a time-translated state $\T{\delta}$, only coherent nodes
should contribute. In fig.~\ref{fig:theta}, the region to the right of
diagonal line~2 does not contribute to the prediction. The lower
limit $\ell$ in the summation over the nodes given in eq.~\ref{eq:pthetatau}
is the position of the diagonal line~2 in fig.~\ref{fig:theta} marking the
position of discontinuity where $\theta_s$ jumps from $+\pi$ to $-\pi$.
In the limit when $c \rightarrow 0$, the $s$ values of neighboring nodes are
very close and the summation can be approximated by an integral. Defining
$x = s \tau_o$ and $y= \tau/\tau_o$ and $v= \delta/\tau_o$, the
above summation can be rewritten as
\begin{equation}
\p{\delta}(\tau,\tau_o) \simeq \frac{\tau_o^{w-2}}{k!^2} \int_{x_{min}}^{x_u} x^{2k+1-w} (y +v)^k e^{-x(1+y+v)} \, dx
\end{equation}
Here $x_{min} = s_N \tau_o$, and $x_u = s_o \tau_o$ for $0< \theta_o< \pi$ and $x_u= \Phi_{max} \tau_o/\delta$ for $\pi< \theta_o <2 \pi$. The integral can be evaluated in terms of lower incomplete gamma functions to be
\begin{eqnarray}
&\p{\delta}&(\tau,\tau_o) \simeq \frac{\tau_o^{w-2}}{k!^2}
\frac{\left[ (\tau+\delta )/\tau_o \right]^k}{[1+ (\tau+\delta)/\tau_o]^C} \,
\times \nonumber \\
& & \left( \Gamma \left[C, (\tau_o+\tau +\delta) U \right] - \Gamma \left[C, (\tau_o+\tau +\delta) s_N \right] \right),
\label{eq:p}
\end{eqnarray}
where $C=2k+2-w$ and $\Gamma[.,.]$ is the lower incomplete gamma function.
For $\theta_o<\pi$ (i.e., when $\delta < \ensuremath{\Phi_{\textnormal{max}}}/s_o$), $U=s_o$ and for
$\theta_o>\pi$ (i.e., when $\delta > \ensuremath{\Phi_{\textnormal{max}}}/s_o$), $U=\ensuremath{\Phi_{\textnormal{max}}}/\delta$.
\begin{figure}
\includegraphics[width=0.5\textwidth]{newfigs/Deltas_1.png}
\caption{Future timeline. Eq.~\ref{eq:p} is plotted as a function of $\delta$. During training, the \textsc{cs} was presented at $\tau_o=3$ before \textsc{a} and $\tau_o=7$ before \textsc{b}. Left: Immediately after presentation of the \textsc{cs}, the predictions for \textsc{a} and \textsc{b} are ordered on the $\delta$
axis. Note that the prediction for \textsc{b} approximates a rescaled
version of that for \textsc{a}. Right: The prediction for \textsc{b}
is shown for varying times after presentation of \textsc{cs}. With the
passage of time, the prediction of \textsc{b} becomes stronger and
more imminent. In this figure,
$\ensuremath{\Phi_{\textnormal{max}}}=10$, $\ensuremath{\Phi_o}=1$, $k=10$, $s_o=10$, $s_N=1$, and $w=1$.
\label{fig:futurepred}
}
\end{figure}
Figure~\ref{fig:futurepred} provides a graphical representation of some key
properties of eq.~\ref{eq:p}. The figure assumes that the \textsc{cs} is
followed by \textsc{a} after $\tau_o=3$ and followed by \textsc{b} after
$\tau_o=7$. The left panel shows the predictions for both \textsc{a} and
\textsc{b} as a function of $\delta$ immediately after presentation of
\textsc{cs}. The prediction for \textsc{a} appears at smaller
$\delta$ and with a higher peak than the prediction for
\textsc{b}. The value of $w$ affects the relative sizes of the peaks. The
right panel shows how the prediction for \textsc{b} changes with the passage
of time after presentation of the \textsc{cs}. As $\tau$ increases from zero
and the \textsc{cs} recedes into the past, the prediction of \textsc{b} peaks
at smaller values of $\delta$, corresponding to more imminent future times.
In particular when $\tau_o$ is much smaller than the largest (and larger than
the smallest) timescale represented by the nodes, then the
shape of $\p{\delta}$ remains the same when $\delta$ and $\tau$ are rescaled by
$\tau_o$. Under these conditions, the timeline of future predictions generated
by $\p{\delta}$ is scale-invariant.
Since $\delta$ is in one-to-one relationship with $\theta_o$, as a predicted
stimulus becomes more imminent, the activity corresponding to that predicted
stimulus should peak at earlier and earlier phases. Hence a timeline of future
predictions can be constructed from $\p{\delta}$ as the phase $\theta_o$ is
swept from $0$ to $2 \pi$. Moreover the cells representing $\p{\delta}$ should
show phase precession with respect to $\theta_o$. Unlike cells representing
$\T{\delta}$, which depend directly on their local theta phase, $\theta_s$,
the phase precession of cells representing $\p{\delta}$ should depend
on the reference phase $\theta_o$ at the dorsal end of the $s$-axis.
We shall further connect this neurophysiology in the next section
(fig.~\ref{fig:MeerRedi}).
\section{Comparisons with Neurophysiology}
\label{sec:nbio}
The mathematical development focused on two entities $\T{\delta}$ and $\p{\delta}$ that change their value based on the theta phase (eqs.~\ref{eq:bigT}~and~\ref{eq:p}). In order to compare these to neurophysiology, we need to have some hypothesis linking them to the activity of neurons from specific brain regions. We emphasize that although the development in the preceding section was done
with respect to time, all of the results generalize to one-dimensional position as well (eq.~\ref{eq:timespace}, \cite{HowaEtal14}). The overwhelming majority of evidence for phase precession comes from studies of place cells (but see \cite{PastEtal08}). Here we compare the properties of $\T{\delta}$ to phase precession in hippocampal neurons and the properties of $\p{\delta}$ to a study showing phase precession in
ventral striatum \cite{MeerEtal11}.\footnote{This is not meant to preclude the possibility that $\p{\delta}$ could be computed at other parts of the brain as well.}
Due to various analytic approximations, the activity of nodes in the $\T{}$
layer as well as the activity of the nodes representing future prediction
(eqs.~\ref{eq:bigT}~and~\ref{eq:p}) are expressed as smooth functions of time
and theta phase. However, neurophysiologically, discrete spikes (action
potentials) are observed. In order to facilitate comparison
of the model to neurophysiology, we adopt a simple stochastic spike-generating
method. In this simplistic approach, the activity of the nodes given by
eqs.~\ref{eq:bigT}~and~\ref{eq:p} are taken to be proportional to the
instantaneous probability for generating a spike. The probability of
generating a spike at any instant is taken to be the instantaneous activity
divided by the maximum activity achieved by the node if the activity is
greater than 60\% of the maximum activity. In addition, we add spontaneous
stochastic spikes at any moment with a probability of 0.05. For all of the
figures in this section, the parameters of the model are set as $k=10$,
$\ensuremath{\Phi_{\textnormal{max}}}=10$, $w=2$, $\Phi_o = 1$, $s_N=1$, $s_o=10$.
This relatively coarse level of realism in spike generation from the analytic
expressions is probably appropriate to the resolution of the experimental
data. There are some experimental challenges associated with exactly
evaluating the model. First, theta phase has to be estimated from a noisy
signal. Second, phase precession results are typically shown as averaged
across many trials. It is not necessarily the case that the average is
representative of an individual trial (although this is the case at least for
phase-precessing cells in medial entorhinal cortex \cite{ReifEtal12}).
Finally, the overwhelming majority of phase precession experiments utilize
extracellular methods, which cannot perfectly identify spikes from individual
neurons.
\subsection{Hippocampal phase precession}
\begin{figure}
\begin{tabular}{lclc}
\textbf{a} && \textbf{b}\\
&
\includegraphics[height=0.12\textheight]{newfigs/MehtaEtal02}
&&
\includegraphics[height=0.12\textheight]{newfigs/Timecell}
\end{tabular}
\caption{
\textbf{a.} Neurophysiological data showing phase precession. Each spike
fired by a place cell is
shown as a function of its position along a linear track (x-axis) and the
phase of local theta (y-axis). After Mehta, et al., 2002.
\textbf{b.}
Simulated spikes from a node in the $\T{}$ layer described by
eq.~\ref{eq:bigT} as a function of $\tau$ and local phase $\theta_s$.
The curvature is a consequence of eq.~\ref{eq:thetas}.
See text for details.
} \label{fig:ph-pr}
\end{figure}
\nocite{MehtEtal02}
It is clear from eq.~\ref{eq:bigT} that the activity of nodes in the $\T{}$
layer depends on both $\theta_s$ and $\tau$. Figure~\ref{fig:ph-pr} shows
phase precession data from a representative cell (Fig.~\ref{fig:ph-pr}a,
\cite{MehtEtal02}) and spikes generated from eq.~\ref{eq:bigT}
(Fig.~\ref{fig:ph-pr}b). The model generates a characteristic curvature for
phase precession, a consequence of the exponential form of the function
$\Phi$ (eq.~\ref{eq:thetas}). The example cell chosen in fig.~\ref{fig:ph-pr}
shows roughly the same form of curvature as that generated by the model.
While it should be noted that there is some variability across cells, careful
analyses have led computational neuroscientists to conclude that the canonical
form of phase precession resembles this representative cell. For instance, a
detailed study of hundreds of phase-precessing neurons \cite{YamaEtal02}
constructed averaged phase-precession plots using a variety of
methods and found a distinct curvature that qualitatively resembles
this neuron.
Because of the analogy between
time and one-dimensional position (eq.~\ref{eq:timespace}), the model yields
the same pattern of phase precession for time cells and place cells.
\begin{figure}
\includegraphics[width=0.9\columnwidth]{newfigs/Place-new}
\caption{ Place cells along the dorsoventral axis of the hippocampus have
place fields that increase in size. \textbf{a.} The three
panels show the activity of place cells recorded at the dorsal, intermediate
and ventral segments of the hippocampus, when a rat runs along an 18~m track.
After Kjelstrup, et al., (2008). Each spike the cell fired is shown as a
function of position and the \emph{local} theta phase at the cell's location
when it fires (recall that theta phase is not constant across the dorsoventral
axis). Regardless of the width of the place field, neurons at all locations
along the dorsoventral axis phase precess through the same range of {local}
theta phases. \textbf{b.}
According to the model, phase precession extends over the same
range of values of local theta $\theta_s$ regardless of
the value of $s$, which sets the scale for a particular
node. As a consequence, cells with different values of
$s$ show time/place fields of different size but phase
precess over the same range of local theta. For the three
figures, $s$ values of the nodes are set to $.1$, $.22$,
and $.7$ respectively, and they are assumed to respond to
landmarks at location $4$, $11$, and $3$ meters
respectively from one end of the track.
\label{fig:Kjel}
}
\end{figure}
The $\T{}$ layer activity represented in fig.~\ref{fig:ph-pr}a is scale-invariant; note that the $x$-axis is expressed in units of the scale of the node ($k/s$).
It is known that the spatial scale of place fields changes systematically
along the dorsoventral axis of the hippocampus. Place cells in the dorsal
hippocampus have place fields of the order of a few centimeters whereas place
cells at the ventral end have place fields as large as a few meters
(fig.~\ref{fig:Kjel}a) \cite{JungEtal94,KjelEtal08}.
However, all of them show the same pattern of precession with respect to
their local theta phase---the phase measured at the same electrode that
records a given place cell (fig.~\ref{fig:Kjel}). Recall that at any given
moment, the local phase of theta oscillation depends on the position along the
dorsoventral axis \cite{LubeSiap09,PateEtal12}, denoted as the $s$-axis in the
model.
Figure~\ref{fig:Kjel}a shows the activity of three different place cells in
an experiment where rats ran down a long track that extended through open
doors connecting three testing rooms \cite{KjelEtal08}. The landmarks
controlling a particular place cell's firing may have been at a variety of
locations along the track. Accordingly, fig.~\ref{fig:Kjel}b shows the
activity of cells generated from the model with different values of $s$ and
with landmarks at various locations along the track (described in the
caption).
From fig.~\ref{fig:Kjel} it can be qualitatively noted that phase precession
of different cells only depends on the local theta phase and is unaffected by
the spatial scale of firing. This observation is perfectly consistent with the
model.
\subsection{Prediction of distant rewards \emph{via} phase precession in the
ventral striatum}
\begin{figure}
\begin{tabular}{lclc}
\textbf{a} && \textbf{b}\\
& \includegraphics[width=0.45\columnwidth]{newfigs/VandermeerData}
&& \includegraphics[width=0.45\columnwidth]{newfigs/Vandermeer-model}
\end{tabular}
\caption{
\textbf{a.} A representative ramping cell in the ventral striatum.
On each trial the animal started the maze at S, made a series
of turns (T1, T2, etc) and received reward at F1 on 75 percent
of trials. The total distance between S and F1 is on the order
of a few meters. Position along the track is represented
linearly on the x-axis for convenience. In the top panel, the
spikes are shown as a function of theta phase at the dorsal
hippocampus and position. The bottom panel shows the firing
rate as a function of position, which is seen to gradually
ramp up towards the reward location. \textbf{b.} The
activity of prediction node generated by the model is plotted
w.r.t. the reference phase $\theta_o$ and position in the top
panel, and the the average activity within a theta cycle is
plotted against position in the bottom panel.
\label{fig:MeerRedi}
}
\end{figure}
We compare the future predictions generated by the model (eq.~\ref{eq:p}) to
an experiment that recorded simultaneously from the hippocampus and nucleus
accumbens, a reward-related structure within the ventral striatum
\cite{MeerEtal11}. Here the rat's task was to learn to make several turns in
sequence on a maze to reach two locations where reward was available.
Striatal neurons fired over long stretches of the maze, gradually ramping up
their firing as a function of distance along the path and terminating at the
reward locations (bottom fig.~\ref{fig:MeerRedi}a). Many striatal neurons
showed robust phase precession relative to the theta phase at the dorsal
hippocampus (top fig.~\ref{fig:MeerRedi}a). Remarkably, the phase of
oscillation in the hippocampus controlled firing in the ventral striatum to a
greater extent than the phase recorded from within the ventral striatum.
On trials where there was not a reward at the expected location (F1), there was
another ramp up to the secondary reward location (F2), accompanied again by
phase precession (not shown in fig.~\ref{fig:MeerRedi}a).
This experiment corresponds reasonably well to the conditions assumed in the
derivation of eq.~\ref{eq:p}. In this analogy, the start of the trial (start
location S) plays the role of the \textsc{cs} and the reward
plays the role of the predicted
stimulus. However, there is a discrepancy between the methods and the
assumptions of the derivation. The ramping cell (fig.~\ref{fig:MeerRedi}a)
abruptly terminates after the reward is consumed, whereas
eq.~\ref{eq:p} would gradually decay back towards zero. This is
because of the way the experiment was set up--there were never two
rewards presented consecutively. As a consequence, having just
received a reward strongly predicts that there will not be a reward in
the next few moments. In light of this consideration, we force the
prediction generated in eq.~\ref{eq:p} to be zero beyond the reward
location and let the firing be purely stochastic.
The top panel of fig.~\ref{fig:MeerRedi}b shows the spikes generated by model
prediction cells with respect to the reference theta phase $\theta_o$, and the
bottom panel shows the ramping activity computed as the average firing
activity within a complete theta cycle around any moment.
The model correctly captures the qualitative pattern observed in the data.
According to the model, the reward starts being predicted at the beginning of the
track. Initially, the reward is far in the future, corresponding to a large
value of $\delta$. As the animal approaches the location of the reward, the
reward moves closer to the present along the $\delta$ axis, reaching zero near
the reward location. The ramping activity is a consequence of the
exponential mapping between $\delta$ and $\theta_o$ in eq.~\ref{eq:thetas}.
Since the proportion of the theta cycle devoted to large values of $\delta$ is
small, the firing rate averaged across all phases will be small, leading to an
increase in activity closer to the reward.
\subsection{Testable properties of the mathematical model}
\begin{figure}
\includegraphics[width=0.8\columnwidth]{newfigs/Phase-entry}
\caption{Changing $\tau_o$ affects the phase at which prediction cells start
firing. At early times, the magnitude of translation required to
predict the $\tau_o=3$ outcome is smaller than that required to
predict the $\tau_o=7$ outcome. Consequently, the cell begins to fire
at a larger $\theta_o$ for $\tau_o=7$. Parameter values are the same
as the other figures as given in the beginning of this section, except
for clarity the background probability of spiking has been set to
zero. \label{fig:phaseentry}}
\end{figure}
Although the model aligns reasonably well with known properties of theta phase
precession, there are a number of features of the model that have, to our
knowledge, not yet been evaluated. At a coarse level, the correspondence
between time and one-dimensional space implies that time cells should exhibit
phase precession with the same properties as place cells. While phase
precession has been extensively observed and characterized in hippocampal
place cells, there is much less evidence for phase precession in hippocampal
time cells (but see
\cite{PastEtal08}).
According to the model, the pattern of phase precession is related to the
distribution of $s$ values represented along the dorsoventral axis. While it
is known that a range of spatial scales are observed along the dorsoventral
axis, their actual distribution is not known. The Weber-Fechner scale of
eq.~\ref{eq:thetas} is a strong prediction of the framework developed here.
Moreover, since $\ensuremath{\Phi_{\textnormal{max}}}/\ensuremath{\Phi_o} = s_o/s_N$, the ratio of the largest to
smallest scales represented in the hippocampus places constraints on the form
of phase precession. The larger this ratio, the larger will be the value of
$b$ in eq.~\ref{eq:thetas}, and the curvature in the phase precession plots
(as in fig.~\ref{fig:ph-pr}) will only emerge at larger values of the local
phase $\theta_s$. Neurophysiological observation of this ratio could help
evaluate the model.
The form of $\p{\delta}$ (eq.~\ref{eq:p}) leads to several distinctive
features in the pattern of phase precession of the nodes representing future
prediction. It should be possible to observe phase precession for cells that
are predicting any stimulus, not just a reward. In addition, the model's
assumption that a timeline of future predictions is aligned with global theta
phase
has interesting measurable consequences. Let's reconsider the thought
experiment from the previous section (fig.~\ref{fig:futurepred}), where a
stimulus predicts an outcome after a delay $\tau_o$. Immediately after the
stimulus is presented, the value of $\delta$ at which the prediction peaks is
monotonically related to $\tau_o$. Since $\delta$ is monotonically related to
the reference phase $\theta_o$, the prediction cells will begin to fire at later
phases when $\tau_o$ is large, and as time passes, they will fire at earlier
and earlier phases all the way untill $\theta_o=0$. In other other words, the
entry-phase (at which the firing activity begins) should depend on $\tau_o$, the
prediction timescale. This is illustrated in fig.~\ref{fig:phaseentry} with
$\tau_o=3$ and $\tau_o=7$, superimposed on the same graph to make visual
comparison easy. The magnitude of the peak activity would in general depend on
the value of $\tau_o$ except when $w=2$ (as assumed here for visual clarity).
Experimentally manipulating the reward times and studying the phase
precession of prediction cells could help test this feature.
\section{Introduction}
The brain encodes externally observed stimuli in real time and represents
information about the current spatial location and temporal history of recent
events as activity distributed over neural networks. Although we are
physically localized in space and time, it is often useful for us to make
decisions based on non-local events, by anticipating events to occur at
distant future and remote locations. Clearly, a flexible access to the
current state of spatio-temporal memory is crucial for the brain to
successfully anticipate events that might occur in the immediate next moment.
In order to anticipate events that might occur in the future after a given
time or at a given distance from the current location, the brain needs to
simulate how the current state of spatio-temporal memory representation will
have changed after waiting for a given amount of time or after moving through
a given amount of distance. In this paper, we propose that the brain can
swiftly and non-destructively perform space/time-translation operations on the
memory state so as to anticipate events to occur at various future moments
and/or remote locations.
The rodent brain contains a rich and detailed representation of current spatial location and temporal history. Some neurons--\emph{place cells}--in the hippocampus fire in circumscribed locations within an environment, referred to as their \emph{place fields}.
Early work excluded confounds based on visual \cite{SaveEtal98} or olfactory cues \cite{MullKubi87}, suggesting that the activity of place cells is a consequence of some form of path integration mechanism guided by the animal's velocity.
Other neurons in the hippocampus---\emph{time cells}---fire during a
circumscribed period of time within a delay interval
\cite{PastEtal08,MacDEtal11,GillEtal11,KrauEtal13,Eich14}. By analogy to
place cells, a set of time cells represents the animal's current temporal
position relative to past events. Some researchers have long hypothesized a
deep connection between the hippocampal representations of place and time
\cite{EichCohe14,Eich00}.
\begin{figure}
\includegraphics[width=0.98\columnwidth]{newfigs/schematic}
\caption{\textbf{a}.
Theta oscillations of 4-8~Hz are observed in the voltage recorded from the hippocampus. \emph{Hypothesis}: Within a theta cycle, a timeline of future translations of magnitude $\delta$ is constructed. \textbf{b}. Two layer network with theta-modulated connections.
The $\ti{}$ layer receives external input $\ensuremath{\mathbf{f}}{}$ in real time and encodes its Laplace transform. The Laplace transform is inverted via a synaptic operator $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ to yield an estimate of the function $\ensuremath{\mathbf{f}}{}$ on the $\T{}$ layer nodes. By periodically manipulating the weights in $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$, the memory state represented in $\T{}$ layer can be translated to represent its future states.
\label{fig:TILT}}
\end{figure}
Motivated by the spatial and temporal memory represented in the hippocampus,
we hypothesize that the translation operation required to anticipate the
events at a distant future engages this part of the brain
\cite{HassEtal07,SchaEtal07}. We hypothesize
that \emph{theta oscillations}, a well-characterized rhythm of 4-8 Hz in the
local field potential observed in the hippocampus may be responsible for the
translation operation. In particular, we hypothesize that sequential
translations of different magnitudes take place at different phases within a
cycle of theta oscillation, such that a timeline of anticipated future events
(or equivalently a spaceline of anticipated events at distant locations) is
swept out in a single cycle (fig.~\ref{fig:TILT}a).
Theta oscillations are prominently observed during periods of navigation
\cite{Vand69}. Critically, there is a systematic relationship between the
animal's position within a neuron's place field and the phase of the theta
oscillation at which that neuron fires \cite{OKeeRecc93}, known as
\emph{phase precession}.
This suggests that the phase of firing of the place cells conveys
information about the anticipated future location of the animal. This provides a
strong motivation for our hypothesis that the phase of theta oscillation would be
linked to the translation operation.
\subsection{Overview}
This paper develops a computational mechanism for the translation operation of
a spatial/temporal memory representation constructed from a two-layer neural
network model \cite{ShanHowa12}, and links it to theta oscillations by
imposing certain constraints based on some neurophysiological observations and
some physical principles we expect the brain to satisfy. Since the focus here
is to understand the computational mechanism of a higher level cognitive
phenomena, the imposed constraints and the resulting derivation should
be viewed at a phenomenological level, and not as emerging from biophysically
detailed neural interactions.
Computationally, we assume that the memory representation is constructed by a
two-layer network (fig.~\ref{fig:TILT}b) where the first layer encodes the
Laplace transform of externally observed stimuli in real time, and the second
layer approximately inverts the Laplace transform to represent a fuzzy
estimate of the actual stimulus history \cite{ShanHowa12}. With access to
instantaneous velocity of motion, this two layer network representing temporal
memory can be straightforwardly generalized to represent one-dimensional
spatial memory \cite{HowaEtal14}. Hence in the context of this two layer
network, time-translation of the temporal memory representation can be
considered mathematically equivalent to space-translation of the spatial
memory representation.
Based on a simple, yet powerful, mathematical observation that translation
operation can be performed in the Laplace domain as an instantaneous
point-wise product, we propose that the translation operation is achieved by
modulating the connection weights between the two layers within each theta
cycle (fig.~\ref{fig:TILT}b). The translated representations can then be used
to predict events at distant future and remote locations. In constructing
the translation operation, we impose two physical principles we expect the
brain to satisfy. The first principle is \emph{scale-invariance}, the
requirement that all scales (temporal or spatial) represented in the memory
are treated equally in implementing the translation. The second principle is
\emph{coherence}, the requirement that at any moment all nodes forming the
memory representation are in sync, translated by the same amount.
Further, to implement the computational mechanism of translation as a neural
mechanism, we impose certain phenomenological constraints based on
neurophysiological observations. First, there exists a dorsoventral axis in
the hippocampus of a rat's brain, and the size of place fields increase
systematically from the dorsal to the ventral end \cite{JungEtal94,KjelEtal08}.
In light of this observation, we hypothesize that the nodes representing
different temporal and spatial scales of memory are ordered along the dorsoventral
axis. Second, the phase of theta oscillation is not uniform along the
dorsoventral axis; phase advances from the dorsal to the ventral end like
a traveling wave \cite{LubeSiap09,PateEtal12} with a phase difference of about
$\pi$ from one end to the other. Third, the synaptic weights change
as a function of phase of the theta oscillation throughout the hippocampus
\cite{WyblEtal00,SchaEtal08}. In light of this observation, we hypothesize
that the change in the connection strengths between the two layers required to
implement the translation operation depend only on the local phase of the theta oscillation at any node (neuron).
In section~\ref{sec:math}, we impose the above mentioned physical principles
and phenomenological constraints to derive quantitative relationships for the
distribution of scales of the nodes representing the memory and the
theta-phase dependence of the translation operation. This yields specific
forms of phase-precession in the nodes representing the memory as well as the
nodes representing future prediction. Section~\ref{sec:nbio} compares these
forms to neurophysiological phase precession observed in the hippocampus and
ventral striatum. Section~\ref{sec:nbio} also makes explicit
neurophysiological predictions that could verify our hypothesis that theta
oscillations implement the translation operation to construct a timeline of
future predictions.
\input{FutureTimePRX-math}
\input{FutureTimePRX-nbio}
\input{FutureTimePRX-discussion}
\input{FutureTimePRX.bbl}
\end{document}
\section*{ Reviewer~A }
Reviewer~A identified two major weaknesses of the previous submission. First,
he or she noted that we did not provide a ``concrete description of how
the learning is done.'' The revision provides a thorough description of what
we mean by Hebbian association in the first two paragraphs of section II.C.
Briefly, simple
Hebbian association modifies the the synaptic weights between the neurons
representing a stimulus and the $\T{}$-layer neurons at theta phase zero
to an extent proportional to the activity of
the pre- and post-synaptic neurons. Because the post-synaptic neuron
corresponding to the stimulus is presumed to be activated at a constant rate
when the stimulus is presented, the change in these synapses is just proportional
to the activity of the presynaptic neuron.
Second, Reviewer~A felt that the paper was missing ``an evaluation of the
performance of the model, i.e., a demonstration that the whole thing works
from a practical point of view.
How good is the model at learning to predict
future events, and how good is it at using these predictions to improve its
behavior?''
Because it includes a more detailed description of neural data and concrete
neurophysiological predictions, the revision makes it easier to evaluate the
neurophysiological plausibility of the model (see figs.~4-6 and especially
subsection III.D). Computationally, fig.~3 in the revision makes it clear
that the prediction contains information about what will happen when and that
this prediction updates with the passage of time. Moreover, in the revision
we make clear the prediction is scale-invariant when $\delta=0$ and $\tau_o$
is not too extreme. While this places strong constraints on behavioral
performance of the model in decision-making, evaluating the model behaviorally
is a tangential problem beyond the scope of this work.
Reviewer~A also made several specific points
\begin{itemize}
\item Reviewer~A noted that the captions of several figures were too terse.
This has been corrected in the revision.
\item In the revision the operator $\mathbf{R}_{\delta}{}$ is explained in more detail
than in the previous submission.
\item The revision clarifies that $n$ is the index for the values of $s$.
\item Reviewer~A notes that our assumption that negative
phase of theta does not contribute to prediction should be justified more
thoroughly. The revision makes clear that this assumption follows from a
relatively large body of work noting that the synaptic conductance
and LTP are out of phase in the hippocampal formation. See especially
footnote~4.
\item
The reviewer correctly points out that the Weber-Fechner scale should be
logarithmic. That is exactly what we obtain in eq.~11. $n$ is related to $s_n$
logarithmically, and when we invert it $s_n$ will be related to ’$n$’
exponentially as seen in eq.~11.
\item In the revision, we clarify the notation regarding $\T{0}$ and
$\T{\delta}$.
\item In the revision, the prediction vector $\p{\delta}$ is defined more
clearly (see text preceding eq.~14).
\end{itemize}
\section*{ Reviewer~B }
Overall, Reviewer~B was positive about the idea behind the previous
submission, writing ``Thinking about the Hippocampal formation as a network
which can be used for predicting the future using the past experiences is an
interesting and important idea particularly given the discovery of the so
called time cells in this area.'' He or she concluded that the paper needed
``a major re~writing'' to make the paper more self-contained and clarify the
new contribution of the paper. In addition, he or she suggested a network
simulation to compare. Finally, Reviewer~B concluded
``\ldots these could be much better accomplished in a paper written for a
journal where space is not a limitation.''
We agree that the longer format of the revision has made for a much more accessible
presentation. The revision ought to be self-contained and accessible to a
broad interdisciplinary audience; we have rewritten the
paper such that the terminology is clear, and we have given a concise
description of our prior work in section II.A. The remainder of the paper
(from II.B onward) is entirely new. After careful consideration, we have
decided not to include a network simulation of the model in the revision.
Rather, the revision makes clear that the neurobiological constraints are
treated as phenomenological constraints rather than emergent properties of a
circuit. The reader can evaluate whether these phenomenological rules are
reasonable or not. Regardless, the phenomenological approach is appropriate
to a paper that attempts to provide a high-level computational description
of multiple interacting brain regions (including at least hippocampus,
striatum, and orbitofrontal cortex).
Reviewer~B raised several specific concerns.
\begin{enumerate}
\item Reviewer~B noted that the previous submission was not sufficiently
accessible because it relied on the reader's familiarity our previous work.
``Given this, I think the current paper is certainly not accessible to a wide
audience of physicists, and probably not to most of those working on neural
networks.'' All of the necessary background from our prior work is presented
in section II.A of the revision. Thus the revision is self-contained and
ought to be accessible to a broad interdisciplinary audience.
\item
Reviewer~B noted that we did not demonstrate a detailed mapping between
circuit level neurophysiological observations and computational-level
outcomes. ``In my view, although changes in synaptic strength during theta has
been observed, whether theta can actually perform the required modifications
of the synaptic weights that the current model needs is doubtful. Furthermore,
the assumption that the parameter $s$ is mapped anatomically on the
dorsoventral axis also requires further explanation: in any case such a
mapping will not be perfect and the authors should discuss how such
imperfections would change their model.'' In the revision it is clear that we
use the neurophysiological data to place phenomenological constraints on the
equations. If the phenomenological constraints are not valid, then the
derivation does not go
forward. It is far beyond the scope of the current paper to evaluate whether, for
instance, the magnitude of phasic changes in {EPSPs} at individual synapses is
sufficient to result in the changes in
these higher-level equations. For instance, properly evaluating this question
ought to take into account phasic changes in post-synaptic excitability,
phasically modulated inputs from a variety of brain regions, and the
persistence of excitation from a variety of conductances. Most of these
parameters are not known with certainty. Our hope is that publishing a
more abstract computational level paper such as this one clarifies the
important problems that can be solved using traditional methods in
computational neuroscience.
\item In the revision, we use a simple spike-generating mechanism to simulate
individual neurons from the higher-level equations for $\T{\delta}$ and
$\p{\delta}$. This should greatly facilitate comparison with the empricial
data.
\item Reviewer~B notes that bats do not exhibit theta (nor obviously theta
phase precession, which ``suggests that theta
phase precession may be just a rodent artifact.'' First, a thorough review
of theta oscillations in mammalian species is beyond the scope
of this paper. Having said that, monkeys exhibit theta during
eye movements (see especially Jutras, et al., 2013, \emph{PNAS}), memory tasks
(Brincat \& Miller, 2015) and virtual navigation (unpublished data from the
Buffalo lab).
There is also evidence that humans show task-related theta during virtual
navigation (a literature that goes back to Kahana, et al., 1999,
\emph{Nature}). While phase precession has not been reported in other
species, the widespread involvement of theta in similar tasks in primates
makes it seem unlikely that rodents are the only species that show phase
precession. It is possible that bats are the anomaly. It is also possible
that there is some as-yet-unidentified experimental problem that make it
difficult to measure (or evoke) theta in bat hippocampus. In either case,
at this stage the empirical story across species is too murky to justify a
discussion of bat theta in this paper.
\end{enumerate}
\end{document}
Rev A
The authors present a general mechanism to simulate a scaleinvariant future
timeline, and thus make predictions about future events. The neural associate
of such mechanism is theta phase precession in the hippocampus, the phenomenon
that neurons in the hippocampus tend to fire action potentials earlier and
earlier during a 48Hz oscillation cycle of neural activity. Their model
builds on previous work by the authors on a twolayer feed forward network
that forms, on the second layer, a scaleinvariant and distributed
representation of a stimulus presented to the first layer. This is done by
approximately antitransforming the Laplace transforms of the stimulus built
in the first layer, wherein each transform depends on a spatial parameter
localizing the node of the second layer.
In this manuscript, the authors build a future prediction of the stimulus as a
timetranslated memory of the same stimulus. They do so by adding a time
translation operation before anti-transforming, in layer II, the
representation of the input in layer I. On each node, the amount of
translation depends on the phase of the theta cycle. This mechanism seems to
explain data collected in the hippocampus (Fig. 3). Moreover, it is claimed
that it can learned by a simple Hebbian mechanism. The activity at each node
is pulled, after multiplication by the learned association strength, to obtain
a prediction of future state in a scaleinvariant manner.
The topic is extremely important and the work seems competently done. I like
the fact that the authors are trying to solve a fundamental cognitive process
by using a construction that, although abstract, makes contact with the
neurobiology in several key points (see e.g. Fig. 3). The paper is also
generally well written and the narrative proceeds in a logical way.
As for the negative aspects, it seems to me that the paper is lacking in two
main parts: 1. the absence of a concrete (if not biologically plausible)
description of how the learning is done, and 2. an evaluation of the
performance of the model, i.e., a demonstration that the whole thing works
from a practical point of view. How good is the model at learning to predict
future events, and how good is it at using these predictions to improve its
behavior? From this point of view, the work is rather abstract and offers no
evaluation. I also must lament a fair amount of technical detail that, because
of the short format, makes a deeper understanding of the model (beyond the
general idea) quite difficult to grasp. Unpacking the technical information
would probably require the paper to be written for a more technical Journal.
Specific points:
The Figure captions (especially for Fig. 1 and 2) should contain more
details. What’s on the axes?
p. 2, eq. 4: can you define the operator R_{\theta} more precisely? I’m not
sure that the approximation of footnote 1 is warranted.
p. 2: what is ’n’ in Eq. 5?
p. 2: sentences like “Following [10, 11] we assume that the negative
phase of theta does not contribute to prediction” deserve a little explanation.
p. 3: “a coherent time translation across the memory nodes requires the
values sn to lie on a WeberFechner scale”. Strictly speaking, isn’t the
WeberFechner scale supposed to be logarithmic instead of powerlaw?
p. 3, section B: I find it confusing that the authors refer to T_0 and T sometimes as ‘states’, sometimes as ‘layers’ (especially in Section A).
p. 4, eq. 11: what is the definition of p?
Rev B
The paper by Shankar and Howard proposes a neural network model for predicting
the future of a stimulus, postulating that theta oscillations in the
hippocampus can modify the synaptic weights required for their model to work.
Thinking about the Hippocampal formation as a network which can be used for
predicting the future using the past experiences is an interesting and
important idea particularly given the discovery of the so called time cells in
this area. The authors of the current paper motivate their work by mentioning
some experimental findings in the Hippocampus regarding timing of spikes
(phase precession) and the time cells and then move on to building the network
model. My comments are listed below:
1. Unfortunately I must say that this paper is not very wellwritten. Most of
the main idea of the model seem to have been presented in another paper by the
same authors in Neural Computation 2012 (ref. 5) and I had to check on that
paper before starting to comprehend the current one (and I am still unsure if
I completely do so). Given this, I think the current paper is certainly not
accessible to a wide audience of physicists, and probably not to most of those
working on neural networks.
2. The novel contribution of the current paper is relating the model in Neural
Computation 2012 to theta oscillations by postulating theta as a mechanism to
modify the synaptic weight as required by the model to translate the memory
into future ( by the way "translate memory into future" is a term that is not
clearly defined anywhere in the paper and still remains a bit mysterious to
me). In my view, although changes in synaptic strength during theta has been
observed, whether theta can actually perform the required modifications of the
synaptic weights that the current model needs is doubtful. Furthermore, the
assumption that the parameter "s" is mapped anatomically on the dorsoventral
axis also requires further explanation: in any case such a mapping will not be
perfect and the authors should discuss how such imperfections would change
their model.
3. The supplementary material are essentially a review of experimental facts.
I think it would be much more useful if the authors actually showed simulated
networks in which the responses of the network was compared to the
experimental observations.
4. Data from the Bat hippocampus (from Nachum Ulanovsky's lab) show that bat
hippocampus does not show theta nor phase precession which suggests that theta
phase precession may be just a rodent artifact. How does lack of these
oscillations in a species like bat can be interpreted in this model which so
heavily relies on theta?
In summary, this paper does deal with an interesting idea but in the current
format it is not suitable for publication. I think first and foremost, the
paper requires a major rewriting to (a) provide the background material from
Neural Computation 2012, defining the terms used in the paper before they are
used (terms such time translation, time translated memory, future timeline,
scale invariance of the timeline) and (b) clearly state what is novel in the
current paper compared to Neural Comp 2012. The authors also need to show
network simulations both to validate their model in the presence of various
sources of noise, but also to give a better feeling of the model they are
studying to the reader. In my view, all of these could be much better
accomplished in a paper written for a journal where space is not a limitation.
—————————Response to comments of Reviewer-B ————————————————
The major concern of the reviewer is that the paper uses some terminologies
that are not clearly defined and the reader has to go to our earlier papers
for clarification. We have rewritten the paper such that the terminology is
clear, and we have given a broader description of the two layer memory network
that is adopted from our prior work. However, we emphasize that the only
concept needed from our prior papers is that the two layer network can
represent memory by encoding and inverting Laplace transform of the past.
Analysis of this network model, its scale invariance properties, and effects
of various sources of noise have been previously discussed in detail (Neural
computation 2012 and JMLR 2013– ref 5 and 9). As suggested by
the reviewer, we have revised the paper to clearly portray a succinct
description of the two-layer memory model from 2012.
Rather than devoting a lengthy discussion of the prior model, we focus here on
the novel idea of time-translating the memory state (in the Laplace domain) by
modulating the synapses and relating it to theta phase precession, while at
the same time giving it a cognitive-utility of instant prediction of future
events in an ordered-fashion. We believe that our revised manuscript would
let most physicists comfortably accept that our two layer network model can
represent temporal memory by encoding and decoding the Laplace transform of
the past, beyond which the readers can smoothly transition into the novel
story of instant future prediction through time-translated memory states and
their neural relationship to theta oscillations.
Next, the reviewer points out two features where our theoretical model may not
match the actual neurobiology.
(1) The synaptic changes observed during theta oscillations may not be enough
to perform the time-translations required in the model. Very correct; the
value of the parameters \ensuremath{\Phi_{\textnormal{max}}}{} and $k$ in the model will essentially
determine the range to which the synaptic strength must oscillate. The change
in the EPSPs (shown in ref 8,10) does not have to be directly interpreted as
synaptic weight changes, and neither does the synaptic weights of the Hebbian
learning rule have to be literally interpreted in terms of the observed EPSP
modulations. In other words, we only derive a motivation from the empirical
observation of theta modulated synaptic strengths, and we do not intend to
justify the model requirements based on the quantitative empirical
observation. (2) The anatomical organization of the decay constants $s$ along
the dorsovental hippocampal axis would have imperfections that could affect
the model. Very correct again; we have revised the discussion to clarify this
point. This reasoning behind this assumption relates to the fact that the
place-cells in the hippocampus have place-fields that are increasingly larger
as we traverse through the dorsoventral axis (which are analogous to the decay
constants $s$). As a theoretical physics paper, we focus on the
big-picture by qualitatively accepting basic properties observed in the
hippocampal network and use physical principles to theorize a functional
utility.
The reviewer also points out that bats don’t have theta oscillations(although
they are mammals), while they can still predict and navigate
in 3-dimensional flight. Honestly, we do not have an answer for this. But
this is an important issue which we have now addressed in our introduction and
discussion. May be the three-dimensional flight and prediction is very
different from two dimensional navigation and prediction.
\section{Discussion}
\label{sec:discussion}
This paper presented a neural hypothesis for implementing translations of
temporal and 1-d spatial memory states so that future events can be quickly
anticipated without destroying the current state of memory. The hypothesis
assumes that time cells and place cells observed in the hippocampus represent
time or position as a result of a two-layer architecture that encodes and
inverts the Laplace transform of external input. It also assumes that
sequential translations to progressively more distant points in the future
occur within each cycle of theta oscillations. Neurophysiological
constraints were imposed as phenomenological rules rather than as emerging
from a detailed circuit model. Further, imposing scale-invariance and
coherence in translation across memory nodes resulted in Weber-Fechner
spacing for the representation of both the past (spacing of $s_n$ in the
memory nodes) and the future (the relationship between $\delta$ and
$\theta_o$). Apart from providing cognitive flexibility in accessing a
timeline of future predictions at any moment, the computational mechanism
described qualitative features of phase precession in the hippocampus and in
the ventral striatum. Additionally, we have also pointed out certain
distinctive features of the model that can be tested with existing technology.
\subsection{Computational Advantages}
The property of the $\T{}$ layer that different nodes represent the stimulus
values from various delays (past moments) is reminiscent of a shift register
(or delay-line or synfire chain). However, the two layer network encoding and
inverting the Laplace transform of stimulus has several significant
computational advantages over a shift register representation.
(i) In the current two-layer network, the spacing of $s$ values of the nodes
can be chosen freely. By choosing exponentially spaced $s$-values
(Weber-Fechner scaling) as in eq.~\ref{eq:thetas}, the $\T{}$ layer can
represent memory from exponentially long timescales compared to a shift
register with equal number of nodes, thus making it extremely
resource-conserving. Although information from longer timescales is more
coarse-grained, it turns out that this coarse-graining is optimal to
represent and predict long-range correlated signals \cite{ShanHowa13}.
(ii) The memory representation of this two layer network is naturally
scale-invariant (eq.~\ref{eq:old_big_T}). To construct a scale-invariant
representation from a shift register, the shift register would have to be
convolved with a scale-invariant coarse-graining function at each moment,
which would be computationally very expensive. Moreover, it turns out that any
network that can represent such scale-invariant memory can be identified with
linear combinations of multiple such two layer networks \cite{Shan15}.
(iii)
Because
translation can be trivially performed when we have access to the Laplace
domain, the two layer network enables translations by an amount
$\delta$ without sequentially visiting the intermediate states $<\delta$. This can be
done by directly changing the connection strengths locally
between the two layers as prescribed by diagonal $\mathbf{R}_{\delta}$ operator for any chosen
$\delta$.\footnote{In this paper we considered sequential translations of
various values of $\delta$, since the aim was to construct an
entire future timeline rather than to discontinuously jump to a
distant future state.} Consequently the physical time taken for the
translation can be decoupled from the magnitude of translation. One could
imagine a shift register performing a translation operation by an amount
$\delta$ either by shifting the values sequentially from one node to the next
for $\delta$ time steps or by establishing non-local connections between far
away nodes. The latter would make the computation very cumbersome because it
would require every node in the register to be connected to every other node
(since this should work for any $\delta$), which is in stark contrast with the
local connectivity required by our two layer network to perform any
translation.
Many previous neurobiological models of phase precession have been proposed
\cite{LismJens13,MehtEtal02,BurgEtal07,Hass12}, and many assume that
sequentially activated place cells firing within a theta cycle result from
direct connections between those cells \cite{JensLism96}, not unlike a synfire
chain. Although taking advantage of the Laplace domain in the two
layer network to perform translations is not the only possibility, it seems
to be computationally powerful compared to the obvious alternatives.
\subsection{Translations without theta oscillations}
Although this paper focused on sequential translation within a theta cycle,
translation may also be accomplished \emph{via} other neurophysiological
mechanisms. Sharp wave ripple (SRW) events last for about 100~ms and are
often accompanied by replay events--sequential firing of place cells
corresponding to locations different from the animal's current
location \cite{DaviEtal09,DragTone11,FostWils06,PfeiFost13,JadhEtal12}. Notably,
experimentalists have also observed preplay events during SWRs, sequential activation of
place cells that correspond to trajectories that have never been
previously traversed, as though the animal is planning a future path
\cite{DragTone11,OlafEtal15}.
Because untraversed trajectories could not have been used to learn and build
sequential associations between the place cells along the trajectory, the
preplay activity could potentially be a result of a translation operation on
the overall spatial memory representation.
Sometimes during navigation, a place cell corresponding to a distant goal
location gets activated \cite{PfeiFost13}, as though a finite distance
translation of the memory state has occurred. More interestingly, sometimes a
reverse-replay is observed in which place cells are activated in reverse order
spreading back from the present location \cite{FostWils06}. This is suggestive
of translation into the past (as if $\delta$ was negative), to implement a
memory search. In parallel, there is behavioral evidence from humans that
under some circumstances memory retrieval consists of a backward scan through
a temporal memory representation \cite{Hack80,Hock84,SingEtal15} (although
this is not neurally linked with SWRs). Mathematically, as long as the
appropriate connection strength changes prescribed by the $\mathbf{R}_{\delta}$ operator can
be specified, there is no reason translations with negative $\delta$ or
discontinuous shift in $\delta$ could not be accomplished in this framework.
Whether these computational mechanisms are reasonable in light of the
neurophysiology of sharp wave ripples is an open question.
\subsection{Multi-dimensional translation}
This paper focused on translations along one dimension. However it would be
useful to extend the formalism to multi-dimensional translations. When a rat
maneuvers through an open field rather than a linear track, phase precessing
2-d place cells are observed \cite{SkagEtal96}. Consider the case of an
animal approaching a junction along a maze where it has to either turn left or
right. Phase precessing cells in the hippocampus indeed predict the direction
the animal will choose in the future \cite{JohnRedi07}. In order to
generalize the formalism to 2-d translation, the nodes in the network model
must not be indexed only by $s$, which codes their distance from a landmark,
but also by the 2-d orientation along which distance is calculated. The
translation operation must then specify not just the distance, but also
the instantaneous direction as a function of the theta phase. Moreover,
if translations could be performed on multiple non-overlapping
trajectories simultaneously, multiple paths could be searched in
parallel, which would be very useful for efficient decision making.
\subsection{Neural representation of predictions}
The computational function of $\p{\delta}$ (eq.~\ref{eq:p}) is to represent an
ordered set of events predicted to occur in the future. Although we focused
on ventral striatum here because of the availability of phase precession data
from that structure, it is probable that many brain regions represent future
events as part of a circuit involving frontal cortex and basal ganglia, as
well as the hippocampus and striatum
\cite{SchuEtal97,FerbShap03,TanaEtal04,FeieEtal06,MainKepe09,TakaRoes11,YounShap11a}.
There is evidence that theta-like oscillations coordinates the activity in
many of these brain regions
\cite{JoneWils05,LansEtal09,vanWEtal10,FujiBuzs11}. For instance, 4~Hz
oscillations show phase coherence between the hippocampus, prefrontal cortex
and ventral tegmental area (VTA), a region that signals the presence of
unexpected rewards \cite{FujiBuzs11}. A great deal of experimental work has
focused on the brain's response to future rewards, and indeed the
phase-precessing cells in fig.~\ref{fig:MeerRedi} appear to be predicting the
location of the future reward. The model suggests that $\p{\delta}$ should
predict any future event, not just a reward. Indeed, neurons that appear to
code for predicted stimuli have been observed in the primate inferotemporal
cortex \cite{SakaMiya91} and prefrontal cortex \cite{RainEtal99}. Moreover,
theta phase coherence between prefrontal cortex and hippocampus are
essential for learning the temporal relationships between stimuli
\cite{BrinMill15}.
So, future predictions could be widely distributed throughout the brain.
\section{Mathematical model}
\label{sec:math}
In this section we start with a basic overview of the two layer memory model
and summarize the relevant details from previous work
\cite{ShanHowa12,ShanHowa13,HowaEtal14} to serve as a background. Following
that, we derive the equations that allow the memory nodes to be coherently
time-translated to various future moments in synchrony with the theta
oscillations. Finally we derive the predictions generated for various future
moments from the time-translated memory states.
\subsection{Theoretical background}
\label{ssec:mathback}
The memory model is implemented as a two-layer feedforward network
(fig.~\ref{fig:TILT}b) where the $\ti{}$ layer holds the Laplace transform of
the recent past and the $\T{}$ layer reconstructs a temporally fuzzy estimate
of past events \cite{ShanHowa12,ShanHowa13}. Let the stimulus at any time
$\tau$ be denoted as $\ensuremath{\mathbf{f}}{}(\tau)$. The nodes in the $\ti{}$ layer are leaky
integrators parametrized by their decay rate $s$, and are all independently
activated by the stimulus. The nodes are assumed to be arranged w.r.t.~their
$s$ values. The nodes in the $\T{}$ layer are in one to one correspondence
with the nodes in the $\ti{}$ layer and hence can also be parametrized by the
same $s$. The feedforward connections from the $\ti{}$ layer into the $\T{}$
layer are prescribed to satisfy certain mathematical properties which are
described below. The activity of the two layers is given by
\begin{eqnarray}
\frac{d}{d \tau} \ti{}(\tau,s) &=& -s \ti{}(\tau,s) + \ensuremath{\mathbf{f}}{}(\tau) \label{eq:teq} \\
\T{}(\tau,s) &=& [\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}} ] \, \ti{}(\tau,s) \label{eq:Teq}
\end{eqnarray}
By integrating eq.~\ref{eq:teq}, note that the $\ti{}$ layer encodes the
Laplace transform of the entire past of the stimulus function leading up to
the present. The $s$ values distributed over the $\ti{}$ layer represent
the (real) Laplace domain variable. The fixed connections between the $\ti{}$
layer and $\T{}$ layer denoted by the operator $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ (in eq.~\ref{eq:Teq}), is
constructed to reflect an approximation to inverse Laplace transform. In
effect, the Laplace transformed stimulus history which is distributed over
the $\ti{}$ layer nodes is inverted by $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ such that a fuzzy (or coarse
grained) estimate of the actual stimulus value from various
past moments is represented along the different $\T{}$ layer nodes.
More precisely, treating the $s$ values nodes as continuous, the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator can be succinctly expressed as
\begin{equation}
\T{}(\tau,s) = \frac{(-1)^k}{k!} s^{k+1} \ti{}^{(k)} (\tau,s) \equiv [\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}] \, \ti{}(\tau,s)
\label{eq:Lk}
\end{equation}
Here $\ti{}^{(k)}(\tau,s)$ corresponds to the $k$-th derivative of
$\ti{}(\tau,s)$ w.r.t.~$s$. It can be proven that $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator executes an
approximation to the inverse Laplace transformation and the approximation
grows more and more accurate for larger and larger values of $k$
\cite{Post30}. Further details of $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ depends on the $s$ values chosen for
the nodes \cite{ShanHowa13}, but these details are not relevant for this
paper as the $s$ values of neighboring nodes are assumed to be close enough
that the analytic expression for $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ given by eq.~\ref{eq:Lk} would be
accurate.
To emphasize the properties of this memory representation, consider the
stimulus $\ensuremath{\mathbf{f}}{}(\tau)$ to be a Dirac delta function at $\tau=0$. From
eq.~\ref{eq:teq}~and~\ref{eq:Lk}, the $\T{}$ layer activity following the
stimulus presentation ($\tau >0$) turns out to be
\begin{equation}
\T{} (\tau, s) = \frac{s}{k!} \left[s\tau \right]^k
e^{-\left[s\tau \right]}
\label{eq:old_big_T}
\end{equation}
Note that nodes with different $s$ values in the $\T{}$ layer peak in activity after different delays following the stimulus; hence the $\T{}$ layer nodes behave like {time cells}. In particular, a node with a given $s$ peaks in activity at a time $\tau=k/s$ following the stimulus. Moreover, viewing the activity of any node as a distribution around its appropriate peak-time ($k/s$), we see that the shape of this distribution is exactly the same for all nodes to the extent $\tau$ is rescaled to align the peaks of all the nodes. In other words, the activity of different nodes of the $\T{}$ layer represent a fuzzy estimate of the past information from different timescales and the fuzziness associated with them is directly proportional to the timescale they represent, while maintaining the exact same shape of fuzziness. For this reason, the $\T{}$ layer represents the past information in a {scale-invariant} fashion.
This two-layer memory architecture is also amenable to represent
one-dimensional spatial memory analogous to the representation of temporal
memory in the $\T{}$ layer \cite{HowaEtal14}. If the stimulus $\ensuremath{\mathbf{f}}{}$ is
interpreted as a landmark encountered at a particular location in a
one-dimensional spatial arena, then the $\ti{}$ layer nodes can be made to
represent the Laplace transform of the landmark treated as a spatial function
with respect to the current location. By modifying eq.~\ref{eq:teq} to
\begin{equation}
\frac{d}{d \tau} \ti{}(\tau,s) = v \left[ -s \ti{}(\tau,s) + \ensuremath{\mathbf{f}}{}(\tau) \right] ,
\label{eq:timespace}
\end{equation}
where $v$ is the velocity of motion, the temporal dependence of the $\ti{}$ layer activity can be converted to spatial dependence.\footnote{ Theoretically, the velocity here could be an animal's running velocity in the
lab maze or a mentally simulated human motion while playing video games.} By employing the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator on this modified $\ti{}$ layer activity (eq.~\ref{eq:timespace}), it is straightforward to construct a layer of nodes (analogous to $\T{}$) that exhibit peak activity at different distances from the landmark. Thus the two-layer memory architecture can be trivially extended to yield {place-cells} in one dimension.
In what follows, rather than referring to translation operations separately on spatial and temporal memory, we shall simply consider time-translations with an implicit understanding that all the results derived
can be trivially extended to 1-d spatial memory representations.
\subsection{Time-translating the Memory state}
\label{sec:newstuff}
The two-layer architecture naturally lends
itself for {time-translations} of the memory state in the $\T{}$ layer,
which we shall later exploit to construct a timeline of future predictions.
The basic idea is that if the current state of memory represented in the
$\T{}$ layer is used to anticipate the present (via some prediction
mechanism), then a time-translated state of $\T{}$ layer can be used to predict
events that will occur at a distant future via the same prediction
mechanism. Time-translation means to mimic the $\T{}$ layer
activity at a distant future based on its current state. Ideally translation
should be {non-destructive}, not overwriting the current activity in
the $\ti{}$ layer.
Let $\delta$ be the amount by which we intend to time-translate the state of
$\T{}$ layer. So, at any time $\tau$, the aim is to access
$\T{}(\tau+\delta,s)$ while still preserving the current $\ti{}$ layer
activity, $\ti{}(\tau,s)$. This is can be easily achieved because the $\ti{}$
layer represents the stimulus history in the Laplace domain. Noting that the
Laplace transform of a $\delta$-translated function is simply the product of
$e^{-s \delta}$ and the Laplace transform of the un-translated function, we see
that
\begin{eqnarray}
\ti{}(\tau+\delta,s) &=& e^{-s\delta} \ti{}(\tau,s) \label{eq:translate}
\end{eqnarray}
Now noting that $\T{}(\tau+\delta,s)$ can be obtained by employing the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$
operator on $\ti{}(\tau+\delta,s)$ analogous to eq.~\ref{eq:Lk}, we obtain the
$\delta$-translated $\T{}$ activity as
\begin{eqnarray}
\T{\delta}(\tau,s) \equiv \T{}(\tau+\delta,s) &=& [ \ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}] \, \ti{}(\tau+\delta ,s) \nonumber \\
&=& \left[\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}} \cdot \mathbf{R}_{\delta} \right] \, \ti{}(\tau,s)
\label{eq:Rd}
\end{eqnarray}
where $\mathbf{R}_{\delta}$ is just a diagonal operator whose rows and columns are indexed by
$s$ and the diagonal entries are $e^{-s \delta}$. The $\delta$-translated
activity of the $\T{}$ layer is now subscripted by $\delta$ as $\T{\delta}$ so
as to distinguish it from the un-translated $\T{}$ layer activity given by
eq.~\ref{eq:Lk} without a subscript. In this notation the un-translated state
$\T{}(\tau,s)$ from eq.~\ref{eq:Lk} can be expressed as $\T{0}(\tau,s)$. The
time-translated $\T{}$ activity can be obtained from the current $\ti{}$ layer
activity if the connection weights between the two layers given by $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ is
modulated by $\mathbf{R}_{\delta}$. This computational mechanism of time-translation can be implemented as a neural mechanism in the brain, by imposing certain phenomenological constraints and physical principles.
{\bf \emph{Observation 1:}}
Anatomically, along the dorsoventral axis of the hippocampus, the width of
place fields systematically increases from the dorsal end to the
ventral end \cite{JungEtal94,KjelEtal08}. Fig.~\ref{fig:theta} schematically
illustrates this observation by identifying the $s$-axis of the two-layer
memory architecture with the dorso-ventral axis of the hippocampus,
such that the scales represented by the nodes are monotonically arranged. Let
there be $N+1$ nodes with monotonically decreasing $s$ values given by $s_o$,
$s_1$, \ldots $s_N$.
\begin{figure}
\includegraphics[width=0.4\textwidth]{newfigs/TravelingWave.png}
\caption{Traveling theta wave along the $s$ axis.
The x-axis is real time. Each point along the dorsoventral axis corresponds
to a different value of $s_n$. The curvy blue lines show the theta
oscillation for several different values of $s$. Lines~1~and~2
connect the positions where the local phases $\theta_s$ are 0 and
$\pi$ respectively.
\label{fig:theta}}
\end{figure}
{\bf \emph{Observation 2:}}
The phase of the theta oscillations along the axis is non-uniform,
representing a traveling wave from the dorsal to ventral part of the hippocampus
with a net phase shift of $\pi$ \cite{LubeSiap09,PateEtal12}.
The oscillations in fig.~\ref{fig:theta} symbolize the local
field potentials at different locations of the $s$-axis. The local phase of
the oscillation at any position on the $s$-axis is denoted by $\theta_s$,
which ranges from $-\pi$ to $+\pi$ by convention. However, as a reference we
denote the phase at the top (dorsal) end as $\theta_o$ ranging from $0$ to $2
\pi$, with the understanding that the range $(\pi, 2 \pi)$ is mapped on to
$(-\pi, 0)$. The $x$-axis in fig.~\ref{fig:theta} is time within a theta
oscillation labeled by the phase $\theta_o$.
In this convention, the value of $\theta_s$ discontinuously jumps from $+\pi$
to $-\pi$ as we move from one cycle of oscillation to the next. In
fig.~\ref{fig:theta}, the diagonal (solid-red) line labeled `2' denotes all
the points where this discontinuous jump happens. The diagonal (dashed) line
labeled `1' denotes all the points where $\theta_s=0$. It is straightforward
to infer the relationship between the phase at any two values of $s$. Taking
the nodes to be uniformly spaced anatomically, the local phase $\theta_s$ of the $n$-th
node is related to $\theta_o$ (for $0<\theta_s<\pi$) by\footnote{ Since the $s$
values of the nodes are monotonically arranged, we can interchangeably use $s$
or $n$ as subscritpts to $\theta$. }
\begin{equation}
\theta_s/\pi= \theta_o/\pi - n/N .
\label{eq:thetasnophi}
\end{equation}
{\bf \emph{Observation 3:}} Synaptic weights in the hippocampus are modulated
periodically in synchrony with the phase of theta
oscillation \cite{WyblEtal00,SchaEtal08}. Based on this observation, we impose
the constraint that the connection strengths between the $\ti{}$ and $\T{}$
layers at a particular value of $s$ depend only on the local phase of the
theta oscillations. Thus the diagonal entries in the $\mathbf{R}_{\delta}$ operator should
only depend on $\theta_s$. We take these entries to be of the form $\exp{
(-\Phi_s (\theta_s)) } $, where $\Phi_s$ is any continuous function of
$\theta_s \in (-\pi, +\pi)$. Heuristically, at any moment
within a theta cycle, a $\T{}$ node with a given $s$ value will be
roughly translated by an amount $\delta = \Phi_s(\theta_s)/s$.
{\bf \emph{Principle 1:}} \emph{Preserve Scale-Invariance}
Scale-invariance is an extremely adaptive property for a memory to have; in
many cases biological memories seem to exhibit scale-invariance
\cite{BalsGall09}.
As the untranslated $\T{}$ layer activity already exhibits scale-invariance, we impose the constraint
that the time-translated states of $\T{}$ should also exhibit scale-invariance. This consideration requires the behavior of every node to follow the same pattern with respect to their local theta phase. This amounts
to choosing the functions $\Phi_s$ to be the same for all $s$, which we shall refer to as $\Phi$.
{\bf \emph{Principle 2:}} \emph{Coherence in translation}
Since the time-translated memory state is going to be used to make predictions
for various moments in the distant future, it would be preferable if all the nodes
are time-translated by the same amount at any moment within a theta
cycle. If not, different nodes would contribute to predictions for different
future moments leading to noise in the prediction. However, such a
requirement of global coherence cannot be imposed consistently along with the
principle~1 of preserving scale-invariance.\footnote{This is easily seen by
noting that each node will have a maximum translation inversely
proportional to its $s$-value to satisfy principle~1.} But in the light of
prior work \cite{HassEtal02,Hass12} which suggest that retrieval of
memory or prediction happens only in one half of the theta cycle,\footnote{
This hypothesis follows from the observation that while both synaptic
transmission and synaptic plasticity are modulated by theta phase,
they are out of phase with one another. That is, while certain synapses are
learning, they are not communicating information and vice versa. This
led to the hypothesis that the phases where plasticity is
optimal are specialized for encoding whereas the phases where transmission is
optimal are specialized for retrieval. }
we
impose the requirement of coherence only to those nodes that are all in the
positive half of the cycle at any moment. That is, $\delta= \Phi(\theta_s)/s$
is a constant along any vertical line in the region bounded between the
diagonal lines~1~and~2 shown in fig.~\ref{fig:theta}. Hence for all nodes
with $0<\theta_s< \pi$, we require
\begin{equation}
\Delta \left(\Phi\left(\theta_s\right) / s \right) = \Delta \left(\Phi\left(\theta_o - \pi n/N\right) / s_n \right) = 0.
\label{eq:deltathetas}
\end{equation}
For coherence as expressed in eq.~\ref{eq:deltathetas} to hold at all values
of $\theta_o$ between 0 and $2 \pi$, $\Phi(\theta_s)$ must be an exponential
function so that $\theta_o$ can be
functionally decoupled from $n$; consequently $s_n$ should also have an
exponential dependence on $n$. So the general solution to
eq.~\ref{eq:deltathetas} when $0<\theta_s<\pi$ can be written as
\begin{eqnarray}
\Phi(\theta_s)&=& \ensuremath{\Phi_o} \exp{[b \theta_s]} \label{eq:thetas} \\
s_n &=& s_o (1+c)^{-n} \label{eq:thetasn}
\end{eqnarray}
where $c$ is a positive number. In this paper, we shall take $c \ll 1$, so
that the analytic approximation for the $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ operator given in terms of the
$k$-th derivative along the $s$ axis in eq.~\ref{eq:Lk} is valid.
Thus the requirement of coherence in time-translation implies that the $s$
values of the nodes---the timescales represented by the nodes---are spaced out
exponentially, which can be referred to as a Weber-Fechner scale, a commonly
used terminology in cognitive science. Remarkably, this result strongly
resonates with a requirement of the exact same scaling when the predictive
information contained in the memory system is maximized in response to
long-range correlated signals \cite{ShanHowa13}. This feature allows this
memory system to represent scale-invariantly coarse grained past information
from timescales exponentially related to the number of nodes.
The maximum value attained by the function $\Phi( \theta_s)$ is at $\theta_s=\pi$, and the maximum value is $\ensuremath{\Phi_{\textnormal{max}}} = \ensuremath{\Phi_o} \exp{[b \pi] }$, such that $\ensuremath{\Phi_{\textnormal{max}}}/\ensuremath{\Phi_o} = s_o/s_N$ and $b=(1/\pi) \log{(\ensuremath{\Phi_{\textnormal{max}}}/\ensuremath{\Phi_o})}$.
To ensure continuity around $\theta_s=0$, we take the eq.~\ref{eq:thetas} to
hold true even for $\theta_s \in (-\pi,0)$. However, since notationally
$\theta_s$ makes a jump from $+\pi$ to $-\pi$, $\Phi(\theta_s)$ would exhibit
a discontinuity at the diagonal line~2 in fig.~\ref{fig:theta} from $\ensuremath{\Phi_{\textnormal{max}}}$
(corresponding to $\theta_s=\pi $) to $\ensuremath{\Phi_{\textnormal{min}}}=\ensuremath{\Phi_o}^2/\ensuremath{\Phi_{\textnormal{max}}}$
(corresponding to $\theta_s=-\pi$).
Given these considerations, at any instant within a theta cycle, referenced by
the phase $\theta_o$, the amount $\delta$ by which the memory state is
time-translated can be derived from eq.~\ref{eq:thetasnophi} and
\ref{eq:thetas} as
\begin{equation}
\delta(\theta_o) = (\ensuremath{\Phi_o}/s_o) \exp{[b \theta_o]} .
\label{eq:deltatheta}
\end{equation}
Analogous to having the past represented on a Weber-Fechner scale, the
translation distance $\delta$ into the future also falls on a Weber-Fechner
scale as the theta phase is swept from 0 to $2 \pi$. In other words, the
amount of time spent within a theta cycle for larger translations is
exponentially smaller.
To emphasize the properties of the time-translated $\T{}$ state, consider the
stimulus to be a Dirac delta function at $\tau=0$. From eq.~\ref{eq:Rd}, we
can express the $\T{}$ layer activity analogous to eq.~\ref{eq:old_big_T}.
\begin{equation}
\T{\delta} (\tau, s) \simeq \frac{s}{k!} \left[s\tau+\Phi\left(\theta_s\right)\right]^k
e^{-\left[s\tau+ \Phi\left(\theta_s\right)\right]}
\label{eq:bigT}
\end{equation}
Notice that eqs.~\ref{eq:thetasnophi} and \ref{eq:deltatheta} specify a unique
relationship between $\delta$ and $\theta_s$ for any given $s$. The
r.h.s.~above is expressed in terms of $\theta_s$ rather than $\delta$ so as to shed
light on the phenomenon of {phase precession}.
Since $\T{\delta}(\tau,s)$ depends on both $\tau$ and $\theta_s$ only via the
sum $[ s\tau+ \Phi\left(\theta_s\right) ] $, a given node will show identical
activity for various combinations of $\tau$ and $\theta_s$.\footnote{While
representing timescales much larger than the period of a theta cycle, $\tau$
can essentially be treated as a constant within a single cycle. In other
words, $\theta_s$ and $\tau$ in eq.~\ref{eq:Rd} can be treated as
independent, although in reality the phases evolve in real time.}
For instance, a node would achieve its peak activity when $\tau$ is
significantly smaller than its timescale $(k/s)$ only when $\Phi(\theta_s)$ is
large---meaning $\theta_s \simeq +\pi$. And as $\tau$ increases towards the
timescale of the node, the peak activity gradually shifts to earlier phases
all the way to $\theta_s \simeq -\pi$. An important consequence of imposing
principle~1 is that the relationship between $\theta_s$ and $\tau$ on any
iso-activity contour is scale-invariant. That is, every node behaves similarly when
$\tau$ is rescaled by the timescale of the node. We shall
further pursue the analogy of this phenomenon of phase precession with
neurophysiological findings in the next section (fig.~\ref{fig:ph-pr}).
\subsection{Timeline of Future Prediction}
At any moment, $\T{\delta}$ (eq.~\ref{eq:bigT}) can be used to predict the
stimuli expected at a future moment. Consequently, as $\delta$
is swept through within a theta cycle, a timeline of future predictions can be
simulated in an orderly fashion, such that predictions for closer events occur
at earlier phases (smaller $\theta_o$) and predictions of distant events occur
at later phases. In order to predict from a time-translated state
$\T{\delta}$, we need a prediction mechanism.
For our purposes, we consider here a very simple form of learning and
prediction, \emph{Hebbian association}. In this view, an event is learned (or
an association formed in long term memory) by increasing the connection
strengths between the neurons representing the currently-experienced
stimulus and the neurons representing the recent past events ($\T{0}$).
Because the $\T{}$ layer activity contains temporal information about the preceding stimuli, simple associations between $\T{0}$ and the current stimulus
are sufficient to encode and express well-timed predictions \cite{ShanHowa12}.
In particular, the term Hebbian implies that the change in each connection
strength is proportional to the product of pre-synaptic activity---in this
case the activity of the corresponding node in the $\T{}$ layer---and
post-synaptic activity corresponding to the current stimulus.
Given that the associations are learned in this way, we define the prediction of a
particular stimulus to be the scalar product of its association strengths with
the current state of $\T{}$. In this way, the scalar product of association
strengths and a translated state $\T{\delta}$ can be understood as the
future prediction of that stimulus.
Consider the thought experiment where a
conditioned stimulus \textsc{cs} is consistently followed by another
stimulus, \textsc{a} or \textsc{b}, after a time $\tau_o$. Later when
\textsc{cs} is repeated (at a time $\tau=0$), the subsequent activity in the
$\T{}$ nodes can be used to generate predictions for the future occurrence of
\textsc{a} or \textsc{b}. The connections to the node corresponding to
\textsc{a} will be incremented by the state of $\T{0}$ when \textsc{a} is
presented; the connections to the node corresponding to
\textsc{b} will be incremented by the state of $\T{0}$ when \textsc{b} is
presented. In the context of Hebbian learning, the prediction for the
stimulus at a future time as a function of $\tau$ and $\tau_o$ is obtained as
the sum of $\T{\delta}$ activity of each node multiplied by the learned
association strength ($\T{0}$):
\begin{equation}
\p{\delta}(\tau,\tau_o) = \sum_{n=\ell}^{N} \T{\delta}\left(\tau,s_n\right)
\ \T{0}\left(\tau_o,s_n\right)/ s_{n}^{w}.
\label{eq:pthetatau}
\end{equation}
The factor $s_n^w$ (for any $w$) allows for differential association
strengths for the different $s$ nodes, while still preserving the scale
invariance property. Since $\delta$ and $\theta_o$ are monotonically related
(eq.~\ref{eq:deltatheta}), the prediction $\p{\delta}$ for various future moments
happens at various phases of a theta cycle.
Recall that all the nodes in the $\T{}$ layer are coherently time-translated
only in the positive half of the theta cycle. Hence for computing future
predictions based on a time-translated state $\T{\delta}$, only coherent nodes
should contribute. In fig.~\ref{fig:theta}, the region to the right of
diagonal line~2 does not contribute to the prediction. The lower
limit $\ell$ in the summation over the nodes given in eq.~\ref{eq:pthetatau}
is the position of the diagonal line~2 in fig.~\ref{fig:theta} marking the
position of discontinuity where $\theta_s$ jumps from $+\pi$ to $-\pi$.
In the limit when $c \rightarrow 0$, the $s$ values of neighboring nodes are
very close and the summation can be approximated by an integral. Defining
$x = s \tau_o$ and $y= \tau/\tau_o$ and $v= \delta/\tau_o$, the
above summation can be rewritten as
\begin{equation}
\p{\delta}(\tau,\tau_o) \simeq \frac{\tau_o^{w-2}}{k!^2} \int_{x_{min}}^{x_u} x^{2k+1-w} (y +v)^k e^{-x(1+y+v)} \, dx
\end{equation}
Here $x_{min} = s_N \tau_o$, and $x_u = s_o \tau_o$ for $0< \theta_o< \pi$ and $x_u= \Phi_{max} \tau_o/\delta$ for $\pi< \theta_o <2 \pi$. The integral can be evaluated in terms of lower incomplete gamma functions to be
\begin{eqnarray}
&\p{\delta}&(\tau,\tau_o) \simeq \frac{\tau_o^{w-2}}{k!^2}
\frac{\left[ (\tau+\delta )/\tau_o \right]^k}{[1+ (\tau+\delta)/\tau_o]^C} \,
\times \nonumber \\
& & \left( \Gamma \left[C, (\tau_o+\tau +\delta) U \right] - \Gamma \left[C, (\tau_o+\tau +\delta) s_N \right] \right),
\label{eq:p}
\end{eqnarray}
where $C=2k+2-w$ and $\Gamma[.,.]$ is the lower incomplete gamma function.
For $\theta_o<\pi$ (i.e., when $\delta < \ensuremath{\Phi_{\textnormal{max}}}/s_o$), $U=s_o$ and for
$\theta_o>\pi$ (i.e., when $\delta > \ensuremath{\Phi_{\textnormal{max}}}/s_o$), $U=\ensuremath{\Phi_{\textnormal{max}}}/\delta$.
\begin{figure}
\includegraphics[width=0.5\textwidth]{newfigs/Deltas_1.png}
\caption{Future timeline. Eq.~\ref{eq:p} is plotted as a function of $\delta$. During training, the \textsc{cs} was presented at $\tau_o=3$ before \textsc{a} and $\tau_o=7$ before \textsc{b}. Left: Immediately after presentation of the \textsc{cs}, the predictions for \textsc{a} and \textsc{b} are ordered on the $\delta$
axis. Note that the prediction for \textsc{b} approximates a rescaled
version of that for \textsc{a}. Right: The prediction for \textsc{b}
is shown for varying times after presentation of \textsc{cs}. With the
passage of time, the prediction of \textsc{b} becomes stronger and
more imminent. In this figure,
$\ensuremath{\Phi_{\textnormal{max}}}=10$, $\ensuremath{\Phi_o}=1$, $k=10$, $s_o=10$, $s_N=1$, and $w=1$.
\label{fig:futurepred}
}
\end{figure}
Figure~\ref{fig:futurepred} provides a graphical representation of some key
properties of eq.~\ref{eq:p}. The figure assumes that the \textsc{cs} is
followed by \textsc{a} after $\tau_o=3$ and followed by \textsc{b} after
$\tau_o=7$. The left panel shows the predictions for both \textsc{a} and
\textsc{b} as a function of $\delta$ immediately after presentation of
\textsc{cs}. The prediction for \textsc{a} appears at smaller
$\delta$ and with a higher peak than the prediction for
\textsc{b}. The value of $w$ affects the relative sizes of the peaks. The
right panel shows how the prediction for \textsc{b} changes with the passage
of time after presentation of the \textsc{cs}. As $\tau$ increases from zero
and the \textsc{cs} recedes into the past, the prediction of \textsc{b} peaks
at smaller values of $\delta$, corresponding to more imminent future times.
In particular when $\tau_o$ is much smaller than the largest (and larger than
the smallest) timescale represented by the nodes, then the
shape of $\p{\delta}$ remains the same when $\delta$ and $\tau$ are rescaled by
$\tau_o$. Under these conditions, the timeline of future predictions generated
by $\p{\delta}$ is scale-invariant.
Since $\delta$ is in one-to-one relationship with $\theta_o$, as a predicted
stimulus becomes more imminent, the activity corresponding to that predicted
stimulus should peak at earlier and earlier phases. Hence a timeline of future
predictions can be constructed from $\p{\delta}$ as the phase $\theta_o$ is
swept from $0$ to $2 \pi$. Moreover the cells representing $\p{\delta}$ should
show phase precession with respect to $\theta_o$. Unlike cells representing
$\T{\delta}$, which depend directly on their local theta phase, $\theta_s$,
the phase precession of cells representing $\p{\delta}$ should depend
on the reference phase $\theta_o$ at the dorsal end of the $s$-axis.
We shall further connect this neurophysiology in the next section
(fig.~\ref{fig:MeerRedi}).
\section{Comparisons with Neurophysiology}
\label{sec:nbio}
The mathematical development focused on two entities $\T{\delta}$ and $\p{\delta}$ that change their value based on the theta phase (eqs.~\ref{eq:bigT}~and~\ref{eq:p}). In order to compare these to neurophysiology, we need to have some hypothesis linking them to the activity of neurons from specific brain regions. We emphasize that although the development in the preceding section was done
with respect to time, all of the results generalize to one-dimensional position as well (eq.~\ref{eq:timespace}, \cite{HowaEtal14}). The overwhelming majority of evidence for phase precession comes from studies of place cells (but see \cite{PastEtal08}). Here we compare the properties of $\T{\delta}$ to phase precession in hippocampal neurons and the properties of $\p{\delta}$ to a study showing phase precession in
ventral striatum \cite{MeerEtal11}.\footnote{This is not meant to preclude the possibility that $\p{\delta}$ could be computed at other parts of the brain as well.}
Due to various analytic approximations, the activity of nodes in the $\T{}$
layer as well as the activity of the nodes representing future prediction
(eqs.~\ref{eq:bigT}~and~\ref{eq:p}) are expressed as smooth functions of time
and theta phase. However, neurophysiologically, discrete spikes (action
potentials) are observed. In order to facilitate comparison
of the model to neurophysiology, we adopt a simple stochastic spike-generating
method. In this simplistic approach, the activity of the nodes given by
eqs.~\ref{eq:bigT}~and~\ref{eq:p} are taken to be proportional to the
instantaneous probability for generating a spike. The probability of
generating a spike at any instant is taken to be the instantaneous activity
divided by the maximum activity achieved by the node if the activity is
greater than 60\% of the maximum activity. In addition, we add spontaneous
stochastic spikes at any moment with a probability of 0.05. For all of the
figures in this section, the parameters of the model are set as $k=10$,
$\ensuremath{\Phi_{\textnormal{max}}}=10$, $w=2$, $\Phi_o = 1$, $s_N=1$, $s_o=10$.
This relatively coarse level of realism in spike generation from the analytic
expressions is probably appropriate to the resolution of the experimental
data. There are some experimental challenges associated with exactly
evaluating the model. First, theta phase has to be estimated from a noisy
signal. Second, phase precession results are typically shown as averaged
across many trials. It is not necessarily the case that the average is
representative of an individual trial (although this is the case at least for
phase-precessing cells in medial entorhinal cortex \cite{ReifEtal12}).
Finally, the overwhelming majority of phase precession experiments utilize
extracellular methods, which cannot perfectly identify spikes from individual
neurons.
\subsection{Hippocampal phase precession}
\begin{figure}
\begin{tabular}{lclc}
\textbf{a} && \textbf{b}\\
&
\includegraphics[height=0.12\textheight]{newfigs/MehtaEtal02}
&&
\includegraphics[height=0.12\textheight]{newfigs/Timecell}
\end{tabular}
\caption{
\textbf{a.} Neurophysiological data showing phase precession. Each spike
fired by a place cell is
shown as a function of its position along a linear track (x-axis) and the
phase of local theta (y-axis). After Mehta, et al., 2002.
\textbf{b.}
Simulated spikes from a node in the $\T{}$ layer described by
eq.~\ref{eq:bigT} as a function of $\tau$ and local phase $\theta_s$.
The curvature is a consequence of eq.~\ref{eq:thetas}.
See text for details.
} \label{fig:ph-pr}
\end{figure}
\nocite{MehtEtal02}
It is clear from eq.~\ref{eq:bigT} that the activity of nodes in the $\T{}$
layer depends on both $\theta_s$ and $\tau$. Figure~\ref{fig:ph-pr} shows
phase precession data from a representative cell (Fig.~\ref{fig:ph-pr}a,
\cite{MehtEtal02}) and spikes generated from eq.~\ref{eq:bigT}
(Fig.~\ref{fig:ph-pr}b). The model generates a characteristic curvature for
phase precession, a consequence of the exponential form of the function
$\Phi$ (eq.~\ref{eq:thetas}). The example cell chosen in fig.~\ref{fig:ph-pr}
shows roughly the same form of curvature as that generated by the model.
While it should be noted that there is some variability across cells, careful
analyses have led computational neuroscientists to conclude that the canonical
form of phase precession resembles this representative cell. For instance, a
detailed study of hundreds of phase-precessing neurons \cite{YamaEtal02}
constructed averaged phase-precession plots using a variety of
methods and found a distinct curvature that qualitatively resembles
this neuron.
Because of the analogy between
time and one-dimensional position (eq.~\ref{eq:timespace}), the model yields
the same pattern of phase precession for time cells and place cells.
\begin{figure}
\includegraphics[width=0.9\columnwidth]{newfigs/Place-new}
\caption{ Place cells along the dorsoventral axis of the hippocampus have
place fields that increase in size. \textbf{a.} The three
panels show the activity of place cells recorded at the dorsal, intermediate
and ventral segments of the hippocampus, when a rat runs along an 18~m track.
After Kjelstrup, et al., (2008). Each spike the cell fired is shown as a
function of position and the \emph{local} theta phase at the cell's location
when it fires (recall that theta phase is not constant across the dorsoventral
axis). Regardless of the width of the place field, neurons at all locations
along the dorsoventral axis phase precess through the same range of {local}
theta phases. \textbf{b.}
According to the model, phase precession extends over the same
range of values of local theta $\theta_s$ regardless of
the value of $s$, which sets the scale for a particular
node. As a consequence, cells with different values of
$s$ show time/place fields of different size but phase
precess over the same range of local theta. For the three
figures, $s$ values of the nodes are set to $.1$, $.22$,
and $.7$ respectively, and they are assumed to respond to
landmarks at location $4$, $11$, and $3$ meters
respectively from one end of the track.
\label{fig:Kjel}
}
\end{figure}
The $\T{}$ layer activity represented in fig.~\ref{fig:ph-pr}a is scale-invariant; note that the $x$-axis is expressed in units of the scale of the node ($k/s$).
It is known that the spatial scale of place fields changes systematically
along the dorsoventral axis of the hippocampus. Place cells in the dorsal
hippocampus have place fields of the order of a few centimeters whereas place
cells at the ventral end have place fields as large as a few meters
(fig.~\ref{fig:Kjel}a) \cite{JungEtal94,KjelEtal08}.
However, all of them show the same pattern of precession with respect to
their local theta phase---the phase measured at the same electrode that
records a given place cell (fig.~\ref{fig:Kjel}). Recall that at any given
moment, the local phase of theta oscillation depends on the position along the
dorsoventral axis \cite{LubeSiap09,PateEtal12}, denoted as the $s$-axis in the
model.
Figure~\ref{fig:Kjel}a shows the activity of three different place cells in
an experiment where rats ran down a long track that extended through open
doors connecting three testing rooms \cite{KjelEtal08}. The landmarks
controlling a particular place cell's firing may have been at a variety of
locations along the track. Accordingly, fig.~\ref{fig:Kjel}b shows the
activity of cells generated from the model with different values of $s$ and
with landmarks at various locations along the track (described in the
caption).
From fig.~\ref{fig:Kjel} it can be qualitatively noted that phase precession
of different cells only depends on the local theta phase and is unaffected by
the spatial scale of firing. This observation is perfectly consistent with the
model.
\subsection{Prediction of distant rewards \emph{via} phase precession in the
ventral striatum}
\begin{figure}
\begin{tabular}{lclc}
\textbf{a} && \textbf{b}\\
& \includegraphics[width=0.45\columnwidth]{newfigs/VandermeerData}
&& \includegraphics[width=0.45\columnwidth]{newfigs/Vandermeer-model}
\end{tabular}
\caption{
\textbf{a.} A representative ramping cell in the ventral striatum.
On each trial the animal started the maze at S, made a series
of turns (T1, T2, etc) and received reward at F1 on 75 percent
of trials. The total distance between S and F1 is on the order
of a few meters. Position along the track is represented
linearly on the x-axis for convenience. In the top panel, the
spikes are shown as a function of theta phase at the dorsal
hippocampus and position. The bottom panel shows the firing
rate as a function of position, which is seen to gradually
ramp up towards the reward location. \textbf{b.} The
activity of prediction node generated by the model is plotted
w.r.t. the reference phase $\theta_o$ and position in the top
panel, and the the average activity within a theta cycle is
plotted against position in the bottom panel.
\label{fig:MeerRedi}
}
\end{figure}
We compare the future predictions generated by the model (eq.~\ref{eq:p}) to
an experiment that recorded simultaneously from the hippocampus and nucleus
accumbens, a reward-related structure within the ventral striatum
\cite{MeerEtal11}. Here the rat's task was to learn to make several turns in
sequence on a maze to reach two locations where reward was available.
Striatal neurons fired over long stretches of the maze, gradually ramping up
their firing as a function of distance along the path and terminating at the
reward locations (bottom fig.~\ref{fig:MeerRedi}a). Many striatal neurons
showed robust phase precession relative to the theta phase at the dorsal
hippocampus (top fig.~\ref{fig:MeerRedi}a). Remarkably, the phase of
oscillation in the hippocampus controlled firing in the ventral striatum to a
greater extent than the phase recorded from within the ventral striatum.
On trials where there was not a reward at the expected location (F1), there was
another ramp up to the secondary reward location (F2), accompanied again by
phase precession (not shown in fig.~\ref{fig:MeerRedi}a).
This experiment corresponds reasonably well to the conditions assumed in the
derivation of eq.~\ref{eq:p}. In this analogy, the start of the trial (start
location S) plays the role of the \textsc{cs} and the reward
plays the role of the predicted
stimulus. However, there is a discrepancy between the methods and the
assumptions of the derivation. The ramping cell (fig.~\ref{fig:MeerRedi}a)
abruptly terminates after the reward is consumed, whereas
eq.~\ref{eq:p} would gradually decay back towards zero. This is
because of the way the experiment was set up--there were never two
rewards presented consecutively. As a consequence, having just
received a reward strongly predicts that there will not be a reward in
the next few moments. In light of this consideration, we force the
prediction generated in eq.~\ref{eq:p} to be zero beyond the reward
location and let the firing be purely stochastic.
The top panel of fig.~\ref{fig:MeerRedi}b shows the spikes generated by model
prediction cells with respect to the reference theta phase $\theta_o$, and the
bottom panel shows the ramping activity computed as the average firing
activity within a complete theta cycle around any moment.
The model correctly captures the qualitative pattern observed in the data.
According to the model, the reward starts being predicted at the beginning of the
track. Initially, the reward is far in the future, corresponding to a large
value of $\delta$. As the animal approaches the location of the reward, the
reward moves closer to the present along the $\delta$ axis, reaching zero near
the reward location. The ramping activity is a consequence of the
exponential mapping between $\delta$ and $\theta_o$ in eq.~\ref{eq:thetas}.
Since the proportion of the theta cycle devoted to large values of $\delta$ is
small, the firing rate averaged across all phases will be small, leading to an
increase in activity closer to the reward.
\subsection{Testable properties of the mathematical model}
\begin{figure}
\includegraphics[width=0.8\columnwidth]{newfigs/Phase-entry}
\caption{Changing $\tau_o$ affects the phase at which prediction cells start
firing. At early times, the magnitude of translation required to
predict the $\tau_o=3$ outcome is smaller than that required to
predict the $\tau_o=7$ outcome. Consequently, the cell begins to fire
at a larger $\theta_o$ for $\tau_o=7$. Parameter values are the same
as the other figures as given in the beginning of this section, except
for clarity the background probability of spiking has been set to
zero. \label{fig:phaseentry}}
\end{figure}
Although the model aligns reasonably well with known properties of theta phase
precession, there are a number of features of the model that have, to our
knowledge, not yet been evaluated. At a coarse level, the correspondence
between time and one-dimensional space implies that time cells should exhibit
phase precession with the same properties as place cells. While phase
precession has been extensively observed and characterized in hippocampal
place cells, there is much less evidence for phase precession in hippocampal
time cells (but see
\cite{PastEtal08}).
According to the model, the pattern of phase precession is related to the
distribution of $s$ values represented along the dorsoventral axis. While it
is known that a range of spatial scales are observed along the dorsoventral
axis, their actual distribution is not known. The Weber-Fechner scale of
eq.~\ref{eq:thetas} is a strong prediction of the framework developed here.
Moreover, since $\ensuremath{\Phi_{\textnormal{max}}}/\ensuremath{\Phi_o} = s_o/s_N$, the ratio of the largest to
smallest scales represented in the hippocampus places constraints on the form
of phase precession. The larger this ratio, the larger will be the value of
$b$ in eq.~\ref{eq:thetas}, and the curvature in the phase precession plots
(as in fig.~\ref{fig:ph-pr}) will only emerge at larger values of the local
phase $\theta_s$. Neurophysiological observation of this ratio could help
evaluate the model.
The form of $\p{\delta}$ (eq.~\ref{eq:p}) leads to several distinctive
features in the pattern of phase precession of the nodes representing future
prediction. It should be possible to observe phase precession for cells that
are predicting any stimulus, not just a reward. In addition, the model's
assumption that a timeline of future predictions is aligned with global theta
phase
has interesting measurable consequences. Let's reconsider the thought
experiment from the previous section (fig.~\ref{fig:futurepred}), where a
stimulus predicts an outcome after a delay $\tau_o$. Immediately after the
stimulus is presented, the value of $\delta$ at which the prediction peaks is
monotonically related to $\tau_o$. Since $\delta$ is monotonically related to
the reference phase $\theta_o$, the prediction cells will begin to fire at later
phases when $\tau_o$ is large, and as time passes, they will fire at earlier
and earlier phases all the way untill $\theta_o=0$. In other other words, the
entry-phase (at which the firing activity begins) should depend on $\tau_o$, the
prediction timescale. This is illustrated in fig.~\ref{fig:phaseentry} with
$\tau_o=3$ and $\tau_o=7$, superimposed on the same graph to make visual
comparison easy. The magnitude of the peak activity would in general depend on
the value of $\tau_o$ except when $w=2$ (as assumed here for visual clarity).
Experimentally manipulating the reward times and studying the phase
precession of prediction cells could help test this feature.
\section{Introduction}
The brain encodes externally observed stimuli in real time and represents
information about the current spatial location and temporal history of recent
events as activity distributed over neural networks. Although we are
physically localized in space and time, it is often useful for us to make
decisions based on non-local events, by anticipating events to occur at
distant future and remote locations. Clearly, a flexible access to the
current state of spatio-temporal memory is crucial for the brain to
successfully anticipate events that might occur in the immediate next moment.
In order to anticipate events that might occur in the future after a given
time or at a given distance from the current location, the brain needs to
simulate how the current state of spatio-temporal memory representation will
have changed after waiting for a given amount of time or after moving through
a given amount of distance. In this paper, we propose that the brain can
swiftly and non-destructively perform space/time-translation operations on the
memory state so as to anticipate events to occur at various future moments
and/or remote locations.
The rodent brain contains a rich and detailed representation of current spatial location and temporal history. Some neurons--\emph{place cells}--in the hippocampus fire in circumscribed locations within an environment, referred to as their \emph{place fields}.
Early work excluded confounds based on visual \cite{SaveEtal98} or olfactory cues \cite{MullKubi87}, suggesting that the activity of place cells is a consequence of some form of path integration mechanism guided by the animal's velocity.
Other neurons in the hippocampus---\emph{time cells}---fire during a
circumscribed period of time within a delay interval
\cite{PastEtal08,MacDEtal11,GillEtal11,KrauEtal13,Eich14}. By analogy to
place cells, a set of time cells represents the animal's current temporal
position relative to past events. Some researchers have long hypothesized a
deep connection between the hippocampal representations of place and time
\cite{EichCohe14,Eich00}.
\begin{figure}
\includegraphics[width=0.98\columnwidth]{newfigs/schematic}
\caption{\textbf{a}.
Theta oscillations of 4-8~Hz are observed in the voltage recorded from the hippocampus. \emph{Hypothesis}: Within a theta cycle, a timeline of future translations of magnitude $\delta$ is constructed. \textbf{b}. Two layer network with theta-modulated connections.
The $\ti{}$ layer receives external input $\ensuremath{\mathbf{f}}{}$ in real time and encodes its Laplace transform. The Laplace transform is inverted via a synaptic operator $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$ to yield an estimate of the function $\ensuremath{\mathbf{f}}{}$ on the $\T{}$ layer nodes. By periodically manipulating the weights in $\ensuremath{\mathbf{L}^{\textrm{\scriptsize{-1}}}_{\textrm{k}}}$, the memory state represented in $\T{}$ layer can be translated to represent its future states.
\label{fig:TILT}}
\end{figure}
Motivated by the spatial and temporal memory represented in the hippocampus,
we hypothesize that the translation operation required to anticipate the
events at a distant future engages this part of the brain
\cite{HassEtal07,SchaEtal07}. We hypothesize
that \emph{theta oscillations}, a well-characterized rhythm of 4-8 Hz in the
local field potential observed in the hippocampus may be responsible for the
translation operation. In particular, we hypothesize that sequential
translations of different magnitudes take place at different phases within a
cycle of theta oscillation, such that a timeline of anticipated future events
(or equivalently a spaceline of anticipated events at distant locations) is
swept out in a single cycle (fig.~\ref{fig:TILT}a).
Theta oscillations are prominently observed during periods of navigation
\cite{Vand69}. Critically, there is a systematic relationship between the
animal's position within a neuron's place field and the phase of the theta
oscillation at which that neuron fires \cite{OKeeRecc93}, known as
\emph{phase precession}.
This suggests that the phase of firing of the place cells conveys
information about the anticipated future location of the animal. This provides a
strong motivation for our hypothesis that the phase of theta oscillation would be
linked to the translation operation.
\subsection{Overview}
This paper develops a computational mechanism for the translation operation of
a spatial/temporal memory representation constructed from a two-layer neural
network model \cite{ShanHowa12}, and links it to theta oscillations by
imposing certain constraints based on some neurophysiological observations and
some physical principles we expect the brain to satisfy. Since the focus here
is to understand the computational mechanism of a higher level cognitive
phenomena, the imposed constraints and the resulting derivation should
be viewed at a phenomenological level, and not as emerging from biophysically
detailed neural interactions.
Computationally, we assume that the memory representation is constructed by a
two-layer network (fig.~\ref{fig:TILT}b) where the first layer encodes the
Laplace transform of externally observed stimuli in real time, and the second
layer approximately inverts the Laplace transform to represent a fuzzy
estimate of the actual stimulus history \cite{ShanHowa12}. With access to
instantaneous velocity of motion, this two layer network representing temporal
memory can be straightforwardly generalized to represent one-dimensional
spatial memory \cite{HowaEtal14}. Hence in the context of this two layer
network, time-translation of the temporal memory representation can be
considered mathematically equivalent to space-translation of the spatial
memory representation.
Based on a simple, yet powerful, mathematical observation that translation
operation can be performed in the Laplace domain as an instantaneous
point-wise product, we propose that the translation operation is achieved by
modulating the connection weights between the two layers within each theta
cycle (fig.~\ref{fig:TILT}b). The translated representations can then be used
to predict events at distant future and remote locations. In constructing
the translation operation, we impose two physical principles we expect the
brain to satisfy. The first principle is \emph{scale-invariance}, the
requirement that all scales (temporal or spatial) represented in the memory
are treated equally in implementing the translation. The second principle is
\emph{coherence}, the requirement that at any moment all nodes forming the
memory representation are in sync, translated by the same amount.
Further, to implement the computational mechanism of translation as a neural
mechanism, we impose certain phenomenological constraints based on
neurophysiological observations. First, there exists a dorsoventral axis in
the hippocampus of a rat's brain, and the size of place fields increase
systematically from the dorsal to the ventral end \cite{JungEtal94,KjelEtal08}.
In light of this observation, we hypothesize that the nodes representing
different temporal and spatial scales of memory are ordered along the dorsoventral
axis. Second, the phase of theta oscillation is not uniform along the
dorsoventral axis; phase advances from the dorsal to the ventral end like
a traveling wave \cite{LubeSiap09,PateEtal12} with a phase difference of about
$\pi$ from one end to the other. Third, the synaptic weights change
as a function of phase of the theta oscillation throughout the hippocampus
\cite{WyblEtal00,SchaEtal08}. In light of this observation, we hypothesize
that the change in the connection strengths between the two layers required to
implement the translation operation depend only on the local phase of the theta oscillation at any node (neuron).
In section~\ref{sec:math}, we impose the above mentioned physical principles
and phenomenological constraints to derive quantitative relationships for the
distribution of scales of the nodes representing the memory and the
theta-phase dependence of the translation operation. This yields specific
forms of phase-precession in the nodes representing the memory as well as the
nodes representing future prediction. Section~\ref{sec:nbio} compares these
forms to neurophysiological phase precession observed in the hippocampus and
ventral striatum. Section~\ref{sec:nbio} also makes explicit
neurophysiological predictions that could verify our hypothesis that theta
oscillations implement the translation operation to construct a timeline of
future predictions.
\input{FutureTimePRX-math}
\input{FutureTimePRX-nbio}
\input{FutureTimePRX-discussion}
\input{FutureTimePRX.bbl}
\end{document}
\section*{ Reviewer~A }
Reviewer~A identified two major weaknesses of the previous submission. First,
he or she noted that we did not provide a ``concrete description of how
the learning is done.'' The revision provides a thorough description of what
we mean by Hebbian association in the first two paragraphs of section II.C.
Briefly, simple
Hebbian association modifies the the synaptic weights between the neurons
representing a stimulus and the $\T{}$-layer neurons at theta phase zero
to an extent proportional to the activity of
the pre- and post-synaptic neurons. Because the post-synaptic neuron
corresponding to the stimulus is presumed to be activated at a constant rate
when the stimulus is presented, the change in these synapses is just proportional
to the activity of the presynaptic neuron.
Second, Reviewer~A felt that the paper was missing ``an evaluation of the
performance of the model, i.e., a demonstration that the whole thing works
from a practical point of view.
How good is the model at learning to predict
future events, and how good is it at using these predictions to improve its
behavior?''
Because it includes a more detailed description of neural data and concrete
neurophysiological predictions, the revision makes it easier to evaluate the
neurophysiological plausibility of the model (see figs.~4-6 and especially
subsection III.D). Computationally, fig.~3 in the revision makes it clear
that the prediction contains information about what will happen when and that
this prediction updates with the passage of time. Moreover, in the revision
we make clear the prediction is scale-invariant when $\delta=0$ and $\tau_o$
is not too extreme. While this places strong constraints on behavioral
performance of the model in decision-making, evaluating the model behaviorally
is a tangential problem beyond the scope of this work.
Reviewer~A also made several specific points
\begin{itemize}
\item Reviewer~A noted that the captions of several figures were too terse.
This has been corrected in the revision.
\item In the revision the operator $\mathbf{R}_{\delta}{}$ is explained in more detail
than in the previous submission.
\item The revision clarifies that $n$ is the index for the values of $s$.
\item Reviewer~A notes that our assumption that negative
phase of theta does not contribute to prediction should be justified more
thoroughly. The revision makes clear that this assumption follows from a
relatively large body of work noting that the synaptic conductance
and LTP are out of phase in the hippocampal formation. See especially
footnote~4.
\item
The reviewer correctly points out that the Weber-Fechner scale should be
logarithmic. That is exactly what we obtain in eq.~11. $n$ is related to $s_n$
logarithmically, and when we invert it $s_n$ will be related to ’$n$’
exponentially as seen in eq.~11.
\item In the revision, we clarify the notation regarding $\T{0}$ and
$\T{\delta}$.
\item In the revision, the prediction vector $\p{\delta}$ is defined more
clearly (see text preceding eq.~14).
\end{itemize}
\section*{ Reviewer~B }
Overall, Reviewer~B was positive about the idea behind the previous
submission, writing ``Thinking about the Hippocampal formation as a network
which can be used for predicting the future using the past experiences is an
interesting and important idea particularly given the discovery of the so
called time cells in this area.'' He or she concluded that the paper needed
``a major re~writing'' to make the paper more self-contained and clarify the
new contribution of the paper. In addition, he or she suggested a network
simulation to compare. Finally, Reviewer~B concluded
``\ldots these could be much better accomplished in a paper written for a
journal where space is not a limitation.''
We agree that the longer format of the revision has made for a much more accessible
presentation. The revision ought to be self-contained and accessible to a
broad interdisciplinary audience; we have rewritten the
paper such that the terminology is clear, and we have given a concise
description of our prior work in section II.A. The remainder of the paper
(from II.B onward) is entirely new. After careful consideration, we have
decided not to include a network simulation of the model in the revision.
Rather, the revision makes clear that the neurobiological constraints are
treated as phenomenological constraints rather than emergent properties of a
circuit. The reader can evaluate whether these phenomenological rules are
reasonable or not. Regardless, the phenomenological approach is appropriate
to a paper that attempts to provide a high-level computational description
of multiple interacting brain regions (including at least hippocampus,
striatum, and orbitofrontal cortex).
Reviewer~B raised several specific concerns.
\begin{enumerate}
\item Reviewer~B noted that the previous submission was not sufficiently
accessible because it relied on the reader's familiarity our previous work.
``Given this, I think the current paper is certainly not accessible to a wide
audience of physicists, and probably not to most of those working on neural
networks.'' All of the necessary background from our prior work is presented
in section II.A of the revision. Thus the revision is self-contained and
ought to be accessible to a broad interdisciplinary audience.
\item
Reviewer~B noted that we did not demonstrate a detailed mapping between
circuit level neurophysiological observations and computational-level
outcomes. ``In my view, although changes in synaptic strength during theta has
been observed, whether theta can actually perform the required modifications
of the synaptic weights that the current model needs is doubtful. Furthermore,
the assumption that the parameter $s$ is mapped anatomically on the
dorsoventral axis also requires further explanation: in any case such a
mapping will not be perfect and the authors should discuss how such
imperfections would change their model.'' In the revision it is clear that we
use the neurophysiological data to place phenomenological constraints on the
equations. If the phenomenological constraints are not valid, then the
derivation does not go
forward. It is far beyond the scope of the current paper to evaluate whether, for
instance, the magnitude of phasic changes in {EPSPs} at individual synapses is
sufficient to result in the changes in
these higher-level equations. For instance, properly evaluating this question
ought to take into account phasic changes in post-synaptic excitability,
phasically modulated inputs from a variety of brain regions, and the
persistence of excitation from a variety of conductances. Most of these
parameters are not known with certainty. Our hope is that publishing a
more abstract computational level paper such as this one clarifies the
important problems that can be solved using traditional methods in
computational neuroscience.
\item In the revision, we use a simple spike-generating mechanism to simulate
individual neurons from the higher-level equations for $\T{\delta}$ and
$\p{\delta}$. This should greatly facilitate comparison with the empricial
data.
\item Reviewer~B notes that bats do not exhibit theta (nor obviously theta
phase precession, which ``suggests that theta
phase precession may be just a rodent artifact.'' First, a thorough review
of theta oscillations in mammalian species is beyond the scope
of this paper. Having said that, monkeys exhibit theta during
eye movements (see especially Jutras, et al., 2013, \emph{PNAS}), memory tasks
(Brincat \& Miller, 2015) and virtual navigation (unpublished data from the
Buffalo lab).
There is also evidence that humans show task-related theta during virtual
navigation (a literature that goes back to Kahana, et al., 1999,
\emph{Nature}). While phase precession has not been reported in other
species, the widespread involvement of theta in similar tasks in primates
makes it seem unlikely that rodents are the only species that show phase
precession. It is possible that bats are the anomaly. It is also possible
that there is some as-yet-unidentified experimental problem that make it
difficult to measure (or evoke) theta in bat hippocampus. In either case,
at this stage the empirical story across species is too murky to justify a
discussion of bat theta in this paper.
\end{enumerate}
\end{document}
Rev A
The authors present a general mechanism to simulate a scaleinvariant future
timeline, and thus make predictions about future events. The neural associate
of such mechanism is theta phase precession in the hippocampus, the phenomenon
that neurons in the hippocampus tend to fire action potentials earlier and
earlier during a 48Hz oscillation cycle of neural activity. Their model
builds on previous work by the authors on a twolayer feed forward network
that forms, on the second layer, a scaleinvariant and distributed
representation of a stimulus presented to the first layer. This is done by
approximately antitransforming the Laplace transforms of the stimulus built
in the first layer, wherein each transform depends on a spatial parameter
localizing the node of the second layer.
In this manuscript, the authors build a future prediction of the stimulus as a
timetranslated memory of the same stimulus. They do so by adding a time
translation operation before anti-transforming, in layer II, the
representation of the input in layer I. On each node, the amount of
translation depends on the phase of the theta cycle. This mechanism seems to
explain data collected in the hippocampus (Fig. 3). Moreover, it is claimed
that it can learned by a simple Hebbian mechanism. The activity at each node
is pulled, after multiplication by the learned association strength, to obtain
a prediction of future state in a scaleinvariant manner.
The topic is extremely important and the work seems competently done. I like
the fact that the authors are trying to solve a fundamental cognitive process
by using a construction that, although abstract, makes contact with the
neurobiology in several key points (see e.g. Fig. 3). The paper is also
generally well written and the narrative proceeds in a logical way.
As for the negative aspects, it seems to me that the paper is lacking in two
main parts: 1. the absence of a concrete (if not biologically plausible)
description of how the learning is done, and 2. an evaluation of the
performance of the model, i.e., a demonstration that the whole thing works
from a practical point of view. How good is the model at learning to predict
future events, and how good is it at using these predictions to improve its
behavior? From this point of view, the work is rather abstract and offers no
evaluation. I also must lament a fair amount of technical detail that, because
of the short format, makes a deeper understanding of the model (beyond the
general idea) quite difficult to grasp. Unpacking the technical information
would probably require the paper to be written for a more technical Journal.
Specific points:
The Figure captions (especially for Fig. 1 and 2) should contain more
details. What’s on the axes?
p. 2, eq. 4: can you define the operator R_{\theta} more precisely? I’m not
sure that the approximation of footnote 1 is warranted.
p. 2: what is ’n’ in Eq. 5?
p. 2: sentences like “Following [10, 11] we assume that the negative
phase of theta does not contribute to prediction” deserve a little explanation.
p. 3: “a coherent time translation across the memory nodes requires the
values sn to lie on a WeberFechner scale”. Strictly speaking, isn’t the
WeberFechner scale supposed to be logarithmic instead of powerlaw?
p. 3, section B: I find it confusing that the authors refer to T_0 and T sometimes as ‘states’, sometimes as ‘layers’ (especially in Section A).
p. 4, eq. 11: what is the definition of p?
Rev B
The paper by Shankar and Howard proposes a neural network model for predicting
the future of a stimulus, postulating that theta oscillations in the
hippocampus can modify the synaptic weights required for their model to work.
Thinking about the Hippocampal formation as a network which can be used for
predicting the future using the past experiences is an interesting and
important idea particularly given the discovery of the so called time cells in
this area. The authors of the current paper motivate their work by mentioning
some experimental findings in the Hippocampus regarding timing of spikes
(phase precession) and the time cells and then move on to building the network
model. My comments are listed below:
1. Unfortunately I must say that this paper is not very wellwritten. Most of
the main idea of the model seem to have been presented in another paper by the
same authors in Neural Computation 2012 (ref. 5) and I had to check on that
paper before starting to comprehend the current one (and I am still unsure if
I completely do so). Given this, I think the current paper is certainly not
accessible to a wide audience of physicists, and probably not to most of those
working on neural networks.
2. The novel contribution of the current paper is relating the model in Neural
Computation 2012 to theta oscillations by postulating theta as a mechanism to
modify the synaptic weight as required by the model to translate the memory
into future ( by the way "translate memory into future" is a term that is not
clearly defined anywhere in the paper and still remains a bit mysterious to
me). In my view, although changes in synaptic strength during theta has been
observed, whether theta can actually perform the required modifications of the
synaptic weights that the current model needs is doubtful. Furthermore, the
assumption that the parameter "s" is mapped anatomically on the dorsoventral
axis also requires further explanation: in any case such a mapping will not be
perfect and the authors should discuss how such imperfections would change
their model.
3. The supplementary material are essentially a review of experimental facts.
I think it would be much more useful if the authors actually showed simulated
networks in which the responses of the network was compared to the
experimental observations.
4. Data from the Bat hippocampus (from Nachum Ulanovsky's lab) show that bat
hippocampus does not show theta nor phase precession which suggests that theta
phase precession may be just a rodent artifact. How does lack of these
oscillations in a species like bat can be interpreted in this model which so
heavily relies on theta?
In summary, this paper does deal with an interesting idea but in the current
format it is not suitable for publication. I think first and foremost, the
paper requires a major rewriting to (a) provide the background material from
Neural Computation 2012, defining the terms used in the paper before they are
used (terms such time translation, time translated memory, future timeline,
scale invariance of the timeline) and (b) clearly state what is novel in the
current paper compared to Neural Comp 2012. The authors also need to show
network simulations both to validate their model in the presence of various
sources of noise, but also to give a better feeling of the model they are
studying to the reader. In my view, all of these could be much better
accomplished in a paper written for a journal where space is not a limitation.
—————————Response to comments of Reviewer-B ————————————————
The major concern of the reviewer is that the paper uses some terminologies
that are not clearly defined and the reader has to go to our earlier papers
for clarification. We have rewritten the paper such that the terminology is
clear, and we have given a broader description of the two layer memory network
that is adopted from our prior work. However, we emphasize that the only
concept needed from our prior papers is that the two layer network can
represent memory by encoding and inverting Laplace transform of the past.
Analysis of this network model, its scale invariance properties, and effects
of various sources of noise have been previously discussed in detail (Neural
computation 2012 and JMLR 2013– ref 5 and 9). As suggested by
the reviewer, we have revised the paper to clearly portray a succinct
description of the two-layer memory model from 2012.
Rather than devoting a lengthy discussion of the prior model, we focus here on
the novel idea of time-translating the memory state (in the Laplace domain) by
modulating the synapses and relating it to theta phase precession, while at
the same time giving it a cognitive-utility of instant prediction of future
events in an ordered-fashion. We believe that our revised manuscript would
let most physicists comfortably accept that our two layer network model can
represent temporal memory by encoding and decoding the Laplace transform of
the past, beyond which the readers can smoothly transition into the novel
story of instant future prediction through time-translated memory states and
their neural relationship to theta oscillations.
Next, the reviewer points out two features where our theoretical model may not
match the actual neurobiology.
(1) The synaptic changes observed during theta oscillations may not be enough
to perform the time-translations required in the model. Very correct; the
value of the parameters \ensuremath{\Phi_{\textnormal{max}}}{} and $k$ in the model will essentially
determine the range to which the synaptic strength must oscillate. The change
in the EPSPs (shown in ref 8,10) does not have to be directly interpreted as
synaptic weight changes, and neither does the synaptic weights of the Hebbian
learning rule have to be literally interpreted in terms of the observed EPSP
modulations. In other words, we only derive a motivation from the empirical
observation of theta modulated synaptic strengths, and we do not intend to
justify the model requirements based on the quantitative empirical
observation. (2) The anatomical organization of the decay constants $s$ along
the dorsovental hippocampal axis would have imperfections that could affect
the model. Very correct again; we have revised the discussion to clarify this
point. This reasoning behind this assumption relates to the fact that the
place-cells in the hippocampus have place-fields that are increasingly larger
as we traverse through the dorsoventral axis (which are analogous to the decay
constants $s$). As a theoretical physics paper, we focus on the
big-picture by qualitatively accepting basic properties observed in the
hippocampal network and use physical principles to theorize a functional
utility.
The reviewer also points out that bats don’t have theta oscillations(although
they are mammals), while they can still predict and navigate
in 3-dimensional flight. Honestly, we do not have an answer for this. But
this is an important issue which we have now addressed in our introduction and
discussion. May be the three-dimensional flight and prediction is very
different from two dimensional navigation and prediction.
|
\section{Introduction}
The neutron rich nuclei of the Segr\'e chart near Ni region has contributed valuable input for
the understanding of nuclear shell evolution. \cite{nature,pcs,Sorlin,sahin,Sieja2,Franco1,Stefanescu1,Dijon} It is demonstrated that
the shell evolution in nuclei mainly have two types, first type one in which
evolution of nuclear shell as a function of $N$ or $Z$ and second type evolution normally occur within the same nucleus.
\cite{otsuka} Recently experimental evidence for the doubly magic ( $Z=20$ and $N=34$) nature of $^{54}$Ca
with the onset of a sizable subshell closure at $N=34$ is reported in Ref. \cite{nature}
In the case of Cu isotopes it has been shown that around $N=40$ the proton single-particle ordering
changes when neutrons start occupying $g_{9/2}$ orbital. \cite{flangan09}
The systematics of the $1/2^-$, $3/2^-$ and $5/2^-$ levels, magnetic
and electric quadrupole moments between $N=40$ to $N=50$ for Ga isotopes is reported in Ref. \cite{cheal10}
For $^{79}$Ga the $\pi f_{5/2}$ orbital dominant in the g.s. While, for $^{81}$Ga, 5/2$^-$ level become ground state. This is due to
emptying of $\pi p_{3/2}$ orbital to $\pi f_{5/2}$ is started as we move from $^{71}$Ga to $^{79}$Ga.
Ultimately, I$^\pi$=5/2$^-$ become ground state for $^{81}$Ga. The systematics of low-lying yrast states
in case of As isotopes is shown in Fig. 1. In case of $^{73-81}$As, the g.s. is $3/2^-$, while for $^{83}$As
again $5/2^-$ become g.s. In the recent work, Porquet et al
~\cite{astier11} suggested that $f_{5/2}pg_{9/2}$ space is not enough to explain quadrupole excitation built on the
5/2$_1^-$ and 9/2$_1^+$ state of $^{81}$As.
Theoretical results for $^{67-79}$As isotopes with projected shell model recently reported in
Ref. \cite{preeti13}
\begin{figure}[h]
\begin{center}
\resizebox{0.8\textwidth}{!}{
\includegraphics{asintro.eps}
}
\caption{\label{f_intro}Low lying yrast states in $^{71}$As to $^{83}$As covering N=40 to N=50 shell closure.
}
\end{center}
\end{figure}
\section{Outline of Calculations}
\label{details}
In the present work we have performed calculations in the $f_{5/2} \,p \,g_{9/2}$ space.
We have performed calculations with recently available effective interactions
JUN45 \cite{Honma09} and jj44b. \cite{brown} In case of JUN45, the single-particle energies for the 1$p_{3/2}$,
0$f_{5/2}$, 1$p_{1/2}$ and 0$g_{9/2}$ orbits are -9.8280, -8.7087, -7.8388, and -6.2617 MeV
respectively. For jj44b interaction they are -9.6566, -9.2859, -8.2695, and -5.8944
MeV, respectively. The core is $^{56}$Ni, i.e. $N = Z = 28$, and the calculations are performed in
this valence space without truncation. The JUN45 interaction is based on Bonn-C potential, the single-particle energies and
two-body matrix elements was modified empirically with A = 63$\sim$69.
Similarly the jj44b interaction was obtained from a fit to binding energies and excitation
energies with 30 linear combinations of the good $J-T$ two-body matrix elements.
The calculations were performed with shell-model codes ANTOINE \cite{Antoine} and NuShellX.\cite{MSU-NSCL}
\section{Spectra analysis}
The shell model results for $^{77,79,81,83}{\rm As}$ isotopes are presented with respect to the experiment in Figs. 2-5.
\subsection{$^{77}{\rm As}$}
Comparison of calculated energy levels of $^{77}$As with experimental data is shown in Fig.~2. The
JUN45 and jj44b interactions predicted $3/2^-$ level as a ground state which is in good agreement with experiment.
Level $1/2_1^-$ is the first excited state predicted by JUN45 and jj44b. The calculations predicting $1/2_1^-$ level lower
than the experiment. The sequence of the
experimentally observed energy
levels $3/2_2^-$ and $5/2_1^-$ are same in jj44b while it is interchanged in JUN45. The overall values
of energy levels of $3/2_2^-$ and $5/2_1^-$ with jj44b is in good agreement with experiment. The
calculated $3/2_2^-$ energy level by two effective interactions are higher than experiment. The JUN45 predict
$1/2_2^-$ level at 411 keV, jj44b at 1181 keV, but experiment predicted it at 2195 keV. The
levels $9/2_1^-$ and $7/2_2^-$ calculated by both interactions are in good agreement with experiment. While
levels $5/2_2^-$ and $13/2_1^-$ predicted by two interactions are lower than the experiment. The $7/2_3^-$ level
calculated by JUN45 is in good agreement with experiment while with jj44b it is 303 keV higher
than the experiment.
The sequence of the calculated high-spin negative parity energy levels by JUN45 and jj44b interactions are in
good agreement with experiment.
In the case of positive parity, JUN45 and jj44b predict $9/2_1^+$ as a lowest positive parity state and $5/2_1^+$ as
a second positive parity state which is in good agreement with experiment but these levels are higher than the experimental values.
The level $7/2_1^+$ calculated by JUN45 and jj44b are higher than the
experiment and $7/2_2^+$ level is lower with both JUN45 and jj44b.
The energy separation between $7/2_1^+$ and $7/2_2^+$ is 197 keV in JUN45, 204 keV in jj44b while in the experiment it is 1194 keV.
Experimentally observed two consecutive levels,
$5/2_2^+$ and $5/2_3^+$ are higher
in both calculations and there are five levels $(7/2_1^+,11/2_1^+,13/2_1^+,7/2_2^+ and 9/2_2^+)$
between them in jj44b.
The JUN45 and jj44b interactions predicted $\pi(p_{3/2}^3)$ configuration for g.s. $3/2_1^-$ with probability 15.7\%
and 13.6\%, respectively.
The structure of negative parity states from $3/2_1^-$ to $19/2_1^-$ is mainly from $\pi(p_{3/2}f_{5/2})^5 $ configuration.
In the case of positive parity, the structure for $9/2_1^+$ state is $\pi(g_{9/2}^1)$, with probability 11.4\% (JUN45)
and 16.9\% (jj44b). The occupancy of $\pi g_{9/2}$ orbital for $9/2_1^+$ state is 1.06 (JUN45) and 1.07 (jj44b).
\subsection{$^{79}{\rm As}$}
In the Fig.~3, we have shown the comparison of the values of the energy levels calculated by
JUN45 and jj44b with experimental data for $^{77}$As. The JUN45 predict $3/2^-$ level
as a ground state which is in good agreement with experiment, where as jj44b predicts $1/2^-$ as a
ground state which is different from experiment. The levels $3/2_1^-$ and $1/2_1^-$ are interchanged in jj44b.
The levels $3/2_2^-$ and $5/2_1^-$ calculated
by jj44b are in good agreement with experiment values. The calculted $3/2_3^-$ level with jj44b is in good
agreement with experiment while with JUN45 it is lower by 414
keV than the experiment.
The calculated
value of level $5/2_3^-$ with both interactions are lower than the experiment.
The calculated positive parity levels are higher with both interactions than the
experiment. For positive parity, the JUN45 and jj44b predicts $9/2^+$ as lowest positive parity state which is also predicted by experiment
but it is higher by 656 keV in JUN45 and by 1011 keV in jj44b than the experiment.
As in the experiment the next positive parity state is $5/2^+$ by JUN45, while it is at higher energy
with jj44b interaction.
In the experiment after $9/2_2^+$ next one is $9/2_3^+$
but in between these two levels JUN45 predict two more levels ($5/2_2^+$ at 2345 keV and $5/2_3^+$ at 2374 keV),
while in jj44b five levels ($1/2_1^+$ at 2129 keV,
$5/2_1^+$ at 2194 keV, $5/2_2^+$ at 2221 keV, $1/2_2^+$ at 2270 keV, $5/2_3^+$ at 2402 keV).
In jj44b calculation the $1/2_1^+$ and $1/2_2^+$ levels are lower than $9/2_3^+$ in comparison
to the experiment value.
The JUN45 and jj44b interactions predicted $\pi(p_{3/2}^1)$ configuration for g.s. ($3/2_1^-$) with probability 12.7\%
and 10.1\%, respectively.
The structure of negative parity states from $3/2_1^-$ to $19/2_1^-$ are mainly from $\pi(p_{3/2}f_{5/2})^5 $ configuration.
In the case of positive parity, the structure for $9/2_1^+$ state is $\pi(g_{9/2}^1)$, with probability 16.4\% (JUN45)
and 12.4\% (jj44b). The occupancy of $\pi g_{9/2}$ orbital for $9/2_1^+$ state is 1.07 (JUN45) and 1.06 (jj44b).
\begin{figure*}
\begin{center}
\resizebox{1.2\textwidth}{!}{
\includegraphics{77As.eps}
}
\caption{\label{f_se78}Comparison of experimental and calculated excitation spectra of $^{77}$As.}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\resizebox{1.2\textwidth}{!}{
\includegraphics{79As.eps}
}
\caption{\label{f_se80}Comparison of experimental and calculated excitation spectra of $^{79}$As.}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\resizebox{1.2\textwidth}{!}{
\includegraphics{81As.eps}
}
\caption{\label{f_se82}Comparison of experimental$^{6}$ and calculated excitation spectra of $^{81}$As.}
\end{center}
\end{figure*}
\begin{figure*}[h]
\begin{center}
\resizebox{1.2\textwidth}{!}{
\includegraphics{83As.eps}
}
\caption{\label{f_se82}Comparison of experimental$^{6}$ and calculated excitation spectra of $^{83}$As.}
\end{center}
\end{figure*}
\subsection{$^{81}{\rm As}$}
In the Fig. 4~, we have shown results for $^{81}$As,
the calculated $3/2_2^-$ and $5/2_1^-$ levels by jj44b are in good agreement with experiment, while with JUN45
these levels are interchanged.
The JUN45 and jj44b predicted higher value of $7/2_1^-$ by 142 keV, 151 keV, respectively,
and lower value of $3/2_3^-$ by 2026 keV, 1689 keV respectively than the experiment. The
level $7/2_2^-$ predicted by JUN45 is higher by 94 keV, but it is further higher by 279 keV in jj44b. The
calculated $9/2_1^-$ and $13/2_1^-$ levels by JUN45 are in good agreement with experiment but it is predicted
at higher value by jj44b. The level $17/2_1^-$ calculated by both interactions are in good agreement with experiment.
The JUN45 and jj44b predicted $21/2_1^-$ level higher by 734 and 882 keV, respectively than the experimental result.
For the positive parity, JUN45 and jj44b predicted $9/2^+$ as a lowest positive parity state
as in the experiment but this level is at higher energy than the experimental value.
Both the calculations predicting higher value of the level $13/2_1^+$ than the experimental value.
As we approach towards $N=50$, it is now important to include $\nu 1d_{5/2}$ orbital in the model space
to study neutron excitation across $N=50$ shell.
\subsection{$^{83}{\rm As}$}
We have shown results for $^{83}$As in Fig.~5. Both the interactions
JUN45 and jj44b predicted g.s. as a $5/2^-$. The levels
$9/2_1^-$ and $11/2_1^-$ are lowered by 132 and 141 keV in JUN45 whereas both levels are
higher by 165 and 173 keV in jj44b.
With JUN45 $13/2_1^-$ level is in good agreement while in
jj44b it is higher by 154 keV than the experiment.
The sequence of energy levels $5/2_1^-$, $9/2_1^-$,
$11/2_1^-$ and $13/2_1^-$ in jj44b are in good agreement with experiment.
The JUN45 and jj44b predicts $9/2^+$ as a lowest positive parity state. While there is no
experimental result is available for positive parity. All the positive parity states predicted by jj44b is lower than JUN45.
\section{ Competition of proton- and neutron-pair breakings analysis}
In Figs. 6-7 we have shown decomposition of the total angular momentum of the selected states for $^{79,81}$As isotopes with JUN45
interaction.
From the analysis of the wavefunctions it is possible to identify which nucleon pairs are broken to obtain the total angular
momentum of the calculated states. The two components for neutrons and protons are $I_{n}$ and $I_{p}$ respectively. These
components are coupled to give the total angular momentum of each states. In the Fig. 6(a) -(d), we have shown results of negative parity
states of $^{79}$As. The dominant component ( 46 \%) of the
$3/2_1^-$ ground state comes from $I_{p}$ = 3/2. The $1/2_1^-$ comes from $I_{p}$ = 1/2 (40 \%), $5/2_1^-$ comes from $I_{p}$ = 5/2 (47 \%),
$7/2_1^-$ comes from $I_{p}$ = 7/2 (28 \%).
The $17/2_1^-$ state has a very peculiar wave function. It shows many different components (35 \% of 5/2$_p^-$ $\otimes$ 6$_{n}^+$, 18 \% of
9/2$_p^-$ $\otimes$ 4$_{n}^+$ , 14 \% of 3/2$_p^-$ $\otimes$ 8$_{n}^+$ , ...), thus resembling a ''collective'' state.
The wave function of $19/2_1^-$ state ( predicted at 3396 keV) have the major one corresponding to
$I_{p}$ = 3/2 $\otimes$ $I_{n}$ = 8 ( 44 \%)
and the minor one to $I_{p}$ = 7/2 $\otimes$ $I_{n}$ = 6 ( 14 \%).
Further the wave function of $21/2_1^-$ state ( predicted at 3776 keV) have the major one corresponding to $I_{p}$ = 5/2 $\otimes$ $I_{n}$ = 8 ( 57 \%)
and the minor one to $I_{p}$ = 9/2 $\otimes$ $I_{n}$ = 6 ( 11 \%).
In the Fig. 6(e) -(h), we have shown results of positive parity
states of $^{79}$As. The major component of $9/2_1^+$ comes from $I_{p}$ = 9/2 (40 \%). Similarly
$5/2_1^+$ comes from $I_{p}$ = 5/2 (30 \%).
\begin{figure}
\resizebox{160mm}{!}{\includegraphics{79As_part.eps}}
\caption{
Decomposition of the total angular momentum of selected states of $^{79}$As into their $I_n \otimes I_p$ components. The percentage above 10\%
are written inside the squares, drawn with an area proportional to it. Percentage below 5\% are not written.}
\label{f_79Aspart}
\end{figure}
In the Fig. 7(a) -(d), we have shown results of negative parity
states of $^{81}$As. The $3/2_1^-$ comes from $I_{p}$ = 3/2 (55 \%). The $17/2_1^-$ comes mainly from $I_{n}$ = 8 ( with
$I_{p}$ = 1/2 - 7/2 ), i.e. due to neutron pair breaking. Similarly the $21/2_2^-$ ( at 4403 keV) from $I_{n}$ = 8 ( with
$I_{p}$ = 5/2 - 9/2 ). The $13/2_1^-$ shows many different components (43 \% of 9/2$_p^-$ $\otimes$ 2$_{n}^+$, 23 \% of
5/2$_p^-$ $\otimes$ 4$_{n}^+$ , 16 \% of 13/2$_p^-$ $\otimes$ 0$_{n}^+$ , ...), thus resembling a ''collective'' state.
The positive parity states results are shown in Fig. 7(e) -(h). The major component of $9/2_1^+$ comes from $I_{p}$ = 9/2 (59 \%).
The $9/2_2^+$ and $17/2_1^+$ is coming from pure neutron breaking. The $1/2_1^+$ shows many different components (65 \%
of 5/2$_p^-$ $\otimes$ 3$_{n}^+$, ...), thus resembling a ''collective'' state.
The above three families are drawn with three different colors, the magenta color is for breaking
of neutron pairs, the green color is for that of protons and blue color is for many components
with various values of $I_{n}$ and $I_{p}$.
\begin{figure}
\resizebox{160mm}{!}{\includegraphics{81As_part.eps}}
\caption{
Decomposition of the total angular momentum of selected states of $^{81}$As into their $I_n \otimes I_p$ components. The percentage above 10\%
are written inside the squares, drawn with an area proportional to it. Percentage below 5\% are not written.}
\label{f_81Aspart}
\end{figure}
\section{\label{ep} Electromagnetic properties and occupation numbers analysis}
\subsection{\label{ep} $E2$ transition probability, quadrupole and magnetic moments}
In Table 1 and 2, we have shown calculated $B(E2)$ and $B(M1)$ values for
different transitions. Although experimental data are very sparse.
For $B(E2$:$5/2_1^+$ $\rightarrow$ $9/2_1^+$), the result of jj44b
interaction is better than JUN45.
It is further improve by increasing effective charges.
The experimental $B(M1)$ values
are only available for $^{79}$As, for BM1($5/2_1^- \rightarrow 3/2_1^-$)
transition the predicted value of JUN45 interaction is close to experimental value.
In Table 3, we have also compared results of quadrupole and
magnetic moments.
The overall results of magnetic moments are in good
agreement with the available experimental data and also the differences between the results obtained by the two adopted
interactions are reasonable. On the other hand it is not possible to draw some
definite conclusions for the $B(E2)$ and quadrupole moments due to lack of
experimental data and contradictory results for the $B(E2)$ values.
The calculated results for $B(E2)$ and quadrupole moments for two interactions are
different. This may be because two interactions were obtained from a fit with experimental data of different set of nuclei.
The JUN45 interaction derived by fitting 400 experimental binding
and excitation energy data out of 69 nuclei in the $A =
63 \sim 96$ mass region while jj44b interaction developed by fitting binding
energies and excitation energies from nuclei with $Z = 28 - 30$
and $N = 48 - 50$. The jj44b results are better near $Z = 28$.
The JUN45 interaction is successful along
the $N \sim 50$ isotone chains, while for the Ni region,
the results are not satisfactory, because of the exclusion of
the effect of $\pi f_{7/2}$ excitations. \cite{Honma09}
\subsection{\label{on} Occupation numbers}
In Fig. 8, we show the proton/neutron occupation numbers.
For the ground state ($3/2^-$),
the occupancy of $ \pi (0f_{5/2})$ orbital increase smoothly, while occupancy of
$\pi(1p_{3/2})$ orbital is decreasing. For $9/2^+$, the occupancy of $\pi (0g_{9/2})$ orbital is significant.
In the case of neutrons orbital the occupancy of the $\nu (0g_{9/2})$ orbital increases drastically
as neutrons number increase from $N=44$ ($^{77}$As)
to $N=50$ ($^{81}$As).
\section{\label{sec5}Summary}
\label{conc}
In summary, comprehensive study for the structure of
neutron-rich odd-even As isotopes have been carried out using large-scale shell-model calculations for
full $f_{5/2}pg_{9/2}$ space with JUN45 and jj44b effective interactions.
The overall results
for the energy levels and magnetic moments are in good
agreement with the available experimental data. The experimental data
for $B(E2)$ and quadrupole moments are not available thus
it is not possible to draw some definite conclusions.
We have also predicted electromagnetic properties for
other transitions as a guide for future experimental work.
Further, the following broad conclusions are:
\begin{itemize}
\item The results of JUN45 interaction is better than jj44b.
\item High-spin states in $^{77,79,81,83}$As isotopes comes from breaking of neutron/proton pairs.
\item Further theoretical development is needed by enlarging model
space by including $\pi 0f_{7/2}$ and $\nu 1d_{5/2}$ orbitals to study simultaneously
proton and neutron excitations across $Z=28$ and $Z=50$ shell,
respectively.
\end{itemize}
\begin{landscape}
\begin{table}
\begin{center}
\title{Table 1. $B(E2)$ reduced transition strength in W.u. Effective charges
$e_p=1.5e$, $e_n=0.5e$ / $e_p=1.5e$, $e_n=1.1e$ were used. Experimental values were taken from
the NNDC database.}
\label{tab:table2}
\begin{tabular}{ c | c | c | c | c } \hline
& $^{77}$As &$^{79}$As &$^{81}$As & $^{83}$As \\ \hline
BE2($5/2_1^- \rightarrow 3/2_1^-$) && & & \\ \hline
Experiment & N/A & N/A & N/A & N/A \\
JUN45 & 5.63 / 9.51 & 1.48 / 2.42 & 0.95 / 1.15 & 0.12 / 0.1818 \\
jj44b &20.92 / 36.90 & 0.06 / 0.11& 1.15 / 1.35& 0.01 / 0.0212 \\
&& & & \\ \hline
BE2($3/2_3^- \rightarrow 3/2_1^-$) && & & \\ \hline
Experiment & $>$ 4.4 & N/A & N/A & N/A \\
JUN45 & 0.42 / 0.66 & 4.19 / 6.67 & 1.56 / 2.62 & 0.89 / 0.8905 \\
jj44b & 0.07 / 0.074 & 1.73 / 3.16 & 1.36 / 2.34 & 4.78 / 4.78 \\
&& & & \\ \hline
BE2($7/2_1^- \rightarrow 5/2_1^-$) && & & \\ \hline
Experiment & N/A & N/A & N/A & N/A \\
JUN45 & 2.73 / 4.31 & 4.10 / 6.52 & 0.04 / 0.0365 & 2.58 / 2.5839 \\
jj44b &5.78 / 9.75 & 0.94 / 1.48 & 0.09 / 0.1162 & 0.17 / 0.0167 \\
&& & & \\ \hline
BE2($9/2_1^- \rightarrow 7/2_1^-$) && & & \\ \hline
Experiment & N/A & N/A & N/A & N/A \\ \hline
JUN45 & 1.62 / 2.52 & 1.72 / 2.39 & 0.40 / 0.4568& 0.52 / 0.5231\\
jj44b & 7.09 /12.33 & 0.11 / 0.18 & 0.08 / 0.0668& 0.58 / 0.5831 \\ \hline
BE2($5/2_1^+ \rightarrow 9/2_1^+$) && & & \\ \hline
Experiment & 80(16) & N/A & N/A & N/A \\
JUN45 & 19.39 / 29.78 & 17.79 /27.426 & 12.10 / 18.14 & 5.83 / 5.83\\
jj44b & 23.24 / 38.70 & 7.39 / 11.59 & 12.20 / 18.03& 6.67 /6.67\\
& & & & \\ \hline
BE2($5/2_3^+ \rightarrow 5/2_1^+$) && & & \\ \hline
Experiment & $>$ 0.00031 & N/A & N/A & N/A \\
JUN45 & 0.004 / 0.006 & 0.22 / 0.36& 0.21 / 0.30 & 0.46 / 0.46 \\
jj44b & 0.011 /0.021 & 0.26 / 0.49& 0.64 / 1.065& 1.53 / 1.53\\
& & & & \\ \hline
\end{tabular}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\begin{center}
\title{Table 2. $B(M1)$ values for different transitions in W.u. In the present calculation
$g_s^{eff}$=$g_s^{free}$ / $g_s^{eff}$ = 0.7 $g_s^{free}$ were used. Experimental values were taken from
the NNDC database.}
\label{tab:table2}
\begin{tabular}{ c | c | c | c | c } \hline
& $^{77}$As &$^{79}$As &$^{81}$As & $^{83}$As \\ \hline
BM1($3/2_2^- \rightarrow 3/2_1^-$) && & & \\ \hline
Experiment & $<$ 0.0041 & N/A & N/A & N/A \\
JUN45 & 0.0976 / 0.4782 & 0.0439 / 0.0215 & 0.0414 / 0.0203 & 0.0017 / 0.00083 \\
jj44b &0.0907 / 0.0444 & 0.0021 / 0.0011 & 0.0441 / 0.0216 & 0.3335 / 0.01633 \\
&& & & \\ \hline
BM1($3/2_3^- \rightarrow 3/2_1^-$) && & & \\ \hline
Experiment & $>$ 0.0070 & N/A & N/A & N/A \\
JUN45 & 0.0606 / 0.0297 & 0.0060 / 0.0029 & 0.0864 / 0.0423& 0.0080 / 0.0079\\
jj44b & 0.0197 / 0.0097 & 0.0027 / 0.0013 & 0.0183 / 0.0090 & 0.0025 / 0.0012 \\
&& & & \\ \hline
BM1($5/2_1^- \rightarrow 3/2_1^-$) && & & \\ \hline
Experiment & 0.00193(4) & N/A & N/A & N/A \\
JUN45 & 0.0016 / 0.00078 & 0.0002 / 0.00011 & 0.0001 / 0.000055 & 0.0002 / 0.00011\\
jj44b &0.0193 / 0.0094 & 0.00005 / 0.0000558 & 0.0006 / 0.000279 & 0.0001 / 0.0000558 \\
&& & & \\ \hline
BM1($5/2_3^+ \rightarrow 5/2_2^+$) && & & \\ \hline
Experiment & $>$ 1.9 x 10$^{-5}$ & N/A & N/A & N/A \\
JUN45 & 0.0660 / 0.0312 & 0.0326 / 0.0159 & 0.000005 / 0.0000 & 0.0012 / 0.0006\\
jj44b & 0.000503 / 0.0002 & 0.1088 / 0.0540 & 0.0227 / 0.0587 & 0.00006 / 0.00005 \\
& & & & \\ \hline
\end{tabular}
\end{center}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}
\begin{center}
\title{Table 3. Comparison of calculated and experimental magnetic and quadrupole moments. The magnetic moments, (in $\mu_N$), with $g_s^{eff}$ = $g_s^{free}$ / $g_s^{eff}$ = 0.7$g_s^{free}$
and electric quadrupole moments, $Q_s$ (in eb), with $e_p$=1.5e, $e_n$=0.5e / $e_p$=1.5e, $e_n$=1.1e. }
\label{tab:table3}
\begin{tabular}{ c | c | c | c | c } \hline
&$^{77}$As &$^{79}$As &$^{81}$As & $^{83}$As\\ \hline
$\mu$ ($3/2_1^-$) && & & \\ \hline
Experiment & +1.2946(13) & N/A & N/A & N/A \\
JUN45 & +1.941 / +1.358 & +2.825 / +1.977 & +3.090 / +2.167 &+3.666 / +2.566\\
jj44b & +1.477 / +1.034 & +2.668 /+1.868 & +2.985 / +2.089 & +3.569 / +2.498\\ \hline
$\mu$($5/2_1^-$) && & & \\ \hline
Experiment & +0.736 (22) & N/A & N/A & N/A \\
JUN45 & +0.473 / 0.331 & +0.580 / +0.406 & +0.788 / +0.552 & +0.915 / +0.641\\
jj44b & +0.406 / +0.284 & +0.422 /+0.296 & +0.541 /+0.379 & +0.875 / +0.612 \\ \hline
$\mu$($9/2_1^+$) & & & & \\ \hline
Experiment & +5.525(9)& N/A & N/A & N/A \\
JUN45 & +5.953 / +4.167 & +6.076 / +4.252 & +6.335 / +4.435 & +6.69 / +4.683\\
jj44b & +5.945 / + 4.1617 & +6.032 / +4.222 & +5.976 / +4.183 & +6.53 / +4.571 \\
& & & &\\ \hline \hline
&$^{77}$As &$^{79}$As &$^{81}$As & $^{83}$As \\ \hline
Q($3/2_1^-$) && & & \\ \hline
Experiment & N/A & N/A & N/A & N/A \\
JUN45 & +0.26 / +0.32 & -0.16 / -0.21 & -0.21 /-0.25 & -0.12 /-0.12\\
jj44b & +0.27 / +0.35 & -0.06 / -0.08 & -0.19 / -0.23 & -0.11 /-0.11\\ \hline
Q($5/2_1^-$) && & &\\ \hline
Experiment & $<$ 0.75 & N/A & N/A & N/A \\ \hline
JUN45 & -0.05 / -0.06 & -0.05 / -0.07 & -0.01 / -0.03 & +0.13 / +0.13\\
jj44b & -0.09 / -0.11 & -0.10 / -0.13 & -0.10 /-0.14 & +0.14 /+0.14 \\ \hline
Q($9/2_1^+$) & & & & \\ \hline
Experiment & N/A & N/A & N/A & N/A \\
JUN45 & -0.74 / -0.97 & -0.68 / -0.83 & -0.538 / -0.612 & -0.387 / -0.386 \\
jj44b &-0.75 / -0.97 & -0.71/ -0.87 & -0.536 / -0.603 & -0.447 / -0.446 \\
& & & &\\ \hline
\end{tabular}
\end{center}
\end{table}
\end{landscape}
\begin{figure}[h]
\begin{center}
\resizebox{120mm}{!}{\includegraphics{oddas-proton-occ.eps}}
\begin{center}
(a) Proton
\end{center}
\resizebox{120mm}{!}{\includegraphics{oddas-neutron-occ.eps}}
\begin{center}
(b) Neutron
\end{center}
\caption{\label{Fig10}(Color online) Proton/Neutron occupation numbers of the JUN45 and jj44b
($p_{3/2}$, $f_{5/2}$, $p_{1/2}$ and $g_{9/2}$ -shell orbits ) interactions -
for two low-lying states in $^{77,79,81,83}$As isotopes. Upper
panel for $3/2_1^-$; lower panel for $9/2_1^+$.
}
\end{center}
\label{f_82ge}
\end{figure}
{\bf Acknowledgment}
VK acknowledges partial financial support from CSIR, India for his PhD thesis work.
\bibliographystyle{elsarticle-num}
|
\section{The combination problem in concept theory\label{intro}}
That concepts exhibit aspects of `contextuality', `vagueness' and `graded typicality' was already known in the seventies since the
investigations of Eleanor Rosch \cite{r1973}. These studies questioned explicitly the traditional view that `concepts are containers of instantiations' and, additionally, although not explicitly stated, there was already the suspicion that `the human mind combines concepts not following the algebraic rules of classical logic even if the combinations are conjunctions or disjunctions'. In particular, conceptual gradeness led scholars to introduce elements of probability theory in structuring and representing concepts. A possible way to at least preserve a set-theoretical basis was the fuzzy set approach \cite{z1989}. According to this proposal, concepts would combine in such a way that the conjunction of two concepts satisfies the `minimum rule of fuzzy set conjunction' and the disjunction of two concepts satisfies the `maximum rule of fuzzy set disjunction'. In this way, one could still maintain that `concepts can be represented as (fuzzy) sets and combine according to set-theoretic rules'. However, a whole set of experimental findings in the last thirty years revealed that the latter does not hold, thus raising the so-called `combination problem'.
(i) `Guppy effect'. Osherson and Smith measured the `typicality' of specific exemplars with respect to the concepts {\it Pet} and {\it Fish} and their conjunction {\it Pet-Fish} \cite{os1981}, and they found that people rate an exemplar such as {\it Guppy} as a very typical example of {\it Pet-Fish}, without rating {\it Guppy} as a very typical example neither of {\it Pet} nor of {\it Fish} (`Pet-Fish problem').\footnote{In a typicality measurement, subjects are asked to choose the exemplar they consider as the most typical example of a given concept, hence they have to pick the best example in a list of items.} Interestingly enough, this guppy effect violates the minimum rule of fuzzy set conjunction.
(ii) `Overextension and underextension effects'. Hampton measured the `membership weight' of several exemplars with respect to specific pairs of concepts and their conjunction \cite{h1988a} and disjunction \cite{h1988b}, finding systematic deviations from fuzzy set modeling. Adopting his terminology, if the membership weight of an exemplar $x$ with respect to the conjunction `$A \ {\rm and} \ B$' of two concepts $A$ and $B$ is higher than the membership weight of $x$ with respect to one concept (both concepts), we say that the membership weight of $x$ is `overextended' (`double overextended') with respect to the conjunction (briefly, $x$ is overextended with respect to the conjunction). If the membership weight of an exemplar $x$ with respect to the disjunction `$A \ {\rm or} \ B$' of two concepts $A$ and $B$ is less than the membership weight of $x$ with respect to one concept (both concepts), we say that the membership weight of $x$ is `underextended' (`double underextended') with respect to the disjunction (briefly, $x$ is underextended with respect to the disjunction). We have recently performed a similar cognitive test on conceptual conjunctions of the form `$A$ and $B$' \cite{s2014b,asv2014}, detecting systematic overextension and also double overextension.\footnote{In a membership measurement, subjects are asked to decide whether a given exemplar $x$ is a member of a given concept $A$. When many subjects are involved in the measurement, a membership weight $\mu(A)$ can be defined for $x$ as a large number limit of the relative frequency of positive answers.}
(iii) `Deviations from classicality in conceptual negation'. More recently, Hampton measured the membership weights of many exemplars with respect to specific pairs of concepts and their conjunction, e.g., {\it Tools Which Are Also Weapons}, and also conjunction when the second concept is negated, e.g., {\it Tools Which Are Not Weapons} \cite{h1997}. He detected overextension in both types of conjunctions, as well as deviations from classical logical behaviour in conceptual negation. We have recently performed a more general cognitive test \cite{s2014b,asv2014}, detecting systematic overextension, double overextension and violation of classical logic negation rules in conceptual conjunctions of the form `$A$ and not $B$', `not $A$ and $B$' and `not $A$ and not $B$'.
(iv) `Borderline contradictions'. Alxatib and Pelletier asked human subjects to estimate the truth value of a sentence such as ``$x$ is tall and not tall'' for a given person $x$ who was showed to the eyes of the subjects \cite{ap2011}. They found that a significant number of subjects estimated such a sentence as true, in particular, for borderline cases.\footnote{A borderline contradiction can be formalized as a sentence of the form $P(x) \land \lnot P(x)$, for a vague predicate $P$ and a borderline case $x$, e.g., the sentence ``Mark is rich and Mark is not rich''.}
Difficulties (i)--(iv) entail, in particular, that the formation and combination rules of human concepts do not generally follow the laws of classical (fuzzy set) logic \cite{os1981,h1988a,h1988b,h1997}. Moreover, the corresponding experimental data cannot generally be modeled in a single classical probability space satisfying the axioms of Kolmogorov, which we proved in various articles \cite{s2014b,a2009a,ags2013,s2014a}.
Our investigation of the above `deviations from classicality'\footnote{By the locution `deviation from classicality' we actually mean that
classical logical and probabilistic structures, i.e. the most traditional models of cognition, cannot account for the experimentally observed patterns} in conceptual combinations can be traced back to our studies on the axiomatic and operational foundations of quantum physics and the origins of quantum probability (see, e.g., \cite{a1986}). We recognized that any decision process, e.g., a typicality measurement, or a membership estimation, involves a `transition from potential to actual', in which an outcome is actualized from a set of possible outcomes as a consequence of a contextual interaction (of a cognitive nature) of the subject with the conceptual situation that is the object of the decision. Hence, human decision processes exhibit deep analogies with what occurs in a quantum measurement process, where the measurement context (of a physical nature) influences the measured quantum particle in a non-deterministic way. Quantum probability -- which is able to formalize this `contextually driven actualization of potential', not classical probability -- which only formalizes lack of knowledge about actuality -- can conceptually and mathematically cope with this situation underlying both quantum and conceptual realms \cite{aa1995}.
The second step of our research was the elaboration of a `State Context Property' (SCoP) formalism to abstractly represent any type of entity, e.g., a conceptual entity, in terms of its states, contexts and properties. In SCoP, a concept is represented as an `entity being in a specific state and changing under the influence of a cognitive context', rather than as a `container of instantiations', and we were able to provide a quantum-theoretic model in Hilbert space that successfully describes the guppy effect \cite{ag2005a,ag2005b} (Section \ref{contextuality}).
The successive development of our research was the employment of the mathematical formalism of quantum theory in Fock space to model the overextension and underextension of membership weights measured in \cite{h1988a,h1988b}. These effects can be described in terms of genuine quantum aspects, like `interference', `superposition' and `emergence' \cite{a2009a,ags2013,a2009b,abgs2013}. This quantum-mechanical model was successfully applied to describe borderline contradictions \cite{s2014a}. More recently, we extended the model to incorporate conceptual negation, thus faithfully representing the above mentioned experiments by ourselves on concept conjunctions and negations \cite{s2014b,asv2014} (Section \ref{fockspace}).
Our results allowed us to put forward a unifying explanatory hypothesis for this whole set of experimental findings in human cognition, namely, that human thought is guided by two simultaneous processes -- `quantum conceptual thought', whose nature is `emergence', and `quantum logical thought', whose nature is `logic' \cite{asvIQSA2}. Our investigations indicate that the former generally prevails over the latter, and that the effects, paradoxes, contradictions, fallacies, experimentally detected in human cognition can be considered as expressions of this dominance, rather than `biases' of the human mind. More recently, we received a further crucial confirmation of this two-layered structure in human thought, namely the stability of the deviation from classical probabilistic rules we detected in \cite{asv2014} (Section \ref{explanation}).
Our quantum-theoretic perspective also accounts for two recent experimental results we obtained, namely, the identification of `quantum entanglement' in the conceptual combination {\it The Animal Acts} \cite{as2011,as2014} (Section \ref{entanglement}) and the detection of `quantum indistinguishability of Bose-Einstein type' in specific combinations of identical concepts, such as {\it Eleven Animals} \cite{asvIQSA1} (Section \ref{identity}). These discoveries are also important, in our opinion, from the point of view of the foundations of quantum physics, since they can shed new light on two mysterious aspects of the microscopic world -- entanglement and indistinguishability.
In this review article, we present the above results by basically following a historical reconstruction, though justified and restructured in a unitary and more general rational framework. We conclude our paper with some epistemological remarks on the role and interpretation of our quantum-theoretic perspective within the domain of cognitive modeling, and with some hints for future developments (Section \ref{conclusions}).
\section{The first steps: potentiality and contextuality in decision processes\label{contextuality}}
The first move towards the development of a quantum-theoretic perspective in cognition came from our former research on the mathematical and conceptual foundations of quantum physics. In particular, we were guided by our studies on:
(i) the identification of quantum structures outside the microscopic world, e.g., in the cognitive situation of the liar paradox \cite{abs1999,aabg2000};
(ii) the recognition of the existence of deep analogies between quantum particles and conceptual entities with respect to `potentiality' and `contextuality';
(iii) the role played by quantum probability in formalizing experimental situations where these aspects of potentiality and contextuality occur.
It is well known from quantum physics that, in a quantum measurement process, the measurement context influences the quantum entity that is measured in a non-deterministic way, actualizing one outcome in a set of possible measurement outcomes, as a consequence of the interaction between the quantum entity and the measurement context. Suppose now that a statistics of measurement outcomes is collected after a sequence of many repeated measurement processes on an arbitrary entity, and such that (i) the measurement actualises properties of the entity that were not actual before the measurement started, (ii) different outcomes and actualisations are obtained probabilistically. What type of probability can formalize such experimental situation? It cannot be classical probability, because classical probability formalizes lack of knowledge about actual properties of the entity that were already actual before the measurement started. We proved many years ago that a situation where context actualizes potential properties can instead be represented in a suitable quantum probabilistic framework \cite{a1986}.
What about a human decision process? Well, we realized that a decision process is generally made in a state of genuine potentiality, which is not of the type of a lack of knowledge of an actuality. The following example may help to illustrate this point. In \cite{aa1995}, we considered a survey including the question ``are you a smoker or not?''. Suppose that 21 participants over a whole sample of 100 participants answered `yes' to this question. We can then consider 0.21 as the probability of finding a smoker in this sample of participants. However, this probability is obviously of the type of a `lack of knowledge about an actuality', because each participant `is' a smoker or `is not' a smoker before the property has been tested, hence before the experiment to test it -- the survey -- starts. Suppose that we now consider the question ``are you for or against the use of nuclear energy?'', and that 31 participants answer they are in favor. In this case, the resulting probability $0.31$ is `not' of the type of `lack of knowledge about an actuality'. Indeed, it is very plausible for this type of question that some of the participants had no opinion about it before the survey, and hence for these participants the outcome was influenced by the context at the time the question was asked, including the specific conceptual structure of how the question was formulated. This is how context plays an essential role whenever the human mind is concerned with outcomes of experiments of a cognitive nature. We showed that the first type of probability, i.e. the type that models a `lack of knowledge about an actuality', is classical, and that the second type is non-classical and, possibly, quantum \cite{a1986}.
The effect due to role that context plays on a conceptual entity is equally fundamental than the effect due to the actualizing of potentialities during a decision process. Exactly as in a quantum measurement the measurement context changes the state of the quantum entity that is measured, in a decision process the cognitive context changes the state of the concept \cite{ag2005a,ag2005b}. For example, in our modeling of the concept {\it Pet}, we considered the context $e$ expressed by {\it Did you see the type of pet he has? This explains that he is a weird person}, and found that when participants in an experiment were asked to rate different exemplars of {\it Pet}, the scores for {\it Snake} and {\it Spider} were very high in this context. In our perspective, this is explained by the existence of different states for the concept {\it Pet}, where we use the notion of `state' in the same way as it is used in quantum theory, but also as it is used in ordinary language, i.e. `the state of affairs', meaning `how the affairs will react on different measurement contexts. We call `the state of {\it Pet} when no specific context is present', its ground state $\hat p$. The context $e$ then changes the ground state $\hat p$ into a new state $p_{weird\ person\ pet}$. Typicality, in our perspective, is an observable semantic quantity, which means that it takes different values in different states. Hence, in our perspective the typicality variations as encountered in the guppy effect are due to changes of state of the concept {\it Pet} under influence of a context. More specifically, the conjunction {\it Pet-Fish} is {\it Pet} under the context {\it Fish}, in which case the ground state $p$ of {\it Pet} is changed into a new state $p_{Fish}$. The typicality of {\it Guppy}, being an observable semantic quantity, will be different depending on the state, and this explains the high typicality of {\it Guppy} in the state $p_{Fish}$ of {\it Pet}, and its normal typicality in the ground state $p$ of {\it Pet} \cite{ag2005a}.
We developed this approach in a formal way, and called the underlying mathematical structure a `State Context Property' (SCoP) system \cite{ag2005a}. To build SCoP for an arbitrary concept $S$ we introduce three sets, namely, the set $\Sigma$ of states, denoting states by $p, q, \ldots$, the set ${\mathcal M}$ of contexts, denoting contexts by $e, f, \ldots$, and the set ${\mathcal L}$ of properties, denoting properties by $a, b, \ldots$. The `ground state' $\hat{p}$ of the concept $S$ is the state where $S$ is not under the influence of any particular context. Whenever $S$ is under the influence of a specific context $e$, a change of the state of $S$ occurs. In case $S$ was in its ground state $\hat{p}$, the ground state changes to a state $p$. The difference between states $\hat{p}$ and $p$ is manifested, for example, by the typicality values of different exemplars of the concept, and the applicability values of different properties being different in the two states $\hat{p}$ and $p$. Hence, to complete the mathematical construction of SCoP, also two functions $\mu$ and $\nu$ are introduced. The function $\mu: \Sigma \times {\mathcal M} \times \Sigma \longrightarrow [0, 1]$ is defined such that $\mu(q,e,p)$ is the probability that state $p$ of concept $S$ under the influence of context $e$ changes to state $q$ of concept $S$. The function $\nu: \Sigma \times {\mathcal L} \longrightarrow [0, 1]$ is defined such that $\nu(p,a)$ is the weight, or normalization of applicability, of property $a$ in state $p$ of concept $S$. With these mathematical structures and tools the SCoP formalism copes with both `contextual typicality' and `contextual applicability'.
We likewise built an explicit quantum-mechanical representation in a complex Hilbert space of the data of the experiment on {\it Pet} and {\it Fish} and different states of {\it Pet} and {\it Fish} in different contexts explored in \cite{ag2005a}, as well as of the concept {\it Pet-Fish} \cite{ag2005b}. In this way, we were able to cope with the pet-fish problem illustrated in Section \ref{intro}, (i).
The analysis above already contained the seeds of our quantum modeling perspective for concept combinations -- in particular, the notion of state of a concept marked the departure from the traditional idea of a concept as a set, eventually fuzzy, that contains instantiations. However, this analysis was still preliminary, and a general quantum-mechanical modeling required further experimental and theoretic steps, as it will be clear from the following section.
\section{Modeling concept combinations in Fock space\label{fockspace}}
We present here our quantum modeling perspective in Fock space for the combination of two concepts. It is successful in describing the classically problematical results illustrated in Section \ref{intro}, (ii) (concept conjunction and disjunction), (iii) (concept negation) and (iv) (borderline contradictions).
Let us firstly consider the membership weights of exemplars of concepts and their conjunctions/disjunctions measured by Hampton \cite{h1988a,h1988b}. He identified systematic deviations from classical (fuzzy) set conjunctions/disjunctions, an effect known as `overextension' or `underextension' (see Section \ref{intro}). We showed in \cite{a2009a} that a large part of Hampton's data cannot be modeled in a classical probability space satisfying the axioms of Kolmogorov \cite{k1933,p1989}. For example, the exemplar {\it Mint} scored in \cite{h1988a} the membership weight
$\mu(A)=0.87$ with respect to the concept {\it Food},
$\mu(B)=0.81$ with respect to the concept {\it Plant},
and $\mu(A \ {\rm and } \ B)=0.9$ with respect to their conjunction {\it Food And Plant}. Thus, the exemplar \emph{Mint} exhibits overextension with respect to the conjunction \emph{Food And Plant} of the concepts \emph{Food} and \emph{Plant}, and no classical probability representation exists for these data. More generally, the membership weights $\mu(A), \mu(B)$ and $\mu(A\ {\rm and}\ B)$ of the exemplar $x$ with respect to concepts $A$, $B$ and their conjunction `$A$ and $B$', respectively, can be represented in a classical Kolmogorovian probability model if and only if they satisfy the following inequalities \cite{s2014b,a2009a}
\begin{eqnarray} \label{ineq01}
&\mu(A\ {\rm and}\ B)-\min(\mu(A),\mu(B)) \le 0 \\
& \mu(A) + \mu(B) - \mu(A\ {\rm and}\ B) \le 1\label{ineq02}
\end{eqnarray}
A violation of (\ref{ineq01}) entails, in particular, that the minimum rule of fuzzy set conjunction does not hold, as in the case of {\it Mint}. A similar situation occurs in the case of disjunctions. We showed in \cite{a2009a} that a large part of Hampton's data cannot be modeled in a classical Kolmogorovian probability space. For example,
the exemplar {\it Sunglasses} scored in \cite{h1988b}
the membership weight $\mu(A)=0.4$ with respect to the concept {\it Sportswear}, $\mu(B)=0.2$ with respect to the concept
{\it Sports Equipment}, and
$\mu(A \ {\rm or} \ B)=0.1$
with respect to their disjunction
{\it Sportswear Or Sports Equipment}. Thus, the exemplar
\emph{Sunglasses} exhibits underextension with respect to the disjunction \emph{Sportswear Or Sports Equipment} of the concepts \emph{Sportswear} and \emph{Sports Equipment}, and no classical probability representation exists for these data. More generally, the membership weights $\mu(A), \mu(B)$ and $\mu(A\ {\rm or}\ B)$ of the exemplar $x$ with respect to concepts $A$, $B$ and their disjunction `$A$ or $B$', respectively, can be represented in a classical Kolmogorovian probability model if and only if they satisfy the following inequalities \cite{a2009a}
\begin{eqnarray} \label{maxdeviation}
\max(\mu(A),\mu(B))-\mu(A\ {\rm or}\ B)\le 0 \\ \label{kolmogorovianfactordisjunction}
0 \le \mu(A)+\mu(B)-\mu(A\ {\rm or}\ B)
\end{eqnarray}
A violation of (\ref{maxdeviation}) entails, in particular, that the maximum rule of fuzzy set disjunction does not hold, as in the case of {\it Sunglasses}.
In a first attempt to elaborate a quantum mathematics model for the data in \cite{h1988a} and \cite{h1988b} we were inspired by the quantum two-slit experiment.\footnote{In the present paper we use for our modeling purposes the standard quantum-mechanical formalism that is presented in modern manuals of quantum physics (see, e.g., \cite{d1958}). A basic summary of this formalism is contained in the volume including this article \cite{thisvolume}.} Consider, for example, the disjunction of two concepts. This led us to suggest the following Hilbert space model. One could represent the concepts $A$ and $B$ by the unit vectors $|A\rangle$ and $|B\rangle$, respectively, of a Hilbert space $\cal H$, and describe the decision measurement of a subject estimating whether the exemplar $x$ is a member of $A$ by means of a dichotomic observable represented by the orthogonal projection operator $M$. The probabilities $\mu(A)$ and $\mu(B)$ that $x$ is chosen as a member of $A$ and $B$, i.e. its membership weights, are given by the scalar products $\mu(A)=\langle A| M|A\rangle$ and $\mu(B)=\langle B| M|B\rangle$, respectively. The concept `$A \ \textrm{or} \ B$' is instead represented by the normalized superposition ${1 \over \sqrt{2}}(|A\rangle+|B\rangle)$ in $\cal H$. If $|A\rangle$ and $|B\rangle$ are chosen to be orthogonal, that is, $\langle A| B\rangle=0$, the membership weights $\mu(A), \mu(B)$ and $\mu(A\ {\rm or} \ B)$ of an exemplar $x$ for the concepts $A$, $B$ and `$A \ {\rm or} \ B$' are given by
\begin{eqnarray}
\mu(A)&=&\langle A | M|A\rangle \\
\mu(B)&=&\langle B | M|B\rangle \\
\mu(A \ {\rm or} \ B)&=&{1 \over 2}(\mu(A)+\mu(B))+\Re\langle A|M|B\rangle
\end{eqnarray}
respectively, where $\Re\langle A|M|B\rangle$ is the real part of the complex number $\langle A|M|B\rangle$. The term $\Re\langle A|M|B\rangle$ is called `interference term' in the quantum jargon, since it produces a deviation from the average ${1 \over 2}(\mu(A)+\mu(B))$ which would have been observed in the quantum two-slit experiment in absence of interference. In this way, the deviation from classicality in \cite{h1988a,h1988b} would be due to quantum interference, superposition and emergence, exactly as quantum interference, superposition and emergence are responsible of the deviation from the classically expected pattern in the two-slit experiment.
This `emergence-based' model in Hilbert space succeeded in describing many non-classical situations in \cite{h1988a} and \cite{h1988b}. However, it did not work for various cases, and these were exactly the cases where logic seemed to play a role in the mechanism of conceptual combination. This led us to work out a more general model in Fock space. We present the model in the following. We omit proofs and technical details in the following, for the sake of brevity, inviting the interested reader to refer to the bibliography quoted in this section.
In the case of two combining entities, a Fock space $\mathcal F$ consists of two sectors: `sector 1' is a Hilbert space $\cal H$, while `sector 2' is a tensor product $\cal H \otimes \cal H$ of two isomorphic versions of $\cal H$. It can be proved that a quantum probability model in Fock space exists for Hampton's data on conjunction and disjunction \cite{a2009a,ags2013}.
Let us start with the conjunction of two concepts. Let $x$ be an exemplar and let $\mu(A)$, $\mu(B)$ and $\mu(A \ {\rm and} \ B)$ be the membership weights of $x$ with respect to the concepts $A$, $B$ and `$A \ \textrm{and} \ B$', respectively. Let ${\cal F}={\cal H} \oplus ({\cal H} \otimes {\cal H})$ be the Fock space where we represent the conceptual entities. The states of the concepts $A$, $B$ and $`A \ \textrm{and} \ B'$ are represented by the unit vectors $|A\rangle, |B\rangle \in {\cal H}$ and $|A \ \textrm{and} \ B\rangle \in {\cal F}$, respectively, where
\begin{eqnarray}
|A \ \textrm{and} \ B\rangle=m e^{i\lambda}|A\rangle\otimes|B\rangle+ne^{i\nu}{1\over \sqrt{2}}(|A\rangle+|B\rangle)
\end{eqnarray}
The superposition vector ${1 \over \sqrt{2}}(|A\rangle+|B\rangle)$ describes `$A$ and $B$' as a new emergent concept, while the product vector $|A\rangle\otimes|B\rangle$ describes `$A$ and $B$' in terms of concepts $A$ and $B$. The positive numbers $m$ and $n$ are such that $m^{2}+n^{2}=1$. The decision measurement of a subject who estimates the membership of the exemplar $x$ with respect to the concept `$A \ \textrm{and} \ B$' is represented by the orthogonal projection operator $M\oplus (M \otimes M)$ on ${\cal F}$, where $M$ is an orthogonal projection operator on ${\cal H}$. Hence, the membership weight of $x$ with respect to `$A \ \textrm{and} \ B$' is given by
\begin{eqnarray} \label{AND}
\mu(A \ \textrm{and} \ B)&=&\langle A \ \textrm{and} \ B|M \oplus (M \otimes M)|A \ \textrm{and} \ B \rangle \nonumber \\
&=&m^2\mu(A)\mu(B)+n^2 \left ( {\mu(A)+\mu(B) \over 2}+\Re\langle A|M|B\rangle \right )
\end{eqnarray}
where $\mu(A)=\langle A|M|A\rangle$ and $\mu(B)=\langle B|M|B\rangle$ as above. The term $\Re\langle A|M|B\rangle$ is again the interference term of quantum theory. A solution of (\ref{AND}) exists in the Fock space ${\mathbb C}^{3} \oplus ({\mathbb C}^{3} \otimes {\mathbb C}^{3})$ where this interference term is given by
\begin{equation}
\Re\langle A|M|B\rangle=
\left\{
\begin{array}{ccc}
\sqrt{1-\mu(A)}\sqrt{1-\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)>1 \\
\sqrt{\mu(A)}\sqrt{\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)\le 1
\end{array}
\right.
\end{equation}
($\theta$ is the `interference angle'). Coming to the example above, namely, the exemplar {\it Mint} with respect to {\it Food}, {\it Plant} and {\it Food And Plant}, we have that (\ref{AND}) is satisfied with $m^2=0.3$, $n^2=0.7$ and $\theta=50.21^{\circ}$.
The previous mathematical representation admits the following interpretation. Whenever a subject is asked to estimate whether a given exemplar $x$ belongs to the concepts $A$, $B$, `$A \ {\rm and} \ B$', two mechanisms act simultaneously and in superposition in the subject's thought. A `quantum logical thought', which is a probabilistic version of the classical logical reasoning, where the subject considers two copies of exemplar $x$ and estimates whether the first copy belongs to $A$ and the second copy of $x$ belongs to $B$, and further the probabilistic version of the conjunction is applied to both estimates. But also a `quantum conceptual thought' acts, where the subject estimates whether the exemplar $x$ belongs to the newly emergent concept `$A \ {\rm and} \ B$'.
The place whether these superposed processes can be suitably structured is Fock space. Sector 1 hosts the latter process, while sector 2 hosts the former, while the weights $m^2$ and $n^2$ measure the `degree of participation' of sectors 2 and 1, respectively, in the case of conjunction.
In the case of {\it Mint}, subjects consider {\it Mint} to be more strongly a member of the concept {\it Food And Plant}, than they consider it to be a member of {\it Food} or of {\it Plant}. This is an effect due to a strong presence of quantum conceptual thought, the newly formed concept {\it Food And Plant} being found to be a better fitting category for {\it Mint} than the original concepts {\it Food} or {\it Plant}.
And indeed, in the case of {\it Mint}, considering the values of $n^2$ and $m^2$, the combination process mainly occurs in sector 1 of Fock space, which means that emergence prevails over logic.
Let us now come to the disjunction of two concepts. Let $x$ be an exemplar and let $\mu(A)$, $\mu(B)$ and $\mu(A \ {\rm or} \ B)$ be the membership weights of $x$ with respect to the concepts $A$, $B$ and `$A \ \textrm{or} \ B$', respectively. Let ${\cal F}={\cal H} \oplus ({\cal H} \otimes {\cal H})$ be the Fock space where we represent the conceptual entities. The concepts $A$, $B$ and `$A \ \textrm{or} \ B$' are represented by the unit vectors $|A\rangle, |B\rangle \in {\cal H}$ and $|A \ \textrm{or} \ B\rangle \in {\cal F}$, respectively, where
\begin{eqnarray}
|A \ \textrm{or} \ B \rangle=m e^{i\lambda}|A\rangle\otimes|B\rangle+ne^{i\nu}{1\over \sqrt{2}}(|A\rangle+|B\rangle)
\end{eqnarray}
The superposition vector $\frac{1}{\sqrt{2}}(|A\rangle+|B\rangle)$ describes `$A$ or $B$' as a new emergent concept, while the product vector $|A\rangle\otimes|B\rangle$ describes `$A$ or $B$' in terms of concepts $A$ and $B$. The positive numbers $m$ and $n$ are such that $m^{2}+n^{2}=1$, and they estimate the `degree of participation' of sectors 2 and 1, respectively, in the disjunction case. The decision measurement of a subject who estimates the membership of the exemplar $x$ with respect to the concept `$A \ \textrm{or} \ B$' is represented by the orthogonal projection operator $M \oplus ( M \otimes \mathbbmss{1}+\mathbbmss{1}\otimes M - M \otimes M)$ on ${\cal F}$, where $M$ has been introduced above. We notice that
\begin{equation}
M \otimes \mathbbmss{1}+\mathbbmss{1}\otimes M - M \otimes M= \mathbbmss{1}-
(\mathbbmss{1}-M)\otimes(\mathbbmss{1}-M)
\end{equation}
that is, we have applied de Morgan's laws of logic in sector 2 of Fock space in the transition from conjunction to disjunction. The membership weight of $x$ with respect to `$A \ \textrm{or} \ B$' is given by
\begin{eqnarray} \label{OR}
\mu(A \ \textrm{or} \ B)=\langle A \ \textrm{or} \ B | M \oplus (M \otimes \mathbbmss{1}+\mathbbmss{1}\otimes M - M \otimes M)
|A \ \textrm{or} \ B \rangle \nonumber \\
m^2 \left (\mu(A)+\mu(B)-\mu(A)\mu(B) \right )+n^2 \left ( {\mu(A)+\mu(B) \over 2}+\Re\langle A|M|B\rangle \right )
\end{eqnarray}
where $\mu(A)=\langle A|M|A\rangle$ and $\mu(B)=\langle B|M|B\rangle$. The term $\Re\langle A|M|B\rangle$ is the interference term. A solution of (\ref{OR}) exists in ${\mathbb C}^{3} \oplus ({\mathbb C}^{3} \otimes {\mathbb C}^{3})$ where this interference term
is given by
\begin{equation}
\Re\langle A|M|B\rangle=
\left\{
\begin{array}{ccc}
\sqrt{1-\mu(A)}\sqrt{1-\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)>1 \\
\sqrt{\mu(A)}\sqrt{\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)\le 1
\end{array}
\right.
\end{equation}
Coming to the example above, namely, the exemplar
{\it Sunglasses} with respect to {\it Sportswear}, {\it Sports Equipment} and {\it Sportswear Or Sports Equipment}, we have that (\ref{OR}) is satisfied with $m^2=0.03$, $n^2=0.97$ and $\theta=155.00^{\circ}$.
The previous mathematical representation admits the following interpretation. Whenever a subject is asked to estimate whether a given exemplar $x$ belongs to the concepts $A$, $B$, `$A \ {\rm or} \ B$', two mechanisms act simultaneously and in superposition in the subject's thought. A `quantum logical thought', which is a probabilistic version of the classical logical reasoning, where the subject considers two copies of exemplar $x$ and estimates whether the first copy belongs to $A$ or the second copy of $x$ belongs to $B$, and further the probabilistic version of the disjunction is applied to both estimates. And also a `quantum conceptual thought' acts, where the subject estimates whether the exemplar $x$ belongs to the newly emergent concept `$A \ {\rm or} \ B$'. The place whether these superposed processes are structured is again Fock space. Sector 1 hosts the latter process, while sector 2 hosts the former, while the weights $m^2$ and $n^2$ measure the `degree of participation' of sectors 2 and 1, respectively, in the case of disjunction.
In the case of {\it Sunglasses}, subjects consider {\it Sunglasses} to be less strongly a member of the concept {\it Sportswear Or Sports Equipment}, than they consider it to be a member of {\it Sportswear} or of {\it Sports Equipment}. This is an effect due to a strong presence of quantum conceptual thought, the newly formed concept {\it Sportswear Or Sports Equipment} being found to be a less well fitting category for {\it Sunglasses} than the original concepts {\it Sportswear} or {\it Sports Equipment}.
And indeed, in the case of {\it Sunglasses}, considering the values of $n^2$ and $m^2$, the combination process mainly occurs in sector 1 of Fock space, which means that emergence aspects prevails over logical aspects in the reasoning process.
Let us then analyze the experiment of Alxatib and Pelletier on borderline contradictions \cite{ap2011}. We proved in \cite{s2014a} that our quantum-theoretic model for the conjunction correctly represents the collected data. Suppose that a large sample of human subjects is asked to estimate the truth values of the sentences ``John is tall'', ``John is not tall'' and ``John is tall and not tall'', for a given subject John showed to the eyes of the subjects. And suppose that the fractions of positive answers are $0.01$, $0.95$ and $ 0.15$, respectively \cite{ap2011}. This `borderline case' is clearly problematical from a classical logical perspective, and can be modeled in terms of overextension. Indeed,
let us denote by $\mu(A)$, $\mu(A')$ and $\mu (A \ {\rm and} \ A')$ the probabilities that the sentences ``John is tall'', ``John is not tall'' and ``John is tall and not tall'' are true, and interpret them as membership weights of the exemplar {\it John} with respect to the concepts {\it Tall}, {\it Not Tall} and {\it Tall And Not Tall}, respectively. Then ({\ref{AND}}) is solved for $m^{2}=0.77$, $n^{2}=0.23$ and $\theta=0^{\circ}$ \cite{s2014a}. The explanation of this behaviour is that the reasoning process of the subject mainly occurs in sector 2 of Fock space, hence
logical reasoning is dominant, although emergent reasoning is also present, and it is its presence which evoked the name `contradiction' for this situation.
Let us finally come to the experiments on conceptual negation. The first studies on the negation of natural concepts were also performed by Hampton \cite{h1997}. He tested membership weights on conceptual conjunctions of the form {\it Tools Which Are Not Weapons}, finding overextension and deviations from
Boolean behaviour in the negation. We recently performed a more general cognitive test inquiring into the membership weights of exemplars with respect to conjunctions of the form {\it Fruits And Vegetables}, {\it Fruits And Not Vegetables}, {\it Not Fruits And Vegetables} and {\it Not Fruits And Not Vegetables} \cite{s2014b,asv2014}. Our
data confirmed significant deviations from classicality
and evidenced a very stable pattern of such deviations from the classicality conditions. The data could very faithfully be represented in two-sector Fock space, thus providing support to our quantum-theoretic modeling. More, they allowed us to attain new fundamental results in concept research and to sustain and corroborate our explanatory hypothesis in Section \ref{explanation}. Hence, it is worth to briefly review our recent results starting from the conditions for classicality of conceptual data sets, i.e. representability of empirical membership weights in a
Kolmogorovian probability space.
Let $\mu(A), \mu(B), \mu(A'), \mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A\ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$, and $\mu(A'\ {\rm and}\ B')$ be the membership weights of an exemplar $x$ with respect to the concepts $A$, $B$, their negations `not $A$', `not $B$' and the conjunctions `$A$ and $B$', `$A$ and not $B$', `not $A$ and $B$' and `not $A$ and not $B$', respectively, and suppose that all these membership weights are contained in the interval $[0,1]$ (which they will be in case they are experimentally determined as limits of relative frequencies of respective memberships). Then, they are `classical conjunction data' if and only if they satisfy the following conditions.
\begin{eqnarray} \label{condbis01}
&\mu(A)=\mu(A\ {\rm and}\ B)+\mu(A \ {\rm and}\ B') \\ \label{condbis02}
&\mu(B)=\mu(A\ {\rm and}\ B)+\mu(A' \ {\rm and}\ B) \\ \label{condbis03}
&\mu(A')=\mu(A'\ {\rm and}\ B')+\mu(A' \ {\rm and}\ B) \\ \label{condbis04}
&\mu(B')=\mu(A'\ {\rm and}\ B')+\mu(A \ {\rm and}\ B') \\ \label{condbis05}
&\mu(A\ {\rm and}\ B)+\mu(A\ {\rm and}\ B')+\mu(A'\ {\rm and}\ B)+\mu(A'\ {\rm and}\ B')=1
\end{eqnarray}
(see \cite{asv2014} for the proof).
A large amount of data collected in \cite{asv2014} violates very strongly and also very systematically (\ref{condbis01})--(\ref{condbis05}), hence these data cannot be generally reproduced in a classical Kolmogorovian probability framework. It can instead be shown that almost all these data can be represented by using our quantum-theoretic modeling in two-sector Fock space, as above. For the sake of simplicity, let us work out separate representations for the two sectors.
Let us start from sector 1 of Fock space, which models genuine emergence. We represent the concepts $A$, $B$ and their negations `not $A$', `not $B$' by the mutually orthogonal unit vectors $|A\rangle$, $|B\rangle$ and $|A'\rangle$, $|B'\rangle$, respectively, in the individual Hilbert space ${\cal H}$. The corresponding membership weights for a given exemplar $x$ are then given by the quantum probabilistic Born rule
\begin{eqnarray}
\mu(A)=\langle A|M|A\rangle & \quad & \mu(B)=\langle B|M|B\rangle \\
\mu(A')=\langle A'|M|A'\rangle& \quad & \mu(B')=\langle B'|M|B'\rangle
\end{eqnarray}
in sector 1. The conjunctions `$A$ and $B$', `$A$ and not $B$', `not $A$ and $B$', and `not $A$ and not $B$' are represented by the superposition vectors $\frac{1}{\sqrt{2}}(|A\rangle+|B\rangle)$, $\frac{1}{\sqrt{2}}(|A\rangle+|B'\rangle)$, $\frac{1}{\sqrt{2}}(|A'\rangle+|B\rangle)$ and $\frac{1}{\sqrt{2}}(|A'\rangle+|B'\rangle)$, respectively, in ${\cal H}$, i.e. sector 1 of Fock space, which expresses the fact `$A$ and $B$', `$A$ and not $B$', `not $A$ and $B$', and `not $A$ and not $B$' are considered as newly emergent concepts in sector 1.
Let us come to sector 2 of Fock space, which models logical reasoning.
Here we introduce a new element,
expressing an insight which we had not
yet in our earlier application of Fock space \cite{s2014b,a2009a,ags2013,s2014a}, and which we explain in detail in \cite{asv2014}. In short it comes to `taking into account that possibly $A$ and $B$ are meaning-connected and hence their probability weights mutually
dependent'. If this is the case, we cannot represent, e.g., the conjunction `$A$ and $B$' by the tensor product vector $|A\rangle \otimes |B\rangle$ of ${\cal H} \otimes {\cal H}$. This would indeed entail that the membership weight for the conjunction is $\mu(A \ {\rm and} \ B)=\mu(A)\mu(B)$ in sector 2, that is, probabilistic independence between the membership estimations of $A$ and $B$. We instead, following this new insight, represent the conjunction `$A$ and $B$' by an arbitrary vector $|C\rangle\in {\cal H} \otimes {\cal H}$, in sector 2, which in general will be entangled if $A$ and $B$ are meaning dependent. If we represent the decision measurements of a subject estimating the membership of the exemplar $x$ with respect to the concepts $A$ and $B$ by the orthogonal projection operators $M\otimes \mathbbm{1}$ and $\mathbbm{1}\otimes M$, respectively, we have
\begin{equation}
\mu(A)=\langle C|M\otimes \mathbbm{1}|C\rangle \quad
\mu(B)=\langle C| \mathbbm{1}\otimes M|C\rangle
\end{equation}
in sector 2. We have now to formalize the fact that this sector 2 has to express logical relationships between the concepts. More explicitly, the decision measurements of a subject estimating the membership of the exemplar $x$ with respect to the negations `not $A$' and `not $B$' should be represented by the orthogonal projection operators $(\mathbbmss{1}-M)\otimes \mathbbm{1}$ and $\mathbbm{1}\otimes (\mathbbmss{1}-M)$, respectively, in sector 2, in such a way that
\begin{equation}
\mu(A')=1-\mu(A)=\langle C|(\mathbbm{1}-M)\otimes \mathbbm{1}|C\rangle \quad
\mu(B')=1-\mu(B)=\langle C| \mathbbm{1}\otimes (\mathbbm{1}-M|C\rangle)
\end{equation}
in this sector.
Interestingly enough, there is a striking connection between logic and classical probability when conjunction and negation of concepts are at stake. Namely, the logical probabilistic structure of sector 2 of Fock space sets the limits of classical probabilistic models, and vice versa. In other words, if the experimentally collected membership weights $\mu(A)$, $\mu(B)$, $\mu(A')$, $\mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ can be represented in sector 2 of Fock space for a given choice of the state vector $|C\rangle$ and the decision measurement projection operator $M$, then the membership weights satisfy (\ref{condbis01})--(\ref{condbis05}), hence they are classical data. Vice versa, if $\mu(A)$, $\mu(B)$, $\mu(A')$, $\mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ satisfy (\ref{condbis01})--(\ref{condbis05}), hence they are classical data, then an entangled state vector $|C\rangle$ and a decision measurement projection operator $M$ can always be found such that $\mu(A)$, $\mu(B)$, $\mu(A')$, $\mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ can be represented in sector 2 of Fock space (see \cite{asv2014} for the proof).
Let us finally come to the general representation in two-sector Fock space. We can now introduce the general form of the vector representing the state of the conjunction of the concepts $A, B$ and their respective negations.
\begin{eqnarray}
| \Psi_{AB} \rangle&=& m_{AB}e^{i \lambda_{AB}} |C\rangle + \frac{n_{AB}e^{i \nu_{AB}}}{\sqrt{2}} (|A\rangle+|B\rangle) \\
| \Psi_{AB'} \rangle&=& m_{AB'}e^{i \lambda_{AB'}} |C\rangle + \frac{n_{AB'}e^{i \nu_{AB'}}}{\sqrt{2}} (|A\rangle+|B'\rangle) \\
| \Psi_{A'B} \rangle&=& m_{A'B}e^{i \lambda_{A'B}} |C\rangle + \frac{n_{A'B}e^{i \nu_{A'B}}}{\sqrt{2}} (|A'\rangle+|B\rangle) \\
| \Psi_{A'B'} \rangle&=& m_{A'B'}e^{i \lambda_{A'B'}} |C\rangle + \frac{n_{A'B'}e^{i \nu_{A'B'}}}{\sqrt{2}} (|A'\rangle+|B'\rangle)
\end{eqnarray}
where $m^2_{XY}+n^2_{XY}=1$, $X=A,A',Y=B,B'$.
The corresponding membership weights $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ can be written as in (\ref{AND}). We proved in \cite{asv2014} that they can be expressed in the Fock space ${\mathbb C}^{8} \oplus ({\mathbb C}^{8} \otimes {\mathbb C}^{8})$ as
\begin{eqnarray}
\mu(A\ {\rm and}\ B)&=&m_{AB}^2 \alpha_{AB}+n_{AB}^2(\frac{\mu(A)+\mu(B)}{2}+\beta_{AB}\cos\phi_{AB})
\label{FockSpaceSolutionAB} \\
\mu(A \ {\rm and} \ B')&=&m_{AB'}^2 \alpha_{AB'}+n_{AB'}^2(\frac{\mu(A)+\mu(B')}{2}+\beta_{AB'}\cos\phi_{AB'}) \label{FockSpaceSolutionAB'} \\
\mu(A' \ {\rm and}\ B)&=&m_{A'B}^2 \alpha_{A'B}+n_{A'B}^2(\frac{\mu(A')+\mu(B)}{2}+\beta_{A'B}\cos\phi_{A'B})
\label{FockSpaceSolutionA'B} \\
\mu(A'\ {\rm and}\ B')&=&m_{A'B'}^2\alpha_{A'B'}+n_{A'B'}^2(\frac{\mu(A')+\mu(B')}{2}+\beta_{A'B'}\cos\phi_{A'B'}) \label{FockSpaceSolutionA'B'}
\end{eqnarray}
where $0\le \alpha_{XY}\le 1$, $-1\le \beta_{XY} \le 1$, $X=A,A',Y=B,B'$.
Let us consider a relevant example, {\it Goldfish}, with respect to ({\it Pets}, {\it Farmyard Animals}) (big overextension in all experiments, but also double overextension with respect to {\it Not Pets And Farmyard Animals}). {\it Goldfish} scored $\mu(A)=0.93$ with respect to {\it Pets}, $\mu(B)=0.17$ with respect to {\it Farmyard Animals}, $\mu(A')=0.12$ with respect to {\it Not Pets}, $\mu(B')=0.81$ with respect to {\it Not Farmyard Animals}, $\mu(A \ {\rm and} \ B)=0.43$ with respect to {\it Pets And Farmyard Animals}, $\mu(A \ {\rm and} \ B')=0.91$ with respect to {\it Pets And Not Farmyard Animals}, $\mu(A' \ {\rm and} \ B)=0.18$ with respect to {\it Not Pets And Farmyard Animals}, and $\mu(A' \ {\rm and} \ B')=0.43$ with respect to {\it Not Pets And Not Farmyard Animals}. A complete modeling in the Fock space satisfying Eqs. (\ref{FockSpaceSolutionAB}), (\ref{FockSpaceSolutionAB'}), (\ref{FockSpaceSolutionA'B}) and (\ref{FockSpaceSolutionA'B'}) is characterized by coefficients:
(i) interference angles $\phi_{AB}=78.9^{\circ}$, $\phi_{AB'}=43.15^{\circ}$, $\phi_{A'B}=54.74^{\circ}$ and $\phi_{A'B'}=77.94^{\circ}$;
(ii) coefficents $\alpha_{AB}=0.12$, $\alpha_{AB'}=0.8$, $\alpha_{A'B}=0.05$ and $\alpha_{A'B'}=0.03$;
(iii) coefficients
$\beta_{AB}=-0.24$,
$\beta_{AB'}=0.10$, $\beta_{A'B}=0.12$ and $\beta_{A'B'}=0.30$;
(iv) convex weights
$m_{AB}=0.45$,
$n_{AB}=0.89$,
$m_{AB'}=0.45$,
$n_{AB'}=0.9$,
$m_{A'B}=0.48$, $n_{A'B}=0.88$, $m_{A'B'}=0.45$, and $n_{A'B'}=0.89$.
Following our interpretation in the case of conjunction and disjunction, we can say that, whenever a subject is asked to estimate whether a given exemplar $x$ belongs to the concepts $A$, $B$, `$A \ {\rm and} \ {\rm not} \ B$', both quantum logical and quantum conceptual thought simultaneously act in the subject's thought. According to the former, the subject considers two copies of $x$ and estimates whether the first copy belongs to $A$ and the second copy of $x$ does not belong to $B$. According to the latter, the subject estimates whether the exemplar $x$ belongs to the newly emergent concept `$A \ {\rm and} \ {\rm not} \ B$'. Fock space naturally captures this two-layered structure.
\section{A unifying explanatory hypothesis\label{explanation}}
The Fock space modeling presented in the previous section suggested us to formulate a general hypothesis which justifies and explains a whole set of empirical results on cognitive psychology under a unifying theoretic scheme \cite{asvIQSA2}. According to our explanatory hypothesis, human reasoning is a specifically structured superposition of two processes, a `logical reasoning' and an `emergent reasoning'. Logical reasoning combines cognitive entities -- concepts, combinations of concepts, or propositions -- by applying the rules of logic, though
generally in a probabilistic way. Emergent reasoning enables instead formation of combined cognitive entities as
newly emerging entities -- in the case of concepts, new concepts, in the case of propositions, new propositions --
carrying new meaning, connected with the meaning of the component cognitive entities, but with a connection not defined by the algebra of logic. These two mechanisms act simultaneously in human thought during a reasoning process, the first one is guided by an algebra of `logic', the second one follows a mechanism of `emergence'.
Human reasoning can be mathematically formalized in the two-sector Fock space presented in Section \ref{fockspace}. The states of conceptual entities are represented
by unit vectors of this Fock space as we have seen in the specific case of concept combinations. More specifically, `sector 1 of Fock space' models `conceptual emergence', hence the combination of two concepts is represented by a superposition vector of
the vectors representing the component concepts in this Hilbert space, allowing `quantum interference' between conceptual entities
to play a role in the process of emergence. `Sector 2 of Fock space' models a conceptual combination from the combining concepts by requiring
the rules of logic for the logical connective used for the combining, i.e. conjunction or disjunction,
to be satisfied in a probabilistic setting. This quantum-theoretic modeling suggested us to call `quantum conceptual thought' the process occurring in sector 1 of Fock space, `quantum logical thought' the process occurring in sector 2. The relative importance of emergence or logic in a specific cognitive process is measured by the `degree of participation' of sectors 1 and 2.
The abundance of evidence of deviations from classical logical reasoning in concrete human decisions (paradoxes, fallacies, effects, contradictions), together with our results in these two sections, led us to draw the conclusion that emergence constitutes the dominant dynamics of human reasoning, while logic is only a secondary structure. Therefore, we put forward the view that the aforementioned deviations from classicality are a consequence of the dominant dynamics and their nature is emergence, while classical logical reasoning is not a default to deviate from but, rather, a consequence of a secondary
structure and its nature is logic.
There is further empirical evidence revealing that what primarily guides human subjects in concrete human decisions is emergent reasoning, but logical aspects are likewise present.
A first element of evidence we identified by comparing the behavior of experimental data of different experiments. Consider, e.g., the exemplar {\it Olive}
and its membership weights with respect to the concepts {\it Fruits}, {\it Vegetables} and their conjunction {\it Fruits And Vegetables}, measured by ourselves \cite{s2014b,asv2014}, and its membership weights with respect to the concepts {\it Fruits}, {\it Vegetables} and their disjunction {\it Fruits Or Vegetables}, measured by Hampton \cite{h1988b}.
{\it Olive} scored $\mu(A)=0.56$ with respect to {\it Fruits}, $\mu(B)=0.63$ with respect to {\it Vegetables} and $\mu(A \ {\rm and} \ B)=0.65$ with respect to {\it Fruits
And Vegetables}, that is, {\it Olive} was double overextended with respect to the conjunction.
But, {\it Olive} was also double overextended with respect to the disjunction, since it scored $\mu(A)=0.5$ with respect to {\it Fruits}, $\mu(B)=0.1$ with respect to {\it Vegetables} and $\mu(A \ {\rm
or} \ B)=0.8$ with respect to {\it Fruits Or Vegetables}.
Our interpretation of these {\it Olive} case is the following. People see {\it Olive} as an exemplar which could be considered to be a fruit, but also could be considered to be a vegetable. Hence it could also, and even more so, be considered to be both, a fruit `and' a vegetable. This explains the double overextension of {\it Olive} with respect to the conjunction. This way of looking at {\it Olive} gives indeed the necessary weight to the conjunction to produce a double overextension.
Equally so, people see {\it Olive} as an exemplar which induces doubt about whether it is a vegetable or whether it is a fruit. Hence its could also, and even more so, be considered to be a 'fruit or a vegetable'. This explains the double overextension of {\it Olive} with respect to the disjunction. This way of looking at {\it Olive} gives indeed the necessary weight to the disjunction to produce a very big double overextension. Let us remark indeed that a double overextention with respect to the disjunction does not necessarily violates the classicality conditions, on the contrary, for a classical probability, the disjunction should be double overextended. For {\it Olive} the overextension is however so big that another one of the classicality conditions, namely the one linked to the Kolmogorovian factor is violated. For a classical probability model we have $\mu(A)+\mu(B)-\mu(A \ {\rm
or} \ B)=\mu(A \ {\rm and}\ B)$, which means $0\le \mu(A)+\mu(B)-\mu(A \ {\rm
or} \ B)$. However, for {\it Olive} we have $\mu(A)+\mu(B)-\mu(A \ {\rm
or} \ B)=0.5+0.1-0.8=-0.2<0$, which shows that the double overextension for the disjunction in the case of {\it Olive} is of a non-classical nature.
How come that {\it Olive} can give such weight to both conjunction and disjunction, although conjunction and disjunction are considered in classical probability to be distinctly different? It is because the meaning of {\it Olive} plays dominantly in sector 1, were quantum conceptual structures exist, and both connectives, `and' and `or' quite well resemble each other in this realm of conceptuality.
The second empirical evidence became manifest when we calculated the deviations from (\ref{condbis01})--(\ref{condbis05}) across all exemplars in our experiment in \cite{s2014b,asv2014}, and we noticed that these deviations have approximately constant numerical values. Indeed, let us introduce the following quantities.
\begin{equation}
I_{ABA'B'}=1-\mu(A\ {\rm and}\ B)-\mu(A\ {\rm and}\ B')-\mu(A'\ {\rm and}\ B)-\mu(A'\ {\rm and}\ B') \label{normalization}
\end{equation}
\begin{eqnarray}
I_{A}&=&\mu(A)-\mu(A\ {\rm and}\ B)-\mu(A \ {\rm and}\ B') \label{negationA} \\
I_{B}&=&\mu(B)-\mu(A\ {\rm and}\ B)-\mu(A' \ {\rm and}\ B) \label{negationB}\\
I_{A'}&=&\mu(A')-\mu(A'\ {\rm and}\ B')-\mu(A' \ {\rm and}\ B) \label{negationA'}\\
I_{B'}&=&\mu(B')-\mu(A'\ {\rm and}\ B')-\mu(A \ {\rm and}\ B') \label{negationB'}
\end{eqnarray}
We were very excited ourselves to find that for every $X=A,A',Y=B,B'$, $I_{X}$, $I_{Y}$ and $I_{ABA'B'}$ are constant functions across all exemplars, because this constitutes a very strong experimental evidence for the non-classical nature of what happens during concept combinations. More concretely, the last four equations give rise to values between $0$, which would be the classical value, and $-0.5$, but substantially closer to $-0.5$ than to $0$, and the fifth equation gives rise to a value between $0$, which again would be the classical value, and $-1$, but closer to $-1$ than to $0$. This is very strong evidence for the presence of non-classicality, indeed, if the classicality conditions are violated in such a strong and systematical way, the underlying structure cannot in any way be classical.
To test
the rugularity of this violation we firstly performed a `linear regression analysis' of the data to check whether these quantities can be represented by a line of the form $y=mx+q$, with $m=0$. This was the case. For $I_{A}$, we obtained $m=3.0 \cdot 10^{-3}$ with $R^{2}=0.94$; for $I_{B}$, we obtained $m=2.9 \cdot 10^{-3}$ with $R^{2}=0.93$; for $I_{A'}$, we obtained $m=2.6 \cdot 10^{-3}$ with $R^{2}=0.96$; for $I_{B'}$, we obtained $m=3.1 \cdot 10^{-3}$ with $R^{2}=0.98$; for $I_{ABA'B'}$, we obtained $m=4 \cdot 10^{-3}$ with $R^{2}=0.92$. Secondly, we computed the $95\%$-confidence interval for these parameters and obtained interval $(-0.51, -0.33)$ for $I_A$, interval $(-0.42, -0.28)$ for $I_{A'}$, interval $(-0.52, -0.34)$ for $I_B$, interval $(-0.40, -0.26)$ for $I_{B'}$, and interval $(-0.97, -0.64)$ for $I_{ABA'B'}$. This means that the measured parameters systematically fall within a narrow band centered at very similar values.
Next to the very strong experimental evidence for the non-classical nature of the underlying structure, the finding of this very stable pattern of violation constitutes also strong evidence for the validity of our Fock space model, and for the dominance of emergent reasoning with respect to logical reasoning when concepts are combined. Indeed, suppose for a moment that we substitute in the place of the experimental values in our equations to test classicality, the values that would be obtained theoretically in case we apply the first sector of Fock space equation of our Fock space model. Since interference in this equation can be negative as well as positive, and there is a priori no reason to suppose that there would be more of the one than the other, we can neglect the interference parts of the equation, since it is reasonable to suppose that they will cancel out when summing on all the terms of the equations of our classicality conditions. This means that we get the following, for every $X=A,A',Y=B,B'$, we have $\mu(X \ {\rm and } \ Y)=\frac{1}{2}({\mu(X)+\mu(Y)})$ (see (\ref{FockSpaceSolutionAB})--(\ref{FockSpaceSolutionA'B'})). A simple calculation shows that, for every $X=A,A',Y=B,B'$, $I_{X}=I_{Y}=-0.5$ and $I_{ABA'B'}=-1$, in this case. These are exactly the values to which our experimental violations are close, which means that are Fock space model captures the underlying structures in a systematic and deep way. The experimental values are in between these values, and $0$. which is the classical value, which means that also logical reasoning is present, but the emergent reasoning is dominant.
We think that these two results confirm, on one side, the general validity of our quantum-theoretic perspective in cognition and, on the other side, they constitute a very strong experimental support to the explanatory hypothesis presented in this section.
The next two sections complete our overview on the identification of quantum structures in concept combination, also shedding new light on the mysteries that surround quantum entanglement and indistinguishability at a microscopic level.
\section{Identification of entanglement\label{entanglement}}
The presence of entanglement is typically revealed in quantum physics by a violation of Bell-type inequalities \cite{b1964,chsh1969}, indicating that the corresponding coincidence measurements exhibit correlations that cannot be modeled in a classical Kolmogorovian probability framework \cite{a1986,p1989}.
We recently measured in a cognitive test statistical correlations in the conceptual combination {\it The Animal Acts}. We experimentally found that this combination violates Bell's inequalities \cite{ags2013,as2011} and elaborated a model that faithfully represents the collected data in complex Hilbert space \cite{as2014}. {\it The Animal Acts} unexpectedly revealed the presence of a `conceptual entanglement' which is only partly due to the component concepts, or `state entanglement', because it is also caused by `entangled measurements' and `entangled dynamical evolutions between measurements' \cite{as2014}. Our analysis shed new light on the mathematical and conceptual foundations of quantum entanglement, revealing that situations are possible where only states are entangled and measurements are products (`customary state entanglement'), but also situations where entanglement appears on the level of the measurements, in the form of the presence of both entangled measurements and entangled evolutions (`nonlocal box situation', `nonlocal non-marginal box situation'), due to the violation of the marginal distribution law, as in {\it The Animal Acts}. More specifically, {\it The Animal Acts} is a paradigmatic example of a `nonlocal non-marginal box situation', that is, an experimental situation where (i) joint probabilities do not factorize, (ii) Bell's inequalities are violated, and (iii) the marginal distribution law does not hold. Whenever these conditions are simultaneously satisfied, a form of entanglement appears which is stronger than the `customarily identified quantum entanglement in the states of microscopic entities'. In these cases, it is not possible to work out a quantum-mechanical representation in
a fixed ${\mathbb C}^2\otimes{\mathbb C}^2$ space which satisfies empirical data and where only the initial state is entangled while the measurements are products. It follows that entanglement is a more complex property than usually thought.
Shortly, if a single measurement is at play, one can distribute the entanglement between state and measurement but, if more measurements are considered, the marginal distribution law imposes limits on the ways to model the presence of the entanglement.
Let us now come to our coincidence measurements $e_{AB}$, $e_{AB'}$, $e_{A'B}$ and $e_{A'B'}$ for the conceptual combination {\it The Animal Acts}. In all measurements, we asked subjects to answer the question `is a good example of' the concept {\it The Animal Acts}. In measurement $e_{AB}$, participants choose among the four possibilities (1) {\it The Horse Growls}, (2) {\it The Bear Whinnies} -- and if one of these is chosen, the outcome is $+1$ -- and (3) {\it The Horse Whinnies}, (4) {\it The Bear Growls} -- and if one of these is chosen, the outcome is $-1$. In measurement $e_{AB'}$, they choose among (1) {\it The Horse Snorts}, (2) {\it The Bear Meows} -- and in case one of these is chosen, the outcome is $+1$ -- and (3) {\it The Horse Meows}, (4) {\it The Bear Snorts} -- and in case one of these is chosen, the outcome is $-1$. In measurement $e_{A'B}$, they choose among (1) {\it The Tiger Growls}, (2) {\it The Cat Whinnies} -- and in case one of these is chosen, the outcome is $+1$ -- and (3) {\it The Tiger Whinnies}, (4) {\it The Cat Growls} -- and in case one of these is chosen, the outcome is $-1$.
Finally, in measurement $e_{A'B'}$, participants choose among (1) {\it The Tiger Snorts}, (2) {\it The Cat Meows} -- and in case one of these is chosen, the outcome is $+1$ -- and (3) {\it The Tiger Meows}, (4) {\it The Cat Snorts} -- and in case one of these is chosen, the outcome is $-1$. We evaluate now the expectation values $E(A,B)$, $E(A, B')$, $E(A', B)$ and $E(A', B')$ associated with the measurements $e_{AB}$, $e_{AB'}$, $e_{A'B}$ and $e_{A'B'}$ respectively, and insert the values into the Clauser-Horne-Shimony-Holt (CHSH) version of Bell's inequality \cite{chsh1969}
\begin{equation} \label{chsh}
-2 \le E(A',B')+E(A',B)+E(A,B')-E(A,B) \le 2
\end{equation}
We performed a test on 81 participants who were presented a questionnaire to be filled out in which they were asked to choose among the above alternatives in $e_{AB}$, $e_{AB'}$, $e_{A'B}$ and $e_{A'B'}$.
Table 1 contains the results of our experiment \cite{as2011}.
If we denote by $P(H,G)$, $P(B,W)$, $P(H,W)$ and $P(B,G)$, the probability that {\it The Horse Growls}, {\it The Bear Whinnies},
{\it The Horse Whinnies} and {\it The Bear Growls}, respectively, is chosen in $e_{AB}$, and so for in the other measurements, the expectation values are, in the large number limits,
\begin{eqnarray}
&E(A,B)=&p(H,G)+p(B,W)-p(B,G)-p(H,W)=-0.7778 \nonumber \\
&E(A',B)=&p(T,G)+p(C,W)-p(C,G)-p(T,W)=0.6543 \nonumber \\
&E(A,B')=&p(H,S)+p(B,M)-p(B,S)-p(H,M)=0.3580 \nonumber \\
&E(A',B')=&p(T,S)+p(C,M)-p(C,S)-p(T,M)= 0.6296 \nonumber
\end{eqnarray}
Hence, (\ref{chsh}) gives
\begin{equation}
E(A',B')+E(A',B)+E(A,B')-E(A,B)=2.4197
\end{equation}
which is significantly greater than 2. This implies that (i) it violates Bell's inequalities, and (ii) the violation is close the maximal possible violation in quantum theory, i.e. $2\cdot\sqrt{2} \approx 2.8284$.
\begin{table} \label{tab}
\centering
\begin{footnotesize}
\begin{tabular}{|c | c | c| c| }
\hline
\emph{Horse Growls} & \emph{Horse Whinnies} & \emph{Bear Growls} & \emph{Bear Whinnies}\\
$p(H,G)=0.049$ & $p(H,W)=0.630$ & $p(B,G)=0.259$ & $p(B,W)=0.062$ \\
\hline
\emph{Horse Snorts} & \emph{Horse Meows} & \emph{Bear Snorts} & \emph{Bear Meows}\\
$p(H,S)=0.593$ & $p(H, M)=0.025$ & $p(B,S)=0.296$ & $p(B,M)=0.086$ \\
\hline
\emph{Tiger Growls} & \emph{Tiger Whinnies} & \emph{Cat Growls} & \emph{Cat Whinnies}\\
$p(T,G)=0.778$ & $p(T, W)=0.086$ & $p(C,G)=0.086$ & $p(C,W)=0.049$ \\
\hline
\emph{Tiger Snorts} & \emph{Tiger Meows} & \emph{Cat Snorts} & \emph{Cat Meows}\\
$p(T,S)=0.148$ & $p(T, M)=0.086$ & $p(C,S)=0.099$ & $p(C,M)=0.667$\\
\hline
\end{tabular}
\caption{The data collected in coincidence measurements on entanglement in concepts \cite{as2011}.}
\end{footnotesize}
\end{table}
Let us now construct a quantum representation in complex Hilbert space for the collected data by starting from an operational description of the conceptual entity {\it The Animal Acts}. The entity {\it The Animal Acts} is abstractly described by an initial state $p$. Measurement $e_{AB}$ has four outcomes $\lambda_{HG}$, $\lambda_{HW}$, $\lambda_{BG}$ and $\lambda_{BW}$, and four final states $p_{HG}$, $p_{HW}$, $p_{BG}$ and $p_{BW}$. Measurement $AB'$ has four outcomes $\lambda_{HS}$, $\lambda_{HM}$, $\lambda_{BS}$ and $\lambda_{BM}$, and four final states $p_{HS}$, $p_{HM}$, $p_{BS}$ and $p_{BM}$. Measurement $A'B$ has four outcomes $\lambda_{TG}$, $\lambda_{CG}$, $\lambda_{TW}$ and $\lambda_{CW}$, and four final states $p_{TG}$, $p_{TW}$, $p_{CG}$ and $p_{CW}$. Measurement $A'B'$ has four outcomes $\lambda_{TS}$, $\lambda_{CS}$, $\lambda_{TM}$ and $\lambda_{CM}$, and four final states $p_{TS}$, $p_{TM}$, $p_{CS}$ and $p_{CM}$.
Then, we consider the Hilbert space ${\mathbb C}^4$ as the state space of {\it The Animal Acts} and represent the state $p$ by the unit vector $|p\rangle \in {\mathbb C}^4$. We assume that $\{|p_{HG}\rangle, |p_{HW}\rangle, |p_{BG}\rangle,$ $ |p_{BW}\rangle \}$, $\{|p_{HS}\rangle, |p_{HM}\rangle, |p_{BS}\rangle, |p_{BM}\rangle\}$, $\{|p_{TG}\rangle$, $ |p_{TW}\rangle$, $|p_{CG}\rangle$, $|p_{CW}\rangle\}$, $\{|p_{TS}\rangle,$ $|p_{TM}\rangle, |p_{CS}\rangle, |p_{CM}\rangle\}$ are orthonormal (ON) bases of ${\mathbb C}^4$. Therefore, $|\langle p_{HG}|\psi\rangle|^2=p(H,G)$, $|\langle p_{HW}|\psi\rangle|^2=p(H,W)$, $|\langle p_{BG}|\psi\rangle|^2=p(B,G)$, $|\langle p_{BW}|\psi\rangle|^2=p(B,W)$, in the measurement $e_{AB}$. We proceed analogously for the other probabilities. Hence, the self-adjoint operators
\begin{eqnarray}
{\cal E}_{AB}&=&\sum_{i=H,B}\sum_{j=G,W}\lambda_{ij}|p_{ij}\rangle \langle p_{ij}| \\
{\cal E}_{AB'}&=&\sum_{i=H,B}\sum_{j=S,M}\lambda_{ij}|p_{ij}\rangle \langle p_{ij}|\\
{\cal E}_{A'B}&=&\sum_{i=T,C}\sum_{j=G,W}\lambda_{ij}|p_{ij}\rangle \langle p_{ij}|\\
{\cal E}_{A'B'}&=&\sum_{i=T,C}\sum_{j=S,M}\lambda_{ij}|p_{ij}\rangle \langle p_{ij}|
\end{eqnarray}
represent the measurements $e_{AB}$, $e_{AB'}$, $e_{A'B}$ and $e_{A'B'}$ in ${\mathbb C}^4$, respectively.
Let now the state $p$ of {\it The Animal Acts} be the entangled state represented by the unit vector $|p\rangle=|0.23e^{i13.93^\circ}, 0.62e^{i16.72^\circ},0.75e^{i9.69^\circ},0e^{i194.15^\circ}\rangle$ in the canonical basis of ${\mathbb C}^{4}$. This choice is not arbitrary, but deliberately `as close as possible to a situation of only product measurements', as we explained in \cite{as2014}. Moreover, we choose the outcomes $\lambda_{HG}=\lambda_{BW}=+1$, $\lambda_{HW}=\lambda_{BG}=-1$, and so on, as in our concrete experiment. We proved that
\scriptsize
\begin{eqnarray}
{\cal E}_{AB}
&=&
\left( \begin{array}{cccc}
0.952 & -0.207-0.030i & 0.224+0.007i & 0.003-0.006i \\
-0.207+0.030i & -0.930 & 0.028-0.001i & -0.163+0.251i \\
0.224-0.007i & 0.028+0.001i & -0.916 & -0.193+0.266i \\
0.003+0.006i & -0.163-0.251i & -0.193-0.266i & 0.895
\end{array} \right)
\\
{\cal E}_{AB'}
&=&
\left( \begin{array}{cccc}
-0.001 & 0.587+0.397i & 0.555+0.434i & 0.035+0.0259i \\
0.587-0.397i & -0.489 & 0.497+0.0341i & -0.106-0.005i \\
0.555-0.434i & 0.497-0.0341i & -0.503 & 0.045-0.001i \\
0.035-0.0259i & -0.106+0.005i & 0.045+0.001i & 0.992 \end{array} \right)
\\
{\cal E}_{A'B}
&=&
\left( \begin{array}{cccc}
-0.587 & 0.568+0.353i & 0.274+0.365i & 0.002+0.004i \\
0.568-0.353i & 0.090 & 0.681+0.263i & -0,110-0.007i \\
0.274-0.365i & 0.681-0.263i & -0.484 & 0.150-0.050i \\
0,002-0.004i & -0,110+0.007i & 0.150+0.050i & 0.981
\end{array} \right)
\end{eqnarray}
\begin{eqnarray}
{\cal E}_{A'B'}
&=&
\left( \begin{array}{cccc}
0.854 & 0.385+0.243i & -0.035-0.164i & -0.115-0.146i \\
0.385-0.243i & -0.700 & 0.483+0.132i & -0.086+0.212i \\
-0.035+0.164i & 0.483-0.132i & 0.542 & 0.093+0.647i \\
-0.115+0.146i & -0.086-0.212i & 0.093-0.647i & -0.697
\end{array} \right)
\end{eqnarray}
\normalsize
in this case \cite{as2014}.
This completes the quantum-theoretic modeling in ${\mathbb C}^{4}$ for our cognitive test. One can then resort to the definitions of entangled states and entangled measurements and to the canonical isomorphisms, ${\mathbb C}^{4}\cong{\mathbb C}^{2}\otimes{\mathbb C}^{2}$ and $L({\mathbb C}^{4})\cong L({\mathbb C}^{2})\otimes L({\mathbb C}^{2})$ ($L({\cal H})$ denotes the vector space of linear operators on the Hilbert space $\cal H$), and one can prove that all measurements $e_{AB}$, $e_{AB'}$, $e_{A'B}$ and $e_{A'B'}$ are entangled with this choice of the entangled state $p$ \cite{as2014}. Moreover, the marginal distribution law is violated by all measurements, e.g., $p(H,G)+p(H,W) \ne p(H,S)+p(H,M)$. Since we are below Tsirelson's bound \cite{tsirelson80}, this modeling is an example of a `nonlocal non-marginal box modeling 1', following the classification we proposed in \cite{asQI2013}.
To conclude the section we remind that we have used the term `entanglement' by explicitly referring to the structure within the theory of quantum physics that a modeling of experimental data takes, if (i) these data are represented, following carefully the rules of standard quantum theory, in a complex Hilbert space, and hence states, measurements, and evolutions, are presented respectively by vectors (or density operators), self-adjoint operators, and unitary operators in this Hilbert space; (ii) a situation of coincidence joint measurement on a compound entity is considered, and the subentities are identified following the tensor product rule of `compound entity description in quantum theory' (iii) within this tensor product description of the compound entity entanglement is identified, as `not being product', whether it is for states (non-product vectors), measurements (non-product self-adjoint operators), or evolutions (non-product unitary transformations).
\section{The quantum nature of conceptual identity\label{identity}}
One of the most mysterious and less understood aspects of quantum entities is the way they behave with respect to `identity', and more specifically their statistical behaviour due to indistinguishability. Indeed, the statistical behaviour of quantum entities is very different from the statistical behaviour of classical objects, which are instead, in principle,
not identical, hence distinguishable, whenever there is more than one. The latter is governed by the Maxwell-Boltzmann (MB) distribution, while the former is described by the Bose-Einstein (BE) distribution for quantum particles with integer spin, and by the Fermi-Dirac (FD) distribution for quantum particles with semi-integer spin (we omit considering fractional statistical particles here, for the sake of brevity) \cite{French2006,Dieks2008}.
What about concepts? Consider, e.g., the linguistic expression ``eleven animals''. This expression, when both ``eleven'' and ``animals'' are looked upon with respect to their conceptual structure, represents the combination of concepts {\it Eleven} and {\it Animals} into {\it Eleven Animals}, which is again a concept.
Each of the {\it Eleven Animals} is then completely identical on this conceptual level, and hence indistinguishable. The same linguistic expression can however also elicit the thought about eleven objects, present in space and time, each of them being an instantiation of {\it Animal}, and thus distinguishable from each other. We recently inquired into experiments on such combinations of concepts, surprisingly finding that BE statistics appears at an empirical level for specific types of concepts, hence finding strong evidence for the hypothesis that indeed there is a profound connection between the behavior of concepts with respect to identity and indistinguishability and the behaviour of quantum entities with respect to these notions \cite{asvIQSA1}.
What is interesting in this respect is that we can intuitively understand the behavior of concepts with respect to identity and indistinguishability, which means that it might well be that an understanding of the behavior of quantum entities with respect to identity and indistinguishability should be searches for by making use of this analogy. In this sense, that identical concepts can be modeled exactly as identical quantum entities, i.e. by using quantum theory, is not only a strong achievement for quantum cognition, but it might also incorporate a new way to reflect about this mysterious behaviour of identical quantum entities.
Let us discuss these aspects both at a theoretic and an empirical level, as follows.
Let us firstly consider the SCoP structure in Section \ref{contextuality} and two states of {\it Animal}, namely {\it Cat} and {\it Dog}, hence the situation where {\it Eleven Animals} can be either {\it Cats} or {\it Dogs}. Then, the conceptual meaning of {\it Eleven Animals}, which can be {\it Cats} or {\it Dogs}, gives rise in a unique way to twelve possible states. Let us denote them by $p_{11,0}$, $p_{10,1}$, \ldots, $p_{1,10}$ and $p_{0,11}$, and they stand respectively for {\it Eleven Cats} (and no dogs), {\it Ten Cats And One Dog}, \ldots, {\it One Cat And Ten Dogs} and {\it Eleven Dogs} (and no cats).
We investigated the `probabilities of change of the ground state $\hat p$ of the combined concept {\it Eleven Animals} into one of the twelve states $p_{11,0}$, $p_{10,1}$, \ldots, $p_{1,10}$ and $p_{0,11}$' in a cognitive experiment on human subjects. The subjects were presented the twelve states and asked to choose their preferred one. The relative frequency arising from their answers was interpreted as the probability of change of the ground state $\hat p$, to the chosen state, i.e. one of the set $\{p_{11,0}$, $p_{10,1}$, \ldots, $p_{1,10}$, $p_{0,11}\}$. The context $e$ involved in this experiment is mainly determined by the
`combination procedure of the concepts {\it Eleven} and {\it Animals}'
and the `meaning contained in the new combination' for participants in the experiment \cite{ag2005a}. Hence, our psychological experiment tested whether participants follow the `conceptual meaning' of {\it Eleven Animals} treating {\it Dogs} ({\it Cats}) as identical, or participants follow the `instantiations into objects meaning' of {\it Eleven Animals} treating {\it Dogs} ({\it Cats}) as distinguishable.
We mathematically represent the conceptual entity {\it Eleven Animals} by the SCoP model $(\Sigma,{\cal M},
\mu)$, where $\Sigma=\{\hat p, p_{11,0}, p_{10,1}, \ldots, p_{1,10}, p_{0,11}\}$, ${\cal M}= \{ e \}$, and our transition probabilities are $\{\mu(q,e,\hat p)\ \vert q\in \{p_{11,0}, p_{10,1}, \ldots, p_{1,10}, p_{0,11}\}\}$. We recognise in the structure of $\mu(q,e,\hat p)$ the situation is analogous to the one in which one has $N=11$ particles that can be distributed in $M=2$ possible states. It is
thus possible, by looking at the relative frequencies obtained in the experiment, to find out whether a classical MB statistics or a quantum-type, i.e. BE statistics, applies to this situation. In case that MB would apply, it would mean that things happen as if there are underlying the twelve states hidden possibilities, namely $T(n,C;11-n,D)=11!/n!(11-n)!$ in number, for the specific state of $n$ {\it Cats} and $11-n$ {\it Dogs}, $n=0, \ldots,11$. Of course, this ``is'' true in case the cats and dogs are real cats and dogs, hence are `objects existing in space and time', which is why for objects in the classical world indeed MB statistics applies. Let us calculate the probabilities involved then.
For the sake of simplicity, we assume that two probability values $P_{Cat}$ and $P_{Dog}$ exist such that $P_{Cat}+P_{Dog}=1$, and that the events of
making actual such an underlying state for {\it Cat} and {\it Dog} are independent. Hence the probability
for $n$ exemplars of {\it Cat} and $11-n$ exemplars of {\it Dog} is
then
\begin{equation}
\mu_{MB}^{P_{Cat},P_{Dog}}(p_{n,11-n},e,\hat p)=T(n,C;11-n,D) P_{Cat}^nP_{Dog}^{11-n}=\frac{11!}{n!(11-n)!} P_{Cat}^nP_{Dog}^{11-n}
\end{equation}
Note that, under the assumption of MB statistics, $\mu_{MB}^{P_{Cat},P_{Dog}}(p,e,\hat p)$ becomes the binomial probability distribution. For example, if $P_{Dog}=P_{Cat}=0.5$, the number of possible arrangements for the state {\it Eleven Cats And Zero Dogs} and for the state {\it Zero Cats And Eleven Dogs} is 1, hence the corresponding probability for these configurations is $\mu_{MB}^{P_{Cat},P_{Dog}}(p_{0,11},e,\hat p)=\mu_{MB}^{P_{Cat},P_{Dog}}(p_{11,0},e,\hat p)=0.0005$. Analogously, the number of possible arrangements for the state {\it Ten Cats And One Dog} and for the state {\it One Cat And Ten Dogs} is 11, hence the corresponding probability for these configurations is $\mu_{MB}^{P_{Cat},P_{Dog}}(p_{10,1},e,\hat p)=\mu_{MB}^{P_{Cat},P_{Dog}}(p_{1,10},e,\hat p)=0.0054$, and so on. When $P_{Cat}$ and $P_{Dog}$ are equal, MB distribution entails a maximum value for such a probability. In this example, this corresponds to the situation of {\it Six Cats And Five Dogs} and {\it Five Cats And Six Dogs} with $\mu_{MB}^{P_{Cat},P_{Dog}}(p_{6,5},e,\hat p)=\mu_{MB}^{P_{Cat},P_{Dog}}(p_{6,5},e,\hat p)=0.2256$.
Let us now make the calculation for BE statistics, where we keep making the exercise of only reasoning on the level of concepts, and not on the level of instantiations. This means that the twelve different states do not admit underlying hidden states, because the existence of such states would mean that we reason on more concrete forms in the direction of instantiations. As above, we suppose that {\it Cat} and {\it Dog} have an independent elicitation probability $P_{Cat}$ and $P_{Dog}$ such that $P_{Cat}+P_{Dog}=1$. Hence, the probability that there are $n$ exemplars of {\it Cat} and $(11-n)$ exemplars of {\it Dog} is
\begin{equation}
\mu_{BE}^{P_{Cat},P_{Dog}}(p_{n,11-n},e,\hat p)= \frac{(nP_{cat}+(11-n)P_{dog})}{(\frac{12\times 11}{2})}
\end{equation}
Note that as $P_{Cat}=1-P_{Dog}$, then $\mu_{BE}^{P_{Cat},P_{Dog}}(p_{n,11-n},e,\hat p)$ is a linear function. Moreover, when
$P_{Cat}=P_{Dog}=0.5$, we have that $\mu_{BE}(p_{n,11-n},e,\hat p)=1/12$ for all values of $n$, thus recovering BE distribution \cite{asvIQSA1}.
Starting from the above theoretic analysis, if one performs experiments on a collection of concepts like {\it Eleven Animals} to estimate the probability of elicitation for each state, then one can establish whether a distribution of MB-type $\mu_{MB}^{P_{Cat},P_{Dog}}(p_{n,11-n},e,\hat p)$, or of BE-type $\mu_{BE}^{P_{Cat},P_{Dog}}(p_{n,11-n},e,\hat p)$, or a different one, holds. However, in case there are strong deviations from a MB statistics, while a quasi-linear distribution is obtained, then this would indicate that, in context $e$, where only {\it Cat} and {\it Dog} are allowed to be states of the concept {\it Animal}, the statistical distribution of the collection of concepts {\it Eleven Animals} is of a BE-type and that concepts present a quantum-type indistinguishability.
We performed a cognitive experiment with 88 participants. We considered a list of concepts $A^i$ of different (physical and non-physical) nature, $i=1, \ldots,14$, and two possible exemplars (states) $p_{1}^i$ and $p_{2}^i$ for each concept. Next we requested participants to choose one exemplar of a combination $N^{i} A^i$ of concepts, where
$N^{i}$ is a natural number. The exemplars of these combinations of concepts
$A^i$ are the states $p^i_{k,N^{i}-k}$ describing the conceptual combination `$k$ exemplars in state $p^i_1$ and $(N^i-k)$ exemplars in state $p^i_2$', where $k$ is an integer such that
$k=0, \ldots, N^i$. For example, the first collection of concepts we considered is $N^1 A^1$ corresponding to the compound conceptual entity {\it Eleven Animals}, with $p^i_{1}$ and $p^i_{2}$ describing the exemplars {\it Cat} and {\it Dog} of the individual concept {\it Animal}, respectively, and $N^1=11$. The exemplars (states) we considered are thus $p^1_{11,0}$, $p^1_{10,1}$, \ldots, $p^{1}_{1,10}$, and $p^{1}_{0,11}$, describing the combination {\it Eleven Cats And Zero Dogs}, {\it Ten Cats And One Dog}, \ldots, {\it One Cat And Ten Dogs}, and {\it Zero Cats And Eleven Dogs}. The other collections of concepts we considered in our cognitive experiment are reported in Table 2.
\begin{table}[H]\label{categories-psych}
\begin{center}
\begin{tabular}{|c|c|c|c|c|} \hline
$i$ & $N^i$ & $A^i$ & $p_{1}^i$ & $p_{2}^i$ \\
\hline
\hline
1& 11 & {\it Animals} & {\it Cat} & {\it Dog} \\ \hline
2& 9 & {\it Humans} & {\it Man} & {\it Woman}\\ \hline
3& 8 & {\it Expressions of Emotion} & {\it Laugh} & {\it Cry} \\ \hline
4& 7& {\it Expressions of Affection} & {\it Kiss} & {\it Hug}\\ \hline
5& 11& {\it Moods} & {\it Happy} & {\it Sad} \\ \hline
6& 8 &{\it Parts of Face} & {\it Nose} & {\it Chin} \\ \hline
7& 9 & {\it Movements} & {\it Step} & {\it Run} \\ \hline
8& 11 & {\it Animals} & {\it Whale} & {\it Condor}\\ \hline
9& 9 & {\it Humans} & {\it Child} & {\it Elder} \\ \hline
10& 8 & {\it Expressions of Emotion} & {\it Sigh} & {\it Moan} \\ \hline
11& 7& {\it Expressions of Affection} & {\it Caress} & {\it Present}\\ \hline
12& 11& {\it Moods} & {\it Thoughtful} & {\it Bored}\\ \hline
13& 8 &{\it Parts of Face} &{\it Eye} & {\it Cheek} \\ \hline
14& 9 &{\it Movements} & {\it Jump} & {\it Crawl} \\ \hline
\end{tabular}
\caption{List of concepts and their respective states for the psychological concept on identity and indistinguishability.}
\end{center}
\end{table}
We computed the parameters $P^{MB}_{p_{1}^i}$ and $P^{BE}_{p_{1}^i}$ that minimize the the R-squared value of the fit using the distributions $\mu_{MB}^{P_{p_{1}^i},P_{p_{2}^i}}$ and $\mu_{BE}^{P_{p_{1}^i},P_{p_{2}^i}}$ for each $i=1, \ldots, 14$. Hence, we fitted the distributions obtained in the psychological experiments using MB and BE statistics
(note that only one
parameter is needed as $P^{MB}_{p_{2}^i}=1-P^{MB}_{p_{1}^i}$ and $P^{BE}_{p_{2}^i}=1-P^{BE}_{p_{1}^i}$).
Next, we used the `Bayesian Information Criterion (BIC)' \cite{KASSBIC} to estimate which model provides the best fit and contrast this criterion with the R-squared value.
Table
3 summarizes the statistical analysis. The first column of the table identifies the concept in question (see Table 2),
the second and third columns show $P^{MB}_{p_{1}^i}$ and the $R^2$ value of the MB statistical fit, the fourth and fifth columns show $P^{BE}_{p_{1}^i}$ and the $R^2$ value of the BE statistical fit. The sixth column shows the $\Delta_{\textrm{BIC}}$ criterion to discern between the $\mu_{MB}^{P_{p_{1}^i}P_{p_{2}^i}}$ and $\mu_{BE}^{P_{p_{1}^i}P_{p_{2}^i}}$, and the seventh column identifies the distribution which best fits the data for concept $A^i$, $i=1, \ldots,14$.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|} \hline
$i$&$P^{MB}_{p_{1}^i}$&$R_{MB}^2$&$P^{BE}_{p_{1}^i}$&$R_{BE}^2$&$\Delta_{\textrm{BIC}}$ &Best Model \\ \hline \hline
1& 0.55 &-0.05 &0.16 &0.78 &19.31 & BE strong \\ \hline
2& 0.57 &{\bf 0.78} &0.42 &0.44 &-9.54 &MB strong \\ \hline
3& 0.82 &0.29 &{\bf 0.96} &0.79 & 10.81&BE strong \\ \hline
4& 0.71 &0.81 &0.53 &0.77 &-1.69 & MB weak\\ \hline
5& 0.25 &{\bf 0.79} &0.39 &0.93 & 14.27&BE strong\\ \hline
6& 0.62 &0.59 &0.61 &0.57 &-0.37 &MB weak\\ \hline
7& 0.72 &0.41 &0.64 &{\bf 0.83} &12.66 &BE strong\\ \hline
8& 0.63 &0.58 &0.47 &0.73 &5.53 & BE positive\\ \hline
9& 0.45 &{\bf 0.87} &0.26 &0.67 &-9.69 &MB strong\\ \hline
10&0.59 &0.50 &0.63 &0.77 &7.17 &BE positive\\ \hline
11&0.86 &0.46 &1.00 &{\bf 0.87} &11.4 &BE strong \\ \hline
12&0.21 &0.77 &0.00 &0.87 &6.68 &BE positive\\ \hline
13&0.62 &0.54 &0.71 &0.67 &2.97 &BE weak \\ \hline
14&0.81 &0.20 &0.91 &{\bf 0.90} &20.68 &BE strong\\ \hline
\end{tabular}
\label{stat-results}
\caption{Results of statistical fit for the psychological experiment. Each column refers to the 14 collections of concepts introduced in Table 2.}
\end{center}
\end{table}
Note that, according to the BIC criteria, negative $\Delta_{\textrm{BIC}}$ values imply that the category is best fitted by a MB distribution, whereas positive $\Delta_{\textrm{BIC}}$ values on row $i$ imply the concept
$A^i$ is best fitted with a BE distribution. Moreover, when $|\Delta_{\textrm{BIC}}|<2$ there is no clear difference between the models, when $2<|\Delta_{\textrm{BIC}}|<6$ we can establish a positive but not strong difference towards the model with smallest value, whereas when $6<|\Delta_{\textrm{BIC}}|$ we are in presence of strong evidence that one of the models provides better fit than the other model~\cite{KASSBIC}. We see that categories $2$ and $9$ show a strong $\Delta_{\textrm{BIC}}$ value towards MB-type of statistics, and that categories $1,3,5,7,11,12$ and $14$ show a strong $\Delta_{\textrm{BIC}}$ value towards BE-type of statistics.
Complementary to the BIC criterion, the $R^2$ fit indicator helps to see whether or not the indications of $\Delta_{\textrm{BIC}}$ can be confirmed with a good fit of the data. Interestingly, the concepts we have identified with strong indication towards one type of statistics have $R^2$ values larger than $0.78$ (such $R^2$ values are marked in bold text), which indicates a fairly good approximation for the data. Moreover, note that in all the cases with strong tendency towards one type of statistics, the $R^2$ of the other type of statistics shows is poor. This confirms the fact that we can discern between the two types of statistics depending on the concept in question.
The interpretation of our results is thus clear. Conceptual combinations exists, like {\it Nine Humans}, whose distribution follows MB statistics. However, also conceptual combinations, like {\it Eleven Animals}, {\it Eight Expressions of Emotion} or {\it Eleven Moods}, whose distribution follows BE statistics exist. The conclusion is that the nature of identity in these concept combinations is of a quantum-type and
in these combinations the human mind treats the two states we consider as identical and indistinguishable. Also the hypothesis that `the more easy the human mind
imagines spontaneously instantiations, e.g., {\it Nine Humans}, the more MB, and the less easy such instantiations are activated in imagination, e.g., {\it Eight Expressions of Emotion}, the more BE statistics appears' is confirmed by our experiment.
We have an intuitive explanation for this empirical difference. Whenever the human mind `imagines' two different combinations of {\it Eleven Animals}, say 2 cats and 9 dogs, and 5 cats and 6 dogs, the human mind does not really take into account that this situation can be about real cats and dogs, in which case there are many more ways to put 5 cats and 6 dogs into a cage, than to put 2 cats and 9 dogs. This is the reason why BE, not MB, appears in this case. Suppose instead that the human mind considers two different combinations of {\it Nine Humans}, say 2 elders and 7 children, and 4 elders and 5 children. Then the human mind is likely to be influenced by real known families with 9 sons and, in the real world, there are much more situations of families with 4 elders and 5 children, than 2 elders and 7 children.
This pattern was confirmed by a second experiment we performed on the World-Wide Web about the nature of conceptual indistinguishability \cite{asvIQSA1}.
\section{Concluding remarks and perspectives\label{conclusions}}
In the previous sections we have provided an overview of our quantum-theoretic perspective on concepts and their combinations. We have expounded the reasons that led us to develop this perspective, namely our former research on operational and axiomatic approaches to quantum physics, the origins of quantum probability and various experimental results in cognition pointing to a deviation of human reasoning from the structures of classical (fuzzy set) and classical probability theory. We have proved that these deviations can be described in terms of genuine quantum effects, such as contextuality, emergence, interference and superposition. We have identified further quantum aspects in the mechanisms of conceptual combination, namely entanglement and quantum-type indistinguishability. And, we have proposed an explanation that allows the unification of these different empirical results under a common underlying principle on the structure of human thought.
Our quantum-theoretic perspective fits the global research domain that applies the mathematical formalisms of quantum theory in cognitive science and has been called `quantum cognition' (we quote here some known books on this flourishing domain \cite{k2010,bb2012,hk2013}). Further, we believe that our findings in cognition may also have, as a feedback, a deep impact on the foundations of microscopic quantum physics. Indeed, let us consider entanglement. We identified in concepts an entanglement situation where Bell's inequalities are violated within Tsirelson's bound \cite{tsirelson80}, the marginal distribution law is violated and there is `no signaling', which implies that entangled measurements, in addition to entangled states, are needed to model this experimental situation. And this completely occurs within a Hilbert space quantum framework, at variance with widespread beliefs. This theoretic scheme with entangled measurements could explain some `anomalies', i.e. deviations from the marginal distribution law, that were recently observed in the typical Bell-type nonlocality tests with entangled photons \cite{ak1,ak2}. Then, let us consider the quantum nature of conceptual indistinguishability. In our perspective, it is due to
the human mind being able to consider specific conceptual entitites
without the need to also imagine instantiations as objects existing in space and time of these conceptual entities.
Hence, it could well be that quantum indistinguishability at a microscopic level
is provoked by the fact that quantum entitites
are not localized as objects in space and time,
and that non-locality would mean non-spatiality, a view that has been put forward by one of us in earlier work for different reasons \cite{aerts1990}.
These insights could for example have implications on quantum statistics and the so-called `spin-statistics theorem' \cite{French2006,Dieks2008}.
We conclude this article with an epistemological consideration. We think that our quantum-theoretic perspective in concept theory constitutes a step towards the construction of a general theory for
the modeling of conceptual entities. In this sense, we distinguish it from what is typically considered as
an ad hoc cognitive model. To understand what we mean by this distinction let us consider an example taken from everyday life. As an example of a theory, we could introduce the theory of `how to make good clothes'. A tailor needs to learn how to make good clothes for different types of people, men, women, children, old people, etc. Each cloth is a model on itself. Then, one can also consider intermediate situations where one has models of series of clothes. A specific body will not fit in any clothes: you need to adjust the parameters (length, size, etc.) to reach the desired fit. We think that
a theory
should be able to reproduce different experimental results by suitably adjusting the involved parameters, exactly as a
theory of clothing. This is
different from a set of ad hoc models, even if the set can cope with a wide range of experimental data. There is a tendency in psychology to be critical for a theory that can cope with all possible situations it applies to. One then often believes that the theory contains too many parameters, and that it is only by allowing all these parameters to attain different values that all the data can be modeled. In case we have to do with an ad hoc model, i.e, a model specially made for the circumstance of the situation it models, this suspicion is grounded. Adding parameters to such an ad hoc model, or stretching the already contained parameters to other values, does not give rise to what we call a theory.
A theory needs to be well defined, its rules, the allowed procedures, its theoretical, mathematical, and internal logical structure, `independent' of the structure of the models describing specific situations that can be coped with by the theory. Hence also the theory needs to contain a well defined description of `how to produce models for specific situations'. Think again of the theory of clothing. If a taylor knows the theory of clothing, obviously he or she can make a cloth for every human body, because the theory of clothing, although its structure is defined independently of a specific clothe, contains a prescription of how to apply it to any possible specific cloth. Other subtle aspects are involved with the differences between ad hoc models and models finding their origin in a theory \cite{aertsrohrlich1998}. Here we mainly wanted to bring up the issue, because we think it does lead to misunderstandings not to pay attention to the difference between an ad hoc model, and a model which is derived from a theory. Intuitive thoughts about the nature of a model differ often depending on whether the model is inspired by a psychology approach, it will then rather automatically been looked upon as an ad hoc model, and that all data can be modelled is the suspicious, or whether it is inspired by a physics approach, where it will rather be looked upon as resulting from a theory, and that all data can be modeled by it is the a positive aspect, validating the theory.
What is the status of the Fock space model for concept combinations? Hilbert space, hence also Fock space when appropriate, for the description of quantum entities provides models that definitely come from a theory, namely quantum theory, and hence are not ad hoc models. Is quantum theory also a theory for concepts and their combinations, and hence, if so, can we consider our models, e.g. the Fock space model, as models coming from this theory? Or is quantum cognition rather still a discipline where ad hoc models are built, making use, also in a rather ad hoc way, of mathematics arising from quantum theory? An answer to this question can not yet been given definitely, but some hypothesis can be formulated with plausibility in respect to it. We believe that, notwithstanding their deep analogies, concept entities are less crystalized and symmetric structures than quantum entities. As a matter of fact some data in \cite{h1988a,h1988b} and \cite{s2014b,asv2014} cannot be modeled in Fock space, and further experimental findings could in the future confirm such impossibility. Notwithstanding this, we believe that emergence is the actual driver also for these data that cannot be modeled in Fock space. But, this type of emergence cannot be represented in a linear Hilbert (Fock) space, and more general structures are needed. The search for more general mathematical structures capturing conceptual emergence will, by the way, constitute an important aspect of our future investigation in concept theory.
On the other hand, we do believe that we have arrived in the realm of building models that come from a theory, and are not ad hoc. Indeed, although we believe that this theory will turn out to be a generalisation of the actual quantum theory, its basic principles -- except linearity most probably -- will be present in the generalised quantum theory too. We have recently worked out an analysis where the view of the status of actual quantum cognition, as describing a quantum-like domain of reality, less crystallised than the micro-world, but containing deep analogies in its foundations, is put forward \cite{as2014b}.
\bigskip
\noindent
{\bf Acknowledgements.} We are greatly indebted to the scientists who have collaborated with us in these years on the fashinating research illustrated in the present article, namely, Sven Aerts, Jan Broekaert, Bart D'Hooghe, Marek Czachor, Liane Gabora, Massimiliano Sassoli de Bianchi, Sonja Smets, Jocelyn Tapia and Tomas Veloz. The results expounded here would not have been possible without our collaborations with them and their contributions.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.